Neuro-critical multimodal Edge-AI monitoring algorithm and IoT system design and development

Abstract

In recent years, with the continuous development of neurocritical medicine, the success rate of treatment of patients with traumatic brain injury (TBI) has continued to increase, and the prognosis has also improved. TBI patients' condition is usually very complicated, and after treatment, patients often need a more extended time to recover. The degree of recovery is also related to prognosis. However, as a young discipline, neurocritical medicine still has many shortcomings. Especially in most hospitals, the condition of Neuro-intensive Care Unit (NICU) is uneven, the equipment has limited functionality, and there is no unified data specification. Most of the instruments are cumbersome and expensive, and patients often need to pay high medical expenses. Recent years have seen a rapid development of big data and artificial intelligence (AI) technology, which are advancing the medical IoT field. However, further development and a wider range of applications of these technologies are needed to achieve widespread adoption. Based on the above premises, the main contributions of this thesis are the following. First, the design and development of a multi-modal brain monitoring system including 8-channel electroencephalography (EEG) signals, dual-channel NIRS signals, and intracranial pressure (ICP) signals acquisition. Furthermore, an integrated display platform for multi-modal physiological data to display and analysis signals in real-time was designed. This thesis also introduces the use of the Qt signal and slot event processing mechanism and multi-threaded to improve the real-time performance of data processing to a higher level. In addition, multi-modal electrophysiological data storage and processing was realized on cloud server. The system also includes a custom built Django cloud server which realizes real-time transmission between server and WeChat applet. Based on WebSocket protocol, the data transmission delay is less than 10ms. The analysis platform can be equipped with deep learning models to realize the monitoring of patients with epileptic seizures and assess the level of consciousness of Disorders of Consciousness (DOC) patients. This thesis combines the standard open-source data set CHB-MIT, a clinical data set provided by Huashan Hospital, and additional data collected by the system described in this thesis. These data sets are merged to build a deep learning network model and develop related applications for automatic disease diagnosis for smart medical IoT systems. It mainly includes the use of the clinical data to analyze the characteristics of the EEG signal of DOC patients and building a CNN model to evaluate the patient's level of consciousness automatically. Also, epilepsy is a common disease in neuro-intensive care. In this regard, this thesis also analyzes the differences of various deep learning model between the CHB-MIT data set and clinical data set for epilepsy monitoring, in order to select the most appropriate model for the system being designed and developed. Finally, this thesis also verifies the AI-assisted analysis model.. The results show that the accuracy of the CNN network model based on the evaluation of consciousness disorder on the clinical data set reaches 82%. The CNN+STFT network model based on epilepsy monitoring reaches 90% of the accuracy rate in clinical data. Also, the multi-modal brain monitoring system built is fully verified. The EEG signal collected by this system has a high signal-to-noise ratio, strong anti-interference ability, and is very stable. The built brain monitoring system performs well in real-time and stability. Keywords: TBI, Neurocritical care, Multi-modal, Consciousness Assessment, seizures detection, deep learning, CNN, IoT.