Op AI gebaseerde beslissingsondersteuning voor het monitoren van epilepsie

Epilepsy is one of the most prevalent neurological disorders worldwide, affecting approximately 50 million people. Despite the availability of several treatments, about 30% of patients remain resistant to pharmacological therapy, continuing to experience seizures throughout their lives, which significantly impacts their quality of life. Typically, patients are monitored in specialized Epileptic Monitoring Units (EMUs), where electroencephalography (EEG) is used to measure brain activity and detect seizures. This process is labor-intensive, as the data must be manually reviewed. Outside the hospital, seizure tracking usually relies on self-reported diaries, which are widely recognized as unreliable and inaccurate. Recent technological advances have enabled the development of wearable devices capable of continuously measuring multiple physiological signals in a non-intrusive manner. However, the longitudinal nature of wearable recordings poses an additional challenge: the large volumes of data generated must be analyzed automatically using artificial intelligence. Although many machine learning (ML) methods have been developed for seizure detection, these methods do not translate directly to wearable data. This is due to electrodes being placed in non-standard scalp locations, lower data quality in ambulatory scenarios, and the presence of noise and artifacts caused by hardware limitations, electrode misplacement, movement, or environmental factors. Therefore, the main objective of this thesis is to design ML frameworks specifically tailored to wearable data and to evaluate their clinical usability as decision-support tools for long-term monitoring of patients with epilepsy. The thesis is divided into two main parts, addressing focal and generalized absence seizures, as these subtypes present substantially different physiological patterns. All methods are based on data recorded with the Sensor Dot device (Byteflies), which measures behind-the-ear EEG (bte-EEG), electrocardiography (ECG), electromyography (EMG), and movement signals. The first part focuses on focal seizure detection based on wearable data. It begins with a comparative study of different ML methodologies trained on bte-EEG and ECG, establishing benchmarking frameworks and evaluation methods. Data-centric approaches are then introduced, demonstrating improved model performance and robustness. These frameworks are subsequently evaluated in clinical validation studies. In-hospital results show that algorithm-generated alarms reviewed by clinicians provide favorable clinical utility and detection yields, while home-based monitoring reveals limitations due to artifacts and the challenging nature of capturing ictal patterns generated by focal seizures using bte-EEG electrodes. The second part adapts and extends the ML frameworks to detect absence seizures, a generalized seizure type characterized by 3 Hz spike-and-wave discharges and transient loss of consciousness. The proposed multimodal pipeline combines bte-EEG-based models with a post-processing stage that incorporates movement data to filter artifacts, reducing false detections. Performance is further enhanced by tailored signal processing techniques, specifically normalization strategies, which reduce variance in noisy bte-EEG data. This results in more compact latent-space representations and improved class separability, thereby increasing robustness in uncontrolled environments. Given that annotated home-based data are scarce in real-world scenarios, this part also introduces a self-supervised learning paradigm that leverages unlabeled home recordings. Using generative approaches, the method embeds domain-specific knowledge in an unsupervised manner. Fine-tuning the pre-trained models with labeled hospital data achieves high sensitivity in detecting seizures. Overall, this thesis makes important contributions toward the integration of wearable devices into clinical practice. By developing ML frameworks and employing strategies to enhance robustness and performance, it demonstrates the feasibility of reliable seizure monitoring outside specialized hospital environments. These advances not only provide a more accurate alternative to seizure diaries but also reduce the clinical burden while improving monitoring yield, ultimately enabling more effective treatment strategies and paving the way for large-scale longitudinal trials using wearable devices.