Finding the Trade-offs between Model Size and Performance in Lightweight ResNet-18 architecture for ECG Classification
Ahmed, Faizan (2025-06-25)
Finding the Trade-offs between Model Size and Performance in Lightweight ResNet-18 architecture for ECG Classification
(25.06.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2025063076135
https://urn.fi/URN:NBN:fi-fe2025063076135
Tiivistelmä
Deep learning approaches have shown remarkable success in electrocardiogram (ECG) classification, but their computational costs are often a bottleneck for deployment on low-resource platforms such as mobile health initiatives and edge devices. The thesis addresses the challenge by investigating exhaustively the trade-offs between performance and computational cost through architectural design modifications to a ResNet-18 base model for multi-label ECG classification. Using the extensive combined dataset of PhysioNet/Computing in Cardiology Challenge 2020 and Shandong Provincial Hospital, we conducted a series of 18 experiments to study the impact of layer ablation, architectural simplification, and their combinations.
Our findings indicate that extreme compression of the model is achievable with a negligible reduction in performance. We managed to compress model size by up to 99.2% (from 33.78 MB to as low as 0.28 MB) without sacrificing useful diagnostic performance. Interestingly, most of the smaller models, like Experiment E17 (2.24 MB, 93.4% compression), not only survived but outperformed the baseline model, with micro AUROC and micro precision of 0.88 and 0.45, respectively, compared to the baseline’s 0.86 AUROC and 0.39 precision. This seemingly counterintuitive effect is attributed to factors such as reduced overfitting, enhanced architectural efficiency, superior signal-to-noise ratio, and superior optimization landscape in the smaller models.
Among the major effective compression techniques are substituting normal convolutions with depthwise separable convolutions, extensive reduction of feature maps, and removal of redundant layers selectively. These results counteract the notion that larger models equal better, showing how well-optimized small architectures are more efficient for some tasks like ECG classification. The contribution of this study is significant, enabling high-performance ECG analysis models to be implemented on hardware with limited resources, enabling real-time monitoring, and enabling the sustainability of AI in healthcare. The research offers empirical proof of the feasibility of highly efficient deep learning solutions, enabling broader accessibility and utilization of AI in clinical settings.
Our findings indicate that extreme compression of the model is achievable with a negligible reduction in performance. We managed to compress model size by up to 99.2% (from 33.78 MB to as low as 0.28 MB) without sacrificing useful diagnostic performance. Interestingly, most of the smaller models, like Experiment E17 (2.24 MB, 93.4% compression), not only survived but outperformed the baseline model, with micro AUROC and micro precision of 0.88 and 0.45, respectively, compared to the baseline’s 0.86 AUROC and 0.39 precision. This seemingly counterintuitive effect is attributed to factors such as reduced overfitting, enhanced architectural efficiency, superior signal-to-noise ratio, and superior optimization landscape in the smaller models.
Among the major effective compression techniques are substituting normal convolutions with depthwise separable convolutions, extensive reduction of feature maps, and removal of redundant layers selectively. These results counteract the notion that larger models equal better, showing how well-optimized small architectures are more efficient for some tasks like ECG classification. The contribution of this study is significant, enabling high-performance ECG analysis models to be implemented on hardware with limited resources, enabling real-time monitoring, and enabling the sustainability of AI in healthcare. The research offers empirical proof of the feasibility of highly efficient deep learning solutions, enabling broader accessibility and utilization of AI in clinical settings.