Machine Learning-driven Patient-specific Early Seizure Detection for Neuromodulation Devices

Epilepsy is a chronic disorder of the brain that predisposes individuals to experiencing recurrent and unprovoked seizures affecting 50 million people worldwide. Recent advances in fundamental neuroscience and implantable electronics have enabled the development of neuromodulation devices for the treatment of epilepsy. Modern neuromodulation devices detect abnormal electrical activity in the brain associated with seizures and activate electrical stimulation to prevent seizures from occurring. Today, there is a growing trend towards integrating machine learning for seizure detection on such devices to improve their efficacy. This thesis assesses the suitability of current machine learning models for neuromodulation devices by evaluating their seizure detection performance and efficiency in resource-constrained environments. Particular emphasis is placed on comparing traditional machine learning to modern deep learning models. This thesis will show that, in the seizure detection context, deep learning models can be implemented in a compact and resource-efficient way despite their computational complexity.M.A.S