Semi-automated patient-specific scalp EEG seizure detection with unsupervised machine learning

Epilepsy is a debilitating neurological disorder for which millions of people worldwide experience seizures. To diagnose these seizures for treatment, a patient typically stays in the hospital for days while a clinical staff must manually screen hours of electroencephalograph (EEG) data for electroencephalographic and behavioral seizures, a very time-consuming, tedious, and subjective process for humans. While decades of research have been dedicated to algorithms for alleviating this clinical procedural burden, there is still room for improvement in seizure detection techniques. Moreover, most seizure detection strategies use supervised machine learning, which requires subjective human involvement for training data. This research examines seizure detection performance for 24 patients using k-means, k-mediod, and hierarchical clustering, a Gaussian mixture model, and a hidden Markov model after using principal component analyses to reduce 45 measures to 2 measures. This work differs from past unsupervised methods, which were tested using very experimentally controlled seizure and non-seizure EEG signals: analyzing continuous EEG spanning 60 seconds pre-seizure to 60 seconds post-seizure in overlapping sliding windows rather than a single point for each seizure and non-seizure epoch. The `best´ combination of classifier accuracy, sensitivity, specificity, selectivity, and rejectivity is found with k-means and k-mediod algorithms.