Automatic Detection of Epilepsy Seizures in Clinical Electroencephalography Records

Static rule-based algorithms to assist medical diagnosis and treatment monitoring have been used in practice for decades (e.g. heart rate monitoring). Research into detecting epileptic seizures from the electrical activity of the brain (Electroencephalography), with the use of machine learning algorithms, has been an active area of research for over 20 years; although currently is not commonly used in clinical practice due to limited accuracy. However, with more recent advancements in "big data", portable bio-sensing technologies, and other computer hardware/software, the feasibility of successfully implementing such algorithms into practice is improving.

This talk is an overview of the techniques that have been used in my research to develop algorithms for detecting generalized epilepsies. Using Bayesian optimization, we have assessed a variety of signal features in the time and frequency domains, as well as multiple classification pipelines that include feature selection and extraction steps. Over a series of studies, classifiers such as k-nearest neighbours, gradient boosted trees (lightGBM), and 1D Convolutional Neural Networks (CNN) have been shown to be the best classical, ensemble, and deep learning methods to mark diagnostic EEG records. Furthermore, the most important features of the signal to determine the presence of a seizure, appears consistent with those used by neurologists in practice.

This work is consistent with the broader literature applying machine learning to diagnostic imaging, in that new algorithms will soon provide physiologists with better quantitative tools to improve workflow and diagnostic accuracy.