COMPUTATIONALLY EFFICIENT
AI IN BRAIN-TUMOUR MRI AND
PHYSIOLOGICAL SIGNALS IN
SLEEP AND EMOTION
ANALYSIS
Dr. Muhammad Irfan
TURUN YLIOPISTON JULKAISUJA – ANNALES UNIVERSITATIS TURKUENSIS
PUBLICATION SERIES INFORMATION SHOULD GO HERE
University of Turku
Faculty of Technology
Department of Computing
Information and Communication Technology
Doctor of Science (Technology)
Supervised by
Prof, Tomi Westerlund
University of Turku, Finland
Prof, Wei Chen
University of Sydney, Australia
Reviewed by
Prof, David Grayden
University of Melbourne, Australia.
Prof, Huiru (Jane) Zheng
Ulster University, UK.
Opponent
Prof, Theodoros Arvanitis
University of Birmingham, UK
The originality of this publication has been checked in accordance with the University
of Turku quality assurance system using the Turnitin OriginalityCheck service.
ISBN XXX-XXX-XX-XXXX-X (PRINT)
ISBN XXX-XXX-XX-XXXX-X (PDF)
ISSN XXX-XXX-XX-YYYY-Y (PRINT)
ISSN XXX-XXX-XX-ZZZZ-Z (ONLINE)
Printhouse, City, Country, Year
Dedication goes here
...and possibly continues
UNIVERSITY OF TURKU
Faculty of Technology
Department of Computing
Information and Communication Technology
IRFAN, MUHAMMAD: Computationally Efficient AI in Brain-Tumour MRI and
Physiological Signals in Sleep and Emotion Analysis
Doctoral dissertation, 111 pp.
Doctor of Science (Technology)
Feb 2026
ABSTRACT
In healthcare, diagnosing neurological and behavioural disorders depends on AI
models that efficiently process complex data. However, many artificial intelligence
(AI) models are too complex and resource-intensive, limiting their use in settings
with limited computational power, battery capacity, or bandwidth. This thesis aims
to develop AI that is both reliable and efficient in resource-limited environments,
thereby maximising patient benefit. This thesis presents three contributions in med-
ical imaging and signal analysis. First, it introduces two convolutional operators,
Fuzzy Atrous Convolution (FAC) and Fuzzy Sigmoid Convolution (FSC), for binary
brain-tumour classification from magnetic resonance imaging (MRI). Both meth-
ods improve feature extraction while reducing the number of parameters compared
to standard CNNs. Despite their compactness, both models achieve state-of-the-art
(SOTA) results. The FSC model achieves 0.9917, 0.9975, and 0.9989 in binary clas-
sification accuracy on three datasets, with 100 times fewer parameters than transfer-
learning models. The FAC model achieves 0.9883, 0.9967, and 0.9956, outperform-
ing SOTA models with 300 times fewer parameters, showing efficiency in inference
time and size. Second, the thesis examines whether accurate neonatal sleep-stage
classification can be achieved with reduced-channel EEG. Through multiview fea-
ture extraction, AdaptiSelect, and a data-reduction module, it reduces data transmis-
sion by 153.6 times while enabling full data reconstruction. On a four-year dataset,
it achieves classification accuracy of 0.8116 (𝐾𝑎𝑝𝑝𝑎 = 0.7217) with one EEG chan-
nel and 0.8279 (𝐾𝑎𝑝𝑝𝑎 = 0.7470) with two channels under cross-validation. Leave-
one-subject-out validation confirms its generalisability. It outperforms SOTA ap-
proaches in neonatal sleep analysis. Third, the thesis explores lightweight emotion-
recognition models using single-modality electrocardiography. Accurate emotion
detection is vital in human-computer interaction, mental health, and neurorehabilita-
tion, where continuous, reliable evaluation improves outcomes. For edge monitoring,
models must be computationally efficient. The approach employs multidomain fea-
ture extraction & fusion, and a voting model for low-power, real-time emotion track-
ing, ideal for wearables. The tailored model achieves an average accuracy of 0.9559
on the POPANE dataset, outperforming the general model’s 0.6992. Comparisons
show consistent improvement over past research.
KEYWORDS: Computational Efficiency, Medical Imaging, Physiological Signals,
Brain Tumour MRI, Neonatal Sleep, Emotion Recognition, Lightweight AI
iv
TURUN YLIOPISTO
Teknillinen tiedekunta
Tietotekniikan laitos
Tietojenkäsittely ja tietoliikenne
IRFAN, MUHAMMAD: Computationally Efficient AI in Brain-Tumour MRI and
Physiological Signals in Sleep and Emotion Analysis
Väitöskirja, 111 s.
Tekniikan tohtori
Kuukausi 2026
TIIVISTELMÄ
Käytännön terveydenhuollossa neurologisten ja käyttäytymiseen liittyvien häir-
iöiden diagnosointi perustuu tekoälymalleihin, jotka kykenevät käsittelemään mon-
imutkaista dataa tehokkaasti. Monet nykyiset tekoälymallit ovat kuitenkin liian mon-
imutkaisia ja resurssi-intensiivisiä estäen niiden soveltamisen rajallisen laskentate-
hon, akkukapasiteetin tai tiedonsiirron kaistanleveyden yhteydessä. Tämän väitöskir-
jan tavoitteena on kehittää tekoälymenetelmiä, jotka ovat samanaikaisesti luotettavia
ja resurssitehokkaita, jotta potilaiden saama hyöty voidaan maksimoida. Väitöskir-
jassa esitetään kolme tieteellistä kontribuutiota lääketieteellisen kuvantamisen ja sig-
naalinkäsittelyn aloilta. Ensinnäkin esitellään kaksi uutta konvoluutio-operaattoria
FAC ja FSC aivokasvainten luokitteluun magneettikuvantamisesta. Kompaktista rak-
enteestaan huolimatta molemmat mallit saavuttavat alan huipputason suorituskyvyn.
FSC-malli saavuttaa tarkkuudet 0,9917, 0,9975 ja 0,9989 kolmessa eri aineistossa
käyttäen sata kertaa vähemmän parametreja kuin siirtovaikutukseen perustuvat mallit.
FAC-malli saavuttaa vastaavasti tarkkuudet 0,9883, 0,9967 ja 0,9956 ja päihittää alan
huipputason verkot noin 300 kertaa pienemmällä parametrimäärällä, mikä osoittaa
sen tehokkuuden sekä inferenssiajan että mallikoon suhteen. Toiseksi väitöskirjassa
kehitetään resurssitehokas vastasyntyneiden univaiheiden luokittelujärjestelmä, joka
hyödyntää yksi- ja kaksikanavaista elektroenkefalografiaa (EEG). Moninäkymäiseen
piirteiden poimintaan, AdaptiSelect-menetelmään ja datanreduktiomoduuliin perus-
tuva ratkaisu vähentää tiedonsiirtotarvetta153,6 kertaisesti mahdollistaen samalla alku-
peräisen datan täydellisen rekonstruoinnin. Neljän vuoden aikana kerätyllä aineis-
tolla järjestelmä saavuttaa ristiinvalidoinnissa yhden EEG-kanavan asetelmassa tarkku-
uden 0,8116 (𝜅 = 0, 7217) ja kahden kanavan asetelmassa 0,8279 (𝜅 = 0, 7470).
Kolmanneksi väitöskirjassa tarkastellaan kevyitä tunteentunnistusmalleja, jotka hyö-
dyntävät yksimodaalista elektrokardiografiaa (ECG). Reaaliaikaista seurantaa varten
mallien on oltava laskennallisesti kevyitä. Menetelmä soveltuu erityisen hyvin puet-
taville laitteille. Räätälöity malli saavuttaa POPANE-aineistolla tarkkuuden 0,9559,
mikä ylittää selvästi yleistetyn mallin tarkkuuden 0,6992. Vertailut aiempaan tutkimuk-
seen osoittavat suorituskyvyn parantuneen johdonmukaisesti.
Avainsanat: laskennallinen tehokkuus; lääketieteellinen kuvantaminen; fysiologiset
signaalit; aivokasvaimen magneettikuvaus (MRI); vastasyntyneiden uni; tunteentun-
nistus; kevyet tekoälymallit.
v
Acknowledgements
Date: 2025.11.18
Dr. Muhammad Irfan
DR. MUHAMMAD IRFAN
I am a senior researcher and principal investigator work-
ing in artificial intelligence for healthcare, with a focus
on AI-enabled health monitoring, treatment support, and
translational medical imaging. My research spans mul-
timodal biomedical signal and image analysis (MRI, CT,
PET), edge and federated learning, and generative AI, with
applications in brain and head-and-neck cancers, stroke,
and dementia care. I hold a PhD in biomedical engineer-
ing from Fudan University and am completing a second
PhD at the University of Turku. I have published articles
in leading IEEE transactions and journals, as well as in El-
sevier journals, and have secured competitive national and
international research funding.
vi
Table of Contents
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Original Publications . . . . . . . . . . . . . . . . . . . . . . xi
Publications not included to Thesis . . . . . . . . . . . . . . . . . xii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Brain-Tumour . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Brain Tumour Assessment Modalities . . . . . . . . 1
1.1.2 Architecture Efficiency in Brain-Tumour MRI . . . . 3
1.2 Wearable technology . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Neonatal Sleep Analysis . . . . . . . . . . . . . . . . 4
1.2.2 EEG . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 Common Neonatal EEG Artefacts . . . . . . . . . . 10
1.2.4 Validation and Comparison to the Gold Standard . . 11
1.2.5 Visual EEG Interpretation and Inter-Rater Agreement 11
1.2.6 Emotion Analysis: Minimal Sensing . . . . . . . . . 16
1.3 Aims of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . 20
2 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1 MRI Datasets for Brain-Tumour Classification . . . . . . . . 22
2.1.1 Dataset I (Publication I & II) . . . . . . . . . . . . . . 22
2.1.2 Dataset II (Publication I & II) . . . . . . . . . . . . . . 22
2.1.3 Dataset III (Publication I & II; Combined) . . . . . . 22
2.2 Neonatal EEG Dataset (Publication III) . . . . . . . . . . . . 22
2.3 POPANE Emotion Dataset (Publication IV) . . . . . . . . . . 23
3 General Workflow in AI for Healthcare . . . . . . . . . . . . . 25
3.0.1 Data Collection . . . . . . . . . . . . . . . . . . . . . 25
3.0.2 Labelling and Annotation . . . . . . . . . . . . . . . . 26
3.0.3 Data Preparation and Pre-processing . . . . . . . . 27
vii
Dr. Muhammad Irfan
3.0.4 Edge vs. Cloud Processing . . . . . . . . . . . . . . 27
3.0.5 Feature Extraction . . . . . . . . . . . . . . . . . . . 28
3.0.6 Feature Fusion and Selection . . . . . . . . . . . . . 28
3.0.7 Model Development . . . . . . . . . . . . . . . . . . . 28
3.0.8 Data Splitting and Validation . . . . . . . . . . . . . . 29
3.0.9 Post-processing . . . . . . . . . . . . . . . . . . . . . 30
3.1 Fuzzy Atrous Convolution for Brain Tumour MRI . . . . . . . 30
3.2 Fuzzy Sigmoid Convolution for Brain Tumour MRI . . . . . . 31
3.3 Neonatal Sleep Analysis using EEG . . . . . . . . . . . . . . 32
3.4 Minimal Wearable Sensing for Emotions Analysis . . . . . . 33
4 Summary of key findings and Discussion . . . . . . . . . . . 35
4.0.1 Publication I: FAC for Brain–Tumour MRI . . . . . . 36
4.0.2 Publication II: FSC for Brain–Tumour MRI . . . . . . 39
4.0.3 Publication III: Smart IoT-based Solutions for Neona-
tal Sleep Analysis . . . . . . . . . . . . . . . . . . . . 41
4.0.4 Publication IV: Emotion Recognition with Minimal Wear-
able Sensing . . . . . . . . . . . . . . . . . . . . . . . 43
5 Contribution of these Studies to the Overall Thesis Theme 45
5.1 Alignment with the Overall Thesis Aims . . . . . . . . . . . . 45
5.2 Contribution of the Brain Tumour MRI Studies (Publications
I–II) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.3 Contribution of the Neonatal Sleep Study (Publication III) . 46
5.4 Contribution of the Emotion Recognition Study (Publication
IV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.5 Synthesis and Overall Impact . . . . . . . . . . . . . . . . . . 48
5.6 Shortcomings and Future Work . . . . . . . . . . . . . . . . 48
List of References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Original Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
viii
Abbreviations
AI Artificial Intelligence
ML Machine Learning
SOTA State-of-the-Art
MRI Magnetic Resonance Imaging
CT Computed Tomography
PET Positron Emission Tomography
EEG Electroencephalogram
EOG Electrooculogram
EMG Electromyogram
ECG Electrocardiogram
PSG Polysomnography
NICU Neonatal Intensive Care Unit
CHFDU Children’s Hospital of Fudan University
WS Wake State
NI Stage N1 Sleep
NII Stage N2 Sleep
NIII Stage N3 Sleep
AS Active Sleep
QS Quiet Sleep
DTCWT Dual-Tree Complex Wavelet Transform
FAWT Flexible Analytic Wavelet Transform
ECOV Eigenvector Covariance Features
MSPCA Multiscale Principal Component Analysis
STREAM Smart Cloud Data Transfer and Reconstruction Module
FS Feature Selection
KBest SelectKBest Univariate Feature Selection
CNN Convolutional Neural Network
HNN Hierarchical Neural Network
MS-HNN Multi-Scale Hierarchical Neural Network
MBCNN Multi-Branch Convolutional Neural Network
BiLSTM Bidirectional Long Short-Term Memory Network
LR Logistic Regression
GNB Gaussian Naive Bayes
ix
Dr. Muhammad Irfan
SVM Support Vector Machine
DT Decision Tree
RF Random Forest
XTRA ExtraTrees Classifier
KNN k-Nearest Neighbours
MLP Multilayer Perceptron
HMM Hidden Markov Model
CV Cross-Validation
LOSO-CV Leave-One-Subject-Out Cross-Validation
Acc. Accuracy
Pre. Precision
Rec. Recall
F1 F1-Score
Kappa Cohen’s Kappa Coefficient
MCC Matthews Correlation Coefficient
AUC Area Under the ROC Curve
x
List of Original Publications
This dissertation is based on the following original publications, which are referred
to in the text by their Roman numerals:
I Irfan, M., Subasi, A., Mehdi, H., Westerlund, T. and Chen, W., 2024, De-
cember. Fuzzy-Based Atrous Convolution for Brain Tumor Detection Using
MRI. In 2024 IEEE International Conference on Progress in Informatics and
Computing (PIC) (pp. 280-289). doi: 10.1109/PIC62406.2024.10892686.
II Irfan, M., A. Nawaz, R. Klén, A. Subasi, T. Westerlund and W. Chen,
"Improved Brain Tumor Detection in MRI: Fuzzy Sigmoid Convolution in
Deep Learning," 2025 International Joint Conference on Neural Networks
(IJCNN), Rome, Italy, 2025, pp. 1-8,
doi: 10.1109/IJCNN64981.2025.11227858.
III Irfan, M., Wang, L., Xu, Y., Subasi, A., Chen, C., Klen, R., Westurlund, T.
and Chen, W., 2025. Smart iot-based solutions for neonatal sleep stratifica-
tion: Single-dual channel eeg, adaptiselect, multview fusion, & rotational
ensemble stacking. IEEE Internet of Things Journal.
doi: 10.1109/JIOT.2025.3558235
IV Irfan, M., Nawaz, A., Bulbul, A.K., Klén, R., Subasi, A., Westerlund, T.
and Chen, W., 2025. Emotion recognition with minimal wearable sens-
ing: Multi-domain feature, hybrid feature selection, and personalized vs.
generalized ensemble model analysis. In: Proceedings of the 2025 IEEE
International Conference on Bioinformatics and Biomedicine (BIBM), pp.
5447–5453. https://doi.org/10.1109/BIBM66473.2025.11356135
V Bulbul, A.K., Irfan, M., Jaakkola, M.K., Subasi, A. and Klén, R., 2025.
Wearable-based Emotion Recognition Using Electrocardiogram and Gal-
vanic Skin Response and Instrumentation Audit: A Systematic Review.
IEEE Transactions on Instrumentation and Measurement, under review.
The original publications have been reproduced with the permission of the copyright
holders.
xi
Publications not included in Thesis
The following publications are thematically related to the work presented in this
thesis, but are not included as part of the compiled dissertation. They either address
adjacent methodological developments, earlier investigations, or parallel research
contributions conducted during the doctoral period.
II Irfan, M., Shun, P., Felix, B.B., Mustafa, N., Abbasi, S.F., Nahli, A., Subasi,
A., Westerlund, T. and Chen, W., 2023. An IoT-based noncontact ECG sys-
tem: sole of the feet/hands palm. IEEE Internet of Things Journal, 10(21),
pp.18718-18732.
I Nawaz, A., Irfan, M., Yu, X., Zou, Z. and Westerlund, T., 2025. "Blockchain-
Enabled Privacy-Preserving Second-Order Federated Edge Learning in Per-
sonalised Healthcare," in IEEE Transactions on Consumer Electronics, doi:
10.1109/TCE.2025.3620115.
III Nawaz, A., Wang, L., Irfan, M. and Westerlund, T., 2024. Hyperledger
Sawtooth-based supply chain traceability system for counterfeit drugs. Com-
puters & Industrial Engineering, 190, p.110021.
IV Siddiqa, H.A., Tang, Z., Xu, Y., Wang, L., Irfan, M., Abbasi, S.F., Nawaz,
A., Chen, C. and Chen, W., 2024. Single-Channel EEG data analysis us-
ing a multi-branch CNN for neonatal sleep staging. IEEE Access, 12,
pp.29910-29925.
V Irfan, M., Subasi, A., Mustafa, N., Westerlund, T. and Chen, W., 2024. An
evaluation of pretrained convolutional neural networks for stroke classifi-
cation from brain CT images. In Applications of Artificial Intelligence in
Healthcare and Biomedicine (pp. 111-135). Academic Press.
VI Nawaz, A., Irfan, M., Sadiqa, H.A. and Westerlund, T., 2023, November.
Edge-based skin cancer decision support system using machine learning
algorithms. In 2023 IEEE Intl Conf on Dependable, Autonomic and Se-
cure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl
Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and
Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) (pp. 0292-
0297). IEEE.
xii
TABLE OF CONTENTS
VII Ferreira, A., et al. (incl. Irfan, M.) (2025). AI and co-design to support de-
mentia: lessons from a first co-creation session. In Proc. International
Conference on Computer-Human Interaction Research and Applications
(CHIRA 2025), pp. 1–1. Available: https://chira.scitevents.org/Home.aspx?y=2025
(Accepted but yet to be published at Springer)
VIII Irfan, M., Subasi, A., Tang, Z., Wang, L., Xu, Y., Chen, C., Westurlund,
T. and Chen, W., 2025. A novel nicu sleep state stratification: Multiper-
spective features, adaptive feature selection and ensemble model. IEEE
Transactions on Biomedical Engineering.
IX Irfan, M., Wang, L., Shahid, H., Xu, Y., Subasi, A., Munawar, A., Mustafa,
N., Chen, C., Westurlund, T. and Chen, W., 2025. Multidomain selec-
tive feature fusion and stacking-based ensemble framework for EEG-based
neonatal sleep stratification. IEEE Journal of Biomedical and Health Infor-
matics.
X Irfan, M., Jawad, H., Felix, B.B., Abbasi, S.F., Nawaz, A., Akbarzadeh,
S., Awais, M., Chen, L., Westerlund, T. and Chen, W., 2021. Non-wearable
IoT-based smart ambient behaviour observation system. IEEE Sensors Jour-
nal, 21(18), pp.20857-20869.
XI Nawaz, A., Irfan, M. and Westerlund, T., 2023, September. Optical char-
acter recognition using optimised convolutional networks. In 2023 Eighth
International Conference on Fog and Mobile Edge Computing (FMEC) (pp.
107-114). IEEE.
XII Sharma, N., Hwang, Y., Irfan, M., Subasi, A. (2025). Lung cancer de-
tection from histopathological imaging using knowledge graphs and graph
neural networks. In: Generative AI, Large Language Models, Graph Neural
Networks and Knowledge Graphs: Applications. Book chapter, accepted
for publication (forthcoming).
xiii
1 Introduction
Artificial intelligence (AI) has made significant improvements in neurological and
behavioural health, including cancer triage in neuroimaging, assessing sleep in new-
borns, and monitoring affective states. Yet, environments where prompt decisions
are crucial, such as busy clinics, neonatal intensive care units (NICUs), and home
settings, function with severely limited resources: limited on-site computing power,
restricted battery life, and narrow or unreliable network connections. In such scenar-
ios, it’s challenging to utilise today’s large-scale multimodal models consistently and
effectively. Modern deep neural networks can achieve good accuracy but are often
too large for resource-constrained clinical and wearable settings. This challenge led
to the development of parameter-efficient Convolutional Neural Networks (CNNs)
families (including SqueezeNet, MobileNetV2, and EfficientNet) that maintain accu-
racy while using significantly fewer parameters [1; 2; 3]; however, their deployment
remains very limited. In neuro-oncology, deep learning has become a common ap-
proach for Magnetic Resonance Imaging (MRI) classification and segmentation. Re-
cent reviews summarise both the promise and the deployment limitations, highlight-
ing the necessity for lightweight, clinically integrable models. [4; 5; 6]. This thesis
explores computationally efficient methodologies that maintain diagnostic effective-
ness while minimising parameters, possibilities that can reduce energy consumption,
and transmission requirements, thereby enhancing accessibility to AI-driven health-
care.
1.1 Brain-Tumour
The human brain serves as the central hub for vital functions such as memory, speech,
cognition, and motor skills [7; 8; 9]. Any irregular or uncontrolled growth of brain
tissue can disrupt these functions, leading to the development of brain tumours [10].
Because these tumours vary widely in type, behaviour, and severity, they present
substantial challenges for both patients and healthcare providers.
1.1.1 Brain Tumour Assessment Modalities
The assessment and characterisation of brain tumours typically involve a combina-
tion of complementary imaging modalities, primarily MRI, computed tomography
1
Dr. Muhammad Irfan
(CT), and positron emission tomography (PET), in conjunction with advanced MRI
techniques such as diffusion, perfusion, and spectroscopy, and, when accessible, hy-
brid (PET and MRI). Fig 1 shows images of PET, CT, and MRI, respectively. MRI
provides detailed images of the brain and other tissues; CT scans quickly identify
and show bleeding, calcification, and bone; and PET scans provide information on
metabolism and function. These methods work together to aid in diagnosis, plan
surgery and radiation therapy, monitor treatment effectiveness, and track patients
over time.
MRI: It is the primary modality in neuro-oncology due to its exceptional soft-
tissue contrast and comprehensive multiparametric readouts. Standard T1-, T2-, and
FLAIR sequences (with or without gadolinium) demonstrate differences between
enhancing tumours, non-enhancing infiltrative components, vasogenic oedema, and
mass effect. Diffusion-weighted imaging (DWI) with apparent diffusion coefficient
(ADC) maps approximates cellularity; dynamic susceptibility contrast and dynamic
contrast-enhanced perfusion, and arterial spin labelling where available, quantify
haemodynamics, such as relative cerebral blood volume, permeability that corre-
late with grade; and proton MR spectroscopy yields a metabolic profile (elevated
choline, reduced N-acetylaspartate, and lipid-lactate peaks) that helps distinguish tu-
mour from treatment effect or radiation necrosis. MRI accomplishes this without ion-
ising radiation, facilitating secure serial follow-up along the disease trajectory [11].
CT: It remains important when rapid imaging is required or when MRI is con-
traindicated. CT is particularly useful for detecting acute intracranial haemorrhage,
visualising calcifications suggestive of oligodendroglioma, assessing osseous involve-
ment, and quantifying mass effect in emergency settings. Although CT lacks MRI’s
soft-tissue contrast, its rapid acquisition and wide availability make it the typical
first-line study in urgent presentations and a useful adjunct for specific tumour fea-
tures [11].
PET: It contributes functional and metabolic context beyond morphology. Amino-
acid tracers such as 18F-FET, 18F-FDOPA, and 11C-methionine exploit upregulated
amino-acid transport in gliomas, yielding high tumour-to-background contrast that
supports precise delineation, biopsy targeting, radiotherapy planning, and differenti-
ation of recurrence from treatment-related change. By contrast, 18F-FDG is less spe-
cific for primary brain tumours because of high cortical glucose uptake. In combina-
tion with MRI, amino-acid PET refines diagnosis and guides management through-
out the tumour lifecycle [12].
CT and PET offer additional structural and metabolic data, respectiveprovide ad-
ditional structural and metabolic data, respectively; however, MRI remains treatment
2
Introduction
planning, and longitudinal monitoring of brain tumours, ensuring a balance between
diagnostic accuracy and safety for repeated imaging.
(a) (b) (c)
Figure 1. Representative neuro-imaging modalities for brain-tumour assessment: (a) PET; (b) CT;
(c) MRI [13]
1.1.2 Architecture Efficiency in Brain-Tumour MRI
Deep learning methods have been extensively explored for brain tumour detection
using MRI data in recent years [14; 15; 16]. Several studies have applied various
deep learning models, achieving different levels of accuracy and offering diverse
findings in brain tumour diagnosis [17; 18]. Most studies have utilised CNNs to im-
prove the precision of brain tumour classification, ranging from binary to multistage
classification [19; 16]. Furthermore, recent studies demonstrate the various meth-
ods and models available for brain tumour classification and segmentation. How-
ever, while high-performing models, such as Resnet50 and EfficientNet, achieve
high precision [19; 20; 21; 22], these performance gains often come at the cost of
substantially increased model complexity. Large transfer-learning architectures such
as ResNet50 and EfficientNetB0 contain millions of trainable parameters, which in-
crease memory use, computational cost, and inference burden. To address this issue,
we propose two approaches, FAC and FSC, along with two modules: top of the fun-
nel (TOFU) and middle of the funnel (MOFU), designed to provide highly accurate
predictions with significantly fewer parameters. Both approaches have their advan-
tages and limitations in terms of performance and computational requirements.
In Publication I, a Fuzzy Atrous Convolution (FAC) architecture with two addi-
tional modules, TOFU and MOFU, has been proposed for binary brain-tumour clas-
sification from MRI with the aim of developing a computationally efficient model
with fewer trainable parameters while maintaining high classification accuracy. A
novel convolutional operator is introduced in the FAC model to enlarge the receptive
field while preserving input information, thereby enabling efficient feature-map re-
duction and accurate tumour detection. Compared with large transfer-learning mod-
3
Dr. Muhammad Irfan
els, the proposed method uses approximately 300 times fewer trainable parameters,
making it particularly suitable for efficient early classification of brain tumors.
In Publication II, a fuzzy sigmoid convolution (FSC) layer is introduced to de-
velop a lightweight and computationally efficient model that reduces the number of
trainable parameters without compromising classification accuracy. The proposed
methodology significantly reduces the number of trainable parameters without com-
promising classification accuracy. A novel convolutional operator is central to this
approach, as it employs dilated convolutions to enlarge the effective receptive field.
Meanwhile, fuzzy sigmoid-based membership functions adaptively reweight high-
and low-activation responses, rather than discarding them. This design enables com-
putationally efficient feature representation and enhances the model’s tumour detec-
tion capability. In the FSC-based model, fuzzy sigmoid activations are directly inte-
grated into the convolutional layers to enhance feature extraction and classification.
The incorporation of fuzzy logic enhances the network’s adaptability and robust-
ness to input variability. Extensive experiments on three benchmark brain tumour
MRI datasets demonstrate the performance and efficiency of the proposed model.
The model employs approximately 100 times fewer trainable parameters than large-
scale transfer-learning architectures, highlighting its computational efficiency and
suitability for early detection of brain tumours. This research presents lightweight,
high-performance deep learning models for brain tumour stratification.
1.2 Wearable technology
Simple and unobtrusive wearables, particularly watches and wristbands, now enable
continuous monitoring with minimal inconvenience and are now mainstream, with
534.6M units shipped in 2024, and older adults are already using them (38 % of U.S.
50+ own a wearable; 71 % use health/wellness apps), making them a practical foun-
dation for emotion-aware care, as reported by IDC [23; 24]. Fig. 2 from (Publication
V) illustrates the general wearable device system, which includes the collection of
physiological signals from body sensors, cloud-based data transmission, and alerts
sent to family, emergency services, and healthcare providers.
1.2.1 Neonatal Sleep Analysis
In 2020, the situation became increasingly dire, with newborn mortality escalating to
a concerning high of 2.4 million deaths during the first month. This corresponds to
over 6,700 daily fatalities, accounting for 47 % of all child deaths under five, a signif-
icant rise from 40 % in 1990 [25; 26]. While these figures are not specifically about
sleep disorders, they emphasise the necessity of enhanced monitoring for newborns
and the need for early intervention.
4
Introduction
Figure 2. General diagram for wearable devices.
Sleep Assessment Modalities
A range of physiological monitoring modalities are used to help healthcare profes-
sionals and researchers assess sleep patterns and cardiac function. These modali-
ties are particularly important for evaluating sleep-related disorders and cardiovas-
cular health in diverse populations, from neonates to older adults. Key methods
include electroencephalography (EEG), electrocardiography (ECG), electrooculog-
raphy (EOG), electromyography (EMG), photoplethysmography (PPG), as well as
audio and video recordings. Each modality provides distinct insights into underlying
physiological processes, supporting the diagnosis and management of various health
conditions.
ECG: ECG is the gold standard for monitoring the heart’s electrical activity, espe-
cially useful for assessing heart rate variability (HRV), arrhythmias, and sleep-related
cardiovascular issues, such as sleep apnea. It directly measures the heart’s electrical
conduction, providing high accuracy for detecting even the most subtle arrhythmias.
It also provides real-time insights into the autonomic regulation of the heart during
sleep, which is crucial for understanding heart health.
EEG: EEG is a core modality for assessing brain function during sleep, offering
real-time, high-temporal-resolution measurements of neural electrical activity. By
capturing subtle and complex brainwave dynamics, EEG enables the identification
of sleep disorders and functional abnormalities that are often difficult to detect using
5
Dr. Muhammad Irfan
other assessment techniques.
Importance of EEG for Sleep Analysis: Although auxiliary signals such as EMG
and EOG provide complementary information, EEG remains the primary modality
for identifying and classifying sleep stages because it directly measures neural ac-
tivity. Many sleep disorders, including sleep apnea, insomnia, and parasomnias, are
associated with alterations in brain dynamics that are most reliably captured through
EEG recordings. EEG is particularly important in neonatal sleep research, where
developmental changes in brainwave patterns are closely linked to neurodevelop-
mental outcomes. In this thesis, a reduced number of EEG channels is employed
to improve efficiency and practicality. Multimodal approaches that combine EEG
with additional physiological sensors such as EMG or EOG are intentionally ex-
cluded. The goal is to simplify data acquisition while preserving diagnostic accuracy
and robustness. Reducing channel count enables less intrusive monitoring, lowers
computational and hardware requirements, and remains effective through the use of
optimised machine learning models.
EOG: EOG assesses eye movements and is essential for detecting rapid eye move-
ment (REM) sleep, characterised by distinctive eye activity. The EOG records elec-
trical potentials produced by eye movements, facilitating the distinction between
sleep stages, especially REM and non-REM phases. Nonetheless, EOG does not
directly measure cerebral activity, thereby offering less comprehensive insights into
the underlying neurological processes compared to EEG. Consequently, EOG is fre-
quently utilised in conjunction with EEG to facilitate a more thorough examination
of sleep.
EMG: EMG measures the activity of muscles and is valuable for evaluating mus-
cle tone and movement during sleep. It is crucial for diagnosing motion-related sleep
disorders, including REM sleep behaviour disorder and periodic limb movement dis-
order, which are marked by irregular or excessive muscle activation. In sleep re-
search, EMG is generally utilised as an additional signal to EEG rather than as the
principal modality for sleep staging.
PPG: PPG quantifies variations in blood volume in peripheral microvascular tissue
by optical sensing, facilitating non-invasive assessment of heart rate and oxygen sat-
uration during sleep. While PPG is easy and extensively utilised in wearable devices,
it offers inferior temporal precision and signal integrity compared to electrocardio-
graphy (ECG) for comprehensive cardiovascular evaluation. As a result, PPG is
predominantly utilised for general wellness monitoring rather than for clinical-grade
heart assessment in sleep research.
6
Introduction
Audio and Video Data: Audio and video recordings have been utilised in sleep re-
search to record snoring, bodily movements, and other observable behaviours linked
to sleep problems. Nonetheless, video-based monitoring is computationally inten-
sive and often yields imprecise results, such as binary classifications of sleep versus
wakefulness, with limited physiological interpretability. In contrast to biosignals like
EEG or ECG, audio-visual data provide diminished granularity for multi-stage sleep
monitoring. Furthermore, ongoing recording presents significant privacy and ethi-
cal concerns, especially in neonatal environments. This study excludes audio- and
video-based methodologies for these reasons. The emphasis is on improving EEG
analysis to attain precise, multi-stage sleep classification while minimising compu-
tational complexity.
1.2.2 EEG
EEG is a non-invasive electrophysiological method commonly utilised to capture
brain activity from the cerebral cortex. Due to its improved temporal resolution,
EEG is exceptionally useful in studying fast brain dynamics and transitory neuro-
logical processes in real time. The central nervous system principally consists of
neurones and glial cells [27]. Neurones provide information transfer via electrical
and chemical signalling, whereas glial cells offer crucial structural, metabolic, and
functional support. Despite not generating action potentials, glial cells are essential
for preserving synaptic integrity, controlling the extracellular milieu, and facilitat-
ing neuronal signaling [28]. Neural communication transpires at synapses, where
the arrival of an action potential at the presynaptic terminal triggers the release of
neurotransmitters into the synaptic cleft. These neurotransmitters bind to specific
receptors on the postsynaptic membrane, leading to the opening and closing of ion
channels that generate postsynaptic potentials. EEG signals are obtained by placing
electrodes on the scalp to record aggregate electrical activity primarily generated by
synchronised postsynaptic potentials from extensive neuronal populations [29].
EEG recordings do not capture single neuronal spikes; instead, they reflect the
aggregate of excitatory and inhibitory currents across broad cortical areas. The am-
plitude of the recorded signal is significantly influenced by the level of temporal syn-
chronisation among neurones: coherent activity increases signal amplitude, while
asynchronous activity leads to signal attenuation. Factors affecting EEG voltage
encompass the spatial distribution, orientation, and intensity of underlying current
sources. Cortical pyramidal neurones are the primary contributors to EEG signals
owing to their prevalence, parallel orientation, and extensive dendritic architecture.
The structured arrangement of these neurones promotes the aggregation of ionic cur-
rents, producing extracellular electric fields that can be detected at the scalp surface.
Thus, the activity of pyramidal cells has a significant influence on the EEG signal
recorded during sleep and various cognitive states.
7
Dr. Muhammad Irfan
EEG Recording Clinical EEG systems comprise numerous critical components,
including scalp electrodes (which are frequently silver/silver chloride), lead cables,
signal amplifiers, and a computer unit responsible for system control, data storage,
and visualisation. The EEG acquisition pipeline is depicted in a simplified man-
ner in Fig. 3. Electrodes serve as the interface between neural electrical activity
and the recording apparatus, measuring voltage differentials at specified scalp lo-
cations. Electrodes are often classified into two types: invasive needle electrodes,
which pierce the subdermal tissue, and non-invasive disc electrodes, which are placed
on the scalp surface.
In neonatal clinical practice, sleep monitoring typically commences shortly af-
ter birth, especially for newborns considered at increased risk of neurological or
sleep-related problems. Ongoing or recurrent observation throughout this first devel-
opmental phase provides significant insight into cerebral maturation and functional
organisation during sleep. Examining newborn sleep architecture is essential for de-
tecting unusual patterns that may indicate underlying neurological or developmental
disorders [30]. Electrode placement is generally conducted by qualified clinical per-
sonnel following established protocols [31].
Neonatal EEG recordings often utilise the worldwide 10–20 method for electrode
placement, which is based on defined cranial landmarks to guarantee reproducibility
and anatomical significance. Based on clinical aims and practical limitations, either
the complete montage or a smaller selection of electrodes may be utilized [32]. In
reduced-channel setups, electrodes are commonly situated over the biparietal areas
(P3 and P4). Recent methodologies have extended limited montages to include cen-
tral (C3, C4) and frontal (F3, F4) sites, thus enhancing sensitivity to sleep-associated
cortical activity while ensuring practicality in newborn environments [33].
In standard clinical EEG recordings, a referential montage is often employed.
In this setup, one electrode serves as the reference point, and the EEG system mea-
sures the voltage differences between this reference and all other electrodes [34]. For
the review and analysis of neonatal EEG data, a bipolar montage is frequently used,
where the voltage difference is calculated between pairs of electrodes, often adja-
cent ones. This method is particularly advantageous for artefact reduction, providing
clearer and more accurate data for clinical interpretation [34].
EEG as a Neonatal Sleep Modality
In 1924, Hans Berger reported the first human electroencephalography (EEG) record-
ings, thereby establishing non-invasive testing of brain function. Loomis and col-
leagues introduced EEG-based sleep pattern analysis in 1937, marking one of the
first systematic applications of EEG in human sleep research [35; 36; 35].
In modern NICUs, EEG is a common technique for neurological assessment,
evaluation of brain maturation, and detection of abnormalities [37]. The standard re-
8
Introduction
Figure 3. A) Schematic overview of the EEG acquisition system. B) Electrode layout based on the
neonatal-adapted international 10–20 system, with the five channels identified as most informative
in this thesis highlighted in red and green.
mains the clinical reference for monitoring brain function, detecting sleep disorders,
and diagnosing newborn encephalopathy [38; 32; 39; 40; 41; 32; 42]. Depending on
the clinical question, EEG recordings may be brief “snapshot” studies lasting several
minutes to a few hours or prolonged recordings lasting many hours to days, often as
part of a polysomnography (PSG) protocol for detailed analysis of sleep architecture.
PSG: PSG is the standard for assessing sleep disorders and behaviour since its
introduction in the 1960s ([43]). PSG uses a polygraph, a piece of equipment that
captures numerous physiological signals nightly in a sleep lab. The EEG, ECG, pulse
rate, and EMG signals obtained during PSG are shown in Fig. 4.
Sleep Disorders Assessed with PSG: PSG plays a central role in diagnosing sev-
eral major sleep disorders:
• Sleep Apnea: Recurrent upper-airway obstruction causes apneas, hypopneas,
and fragmented sleep. PSG detects these events, snoring, and associated changes
in oxygen saturation, heart rate, and respiratory effort, enabling severity grad-
ing.
• Insomnia: Defined by difficulty initiating or maintaining sleep, PSG helps
identify prolonged sleep latency, reduced sleep efficiency, and altered REM
architecture, supporting the diagnosis and management of chronic insomnia.
9
Dr. Muhammad Irfan
Figure 4. An Overview of EEG, ECG, and EMG Signals Collected via PSG Using a Multichannel
Approach
• Periodic Leg Movements Disorder: Involves stereotyped, repetitive leg move-
ments during sleep. PSG quantifies these movements and their associated
arousals, clarifying their impact on sleep continuity and quality.
1.2.3 Common Neonatal EEG Artefacts
Neonatal EEG recordings are highly susceptible to artefacts from both physiologi-
cal sources, such as movement, ECG, muscle activity, sweat and non-physiological
sources such as electrode problems, electrical interference, which can obscure true
cerebral activity and complicate interpretation [44; 45; 46; 47; 48; 49]. Recognis-
ing these artefacts and understanding their typical appearance is essential for reliable
clinical use and for the development of automated analysis methods.
Movement artefacts are typically caused by limb or head motion and can produce
high-amplitude transients that overshadow the underlying EEG signal. ECG artefacts
appear as regular, periodic deflections that are particularly prominent when brain
activity is suppressed. Muscle activity generates small-amplitude, high-frequency,
stochastic noise superimposed on the EEG. Electrical interference from equipment
and power lines produces spurious spikes or rhythmic oscillations. Electrode-related
artefacts arise from poor contact or electrode movement and manifest as abrupt, of-
ten high-frequency disturbances, while sweat artefacts alter scalp conductivity and
frequently introduce slow-wave components [44; 45; 46; 47; 48; 49]. Mitigation
10
Introduction
strategies include careful infant positioning, maintaining a calm environment, proper
skin preparation and electrode placement, regular impedance checks, appropriate fil-
tering, and minimising environmental electrical noise.
Artefacts present a major obstacle for automated background classification in
neonatal EEG, especially in infants with pathological backgrounds, because they can
both mimic and mask clinically relevant patterns [45; 46; 47; 49; 50; 51]. Most ex-
isting artefact-handling methods are designed for specific artefact types and do not
generalise well to the heterogeneous and dynamic artefacts encountered in long-term
neonatal monitoring, which can degrade system performance [52; 53]. A recent study
by Webb et al. [49] proposed a deep-learning-based approach to classify EEG seg-
ments as artefact-free or artefact-contaminated, highlighting a promising direction
toward more robust, artefact-aware neonatal EEG analysis.
1.2.4 Validation and Comparison to the Gold Standard
A critical step in developing any automated tool for clinical practice involves thor-
ough validation and comparison to the gold standard. In the context of neonatal
sleep monitoring, the gold standard typically refers to expert annotations or scores
derived from the visual interpretation of EEG data, as discussed in Section 1.2.5. Vi-
sual interpretation represents the upper limit of classifier performance, as it captures
the inherent variability in human annotations, setting the benchmark against which
automated classifiers are evaluated.
1.2.5 Visual EEG Interpretation and Inter-Rater Agreement
Visual interpretation of neonatal EEG by expert clinicians remains the clinical gold
standard but is inherently subjective, as complex and imperfect scoring systems do
not fully account for medication effects, pathological conditions, or artefacts, lead-
ing to substantial intra- and inter-rater variability [54; 55; 56]. To quantify consis-
tency between reviewers, inter-rater agreement is commonly assessed using Cohen’s
Kappa coefficient (𝜅), which measures agreement beyond chance [57]:
𝜅 =
𝑃𝑜 − 𝑃𝑒
1− 𝑃𝑒 , (1)
where 𝑃𝑜 is the observed agreement and 𝑃𝑒 is the expected agreement by chance,
given by
𝑃𝑒 =
1
𝑁2
∑︁
𝑘
𝑛𝑘1𝑛𝑘2. (2)
Kappa values range from−1 (less agreement than expected by chance) to+1 (perfect
agreement), but are sensitive to class prevalence and sample composition, which can
complicate their interpretation in EEG studies [58].
11
Dr. Muhammad Irfan
Clinically Significant EEG Characteristics
EEG Pattern Analysis: The EEG activity in newborns is characterised by sev-
eral essential features, including amplitude, frequency, spatial distribution, and tem-
poral evolution [59]. In preterm newborns, the EEG predominantly exhibits high-
amplitude delta activity (below 4 Hz) alongside some theta activity (4–8 Hz), whereas
term infants demonstrate heightened theta and alpha activity (8–12 Hz) as brain mat-
uration occurs [60]. EEG patterns evolve with development, and characteristics such
as continuity, reactivity, symmetry, synchronisation, and sleep-wake cycling (SWC)
are critical indications of brain maturation. EEG continuity signifies a consistent
amplitude, indicating more sophisticated neurological maturation, whereas disconti-
nuity characterized by fluctuating high and low-amplitude bursts, generally suggests
an immature, developing brain. The reactivity of the EEG is manifested as signal
modulation in reaction to external stimuli, providing insights into the brain’s func-
tioning status [61].
The emergence of SWC, which becomes distinctly recognisable at 30 weeks
postmenstrual age (PMA), is a vital indicator of cerebral maturation, although pre-
liminary, less defined cycling may be noted as early as 24–25 weeks. The primary
objective of this dissertation is to differentiate between various sleep states, active
sleep I (N1), active sleep II (REM), quiet sleep I (NII), quiet sleep II (NIII), and wake-
fulness (WS), utilising just a few of EEG channels for sleep analysis.
EEG Relation with Sleep States: Newborns spend a large fraction of their early
life asleep, with healthy full-term infants typically sleeping 16 to 18 hours per day
[62]. This prolonged sleep is far from a passive resting period; it is accompanied
by intense neurological and physiological activity [63; 64]. Numerous studies em-
phasize that neonatal sleep plays a fundamental role in the maturation of cognitive,
psychomotor, and behavioural capacities [65; 66; 67]. Accordingly, systematic eval-
uation of sleep in newborns is vital for assessing brain functional integrity and for
characterizing sleep patterns, particularly in infants at increased risk for conditions
such as sudden infant death syndrome (SIDS) and sleep apnea [68; 69].
EEG activity in neonates differs markedly from that observed in adults. In par-
ticular, the amplitude of neonatal EEG signals is generally much lower. Over the first
year of life (1–12 months), the EEG background undergoes notable developmental
changes. From around two months of age, the posterior dominant rhythm, consid-
ered an early form of the alpha rhythm, begins to appear. Across development, the
proportion of electroencephalographic power in the alpha range increases from es-
sentially 0 % at birth to approximately 70 % in adulthood. The dominant frequency is
initially around 3– Hz, increases to 4–5 Hz by six months, and reaches about 8.5 Hz
by roughly 16 years of age. Asynchronous sleep spindles usually become evident be-
tween 2 and 3 months of age and persist, with ongoing maturation, between 6 and 12
12
Introduction
months. These spindles typically last 10–15 seconds, and the period of asynchrony
between them ranges from 1 to 5 seconds. K-complexes and V-waves emerge be-
tween 2 and 5 months, followed by additional maturational refinements throughout
the first three years of life [70].
Bio-Insights of Neonatal Sleep EEG: The neonatal sleep EEG patterns, as clas-
sified by the American Academy of Sleep Medicine (AASM) Manual, provide key
insights into different sleep stages in newborns, including 𝑄𝑆and Wake/Active Sleep
(AS). The key EEG patterns associated with each stage are outlined below:
Quiet Sleep:
• Trace Alternant (TA): This pattern is characterised by bursts of high-voltage
delta activity (50 𝜇V) alternating with periods of lower amplitude (25–50𝜇V)
theta activity (4–7 Hz). The intervals between bursts, known as inter-burst
intervals (IBIs), increase with age, signalling developmental changes in the
brain.
• High Voltage Slow (HVS): This pattern involves continuous high-voltage
delta activity (100–150 𝜇V at 1–3 Hz) that often shows a central or occipital
predominance. During this stage, there is significant development of the au-
tonomous nervous system (ANS), which is crucial for physiological regulation
during Quiet Sleep.
Wake/Rarely AS:
• Mixed (M): A pattern in which the background is low-voltage and mixed-
frequency but contains intermittent higher-amplitude slow activity. Practically,
the baseline may look LVI-like (often< 50𝜇V much of the time), while super-
imposed delta components can intermittently exceed 100𝜇V (such as 2–4Hz),
especially around arousals/transitions or in some active-sleep epochs.
Wake/AS:
• Low Voltage Irregular (LVI): This stage involves continuous low-voltage,
mixed-frequency activity dominated by delta and theta waves. The irregular
theta activity (25–50 𝜇V, 4–7 Hz) is combined with low-amplitude delta waves
(1–3 Hz). This stage reflects the ongoing development of the central nervous
system (CNS) during AS, highlighting its importance in brain maturation.
Each EEG pattern provides a valuable window into the neurological development
of neonates. Monitoring stages such as 𝑄𝑆 and 𝐴𝑆 is essential, as these stages are
linked to the maturation of critical systems, including the ANS and CNS. Under-
standing these EEG patterns provides key insights into the functional and develop-
mental status of newborns.
13
Dr. Muhammad Irfan
Clinical Significance and Interpretability: Neonatal sleep stages are often cate-
gorised into AS, QS, and WS. AS and QS are further categorised into N1, REM-II,
NII, and NIII. In newborn growth and development, these stages serve as a means
to evaluate brain maturity, diagnose health conditions, and strategise interventions.
Transitional sleep, characterised by a combination of AS and QS, typically arises dur-
ing state transitions and diminishes in frequency with a child’s development. WS is
characterised by wide eyes, irregular respiration, and higher muscle tone. It facili-
tates sensory exploration, attachment, and nourishment, which are essential for early
cognitive and physical growth.
AS, specifically REM-II, is characterised by irregular respiration, rapid eye move-
ments, and diminished muscular tone. This phase is essential for brain development,
the integration of sensory and motor functions, and the formation of neural connec-
tions that support cognitive and motor advancement. Clinically, disturbances in AS
may signify developmental delays or neurological disorders, emphasising its signifi-
cance in the surveillance of neonatal health [71].
QS, commonly referred to as non-rapid eye movement (NREM) sleep, is iden-
tifiable by rhythmic respiration, a consistent heart rate, and high-voltage slow elec-
troencephalography (EEG) patterns, including alternating traces. This phase facili-
tates restorative processes, energy preservation, and immunological function, acting
as an indicator of the brain’s and thalamocortical connections. Irregularities in QS,
including disrupted sleep or erratic respiration, may correlate with hypoxia, cerebral
damage, or other medical conditions [71; 72].
Transitional (T) sleep transpires during the transitions between AS and QS, ex-
hibiting attributes of both states. It manifests in the early stages of newborn life and
progressively diminishes with maturation. Monitoring T sleep reflects the dynamic
characteristics of newborn sleep and its continuous maturation [71; 72].
In clinical practice, newborn sleep phases can serve as diagnostic indicators of
neurological and systemic health. Abnormalities in standard sleep patterns, including
excessive AS or interrupted QS, may indicate developmental disorders, cerebral in-
juries, or systemic malfunctions. Monitoring sleep stages helps healthcare providers
evaluate brain maturation, guide therapeutic actions, and enhance neurodevelopmen-
tal outcomes, particularly for neonates in critical care [71; 72].
Motivation and Challenges: The analysis of neonatal EEG signals presents unique
challenges due to these recordings' high variability and noise. Neonatal sleep differs
significantly from adult sleep, not only in its stages but also in its physiological
characteristics. In contrast, adult sleep stages are typically divided into REM and
non-REM (NREM), with NREM further classified into stages NREM1, NREM2,
and NREM3 (recently simplified from NREM3 and NREM4) [73], neonatal sleep
is categorised primarily into AS (𝐴𝑆) and quiet sleep (𝑄𝑆). A transitional state,
referred to as indeterminate sleep (𝐼𝑆), occurs between these stages, further compli-
14
Introduction
cating the classification process [73]. These complexities in accurately classifying
neonatal sleep stages due to signal variability and noise serve as key motivations for
this research, driving the need for improved analysis techniques.
Key differences between neonatal and adult sleep include:
• Sleep Duration: Neonates typically sleep for 17-18 hours a day, accounting
for 70 % of their daily activity, compared to adults who sleep for 7-9 hours,
or about 16-29 % of their day. Unlike adults, neonates have no established
circadian rhythms to regulate their sleep patterns.
• Sleep Spindles: Sleep spindles, which play a crucial role in memory consoli-
dation in adults, are absent in neonates and only begin to appear after 42 weeks
of gestational age (GA) [74].
• Sleep Cycle Duration: The sleep cycle duration in neonates is approximately
50-60 minutes, significantly shorter than the 80-110 minutes typical of adult
sleep cycles.
• REM and NREM Sleep Distribution: At 28 weeks GA, 𝐴𝑆 constitutes 80 %
of the total sleep time (TST), 𝑄𝑆 18 %, and indeterminate sleep (𝐼𝑆) 2 %. By
40 weeks GA, these distributions shift to 57 % 𝐴𝑆 , 32 % 𝑄𝑆 , and 11 % 𝐼𝑆 .
• Brain Development and EEG Artefacts: Neonates are in a rapid and dy-
namic brain development phase, making it difficult to establish consistent clas-
sification features. Furthermore, EEG data in neonates are often contaminated
by artefacts from muscle movements, eye blinks, and equipment noise, which
complicates the analysis and can lead to classification errors [75].
• Standardisation of EEG Recording and Interpretation: The lack of stan-
dardised protocols for recording and interpreting neonatal EEG signals con-
tributes to variability in data quality and hinders the development of reliable
classification models [76; 57]. Visual interpretation of EEG signals remains
the gold standard in clinical practice. However, this process is inherently com-
plex and subject to variability in expertise, experience, and fatigue, which can
introduce subjectivity and affect the accuracy of interpretations. Addition-
ally, the need for round-the-clock availability of EEG experts at every NICU
presents a significant logistical challenge.
Current methods for staging neonatal sleep rely heavily on multiple EEG chan-
nels. Neonatal PSG often uses multiple EEG derivations (and other channels) be-
cause infant EEG is discontinuous, and staging relies on combining visual/phys-
iological cues. More channels mean more wires/adhesives (practical burden) and
more data to transmit/store. Skin irritation with prolonged EEG is a documented
15
Dr. Muhammad Irfan
risk, making fewer-lead solutions attractive when feasible [71; 9]. Furthermore, ex-
isting approaches to neonatal sleep stage classification are hindered by their inherent
subjectivity, time-consuming nature, and inter- and intra-rater variability. Analysing
neonatal EEG signals presents unique challenges compared to those of adults. A
combination of advanced signal processing with machine learning and the Internet
of Medical Things (IoMT) offers promising solutions in various domains [30; 74],
particularly revolutionising healthcare [77; 78; 79]. The application of IoMT for
continuous, non-invasive monitoring of physiological signals, when coupled with
edge-cloud computing [80; 81], offers a novel approach to the early detection and
analysis of health parameters. This technological advancement aligns with the crit-
ical need for improved neonatal healthcare strategies [82]. However, an IoT-based
system requires low computational demand, data efficiency, energy efficiency, and
streamlined data acquisition.
To address these requirements, in publication III, we propose an innovative au-
tomated classification approach that integrates multi-view feature fusion, AdaptiS-
elect-based feature optimisation, the smart cloud data transfer and reconstruction
(STREAM) module, and a rotational ensemble stacking model. The data reduc-
tion module significantly enhances edge-cloud systems' performance in IoT-based
healthcare environments by reducing data transmission by a factor of 153.6 through
efficient feature selection and compact data packet formation. This module ensures
minimal bandwidth usage, reduces the computational load on resource-constrained
edge devices, and lowers cloud storage requirements while maintaining full data re-
construction. The dataset used in this research combines two large Datasets collected
over a four-year period from the Children’s Hospital, Fudan University, Shanghai. A
unique set of 315 features was extracted from each epoch of a single channel using
flexible analytical wavelet transform (FAWT), dual-tree complex wavelet transform
(DTCWT), enhanced covariance (ECOV), and spectral features based on 𝛼, 𝛽, 𝜃,
and 𝛿 brain waves. These features were refined using AdaptiSelect, which enabled
achieving higher performance than previous published work through 10-fold cross-
validation. Additionally, Leave-One-Subject-Out cross-validation (LOSO-CV) fur-
ther demonstrates the effectiveness of the proposed approach as a generalised solu-
tion. Using both single and multichannel setups, the proposed approach outperforms
the most significant state-of-the-art methods in neonatal sleep analysis.
1.2.6 Emotion Analysis: Minimal Sensing
Emotion is defined as a psychological reaction generated by both external stimuli
and internal cognitive processes, which is accompanied by a variety of physiological
changes in the human body. Also, mood is generally characterised as a dominant,
conscious emotional state at a specific moment in time [83]. While there are many
approaches to classifying emotions, one widely accepted and useful framework in
16
Introduction
psychology is the two-dimensional model of emotions. Despite its simplicity, this
model offers a valuable tool for analysing emotional states and understanding their
functional roles in everyday life. The relationship between basic emotions and the
valence–arousal dimensions is illustrated in Fig. 5 . Emotions strongly influence
both the physiological and psychological aspects of human life. Positive emotions
promote improved health and productivity, but negative emotions can lead to a vari-
ety of health issues, especially if prolonged [84].
Arousal (High)
Arousal (Low)
Valence (Negative) Valence (Positive)Neutral
I
High-Arousal
Positive-Valence
II
High-Arousal
Negative-Valence
III
Low-Arousal
Negative-Valence
IV
Low-Arousal
Positive-Valence
Frustrated
Tense
Angry
Bored
Tired
Depressed
Delighted
Excited
Happy
Content
Relaxed
Calm
Figure 5. 2D Valence and arousal model together with several basic emotions. (Publication V)
To assess human emotion with minimal sensing, prioritisation is given to phys-
iological signals that are information-rich, low power, and easy to acquire in am-
bulatory settings, most commonly single–lead ECG for cardiac dynamics and elec-
trodermal activity (EDA; often called galvanic skin response, GSR) for sympathetic
arousal. Recent review studies [85; 86; 87; 88], have highlighted the importance of
physiological signals for emotion recognition, pointing out their reliability in com-
parison to speech- or facial-based techniques. For this reason, many studies have
focused on physiological signals, particularly multimodal approaches that combine
various physiological data obtained from biosensors, such as ECG, EEG, EMG,
EDA, or GSR, PPG, blood volume pressure, and respiratory inductive PSG. While
multimodal recognition of emotions often produces better results, the unimodal tech-
nique has the advantages of faster processing and easier data collection [89]. These
modalities are well–established in affective computing and clinical psychophysiol-
17
Dr. Muhammad Irfan
ogy, and have been validated across laboratory and real–world contexts [90; 91; 92;
93; 94; 95; 96; 97; 98; 99; 100].
Importance of Emotions Analysis:
Negative emotions are linked to the onset of neurodegenerative diseases and demen-
tia. In Alzheimer’s disease, prolonged negative patterns (repetitive negative thinking)
correlate with increased amyloid/tau accumulation and accelerated cognitive deteri-
oration, indicating a possibly alterable risk for dementia [101]. Behavioural and psy-
chological symptoms of dementia, including sadness, apathy, agitation, delusions,
and anxiety, impact up to 90 % of patients and frequently precipitate institutionalisa-
tion, highlighting the necessity for continuous monitoring of emotional states [102].
Positive emotions, such as gratitude, amusement, and tenderness, activate reward
pathways and neuromodulators like serotonin and dopamine, thereby enhancing cog-
nitive flexibility and resilience. The broaden-and-build theory posits that this leads
to an expansion of thought–action repertoires and enduring resources [103]. To
complement this perspective, emotion neuroscience studies show that positive affect
increases functional connectivity in prefrontal and default-mode networks, thereby
enhancing cognitive reserve, whereas negative affect reduces prefrontal regulatory
control and increases amygdala responsivity. In contrast, prolonged negative emo-
tions activate amygdala/insula circuits and stress-related physiology; persistently
increased stress hormones hinder cognitive function and heighten susceptibility to
mental health issues [104]. Together, these findings emphasise the clinical relevance
of continuously monitoring emotional states, especially in populations vulnerable to
cognitive decline.
Challenges: Although wearables provide non-invasive, continuous physiological
monitoring, which is essential when self-reporting is inconsistent, especially in cases
of dementia, practical limitations persist. Power constraints and Bluetooth Low
Energy bandwidth restrict high-rate, multi-stream sensing. Simultaneous high fre-
quency streams may result in packet loss and battery depletion, as evidenced by pre-
vious studies and commercial devices [105; 106; 80]. These limitations directly af-
fect the feasibility of real-time affective monitoring, where continuous, uninterrupted
data are required for reliable prediction. Thus, reducing sensing and computing is
essential for reliable, real-time applications and becomes a key factor in designing
solutions that can operate robustly in daily life settings.
Furthermore, in the case of dementia diseases, there are over hundred of types
of dementia, and each person living with dementia has different experiences and
symptom progressions. This variability reinforces the need for personalised solu-
tions, where models learn from a subject’s own physiological patterns rather than
relying solely on population-level generalisations. Personalisation is particularly rel-
18
Introduction
evant in emotion recognition because autonomic physiology varies strongly across
individuals; baseline cardiac variability, stress reactivity, and emotional expressivity
differ widely in dementia populations. In this context, personalised models could be
more optimal and effective, enabling the model to adapt to individual differences and
predict emotional or other health outcomes with higher reliability.
Considering these points, personalised and generalised machine-learning models
were evaluated for binary emotion classification (positive vs. negative) via wear-
able biosignals, aiming for a minimal-sensing architecture that employs a single
physiological channel. Our pipeline prioritises multi-domain feature extraction with
lightweight hybrid feature selection to optimise the feature set while minimising
computational expense. We assess interpretable, resource-efficient classifiers de-
signed for embedded and IoT-based deployment, which is particularly important
for cognitively disabled populations, where unobtrusive and simpler solutions are
crucial. By linking emotional neuroscience, dementia-specific variability, and real-
world wearable constraints, this work addresses both the clinical relevance and the
technical feasibility of emotion recognition under minimal sensing conditions.
In publication IV, we propose a lightweight, resource-efficient machine learn-
ing approach for binary emotion classification, distinguishing between negative (sad-
ness, disgust, anger) and positive (amusement, tenderness, gratitude) affective states
using only ECG signals. The method is designed for deployment in resource con-
strained systems, such as for the IoT devices, by reducing battery consumption and
cloud data transmission through the avoidance of computationally expensive mul-
timodal inputs. Our approach integrates multidomain feature extraction, selective
feature fusion, and a voting classifier. We evaluated it using a participant-exclusive
generalised model and a participant-inclusive personalised model. The personalised
model achieved the best performance, outperforming the generalised model. Com-
parisons on the POPANE dataset demonstrate that our method consistently outper-
forms the standard models commonly used in this domain. This work underscores
the effectiveness of personalised models for emotion recognition and their suitability
for wearable devices that demand accurate, low-power, real-time emotion tracking.
1.3 Aims of the Thesis
The central objective of this thesis is to design and assess computationally efficient
AI techniques for analysing neuroimaging and physiological signals, tailored to clin-
ically relevant scenarios with limited resources. The research centres on three il-
lustrative applications spanning the neuro–cardio–sleep spectrum: brain tumour as-
sessment using MRI, neonatal sleep stage classification from EEG, and emotion de-
tection based on wearable ECG. More specifically, the thesis pursues the following
aims:
19
Dr. Muhammad Irfan
• Aim 1: To design lightweight, lesion-aware deep learning architectures for
brain tumour classification from MRI, balancing diagnostic performance with
computational efficiency and inference-time feasibility for real-world deploy-
ment (Publications I-II).
• Aim 2: To develop multiview feature-fusion and channel-reduction methods
for neonatal EEG sleep staging, utilising spectral and time–frequency repre-
sentations to enable reliable four-stage classification with fewer electrodes,
making the montages more feasible for routine NICU use (Publication III).
• Aim 3: To develop emotion recognition from minimal wearable sensing using
the POPANE dataset, which provides single-lead ECG recorded in a modified
Lead II configuration at 1000 Hz; to build a low-complexity pipeline com-
bining multidomain feature extraction, hybrid feature selection, and ensem-
ble learning; and to compare subject-independent (generalised) and subject-
specific (personalised) models for the real-time emotion-aware applications
(Publication IV).
• Aim 4: Highlight the trade-offs between performance and efficiency, and prac-
tical considerations for edge–cloud deployment.
1.4 Structure of the Thesis
This thesis is organised as a compilation of five peer-reviewed publications, framed
by introductory and synthesis chapters.
• Chapter 2 describes the datasets used throughout the thesis. It presents the
publicly available MRI datasets for brain tumour analysis (Publication I and
II), the neonatal EEG dataset (Publication III) for four-stage sleep classifi-
cation from Children’s Hospital, Fudan University, Shanghai, and the ECG-
based emotion dataset (POPANE dataset used in Publication IV). Data col-
lection settings, key signal characteristics, and relevant preprocessing consid-
erations are summarised.
• Chapter 3 outlines a generic AI workflow for healthcare applications. It re-
views typical pipelines for conventional machine learning and deep learning
in medical imaging and physiological signal analysis, with a particular focus
on minimal wearable sensing. The chapter also provides concise overviews of
Publications I–IV, linking each study to the overarching thesis aims.
• Chapter 4 presents and synthesises the main results from Publications I–IV.
It summarises the quantitative findings on brain tumour MRI classification,
neonatal sleep staging with multiview EEG features and reduced channel sets,
20
Introduction
and ECG-based emotion recognition. The chapter compares generalised ver-
sus personalised models, analyses robustness to noise and artefacts, and dis-
cusses computational complexity and inference-time constraints across the dif-
ferent case studies.
• Chapter 5 discusses the overall scientific and practical contributions of the
thesis. It integrates insights from all four publications, reflects on their impli-
cations for deployable AI in brain imaging and physiological monitoring, and
outlines limitations and directions for future research, including edge–cloud
architectures and translational pathways toward clinical and real-world use.
21
2 Datasets
2.1 MRI Datasets for Brain-Tumour Classification
2.1.1 Dataset I (Publication I & II)
The Kaggle “Brain MRI data for Diagnosing Brain Tumour” collection [107] pro-
vides a baseline for binary classification (tumour vs. non-tumour). It contains 3060
2D slices (1500 tumour, 1500 non-tumour; 60 unlabelled excluded). We used a strat-
ified split: 70% training and 30% held out for validation/testing. Image names were
organised by class prefix: “y*” (tumour) and “no*” (non-tumour).
2.1.2 Dataset II (Publication I & II)
Dataset II [108] was originally curated for segmentation with masks; for classifi-
cation, the masks were omitted. The set contains 2501 images (950 tumour, 1551
non-tumour). We used a stratified 70%/30% train/holdout split and the same class
prefixing (“y*”, “no*”).
2.1.3 Dataset III (Publication I & II; Combined)
Dataset III merges Datasets I and II to increase diversity, yielding 5501 images (2450
tumour, 3051 non-tumour). A stratified 70% / 30% split was applied, explicitly bal-
ancing contributions from the two sources across training, validation, and test. The
final test set contained 900 images (450 from Dataset I and 450 from Dataset II).
2.2 Neonatal EEG Dataset (Publication III)
Table 1 summarises cohort and signal specifications. These data were used in Publi-
cation III.
Participant information On average, neonates were born at 38.3 ± 1.8 weeks of
gestation; data collection occurred at a postmenstrual age of 40.5±1.7 weeks. Mean
weight was 3.3 ± 0.6 kg. Mean sleep duration was 86.5 ± 34.4 minutes, and wake-
fulness averaged 43.0± 34.5 minutes.
22
Datasets
Table 1. Neonatal cohort and recording specifications.
Term Detail
Sex (boys:girls) 32:32
Gestational age (weeks) 38.3± 1.8
Postmenstrual age (weeks) 40.5± 1.7
Weight (kg) 3.3± 0.6
Sleep stage counts
Awake (𝑊S) 5514
Active Sleep I (𝑁1) 4785
Active Sleep II (REM–II) 755
Quiet Sleep I (𝑁𝐼𝐼 ) 1247
Quiet Sleep II (𝑁𝐼𝐼𝐼 ) 4502
Reason for inclusion Pneumonia, jaundice, etc.
EEG channels used F3, F4, C3, C4, T3, T4, P3, P4
Sampling rate 500Hz
Notes: boys:girls = biological sex counts; weeks reported as mean ±SD. Stage labels follow
the study’s internal convention.
Inclusion criteria Clinical indications included hyperbilirubinaemia, sepsis, and
other routine NICU admission reasons.
Data collection Physiological signals (EEG, EOG, EMG, ECG) were acquired on
a Nicolet system at 500Hz. EEG electrodes included Fp1, Fp2, F3, F4, C3, C4, P3,
P4, T3, and T4 with Cz as reference; data were resampled to 128Hz for analysis to
reduce redundancy.
Sleep stage annotation Neonatologists specialised in neonatal sleep staged the
recordings manually; these labels served as the reference standard for evaluation.
Ethical approval The study protocol was approved by the Research Ethics Com-
mittee of Fudan University, Shanghai, China, approval number (2017) 89.
2.3 POPANE Emotion Dataset (Publication IV)
We utilise the publicly available POPANE dataset [109], which comprises approxi-
mately 725 hours of multimodal psychophysiological recordings from 1157 healthy
young adults collected across seven independent studies. In line with affective sci-
ence, physiological activity is a core component of emotional responses alongside
23
Dr. Muhammad Irfan
subjective experience and behaviour. POPANE was designed to characterise both
positive such as amusement, tenderness, and gratitude. and negative such as anger,
disgust, fear, sadness, threat) affective states during rest and controlled emotion elic-
itation.
Across studies, affect was recorded continuously during a resting baseline and
during tasks that reliably evoke emotions, including film clips, static pictures, speech
preparation, and expressive writing. Physiological signals include ECG, impedance
cardiography (ICG), EDA, PPG/blood volume pulse (BVP), respiration, and skin
temperature. Each study is organised in a separate folder; within a study, data
for each participant are provided as: (a) a full laboratory recording, for example
123_All.csv); (b) a separate baseline file, such as 123_Baseline.csv); and
(c) one or more emotion-elicitation files, such as 123_Manipulation1.csv).
Files contain multiple channels, such as valence annotations, ECG, EDA, PPG/BVP,
and a marker channel. The accompanying metadata.xlsx file maps markers to
specific emotion conditions.
In this thesis, we analyse ECG only to perform classification of (positive vs.
negative). This minimal-sensing choice aligns with our deployment goals: it reduces
battery consumption on wearables and lowers data transmission to the cloud while
retaining discriminative autonomic information relevant to affect.
24
3 General Workflow in AI for Healthcare
A typical healthcare-oriented AI research pipeline comprises numerous organised
phases to ensure clinically useful, technically robust, and deployable models under
real-world constraints, as shown in Figures 6 and 7. This technique is applied to
brain tumour MRI classification, newborn sleep staging, and minimal-sensing emo-
tion recognition in this thesis.
Data Collection
Sensors, MRI, PSG, Surveys
Annotation
Clinician Scoring, AASM
Pre-processing
Filtering, Artefact Removal, Epochs
Feature Extraction
Time/Freq, DTCWT, Wavelets
Feature Fusion
Early (Concat) / Late
Feature Selection
PCA, RFE, AdaptiSelect
Normalization
StandardScaler, MinMax
Model Development
SVM, RF, XGB, Ensemble
Validation Strategy
k-Fold, LOSO, LOGO
Post-processing
Smoothing, Majority Vote
Deployment
Edge Device / Cloud API
Figure 6. Flow diagram for models performing preprocessing and feature extraction.
3.0.1 Data Collection
The workflow begins with data acquisition, either through new recordings from hu-
man subjects or publicly available Datasets. In this thesis, for example:
• Brain tumour classification uses publicly available MRI Datasets to validate
the fuzzy-logic CNN models.
• Neonatal sleep staging relies on multichannel EEG recorded through PSG.
• Emotion recognition uses wearable-sensing modalities such as ECG and GSR,
where ECG captures cardiac autonomic dynamics, and GSR reflects sympa-
thetic arousal.
25
Dr. Muhammad Irfan
Data Collection
MRI /CT/PET
Annotation
Clinician Labels
Minimal Pre-processing
Filtering, Normalization
Model Input Prep
Reshaping, Augmentation
Feature Learning Layers
CNN / Blocks
Specialized Modules
FAC / FSC Integration
Classification Head
Dense Layers, Softmax
Deep Neural Network
Training Loop
Loss, Optimizer, Backprop
Validation
Subject-wise, k-Fold
Deployment
Edge / Cloud Inference
Figure 7. Deep Learning Pipeline integrated with specialised feature learning (FAC/FSC)
Each physiological signal has unique characteristics:
• EEG provides rich frequency-specific information and is preferred for sleep
staging.
• ECG reflects heart-rate modulation strongly linked to emotional arousal.
• MRI/CT reveal anatomical brain structures important for tumour localisation.
Together, these modalities illustrate the diversity of data types used in AI for
healthcare.
3.0.2 Labelling and Annotation
Following data collection, expert annotation is essential. In MRI, radiologists label
the presence or type of a tumour. When annotations from multiple experts are avail-
able, inter-rater agreement can be evaluated using measures such as Cohen’s Kappa
(𝜅). Clinicians annotate 30-second EEG epochs for sleep staging in accordance with
AASM recommendations, although length may vary based on application. In emo-
tion datasets, labels are derived from stimulus metadata, experimental techniques,
or subjective self-reports. The model’s top performance constraint is directly deter-
mined by high-quality labels, whereas noisy labels propagate errors throughout the
pipeline.
26
General Workflow in AI for Healthcare
3.0.3 Data Preparation and Pre-processing
Pre-processing converts raw data into a form suitable for modelling. Depending on
the modality, this may include:
• Filtering and signal pre-processing: low-pass, high-pass, and notch filtering
for ECG/EEG; MSPCA for multiscale EEG denoising; and baseline correction
for wearable signals.
• Artefact detection: removal of segments affected by movement or sensor
noise.
• MRI pre-processing: timage augmentation using random horizontal and ver-
tical flips, random rotations (up to 20∘), and color jitter. All MRI images
are normalised using ImageNet mean and standard deviation values, resized
to 224×224 or 128×128 pixels depending on GPU memory, and reflection
padding is applied before resizing to preserve edge information. Validation
and test images are not augmented; they are only resized and normalised.
• Segmentation into epochs:
– Sleep staging typically uses epochs of 30 s.
– Real-time emotion recognition advantages from short windows, such as
( 10 s) to reduce alerting latency.
– MRI analysis may use patches or slices, depending on the tumour’s res-
olution requirements.
Shorter windows may compromise accuracy, but they are crucial for prompt re-
sponses, particularly in neonatal care, dementia surveillance, and wearable edge-
computing contexts.
3.0.4 Edge vs. Cloud Processing
Modern healthcare AI often uses edge–cloud architectures:
• Edge computing processes sensitive or time-critical data locally, reducing la-
tency and bandwidth requirements. This demands computationally efficient
models such as the lightweight fuzzy CNN architectures (FAC and FSC) pro-
posed in this thesis.
• Cloud computing handles tasks needing greater computational resources, and
in this case, data are transmitted to a remote server for inference or training.
27
Dr. Muhammad Irfan
• Hybrid systems split the pipeline such that filtering, feature extraction, and
packet formation occur at the edge, while model inference or extended analyt-
ics run in the cloud.
The models (FAC, FSC, ensemble) developed in this thesis are well-suited to both
edge and cloud-oriented workflows, owing to their compact architectures, parameter-
efficient design, and reduced requirements for energy, memory, and data transmission
3.0.5 Feature Extraction
In feature-based machine learning pipelines, meaningful features are typically com-
puted using signal processing methods [? 106]. Examples used in this thesis in-
clude:
• Time-domain features: mean, standard deviation, HRV metrics.
• Frequency-domain features: PSD bands, LF/HF components.
• Wavelet-based transforms: FAWT, DTCWT, MSPCA.
• Autoregressive descriptors: ECOV spectral features.
Each of these captures complementary physiological information: wavelets de-
liver time–frequency localisation; HRV quantifies autonomic nervous system activ-
ity; and spectral features summarise rhythmic structure.
3.0.6 Feature Fusion and Selection
Feature fusion can follow two strategies:
• Early fusion: concatenate all feature groups and then perform feature selec-
tion.
• Late fusion: apply feature selection to each feature group independently, then
combine the reduced sets.
Both approaches aim to reduce redundancy and computational load, particularly
important for edge deployment. A key methodological requirement is that feature
selection must be fitted only on the training data. The learned transformation is
then applied to validation and test sets to avoid data leakage.
3.0.7 Model Development
Common machine-learning models used in healthcare include:
28
General Workflow in AI for Healthcare
• Support Vector Machines (SVM),
• Random Forests and Extra Trees,
• Gradient Boosting Machines (GBM, XGBoost, LightGBM),
• Naïve Bayes,
• k-Nearest Neighbours (KNN),
• Multilayer Perceptrons (MLP).
Ensemble techniques such as stacking, soft/hard voting, and bagging often out-
perform individual models by combining complementary decision boundaries. In
Publication IV, a custom ensemble (EnsemNet) achieved the best performance. Al-
though deep learning models automatically extract informative features, they typi-
cally depend on large datasets and substantial computational capacity. Their reduced
interpretability remains a major concern, reinforcing the need for explainable tech-
niques and, in some applications, favouring simpler machine-learning models that
offer greater transparency.
3.0.8 Data Splitting and Validation
Robust evaluation involves carefully designed splits, such as:
• Train/validation/test splits for hyperparameter optimisation,
• k-Fold cross-validation for stable performance estimation,
• Leave-One-Subject-Out (LOSO) for personalised modelling (subject indepen-
dent),
• Leave-One-Group-Out (LOGO) for cross-site oe group generalisation,
• Random splits, used only when subject leakage is not a concern.
Different publications in this thesis employ validation strategies tailored to the
dataset structure and research objective. In Publications I and II, stratified train/validation/test
splits are used for binary brain-tumour classification from MRI. In Publication III,
both cross-validation and leave-one-subject-out validation are used to assess neonatal
sleep-stage classification performance and subject-level generalisability. In Publica-
tion IV, separate validation settings were used for personalised and generalised ECG-
based emotion recognition. These strategies were selected to ensure fair evaluation
and to match the intended application of each study.
29
Dr. Muhammad Irfan
3.0.9 Post-processing
Post-processing can include probability smoothing, threshold calibration, majority
voting, or temporal filtering. However, heavy post-processing may hinder real-time
applicability. For edge-based use cases such as emotion alerts or neonatal monitor-
ing, minimal or no post-processing is preferred.
These steps form a complete pipeline from raw data to clinically meaningful AI
output. Across all publications, this thesis emphasises low-computational-cost mod-
els, robustness across subjects and modalities, and compatibility with edge–cloud
deployment, ensuring the feasibility of these methods in real healthcare environ-
ments.
3.1 Fuzzy Atrous Convolution for Brain Tumour MRI
In Publication I, we address binary brain-tumour classification in MR images using
a lightweight convolutional architecture (FAC) that enlarges the receptive field (via
dilations) without discarding input samples, using fuzzy memberships at the zero en-
tries of dilated kernels. The full pipeline covers dataset preparation, augmentation,
FAC-based feature extraction (TOFU/MOFU), and final classification. By incorpo-
rating a fuzzy membership function into the existing convolution process, the FAC
can preserve an identity mapping from the input data [110]. It enables unrestricted
data flow and maximises gradient transfer. A convolution function is an operation
that multiplies, adds, or accumulates. When a CNN model is convoluted, the input is
split into the corresponding sliding windows. At every sliding step, every element is
accumulated or added after the convolution kernel is multiplied element-wise [111].
TOFU (Top of the Funnel). A dense dilated block extracts high-resolution fea-
tures using three dilated convolutions with concatenation and a 1 × 1 compressor,
producing rich early representations with minimal depth. The implementation of
this block is given in Algorithm II in Publication I.
MOFU (Middle of the Funnel). Three FAC layers downsample aggressively
with matched stride/dilation values {2, 3, 5}, reducing maps from 512×512→15×15
or 224×224→5×5, while preserving context via FAC. A random channel permutation
between TOFU and MOFU increases diversity. The implementation of this block is
given in Algorithm III in Publication I.
Classifier (Bottom). Flattened MOFU output feeds a fully connected layer for
binary prediction (tumour vs. non-tumour). The implementation of this block is
given in Algorithm IV in Publication I.
30
General Workflow in AI for Healthcare
3.2 Fuzzy Sigmoid Convolution for Brain Tumour MRI
In Publication II, we introduce FSC, an innovative convolutional layer that inte-
grates fuzzy logic principles with convolutional neural networks. FSC improves the
feature extraction abilities of traditional convolution layers by dynamically adjusting
the contributions of various feature activation's through fuzzy membership functions
based on sigmoid activation. The overall network structure is composed of three
major components: TOFU blocks, MOFU blocks, and fully connected layers.
The FSC layer applies fuzzy membership functions to the convolution process,
modulating each feature map through two sigmoid-based membership functions:
high membership and low membership. These functions adjust the importance of
individual features based on their activation values.
Given an input 𝑥 passed through a convolution operation and batch normalisa-
tion, the fuzzy membership functions are defined as follows:
high_membership = 𝜎(𝑥), (3)
low_membership = 𝜎(−𝑥), (4)
where 𝜎(·) is the sigmoid activation function.
The output of the FSC layer, denoted as 𝑜fsc, is computed by combining the
feature map 𝑥 with the high and low membership functions as follows:
𝑜fsc = high_membership · 𝑥+ low_membership · (1− 𝑥). (5)
This mechanism allows the network to emphasise more critical features while
suppressing less relevant ones dynamically. After computing 𝑜fsc, the result is passed
through a ReLU activation function:
𝑜relu = ReLU(𝑜fsc). (6)
TOFU: The TOFU block is designed to gradually enhance feature extraction by
stacking multiple FSC layers with progressively larger receptive fields. Each FSC
layer collects features from the previous layers. The final output of the TOFU block
is produced by concatenating these feature maps and applying a compression step
via a 1× 1 convolution. The implementation of this block is given in Algorithm II in
Publication II.
MOFU: In the MOFU block, the features extracted by the TOFU block are further
refined. It employs downsampling and captures multi-scale features by applying sev-
eral FSC layers with varying dilation rates, as given in Algorithm III in Publication
II.
31
Dr. Muhammad Irfan
Bottom Layer: After feature extraction via the TOFU and MOFU blocks, the
model uses global average pooling to reduce the feature maps to a 256-dimensional
vector. This vector is passed through two fully connected layers with ReLU activa-
tion and dropout to prevent overfitting. The final layer then maps the output to the
target binary class, completing the prediction as given in algorithm IV in Publication
II.
3.3 Neonatal Sleep Analysis using EEG
In Publication III, we examine neonatal sleep analysis using EEG recorded via a
PSG device, with the aim of designing a lightweight, deployment-minded pipeline
that supports accurate sleep staging while remaining suitable for resource-constrained
clinical environments. The diagrams in Fig. 8 illustrate the cloud implementation,
while Fig. 9 presents the proposed IoMT architecture for edge-cloud structures. In
the cloud implementation, data is transferred to the cloud, similar to our previous
studies [112; 80]. However, to minimise data transfer, data packet formation is im-
plemented in the edge-cloud structure, as shown in Fig. 9.
Figure 8. Implemented Cloud-Based Multiview EEG Signal Processing Framework for Neonatal
Sleep Analysis Using AdaptiSelect and Ensemble Models
Raw EEG is preprocessed with finite impulse response (FIR) filtering (0.3–35 Hz),
artefact screening, and segmented into 30 s epochs. To attenuate structured noise
while preserving morphology, we apply multi-scale principle component analysis
(MSPCA; Symlet-4, level = 8). We then extract a multiview feature set per epoch,
comprising the Flexible Analytical Wavelet Transform (FAWT), Dual-Tree Complex
Wavelet Transform (DTCWT), enhanced-covariance (ECOV) AR spectral features,
and band-limited spectral statistics, which capture complementary time–frequency,
autoregressive, and power-spectrum cues. From 315 features per channel epoch (95
FAWT, 130 DTCWT, 50 ECOV, 40 spectral), the AdaptiSelect module selects the
most effective 122 features for single-channel,>900 across eight channels. We have
32
General Workflow in AI for Healthcare
also provided a comparison of the feature importance values of Multi-Scale Hierar-
chical Neural Network (MSHNN), SWT, DTCWT, and multiview methods.
Figure 9. Proposed Edge-cloud-Based Multiview EEG Signal Processing Framework for Neonatal
Sleep Analysis Using AdaptiSelect and Ensemble Models
Classification employs a rotation ensemble stacking design that combines Rota-
tion Forest Extra Trees, Gradient Boosting, and Random Forest (as a meta-learner).
Validation employs a stratified 10-fold CV for model development and LOGO hold-
outs for subject-wise generalisation. For edge/cloud efficiency, the STREAM mod-
ule packages selected features into compact 5-feature packets, yielding∼153.6× re-
duction versus transmitting raw samples per epoch while retaining epoch identity for
lossless reconstruction in the cloud. Overall, the multiview + AdaptiSelect + rotation-
ensemble stack delivers excellent performance for the four-stage neonatal sleep clas-
sification with low computational and transmission cost and outperforms previous
model development in the same domain.
3.4 Minimal Wearable Sensing for Emotions Analysis
In Publication IV, the methodology focuses on using ECG signals to classify emo-
tions into positive and negative classes. To enable real-time and low-latency emo-
tion detection, we chose a 10-second epoch length for segmentation. This decision
enables a balance between the need for timely classification and the constraints of
edge computing and wearable devices. Although shorter epochs can slightly reduce
classification performance compared to longer windows, they enable near-real-time
alerts, which are crucial in scenarios where immediate feedback is required. By seg-
menting into 10-second epochs, the system can deliver emotion classification and
potential emergency alerts almost instantly, rather than waiting for longer windows
that could delay intervention. The workflow involves a multi-stage preprocessing
pipeline to clean the ECG data, including notch, high-pass, and low-pass filtering to
33
Dr. Muhammad Irfan
remove noise and baseline drifts.
In terms of model development, we designed a streamlined pipeline that includes
multi-domain feature extraction from time, frequency, and HRV domains, followed
by a hybrid feature selection process to maintain computational efficiency. We ex-
perimented with several feature selection methods and proposed a hybrid feature
selection approach to select the optimal method for emotion detection. We tested a
range of machine learning classifiers and ultimately developed an ensemble model
for its robust performance and adaptability. This approach is designed for both per-
sonalised and generalised models, recognising that in the case of dementia patients,
there are multiple types of dementia, and each patient’s experience can vary signif-
icantly. By incorporating both personalised and generalised modelling approaches,
we can better tailor the emotion recognition system to individual variability, espe-
cially in populations with diverse and changing emotional responses. The proposed
methodology offers a lightweight, real-time capable solution that can be deployed
on wearable and edge devices, ensuring both efficiency and the flexibility to adapt to
individual needs.
34
4 Summary of key findings and
Discussion
Across Publications I and II, we performed systematic ablation studies to under-
stand how architectural choices influence model behaviour and to evaluate perfor-
mance under conditions that approximate real clinical variability. For brain–tumour
MRI classification, both FAC and FSC were tested under three stress scenarios: (i)
biased class splits, (ii) additive noise, and (iii) partial occlusion. The ablations varied
in support parameters and model depth, such as the number of fuzzy blocks. These
experiments demonstrated that moderate architectural depth, combined with appro-
priately tuned support values, offers the best balance between accuracy, stability, and
robustness. Both FAC and FSC maintained high accuracy and F1 scores across all
testing conditions, with only marginal performance degradation, indicating that the
proposed fuzzy–convolutional modules can tolerate imbalance, noise, and missing
information. We also benchmarked the proposed models against strong baselines, in-
cluding transfer-learning backbones and classical machine-learning pipelines. FAC
and FSC matched or exceeded state-of-the-art performance while using substantially
fewer parameters, supporting their suitability for deployment in edge-cloud pipelines
where computational efficiency is essential.
For Publication III, we performed a comprehensive analysis of feature impor-
tance, channel-reduction strategies, and ablation studies for EEG-based neonatal
sleep staging. We evaluated multiple machine-learning models on complete PSG
data (eight channels) and then progressively reduced the channel set to single- and
dual-channel configurations. Using the stream module, we quantified the data re-
duction ratio. We demonstrated that the proposed method can significantly reduce
transmission to the cloud while still allowing full reconstruction at the cloud side.
The proposed feature-fusion strategy outperformed existing approaches, and we also
reported execution time, feature-fusion latency, and packet-formation time as key
factors for practical neonatal monitoring systems.
In Publication IV, we compared several feature-selection techniques with the
proposed method and evaluated multiple machine-learning classifiers under both per-
sonalised and generalised settings for ECG-based emotion recognition. Across both
settings, the proposed pipeline consistently outperformed standard approaches typi-
cally used for emotion-classification tasks, demonstrating its suitability for real-time,
low-latency affective monitoring on wearable and edge devices.
35
Dr. Muhammad Irfan
4.0.1 Publication I: FAC for Brain–Tumour MRI
In Publication I, we evaluate the FAC architecture for binary brain tumour classi-
fication using three publicly available MRI Datasets (Datasets I–III). We examine
robustness under biased test splits, additive noise, and partial occlusion, and conduct
an ablation on the support parameter and network depth. Publication I provides
comprehensive descriptions of the experimental technique and dataset. Performance
under several conditions is provided in Fig. 10.
Original Biased Noise Occlusion
0.9
0.95
1
P
er
fo
rm
an
ce
S
co
re
accuracy precision recall F1
Figure 10. Performance of the FAC model under four evaluation conditions (original, biased split,
Gaussian noise, partial occlusion). The consistently high accuracy, precision, recall, and F1-score
highlight the robustness of the architecture to data imbalance and perturbations.
Dataset I-Original: With standard augmentation and early stopping, FAC achieved
0.9883 accuracy, an F1 score of 0.9884, and a Cohen’s 𝜅 of 0.9767. These results
were obtained using mixed-precision training and gradient accumulation, demon-
strating an efficient and resource-aware training strategy.
Biased Split of Dataset I: As shown in Fig. 11, despite the imbalance, the model
performed equally well as in the balanced data scenario, achieving a test loss of
0.0468, an accuracy of 0.9895, a precision of 0.9922, a recall of 0.9844, an F1 of
0.9883, and a 𝜅 of 0.9788.
Added Noise in Dataset I: Gaussian noise was introduced to evaluate the model’s
robustness under degraded imaging conditions. Using a 128 × 128 input and an
effective batch size of 52 (via gradient accumulation), the training was capped at 200
36
Summary of key findings and Discussion
(a) accuracy over epochs (b) Loss over epochs (c) Confusion matrix
Figure 11. FAC under biased split on Dataset I . Training dynamics and final confusion matrix
illustrate robustness to class imbalance.
epochs, and the model achieved an accuracy of 0.9684, precision of 0.9785, recall
of 0.9579, F1 score of 0.9681, and 𝑘𝑎𝑝𝑝𝑎 of 0.9368. These results are reported in
Publication I. A second experiment, using a larger 224 × 224 input, batch size of
32, and a patience value of 50, was trained for 139 epochs (early stopping, with a
maximum limit of 200) and yielded improved performance, achieving a test accuracy
of 0.9807, F1 score of 0.9807, and 𝜅 of 0.9614. The corresponding confusion matrix,
learning curves and metric trajectories are presented in Fig. 12.
(a) accuracy over epochs (b) Loss over epochs (c) Confusion matrix
Figure 12. FAC evaluation under added Gaussian noise on Dataset I. Training and validation
curves show stable convergence, and the confusion matrix reflects strong generalisation despite
noise injection, demonstrating robustness across resolutions.
Performance on Dataset I with Simulated Partial Occlusion: To assess robust-
ness to missing visual information, we introduced partial occlusions to images, sim-
ulating real-world conditions where parts of the image might be obscured during
training, validation, and testing. Using a 224 × 224 input, a batch size of 52 (with
gradient accumulation), and 26 for validation and testing, along with early stopping,
the model achieved an accuracy of 0.9883, precision of 0.9867, recall of 0.9900, F1
37
Dr. Muhammad Irfan
score of 0.9884, and a 𝜅 of 0.9767. These results indicate that FAC maintains stable
performance even when relevant image regions are partially obscured, as illustrated
in Fig. 13.
(a) accuracy over epochs (b) Loss over epochs (c) Confusion matrix
Figure 13. FAC evaluation under Simulated Partial Occlusion on Dataset I. Training and validation
curves show stable convergence, and the confusion matrix reflects strong generalisation despite
simulated Partial Occlusion injection, demonstrating robustness across resolutions.
In general, performance remains largely the same across class imbalance and
occlusion, with only a limited drop under noise, aligning with the inductive bias of
FAC towards prominent, spatially localisable signals.
Ablation: Support Parameter & Depth (Dataset III). A three-block FAC offers
the most stable trade-off between accuracy and over-downsampling, achieving an ac-
curacy of 0.9956 with three layers compared with 0.9936 for two layers. The choice
of support parameter also matters: fixed values of 𝜇 ∈ {0.5, 0.7} remain competitive
(accuracy between 0.9946 and 0.9956), while VEOC yields an accuracy of 0.9950,
suggesting that per-channel adaptability is beneficial but not strictly necessary when
the network depth is well chosen.
Comparison to Baselines (Dataset III). Against state-of-the-art transfer baselines
trained at 224×224, such as VGG19 (0.9880 accuracy) and ResNet101 (0.9775 accu-
racy), FAC reaches an accuracy of 0.9956 with only 47.82K parameters, smaller by
several orders of magnitude than typical transfer-learning backbones (6.47–25.89M
parameters). This closes the accuracy gap and improves computational efficiency,
positioning FAC for edge/cloud deployment.
Operational Significance. (i) Stability to class imbalance and occlusion supports
clinical use where sampling is imperfect, and artefacts occur; (ii) the controlled
loss under noise motivates integrating denoise/augment or test-time ensembling in
38
Summary of key findings and Discussion
pipelines with known acquisition variability; (iii) parameter efficiency enables on-
device triage and cost-effective scaling.
Computational efficiency and deployability. To evaluate deployment feasibility,
we benchmarked inference speed and memory footprint using an input tensor,which
represents a single image passed through the model during forward propagation.
Across all compared architectures, the proposed FAC model exhibited the small-
est memory footprint (0.19 MB) and the fastest per–image inference time (3.7 ms;
≈270 FPS). In contrast, widely used transfer learning backbones such as VGG16
(512 MB), VGG19 (532 MB), and ResNet50 (89 MB) are two to three orders of
magnitude larger and substantially slower (7–12 ms per image).
For throughput benchmarking, the proposed model achieved 269.8 FPS, out-
performing both lightweight networks (MobileNetV2: 134 FPS; EfficientNet-B0:
83.8 FPS) and heavier backbones (ResNet50: 130 FPS; VGG16: 105 FPS; DenseNet-
121: 51.9 FPS). This represents approximately a 2× speed-up over MobileNetV2,
3× over VGG16, and more than 5× over DenseNet121, despite requiring signifi-
cantly fewer parameters. These results demonstrate that the FAC architecture delivers
state-of-the-art accuracy while meeting real-time constraints, making it highly suit-
able for edge, embedded, and IoT deployments where memory, latency, and power
budgets are limited.
4.0.2 Publication II: FSC for Brain–Tumour MRI
Publication II evaluates the FSC architecture for binary brain tumour classification
across three publicly available MRI Datasets (Datasets I–III), following the same re-
liability assessment protocol described in Section 4.0.1. The FSC model contains
0.21627 million parameters, which is larger than FAC (47,820 parameters), yet it
consistently achieves higher classification performance. For example, on the original
Dataset I, FSC achieves an accuracy of 0.9917, compared with 0.9883 for FAC. Un-
der noisy conditions, FSC reaches 0.9702 versus 0.9684 for FAC, and in the biased
split and partial occlusion scenarios, FSC achieves 0.9825 and 0.9877, respectively,
outperforming FAC in all cases as provided in Publications I and II. A similar trend is
observed for Datasets II and III. Figures 14, 15, and 16 provide additional visualisa-
tions of the FSC training behaviour, including accuracy and loss trends over epochs
and the corresponding confusion matrices for each perturbation setting. Notably, the
results in Fig. 14 show a slight performance improvement compared to those in Ta-
ble II of Publication II. This improvement is due to using a higher input resolution
(128 × 128) and increasing the early-stopping patience from 20 to 30 epochs, both
of which helped stabilise convergence and reduce over-regularisation.
When comparing both proposed approaches from Publications I and II, FAC
and FSC use substantially fewer parameters than state-of-the-art transfer learning
39
Dr. Muhammad Irfan
(a) accuracy over epochs (b) Loss over epochs (c) Confusion matrix
Figure 14. FSC under biased split on Dataset I . Training dynamics and final confusion matrix
illustrate robustness to class imbalance.
models such as VGG19, EfficientNet, and ResNet. Despite their compact size, both
architectures match or surpass these large models in accuracy while demonstrating
strong robustness to distribution shifts. Their small parameter footprint and stability
under perturbations highlight their suitability for deployment in edge or IoT environ-
ments.
(a) accuracy over epochs (b) Loss over epochs (c) Confusion matrix
Figure 15. FSC evaluation under added Gaussian noise on Dataset I.
Computational efficiency and deployability. The FSC architecture is markedly
compact, using only 0.216M parameters (0.83 MB), about 70–600× fewer than trans-
fer learning models, such as MobileNetV2, ResNet50, EfficientNet, DenseNet121,
VGG16 and VGG19. Despite this footprint, FSC attains exceptional accuracy on
Dataset I, matching or exceeding transfer-learning baselines that span ∼14–143M
parameters. Runtime measurements at a 224 × 224 resolution show a throughput
of 52.76 FPS, which is comparable to DenseNet-121 (54.54 FPS) and meets real-
time requirements. While MobileNetV2 and ResNet50 reach higher frame rates
(≈140 FPS), they incur ≥10–100× larger model sizes and memory budgets, such
40
Summary of key findings and Discussion
(a) accuracy over epochs (b) Loss over epochs (c) Confusion matrix
Figure 16. FSC evaluation under Simulated Partial Occlusion on Dataset I. Training and validation
curves show stable convergence, and the confusion matrix reflects strong generalisation despite
simulated Partial Occlusion injection, demonstrating robustness across resolutions.
as 8.49–89.68 MB vs. 0.83 MB for FSC. Overall, FSC offers a favourable accu-
racy–efficiency–latency trade-off, making it well-suited for edge–cloud deployments
where on-device inference, low memory, and robust performance are required.
4.0.3 Publication III: Smart IoT-based Solutions for Neonatal
Sleep Analysis
The findings reported in Publication III bring together the key advances in neonatal
sleep analysis, highlighting the efficacy, robustness, and computational efficiency of
the proposed methods. The experiments conducted across the datasets collected over
four years, various channel setups, and different perturbation conditions consistently
demonstrate that the combination of multiview feature extraction, AdaptiSelect fea-
ture selection, the STREAM module, and the rotational stacking classifier constitutes
a robust and scalable framework for neonatal sleep staging.
10-Fold-CV: WS vs. NI vs. REM-II vs. QS: Across both 10-fold and LOSO
cross-validation, the results consistently show that increasing the number of EEG
channels improves classification performance, although the largest gain arises when
moving from a single-channel to a dual-channel configuration. In a 10-fold CV,
the model achieves an accuracy, F1, and Kappa of 0.8116, 0.8033, 0.7217 using
one channel, improving to 0.8279, 0.8199, 0.7470 with two channels, and reach-
ing 0.8774, 0.8735, 0.8214 with eight channels, respectively. Channel-pair analysis
highlights that combinations involving F3, particularly F3-T4, provide the strongest
dual-channel performance. The LOSO-CV for (WS vs. NI vs. REM-II vs. QS))
follows the same trend, with accuracies of 0.8047 (1-channel), 0.8165 (2-channel),
and 0.8395 (8-channel), demonstrating the robustness of the approach under subject-
independent evaluation. These results confirm that while eight channels yield the
41
Dr. Muhammad Irfan
highest absolute performance, a two-channel configuration provides a strong accu-
racy–complexity trade-off that is well suited for clinically practical neonatal EEG
systems.
10-fold CV: WS vs. NII vs. NIII vs. AS: For the (WS, NII, NIII, AS), the proposed
methodology again shows a consistent performance gain as the number of EEG chan-
nels increases. In the 10-fold CV, accuracy/F1/Kappa improve from 0.7905/0.7737/-
0.6988 (1 channel) to 0.8055/0.7956/0.7183 (2 channels), and further to 0.8656/-
0.8589/0.8073 using all 8 channels. The LOSO-CV for (WS vs. NII vs. NIII vs.
AS), the results follow the same trend, with the eight-channel configuration achiev-
ing the highest metrics (accuracy 0.8274, F1 0.8167, Kappa 0.7473), outperforming
both the dual-channel and single-channel settings. These results confirm that richer
multichannel information substantially boosts classification performance, while the
two-channel configuration still provides a competitive, lower-complexity alternative
suitable for practical neonatal monitoring systems.
Impact of Three Different Noises on Model Performance: The robustness anal-
ysis demonstrates that neonatal EEG is highly sensitive to noise and motion arte-
facts, which can adversely affect classification performance when uncorrected. In our
experiments, unfiltered motion artefacts reduced multichannel accuracy by 2.21 %,
whereas applying the preprocessing pipeline (MSPCA and filtering) limited this re-
duction to only 0.44 %, confirming the effectiveness of the denoising strategy. Label
purity also influenced performance: ensuring clean 30-second epochs improved ac-
curacy by 1.3 % in 10-fold CV. Across the three noise types evaluated, Gaussian
noise caused the largest degradation, with accuracy dropping by up to 3.21 % in the
8-channel configuration. Salt-and-pepper and dropout noise produced smaller reduc-
tions, remaining below 1.3 % across all channel setups. These findings confirm that
incorporating robust preprocessing and feature selection steps (MSPCA and “Adap-
tiSelect”) substantially stabilises model performance and enhances robustness under
realistic neonatal EEG noise conditions.
Comparison with SOTA: A comparison with leading SOTA techniques for neona-
tal sleep staging demonstrates that the proposed method consistently achieves su-
perior performance on all major evaluation metrics. Earlier CNN-based methods
such as Conv-2D [113; 114] report relatively low accuracies (0.51–0.53), while
feature-engineered or hybrid approaches including Stacking [115], BiLSTM [116],
and MBCNN [117] achieve moderate improvements (0.66–0.71). Deep learning ar-
chitectures adapted for neonatal data, such as DeepSleepNet [118] and MS-HNN [30],
achieve accuracies around 0.72, with MS-HNN providing the strongest baseline (F1
= 0.73, 𝜅 = 0.62).
42
Summary of key findings and Discussion
In Scenario S1, the proposed methodology achieves superior performance (Acc.
= 0.81, F1 = 0.805, 𝜅 = 0.72), outperforming all SOTA benchmarks, including
the best-performing MS-HNN model. In Scenario S2, the proposed method also
exceeds the strongest existing deep learning models, achieving Acc. = 0.7745 and
𝜅 = 0.6788, compared with MS-HNN (Acc. = 0.702, 𝜅 = 0.5992) and DeepSleep-
Net (Acc. ≈ 0.70, 𝜅 = 0.5802). These consistent gains across scenarios underscore
the robustness, generalisability, and discriminative strength of the proposed multi-
view feature and stacking-based framework for neonatal sleep staging.
Computational Complexity and Resource Requirements: The computational as-
sessment shows that the stacking model remains lightweight and efficient across
all channel configurations. Model size increases only modestly from 50.57 MB (1-
channel) to 58.84 MB (8-channels), while inference time remains well within real-
time limits (0.0979–0.1493 s). Peak memory usage scales with channel count, reach-
ing 6004.95 KB for eight channels, with a corresponding rise in energy consumption
(1.0481 J). For multiview feature extraction, DTCWT requires the highest memory
(691 KB) and longest execution time (0.0397 s), whereas FAWT is the most efficient
(0.0181 s). Feature fusion is computationally inexpensive (0.0050 s for one channel;
0.0401 s for eight channels), and STREAM processing adds negligible overhead. To-
gether, these results indicate that the proposed pipeline is computationally viable for
real-time neonatal sleep staging. The memory and timing requirements are compat-
ible with edge devices (e.g., Raspberry Pi 3), supporting a future hybrid edge–cloud
deployment for low-latency clinical monitoring.
4.0.4 Publication IV: Emotion Recognition with Minimal Wear-
able Sensing
Publication IV evaluates whether a single physiological signal (ECG) can reliably
distinguish between positive and negative emotional states using a lightweight ma-
chine learning framework. Across all experiments, personalised models substantially
outperform generalised ones. The best personalised ensemble model achieves an ac-
curacy of 0.9559, compared with 0.6992 obtained by the top generalised model, un-
derscoring the strong influence of inter-subject variability in physiological responses.
An ablation study examines the role of different feature-selection strategies. While
SelectKBest improves performance for several models, the hybrid method (Extra-
Trees + SelectKBest) consistently produces the highest accuracies. In personalised
settings, models such as ExtraTrees, Bagging, and Random Forest achieve accura-
cies in the range of 0.90–0.95 with hybrid feature selection, confirming their ef-
fectiveness in capturing subject-specific discriminatory features. In generalised set-
tings, the hybrid method also yields the greatest improvements, with the ensemble
model reaching 0.6992 and outperforming all baselines. These results collectively
43
Dr. Muhammad Irfan
show that ECG alone provides a sufficiently rich signal for emotion classification
when paired with appropriate feature engineering and personalised modelling. They
also highlight the importance of hybrid feature selection in improving both robust-
ness and interpretability. Publication IV supports the viability of minimal-sensing,
low-complexity emotion-recognition systems suitable for wearable and resource-
constrained environments.
Practical Significance for Real-World Deployment Beyond reporting relative re-
ductions in model size and speed, the practical importance of computational effi-
ciency should be assessed in the target deployment scenario. In Publications I and II,
millisecond-level inference times and high throughput demonstrate that the proposed
MRI models are fast enough for real-time or near-real-time decision support, while
their substantially lower parameter counts enhance the feasibility of deployment on
hardware with limited resources. In Publication III, acceptable performance must be
evaluated alongside reduced sensing requirements, efficient data transmission, and
robustness at the subject level, because neonatal sleep staging is designed for con-
tinuous monitoring rather than one-off predictions. In Publication IV, the usefulness
of emotion recognition systems is determined not only by classification accuracy but
also by their ability to operate reliably with short ECG segments and to incur low
computational overhead in wearable applications. Thus, across all three domains,
practical value hinges on striking an appropriate balance between accuracy, robust-
ness, latency, and deployment feasibility.
44
5 Contribution of these Studies to the
Overall Thesis Theme
This thesis aimed to investigate how computationally efficient AI methods can be
designed for three clinically important, yet technically distinct, problems: brain tu-
mour analysis from MRI, neonatal sleep staging from EEG, and emotion recognition
from physiological signals. Although each publication focuses on its own dataset,
modality, and task, together they form a coherent line of work on lightweight, robust,
and clinically aware machine learning for health monitoring. In what follows, the
main contributions of the four publications are summarised in relation to the overall
thesis theme rather than repeating the full paper-specific result sections.
5.1 Alignment with the Overall Thesis Aims
Across all four studies, the thesis contributes to three overarching goals:
• designing resource-efficient AI models that remain feasible for edge or con-
strained clinical hardware, such as reduced model size, fast inference, minimal
sensing;
• improving robustness and generalisability through subject-aware evaluation,
such as LOSO validation, noise analysis, and principled feature selection;
• demonstrating that carefully engineered “classical” or hybrid approaches (op-
timised features + ensembles) can perform competitively with, or better than,
heavier end-to-end deep networks in several health-related tasks.
The following sections provide a brief overview of each publication within this
broader context.
5.2 Contribution of the Brain Tumour MRI Studies (Pub-
lications I–II)
The MRI-based studies address the first part of the thesis theme: efficient AI for
brain tumour analysis.
45
Dr. Muhammad Irfan
• Both publications introduce and evaluate computationally lean CNN architec-
tures, specifically FAC and FSC-based architectures, that operate on 2D MRI
slices, reducing memory and computational demands while retaining high di-
agnostic performance.
• Through systematic perturbation experiments, such as adding noise, geometric
and intensity variations, and monitoring of training dynamics (accuracy, loss,
and confusion matrices across conditions), the studies quantify robustness and
demonstrate how architectural choices, such as fuzzy atrous modules, improve
stability under clinically realistic image degradations.
• Subject-level aggregation strategies and carefully designed cross-validation,
for example, subject-stratified folds are used to reduce information leakage,
providing a more realistic estimate of how models would behave in real de-
ployments.
These contributions demonstrate that brain tumour classification can progress to-
wards practical application without relying on excessively large, dense models, and
they outline the design principles (efficiency, robustness, and subject-aware evalua-
tion) that are subsequently applied in physiological signal research.
5.3 Contribution of the Neonatal Sleep Study (Publica-
tion III)
The neonatal sleep work extends the thesis from neuroimaging to EEG-based moni-
toring, while keeping a strong focus on computational efficiency and clinical practi-
cality.
• A multiview feature fusion strategy is proposed, combining DTCWT, FAWT,
ECOV, and spectral features into a rich representation of neonatal EEG. This
is coupled with the proposed AdaptiSelect method, which reduces the dimen-
sionality from thousands of raw features to a compact subset, such as from
2520 to 900 features for eight channels, and to 122 per single channel, while
improving performance and stability.
• The work systematically analyses the trade-off between the number of EEG
channels and classification accuracy. 10-fold, and LOSO-CV experiments
show that two-channel configurations, such as F3 paired with temporal or cen-
tral channels–T4, C3, or C4, achieve performance very close to that of eight-
channel setups, while single-channel models remain clinically usable. This is
central to the thesis theme of minimal, clinically feasible sensing.
46
Contribution of these Studies to the Overall Thesis Theme
• Robustness is explicitly studied by injecting Gaussian, salt-and-pepper, and
dropout noise into the signals. The analysis shows that Gaussian noise has the
most detrimental effect on the average performance drop across eight channels.
In contrast, salt-and-pepper and dropout noise lead to smaller performance
reductions. The denoising pipeline (bandpass filtering + MSPCA) reduces the
impact of motion artefacts in ten-fold CV, and label purity analysis shows that
cleaner epoch labelling can improve accuracy.
• Computational profiling demonstrates that feature extraction per epoch re-
mains fast (e.g., FAWT ≈ 0.0181 s, DTCWT ≈ 0.0397 s), and that the stack-
ing classifier has modest model size (up to ≈ 58.8 MB) and short inference
times (≈ 0.10–0.15 s per test fold), supporting future edge-cloud deployment
scenarios.
• Compared with SOTA neonatal sleep methods such as CNN- and BiLSTM-
based models and multi-scale deep models, the proposed approach achieves
higher four-stage classification performance, demonstrating that multiview fea-
tures combined with ensembles can outperform several end-to-end deep learn-
ing baselines.
This research anchors the thesis in neonatal care, demonstrating how the same de-
sign ideas, multiview features, feature selection, channel reduction, robustness anal-
ysis, and computational profiling can be applied to transition from “research-grade”
PSG to more scalable, NICU-friendly solutions.
5.4 Contribution of the Emotion Recognition Study (Pub-
lication IV)
The fourth publication completes the thesis by transitioning to emotion recognition
from minimally sensed physiological signals, with a focus on ECG.
• It demonstrates that a single physiological modality (ECG) can be used to dis-
tinguish between positive and negative emotional states with high accuracy
when combined with multidomain features and appropriate machine learning
models. This aligns with the thesis goal of minimal sensing for real-time mon-
itoring.
• An extensive comparison between generalised and personalised models shows
a substantial performance gap of 0.2567, with personalised models outper-
forming their generalised counterparts. This highlights inter-subject variability
as a key challenge and motivates the development of user-specific or adaptive
models for affective computing.
47
Dr. Muhammad Irfan
• A hybrid feature selection pipeline is proposed and systematically evaluated in
ablation studies. For generalised models, some classifiers, such as ensemble
(Voting_Soft), benefit most from hybrid feature selection, while others favour
KBest. For personalised models, hybrid feature selection consistently yields
the best performance across top classifiers, confirming its role in improving
accuracy without sacrificing computational efficiency.
• The analysis indicates that relatively simple, regularised methods, including
KNN with distance weighting, ExtraTrees, random forest, and Multilayer Per-
ceptron, can attain robust performance in a personalised context, supporting
previous findings where ensembles with carefully designed features matched
or surpassed deep learning models in physiological signal analysis.
5.5 Synthesis and Overall Impact
While the thesis presents four studies in different healthcare domains, they are uni-
fied by a consistent research theme: the development of computationally efficient AI
methods for medical imaging and physiological signal analysis in resource-constrained
settings. Across these studies, the work begins from practical and clinical constraints,
including limited sensing channels, noisy data, heterogeneous subjects, and deploy-
ment limitations, and then develops methods that are explicitly efficient, robust, and
deployable, rather than assuming that larger models and more data alone will solve
these challenges. The novelty of this thesis lies not only in the individual meth-
ods proposed in each publication but also in demonstrating a unified design strat-
egy for computationally efficient AI across heterogeneous healthcare tasks, includ-
ing MRI-based brain-tumour classification, neonatal sleep-stage analysis from EEG,
and ECG-based emotion recognition.
5.6 Shortcomings and Future Work
This thesis has several limitations, including the neonatal sleep EEG dataset origi-
nating from a single site, which restricts population diversity and external validity.
Although the clinical scoring scheme contains five neonatal sleep stages, the present
work focuses on a four-stage classification, and transitional or additional stages are
not modelled. Moreover, a fixed 30 s epoch length is used throughout; the impact of
shorter, longer, or multi-scale epoching on performance and real-time responsiveness
remains unexplored.
For brain tumour MRI, the datasets were collected from multiple sites, but the
tasks are restricted to binary classification (tumour vs. non-tumour), without address-
ing finer-grained grading or subtyping. This simplifies the clinical problem and may
limit direct applicability in real-world neuro-oncology workflows, where tumour
48
Contribution of these Studies to the Overall Thesis Theme
grading, subtype differentiation, and treatment-related changes are also important.
Thus, the present study should be interpreted as a proof-of-concept showing that
computationally efficient architectures can retain strong classification performance
while substantially reducing model complexity. In addition, although the proposed
method performed favourably compared with previous studies, formal statistical sig-
nificance testing was not conducted. Therefore, small differences in performance
metrics should be interpreted with caution, particularly regarding their clinical sig-
nificance. Likewise, the ECG-based emotion analysis is limited to binary (positive
vs. negative) affect and does not consider neutral or multi-class emotional states.
Pathways to clinical application Future work should further evaluate the pro-
posed methods on larger and more diverse external datasets to improve robustness
and clinical generalisability. For brain-tumour MRI (Publication I and II), an im-
portant next step is to move beyond binary classification towards tumour grading and
subtype classification, which are more clinically challenging due to higher inter-class
similarity and data imbalance. Although Publication III demonstrates strong poten-
tial for compact and IoT-compatible neonatal sleep staging, further work is needed
before clinical use. A practical translation pathway would include external valida-
tion across independent, multicentre datasets, prospective testing in real neonatal
monitoring environments, assessment of robustness to signal artefacts and device
variability, and integration with neonatal care workflows. In addition, future studies
should evaluate how such systems can support clinicians as decision-support tools
rather than fully automated diagnostic systems. For ECG-based emotion recogni-
tion (Publication IV), extending the framework from binary affect classification to
neutral and multi-class emotional states remains challenging due to stronger class
overlap and greater inter-subject variability. In addition, exploring alternative epoch
lengths, multi-scale temporal modelling, and validating the proposed methods on
larger, more diverse cohorts will be important steps towards more robust, clinically
generalisable AI solutions.
49
List of References
[1] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt
Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb model
size. arXiv preprint arXiv:1602.07360, 2016.
[2] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen.
Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference
on computer vision and pattern recognition, pages 4510–4520, 2018.
[3] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural
networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019.
[4] Felix J Dorfner, Jay B Patel, Jayashree Kalpathy-Cramer, Elizabeth R Gerstner, and Christo-
pher P Bridge. A review of deep learning for brain tumor analysis in mri. NPJ Precision
Oncology, 9(1):2, 2025.
[5] Tarek Berghout. The neural frontier of future medical imaging: a review of deep learning for
brain tumor detection. Journal of Imaging, 11(1):2, 2024.
[6] Amreen Batool and Yung-Cheol Byun. A lightweight multi-path convolutional neural network
architecture using optimal features selection for multiclass classification of brain tumor using
magnetic resonance images. Results in Engineering, 25:104327, 2025.
[7] Annisa Wulandari, Riyanto Sigit, and Mochamad Mobed Bachtiar. Brain tumor segmentation
to calculate percentage tumor using mri. In 2018 International Electronics Symposium
on Knowledge Creation and Intelligent Computing (IES-KCIC), pages 292–296.
IEEE, 2018.
[8] Muhammad Irfan, Hafza Ayesha Siddiqa, Abdelwahed Nahliis, Chen Chen, Yan Xu, Laishuan
Wang, Anum Nawaz, Abdulhamit Subasi, Tomi Westerlund, and Wei Chen. An ensemble voting
approach with innovative multi-domain feature fusion for neonatal sleep stratification. IEEE
Access, 12:206–218, 2024. doi: 10.1109/ACCESS.2023.3346059.
[9] Muhammad Irfan, Laishuan Wang, Husnain Shahid, Yan Xu, Abdulhamit Subasi, Adnan Mu-
nawar, Noman Mustafa, Chen Chen, Tomi Westurlund, and Wei Chen. Multidomain se-
lective feature fusion and stacking based ensemble framework for eeg-based neonatal sleep
stratification. IEEE Journal of Biomedical and Health Informatics, pages 1–10, 2025. doi:
10.1109/JBHI.2025.3530107.
[10] Duojiao Wu, Catherine M Rice, and Xiangdong Wang. Cancer bioinformatics: A new approach
to systems clinical medicine. BMC bioinformatics, 13(1):1–4, 2012.
[11] Javier E Villanueva-Meyer, Marc C Mabray, and Soonmee Cha. Current clinical brain tumor
imaging. Neurosurgery, 81(3):397–415, 2017.
[12] Ian Law, Nathalie L Albert, Javier Arbizu, Ronald Boellaard, Alexander Drzezga, Norbert
Galldiks, Christian la Fougère, Karl-Josef Langen, Egesta Lopci, Val Lowe, et al. Joint
eanm/eano/rano practice guidelines/snmmi procedure standards for imaging of gliomas using
pet with radiolabelled amino acids and [18f] fdg: version 1.0. European journal of nuclear
medicine and molecular imaging, 46(3):540–557, 2019.
[13] Sapna Singh Kshatri and Deepak Singh. Convolutional neural network in medical image analy-
sis: a review. Archives of Computational Methods in Engineering, 30(4):2793–2810, 2023.
[14] Novsheena Rasool and Javaid Iqbal Bhat. Brain tumour detection using machine and deep learn-
ing: a systematic review. Multimedia tools and applications, 84(13):11551–11604, 2025.
50
LIST OF REFERENCES
[15] Surajit Das and Rajat Subhra Goswami. Advancements in brain tumor analysis: a comprehensive
review of machine learning, hybrid deep learning, and transfer learning approaches for mri-based
classification and segmentation. Multimedia Tools and Applications, 84(23):26645–26682, 2025.
[16] Sara Bouhafra and Hassan El Bahi. Deep learning approaches for brain tumor detection and clas-
sification using mri images (2020 to 2024): a systematic review. Journal of Imaging Informatics
in Medicine, 38(3):1403–1433, 2025.
[17] Jyotismita Chaki and Marcin Woz´niak. A deep learning based four-fold approach to classify
brain mri: Btscnet. Biomedical Signal Processing and Control, 85:104902, 2023.
[18] Santhos A. Kumar, Anil Kumar, Varun Bajaj, and Girish Kumar Singh. An improved fuzzy
min–max neural network for data classification. IEEE Transactions on Fuzzy Systems,
28(9):1910–1924, 2020. doi: 10.1109/TFUZZ.2019.2924396.
[19] Muhammad Irfan, Abdulhamit Subasi, Hassan Mehdi, Tomi Westerlund, and Wei Chen. Fuzzy-
based atrous convolution for brain tumor detection using mri. In 2024 IEEE International Con-
ference on Progress in Informatics and Computing (PIC), pages 280–289. IEEE, 2024.
[20] Muhammad Irfan, Anum Nawaz, Riku Klén, Abdulhamit Subasi, Tomi Westerlund, and Wei
Chen. Improved brain tumor detection in mri: Fuzzy sigmoid convolution in deep learning.
In 2025 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2025. doi:
10.1109/IJCNN64981.2025.11227858.
[21] Mohsen Asghari Ilani and Yaser M Banad. Brain tumor detection through diverse cnn archi-
tectures in iot healthcare industries: Fast r-cnn, u-net, transfer learning-based cnn, and fully
connected cnn. arXiv preprint arXiv:2509.05821, 2025.
[22] Ishak Pacal, Omer Celik, Bilal Bayram, and Antonio Cunha. Enhancing efficientnetv2 with
global and efficient channel attention mechanisms for accurate mri-based brain tumor classifica-
tion. Cluster Computing, 27(8):11187–11212, 2024.
[23] International Data Corporation. Worldwide quarterly wearable device tracker. https://www.
idc.com/promo/wearablevendor/, 2025. Accessed: 2025-09-09.
[24] AARP. Aarp research insights on technology. https://www.
aarp.org/pri/topics/technology/internet-media-devices/
aarp-research-insights-technology/, 2025. Published: July 3, 2025. Ac-
cessed: September 9, 2025.
[25] World Health Organization. Levels and trends in child mortality report 2021. World Health Orga-
nization, 2021. URL https://www.who.int/news-room/fact-sheets/detail/
levels-and-trends-in-child-mortality-report-2021. Accessed: 15 March
2024.
[26] World Health Organization. Newborn mortality, report march 2024. World Health Or-
ganization, 2024. URL https://www.who.int/news-room/fact-sheets/
detail/newborn-mortality#:~:text=The%20world%20has%20made%
20substantial,%2Dneonatal%20under%2D5%20mortality. Accessed: 25
March 2024.
[27] Kritiprasanna Das, Vivek Kumar Singh, and Ram Bilas Pachori. Introduction to eeg signal
recording and processing. In Artificial Intelligence Enabled Signal Processing based Models for
Neural Information Processing, pages 1–19. CRC Press, 2024.
[28] Dale Purves, George J Augustine, David Fitzpatrick, Lawrence C Katz, Anthony-Samuel
LaMantia, James O McNamara, and S Mark Williams. Neuroglial cells. Neuroscience, 2001.
[29] Jeffrey W Britton, Lauren C Frey, Jennifer L Hopp, Pearce Korb, Mohamad Z Koubeissi,
William E Lievens, Elia M Pestana-Knight, and Erk K St Louis. Electroencephalography (eeg):
An introductory text and atlas of normal and abnormal findings in adults, children, and infants.
2016.
[30] Hangyu Zhu, Laishuan Wang, Ning Shen, Yonglin Wu, Shu Feng, Yan Xu, Chen Chen, and Wei
Chen. Ms-hnn: Multi-scale hierarchical neural network with squeeze and excitation block for
neonatal sleep staging using a single-channel eeg. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 31:2195–2204, 2023. doi: 10.1109/TNSRE.2023.3266876.
51
Dr. Muhammad Irfan
[31] Britta Hüning, Tobias Storbeck, Nora Bruns, Frauke Dransfeld, Julia Hobrecht, Julia Karpien-
ski, Selma Sirin, Bernd Schweiger, Christel Weiss, Ursula Felderhoff-Müser, et al. Relationship
between brain function (aeeg) and brain structure (mri) and their predictive value for neurode-
velopmental outcome of preterm infants. European journal of pediatrics, 177:1181–1189, 2018.
[32] Renée A Shellhaas, Taeun Chang, Tammy Tsuchida, Mark S Scher, James J Riviello, Nicholas S
Abend, Sylvie Nguyen, Courtney J Wusthoff, and Robert R Clancy. The american clinical neu-
rophysiology society’s guideline on continuous electroencephalography monitoring in neonates.
Journal of clinical neurophysiology, 28(6):611–617, 2011.
[33] Linda GM van Rooij, Linda S de Vries, Alexander C van Huffelen, and Mona C Toet. Additional
value of two-channel amplitude integrated eeg recording in full-term infants with unilateral brain
injury. Archives of Disease in Childhood-Fetal and Neonatal Edition, 95(3):F160–F168, 2010.
[34] Jayant N Acharya and Vinita J Acharya. Overview of eeg montages and principles of localiza-
tion. Journal of Clinical Neurophysiology, 36(5):325–329, 2019.
[35] Alfred L Loomis, E Newton Harvey, and Garret A Hobart. Cerebral states during sleep, as
studied by human brain potentials. Journal of experimental psychology, 21(2):127, 1937.
[36] Pierre Gloor. Hans berger on electroencephalography. American Journal of EEG Technology, 9
(1):1–8, 1969.
[37] Mohamed El-Dib, Taeun Chang, Tammy N Tsuchida, and Robert R Clancy. Amplitude-
integrated electroencephalography in neonates. Pediatric neurology, 41(5):315–326, 2009.
[38] Andreea M Pavel, Janet M Rennie, Linda S de Vries, Mats Blennow, Adrienne Foran, Divyen K
Shah, Ronit M Pressler, Olga Kapellou, Eugene M Dempsey, Sean R Mathieson, et al. A
machine-learning algorithm for neonatal seizure recognition: a multicentre, randomised, con-
trolled trial. The Lancet Child & Adolescent Health, 4(10):740–749, 2020.
[39] Faye S Silverstein, Frances E Jensen, Terrie Inder, Lena Hellstrom-Westas, Deborah Hirtz, and
Donna M Ferriero. Improving the treatment of neonatal seizures: National institute of neurolog-
ical disorders and stroke workshop report. The Journal of pediatrics, 153(1):12–15, 2008.
[40] Hafza Ayesha Siddiqa, Zhenning Tang, Yan Xu, Laishuan Wang, Muhammad Irfan, Saadul-
lah Farooq Abbasi, Anum Nawaz, Chen Chen, and Wei Chen. Single-channel eeg data analysis
using a multi-branch cnn for neonatal sleep staging. IEEE Access, 12:29910–29925, 2024.
doi: 10.1109/ACCESS.2024.3365570.
[41] GB Boylan, L Burgoyne, C Moore, B O’flaherty, and JM Rennie. An international survey of eeg
use in the neonatal intensive care unit. Acta paediatrica, 99(8):1150–1155, 2010.
[42] Robertino Dilena, Federico Raviglione, Gaetano Cantalupo, Duccio M Cordelli, Paola De Liso,
Matteo Di Capua, Raffaele Falsaperla, Fabrizio Ferrari, Monica Fumagalli, Silvia Lori, et al.
Consensus protocol for eeg and amplitude-integrated eeg assessment and monitoring in neonates.
Clinical Neurophysiology, 132(4):886–903, 2021.
[43] V. Gerla, K. Paul, L. Lhotska, and V. Krajca. Multivariate analysis of full-term neonatal
polysomnographic data. IEEE Transactions on Information Technology in Biomedicine, 13(1):
104–110, 2009. doi: 10.1109/TITB.2008.2007193.
[44] Nicholas S Abend, Ram Mani, Tammy N Tschuda, Tae Chang, Alexis A Topjian, Maureen
Donnelly, Denise LaFalce, Margaret C Krauss, Sarah E Schmitt, and Joshua M Levine. Eeg
monitoring during therapeutic hypothermia in neonates, children, and adults. American journal
of electroneurodiagnostic technology, 51(3):141–164, 2011.
[45] Maarten De Vos, Wouter Deburchgraeve, PJ Cherian, Vladimir Matic, RM Swarte, Paul Govaert,
Gerhard Henk Visser, and Sabine Van Huffel. Automated artifact removal as preprocessing
refines neonatal seizure detection. Clinical Neurophysiology, 122(12):2345–2354, 2011.
[46] De Maarten Vos, Stephanie Riès, Katrien Vanderperren, Bart Vanrumste, Francois-Xavier Alario,
Van Sabine Huffel, and Boris Burle. Removal of muscle artifacts from eeg recordings of spoken
language production. Neuroinformatics, 8:135–150, 2010.
[47] Vitaly Schetinin and Joachim Schult. The combined technique for detection of artifacts in clinical
electroencephalograms of sleeping newborns. IEEE Transactions on Information Technology in
Biomedicine, 8(1):28–35, 2004.
52
LIST OF REFERENCES
[48] Mona C Toet and Petra MA Lemmers. Brain monitoring in neonates. Early human development,
85(2):77–84, 2009.
[49] Lachlan Webb, Minna Kauppila, James A Roberts, Sampsa Vanhatalo, and Nathan J Steven-
son. Automated detection of artefacts in neonatal eeg with residual neural networks. Computer
Methods and Programs in Biomedicine, 208:106194, 2021.
[50] M. Kauppila, S. Vanhatalo, and N.J. Stevenson. Artifact detection in neonatal eeg using gaus-
sian mixture models. In EMBEC & NBC 2017: Joint Conference of the European Medical and
Biological Engineering Conference (EMBEC) and the Nordic-Baltic Conference on Biomedi-
cal Engineering and Medical Physics (NBC), Tampere, Finland, June 2017, pages 221–224.
Springer Singapore, 2018.
[51] Nathan J Stevenson, John M O’Toole, Irina Korotchikova, and Geraldine B Boylan. Artefact de-
tection in neonatal eeg. In 2014 36th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society, pages 926–929. IEEE, 2014.
[52] Vladimir Matic, Perumpillichira J Cherian, Ninah Koolen, Gunnar Naulaers, Renate M Swarte,
Paul Govaert, Sabine Van Huffel, and Maarten De Vos. Holistic approach for automated back-
ground eeg assessment in asphyxiated full-term infants. Journal of neural engineering, 11(6):
066007, 2014.
[53] Saeed Montazeri Moghadam, Elana Pinchefsky, Ilse Tse, Viviana Marchi, Jukka Kohonen,
Minna Kauppila, Manu Airaksinen, Karoliina Tapani, Päivi Nevalainen, Cecil Hahn, et al. Build-
ing an open source classifier for the neonatal eeg background. 2021.
[54] Pierre Bourgoin, Victoire Barrault, Gauthier Loron, Alexandre Roger, Emmanuelle Bataille,
Laurène Leclair-Visonneau, Nicolas Joram, and Alexis Chenouard. Interrater agreement between
critical care providers for background classification and seizure detection after implementation
of amplitude-integrated electroencephalography in neonates, infants, and children. Journal of
Clinical Neurophysiology, 37(3):259–262, 2020.
[55] Nathan J Stevenson, Maria-Luisa Tataranno, Anna Kaminska, Elena Pavlidis, Robert R Clancy,
Elke Griesmaier, James A Roberts, Katrin Klebermass-Schrehof, and Sampsa Vanhatalo. Relia-
bility and accuracy of eeg interpretation for estimating age in preterm infants. Annals of Clinical
and Translational Neurology, 7(9):1564–1573, 2020.
[56] Hangyu Zhu, Yao Guo, Yonglin Wu, Yiyuan Zhang, Ning Shen, Yan Xu, Laishuan Wang, Chen
Chen, and Wei Chen. Magsleepnet: Adaptively multi-scale temporal focused sleep staging
model for multi-age groups. Expert Systems with Applications, 240:122549, 2024.
[57] Saeed Montazeri Moghadam. Advancing automated eeg analysis for neonatal intensive care:
From engineering to bedside solutions. 2023.
[58] Joseph L Fleiss, Bruce Levin, and Myunghee Cho Paik. Statistical methods for rates and pro-
portions. john wiley & sons, 2013.
[59] Giuseppe Buonocore and Rodolfo Bracci Michael Weindling. neonatology a practical approach
to neonatal diseases. Springer, 2012.
[60] Mark S Scher, Mingui Sun, Doris A Steppe, Robert D Guthrie, and Robert J Sclabassi. Com-
parisons of eeg spectral and correlation measures between healthy term and preterm infants.
Pediatric neurology, 10(2):104–108, 1994.
[61] Joseph J Volpe. Neurology of the newborn e-book. 2008.
[62] Monique André, M-D Lamblin, Anne-Marie d’Allest, Lilia Curzi-Dascalova, FTSNT Moussalli-
Salefranque, S Nguyen The Tich, M-F Vecchierini-Blineau, Fabrice Wallois, E Walls-Esquivel,
and P Plouin. Electroencephalography in premature and full-term infants. developmental features
and glossary. Neurophysiologie clinique/Clinical neurophysiology, 40(2):59–124, 2010.
[63] Danièle Jouvet-Mounier, Liliane Astic, and Daniel Lacote. Ontogenesis of the states of sleep
in rat, cat, and guinea pig during the first postnatal month. Developmental Psychobiology: The
Journal of the International Society for Developmental Psychobiology, 2(4):216–239, 1969.
[64] Gerald A Marks, James P Shaffery, Arie Oksenberg, Samuel G Speciale, and Howard P Rof-
fwarg. A functional role for rem sleep in brain maturation. Behavioural brain research, 69(1-2):
1–11, 1995.
53
Dr. Muhammad Irfan
[65] Patricio Peirano, Cecilia Algarín, and Ricardo Uauy. Sleep-wake states and their regulatory
mechanisms throughout early human development. The Journal of pediatrics, 143(4):70–79,
2003.
[66] William P Fifer, Dana L Byrd, Michelle Kaku, Inge-Marie Eigsti, Joseph R Isler, Jillian Grose-
Fifer, Amanda R Tarullo, and Peter D Balsam. Newborn infants learn during sleep. Proceedings
of the National Academy of Sciences, 107(22):10320–10323, 2010.
[67] Elizabeth Hennevin, Bernard Hars, Catherine Maho, and Vincent Bloch. Processing of learned
information in paradoxical sleep: relevance for memory. Behavioural brain research, 69(1-2):
125–135, 1995.
[68] Catherine Maho and Vincent Bloch. Responses of hippocampal cells can be conditioned during
paradoxical sleep. Brain research, 581(1):115–122, 1992.
[69] H Denenberg Victor and B Thoman Evelyn. Evidence for a functional role for active (rem) sleep
in infancy. Sleep, 4(2):185–191, 1981.
[70] Jeffrey W Britton, Lauren C Frey, Jennifer L Hopp, Pearce Korb, Mohamad Z Koubeissi,
William E Lievens, Elia M Pestana-Knight, and Erk K St Louis. Electroencephalography (eeg):
An introductory text and atlas of normal and abnormal findings in adults, children, and infants.
2016.
[71] Madeleine M Grigg-Damberger. The visual scoring of sleep in infants 0 to 2 months of age.
Journal of clinical sleep medicine, 12(3):429–445, 2016.
[72] Lilia Curzi-Dascalova, Patricio Peirano, and Françoise Morel-Kahn. Development of sleep states
in normal premature and full-term newborns. Developmental Psychobiology: The Journal of the
International Society for Developmental Psychobiology, 21(5):431–444, 1988.
[73] Saadullah Farooq Abbasi, Awais Abbas, Iftikhar Ahmad, Mohammed S Alshehri, Sultan Al-
makdi, Yazeed Yasin Ghadi, and Jawad Ahmad. Automatic neonatal sleep stage classification:
A comparative study. Heliyon, 2023.
[74] Awais Abbas, Hafiz Sheraz Sheikh, Haris Ahmad, and Saadullah Farooq Abbasi. An iot
and machine learning-based neonatal sleep stage classification. Manuscript on Research-
Gate. https://www. researchgate. net/publication/369761857_ An_IoT_and_Machine_Learning-
based_Neonatal_Sleep_Stage_Classification (accessed on 08/05/2023).
[75] Saeed Montazeri Moghadam, Päivi Nevalainen, Nathan J Stevenson, and Sampsa Vanhatalo.
Sleep state trend (sst), a bedside measure of neonatal sleep state fluctuations based on single eeg
channels. Clinical Neurophysiology, 143:75–83, 2022.
[76] Amir H Ansari, Ofelie De Wel, Kirubin Pillay, Anneleen Dereymaeker, Katrien Jansen,
Sabine Van Huffel, Gunnar Naulaers, and Maarten De Vos. A convolutional neural network
outperforming state-of-the-art sleep staging algorithms for both preterm and term infants. 17(1):
016028, 2020. doi: 10.1088/1741-2552/ab5469. URL https://dx.doi.org/10.1088/
1741-2552/ab5469.
[77] Saadullah Farooq Abbasi, Jawad Ahmad, Ahsen Tahir, Muhammad Awais, Chen Chen, Muham-
mad Irfan, Hafiza Ayesha Siddiqa, Abu Bakar Waqas, Xi Long, Bin Yin, et al. Eeg-based
neonatal sleep-wake classification using multilayer perceptron neural network. IEEE Access,
8:183025–183034, 2020.
[78] Yue Wu, Jiaming Liu, Maoguo Gong, Zhixiao Liu, Qiguang Miao, and Wenping Ma. Mpct:
Multiscale point cloud transformer with a residual network. IEEE Transactions on Multimedia,
26:3505–3516, 2024. doi: 10.1109/TMM.2023.3312855.
[79] Yongzhe Yuan, Yue Wu, Xiaolong Fan, Maoguo Gong, Wenping Ma, and Qiguang Miao. Egst:
Enhanced geometric structure transformer for point cloud registration. IEEE Transactions on
Visualization and Computer Graphics, 30(9):6222–6234, 2024. doi: 10.1109/TVCG.2023.
3329578.
[80] Muhammad Irfan, Husnain Jawad, Barkoum Betra Felix, Saadullah Farooq Abbasi, Anum
Nawaz, Saeed Akbarzadeh, Muhammad Awais, Lin Chen, Tomi Westerlund, and Wei Chen.
Non-wearable iot-based smart ambient behavior observation system. IEEE Sensors Journal, 21
(18):20857–20869, 2021. doi: 10.1109/JSEN.2021.3097392.
54
LIST OF REFERENCES
[81] Muhammad Irfan, Laishuan Wang, Husnain Shahid, Yan Xu, Abdulhamit Subasi, Adnan Mu-
nawar, Noman Mustafa, Chen Chen, Tomi Westurlund, and Wei Chen. Multidomain selective
feature fusion and stacking based ensemble framework for eeg-based neonatal sleep stratifica-
tion. IEEE Journal of Biomedical and Health Informatics, 2025.
[82] I Dewa Gede Hari Wisana Triwiyanto and Bedjo Utomo Sumber. Oxigen concentration detector
in the continuous positive airway pressure for sleep apnea therapy based on iot technology. In
2021 International Seminar on Application for Technology of Information and Communication
(iSemantic), pages 156–160, 2021. doi: 10.1109/iSemantic52711.2021.9573207.
[83] Elizabeth A Kensinger. Remembering emotional experiences: The contribution of valence and
arousal. Reviews in the Neurosciences, 15(4):241–252, 2004.
[84] Ed Diener, Stuti Thapa, and Louis Tay. Positive emotions at work. Annual review of organiza-
tional psychology and organizational behavior, 7(1):451–477, 2020.
[85] Evgenia Gkintoni, Anthimos Aroutzidis, Hera Antonopoulou, and Constantinos Halkiopoulos.
From neural networks to emotional networks: A systematic review of eeg-based emotion recog-
nition in cognitive neuroscience and real-world applications. Brain Sciences, 15(3):220, 2025.
[86] Shiqing Zhang, Yijiao Yang, Chen Chen, Xingnan Zhang, Qingming Leng, and Xiaoming Zhao.
Deep learning-based multimodal emotion recognition from audio, visual, and text modalities: A
systematic review of recent advancements and future prospects. Expert Systems with Applica-
tions, 237:121692, 2024.
[87] Huan Liu, Tianyu Lou, Yuzhe Zhang, Yixiao Wu, Yang Xiao, Christian S Jensen, and Dalin
Zhang. Eeg-based multimodal emotion recognition: A machine learning perspective. IEEE
Transactions on Instrumentation and Measurement, 2024.
[88] Runfang Guo, Hongfei Guo, Liwen Wang, Mengmeng Chen, Dong Yang, and Bin Li. Develop-
ment and application of emotion recognition technology—a systematic literature review. BMC
psychology, 12(1):95, 2024.
[89] Maja Pantic, George Caridakis, Elisabeth André, Jonghwa Kim, Kostas Karpouzis, and Stefanos
Kollias. Multimodal emotion recognition from low-level cues. In Emotion-oriented systems:
The humaine handbook, pages 115–132. Springer, 2010.
[90] Loic Kessous, Ginevra Castellano, and George Caridakis. Multimodal emotion recognition in
speech-based interaction using facial expression, body gesture and acoustic analysis. Journal on
Multimodal User Interfaces, 3:33–48, 2010.
[91] Sidney K D’mello and Jacqueline Kory. A review and meta-analysis of multimodal affect detec-
tion systems. ACM computing surveys (CSUR), 47(3):1–36, 2015.
[92] Moataz El Ayadi, Mohamed S Kamel, and Fakhri Karray. Survey on speech emotion recognition:
Features, classification schemes, and databases. Pattern recognition, 44(3):572–587, 2011.
[93] Daniel Canedo and António JR Neves. Facial expression recognition using computer vision: A
systematic review. Applied Sciences, 9(21):4678, 2019.
[94] Gloria Cosoli, Angelica Poli, Lorenzo Scalise, and Susanna Spinsante. Heart rate variability
analysis with wearable devices: Influence of artifact correction method on classification accu-
racy for emotion recognition. In 2021 IEEE International Instrumentation and Measurement
Technology Conference (I2MTC), pages 1–6, 2021. doi: 10.1109/I2MTC50364.2021.9459828.
[95] Yixiang Dai, Xue Wang, Xuanping Li, and Pengbo Zhang. Reputation-driven multimodal emo-
tion recognition in wearable biosensor network. In 2015 IEEE International Instrumentation
and Measurement Technology Conference (I2MTC) Proceedings, pages 1747–1752, 2015. doi:
10.1109/I2MTC.2015.7151544.
[96] Guanglong Du, Qinglin Tan, Chunquan Li, Xueqian Wang, Shaohua Teng, and Peter X. Liu. A
noncontact emotion recognition method based on complexion and heart rate. IEEE Transactions
on Instrumentation and Measurement, 71:1–14, 2022. doi: 10.1109/TIM.2022.3194858.
[97] Yueyang Li, Weiming Zeng, Wenhao Dong, Di Han, Lei Chen, Hongyu Chen, Zijian Kang,
Shengyu Gong, Hongjie Yan, Wai Ting Siok, and Nizhuan Wang. A tale of single-channel
electroencephalography: Devices, datasets, signal processing, applications, and future di-
55
Dr. Muhammad Irfan
rections. IEEE Transactions on Instrumentation and Measurement, 74:1–20, 2025. doi:
10.1109/TIM.2025.3556900.
[98] Yun-Kai Li, Qing-Hao Meng, Tian-Hao Yang, Ya-Xin Wang, and Hui-Rang Hou. Touch gesture
and emotion recognition using decomposed spatiotemporal convolutions. IEEE Transactions on
Instrumentation and Measurement, 71:1–9, 2022. doi: 10.1109/TIM.2022.3147338.
[99] Andrea Apicella, Pasquale Arpaia, Francesco Isgrò, Giovanna Mastrati, and Nicola Moccaldi.
A survey on eeg-based solutions for emotion recognition with a low number of channels. IEEE
Access, 10:117411–117428, 2022. doi: 10.1109/ACCESS.2022.3219844.
[100] Karim Alghoul, Hussein Al Osman, and Abdulmotaleb El Saddik. Enhancing generalization
in ppg-based emotion measurement with a cnn-tcn-lstm model. In 2025 IEEE International
Instrumentation and Measurement Technology Conference (I2MTC), pages 1–6, 2025. doi: 10.
1109/I2MTC62753.2025.11079085.
[101] Marco Schlosser, Harriet Demnitz-King, Tim Whitfield, Miranka Wirth, and Natalie L Marchant.
Repetitive negative thinking is associated with subjective cognitive decline in older adults: a
cross-sectional study. BMC psychiatry, 20:1–10, 2020.
[102] Rajal Devshi, Sarah Shaw, Jordan Elliott-King, Eef Hogervorst, Avinash Hiremath, Latha Ve-
layudhan, Satheesh Kumar, Sarah Baillon, and Stephan Bandelow. Prevalence of behavioural
and psychological symptoms of dementia in individuals with learning disabilities. Diagnostics,
5(4):564–576, 2015.
[103] Xudong Yang, Hongli Yan, Anguo Zhang, Pan Xu, Sio Hang Pan, Mang I Vai, and Yueming
Gao. Emotion recognition based on multimodal physiological signals using spiking feed-forward
neural networks. Biomedical Signal Processing and Control, 91:105921, 2024.
[104] Caroline Cobb (Amey). Emotions, aggression, and stress. https://pressbooks.
library.vcu.edu/psyc629/chapter/emotions-aggression-stress/, 2024.
Accessed: 2025-05-23.
[105] Polar Electro Oy. Polar BLE SDK. https://github.com/polarofficial/
polar-ble-sdk, 2025. Accessed: 2025-05-23.
[106] Muhammad Irfan, Laishuan Wang, Yan Xu, Abdulhamit Subasi, Chen Chen, Riku Klen, Tomi
Westurlund, and Wei Chen. Smart iot-based solutions for neonatal sleep stratification: Single-
dual channel eeg, adaptiselect, multview fusion, & rotational ensemble stacking. IEEE Internet
of Things Journal, 2025.
[107] Abhranta Panigrahi. Brain tumor detection mri. https://www.kaggle.com/datasets/
abhranta/brain-tumor-detection-mri, 2021. Accessed: December 14, 2024.
[108] Maciej A Mazurowski, Kal Clark, Nicholas M Czarnek, Parisa Shamsesfandabadi, Kather-
ine B Peters, and Ashirbani Saha. Radiogenomics of lower-grade glioma: algorithmically-
assessed tumor shape is associated with tumor genomic subtypes and patient outcomes in a
multi-institutional study with the cancer genome atlas data. Journal of neuro-oncology,
133:27–35, 2017.
[109] Maciej Behnke, Mikołaj Buchwald, Adam Bykowski, Szymon Kupin´ski, and Lukasz D Kacz-
marek. Psychophysiology of positive and negative emotions, dataset of 1157 cases and 8 biosig-
nals. Scientific Data, 9(1):10, 2022.
[110] Amitava Halder and Debangshu Dey. Atrous convolution aided integrated framework for lung
nodule segmentation and classification. Biomedical Signal Processing and Control, 82:
104527, 2023.
[111] Rangan Das, Samayan Bhattacharya, Swadesh Jana, Ujjwal Maulik, and Sanghamitra Bandy-
opadhyay. Improving lung ct analysis through fuzzy dilated convolution attention. In 2023
IEEE 3rd Applied Signal Processing Conference (ASPCON), pages 71–76, 2023.
doi: 10.1109/ASPCON59071.2023.10396336.
[112] Muhammad Irfan, Peng Shun, Barkoum Betra Felix, Noman Mustafa, Saadullah Farooq Abbasi,
Abdelwahed Nahli, Abdulhamit Subasi, Tomi Westerlund, and Wei Chen. An iot-based non-
contact ecg system: Sole of the feet/hands palm. IEEE Internet of Things Journal, 2023.
56
[113] Amir Hossein Ansari, Ofelie De Wel, Mario Lavanga, Alexander Caicedo, Anneleen Derey-
maeker, Katrien Jansen, Jan Vervisch, Maarten De Vos, Gunnar Naulaers, and Sabine Van Huf-
fel. Quiet sleep detection in preterm infants using deep convolutional neural networks. Journal
of neural engineering, 15(6):066006, 2018.
[114] Amir H Ansari, Ofelie De Wel, Kirubin Pillay, Anneleen Dereymaeker, Katrien Jansen, Sabine
Van Huffel, Gunnar Naulaers, and Maarten De Vos. A convolutional neural network outperform-
ing state-of-the-art sleep staging algorithms for both preterm and term infants. Journal of Neural
Engineering, 17(1):016028, 2020.
[115] Saadullah Farooq Abbasi, Harun Jamil, and Wei Chen. Eeg-based neonatal sleep stage classifi-
cation using ensemble learning. Comput. Mater. Contin, 70:4619–4633, 2022.
[116] Hojat Ghimatgar, Kamran Kazemi, Mohammad Sadegh Helfroush, Kirubin Pillay, Anneleen
Dereymaker, Katrien Jansen, Maarten De Vos, and Ardalan Aarabi. Neonatal eeg sleep stage
classification based on deep learning and hmm. Journal of neural engineering, 17(3):036031,
2020.
[117] Hafza Ayesha Siddiqa, Zhenning Tang, Yan Xu, Laishuan Wang, Muhammad Irfan, Saadul-
lah Farooq Abbasi, Anum Nawaz, Chen Chen, and Wei Chen. Single-channel eeg data analysis
using a multi-branch cnn for neonatal sleep staging. IEEE Access, 2024.
[118] Akara Supratak, Hao Dong, Chao Wu, and Yike Guo. Deepsleepnet: A model for automatic
sleep stage scoring based on raw single-channel eeg. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 25(11):1998–2008, 2017.