Unsupervised Online Learning for Long-Term High Sensitivity Seizure Detection

Current seizure detection systems rely on machine learning classifiers that are trained offline and subsequently require manual retraining to maintain high detection accuracy over long periods of time. For a true deploy-and-forget implantable seizure detection system, a low power, at-the-edge, online learning algorithm can be employed to dynamically adapt to the neural signal drifts over time. This work proposes SOUL: Stochastic-gradient-descent-based Online Unsupervised Logistic regression classifier, which provides continuous unsupervised online model updates that was initially trained with labels offline. SOUL was tested on two datasets, the CHB-MIT scalp EEG dataset, and a long (>250 hours) human ECoG dataset from the University of Melbourne. SOUL achieves an average cumulative sensitivity of 97.5% and 97.9% for the two datasets respectively, while maintaining <1.2 false alarms per day. When compared with state-of-the-art, a moderate sensitivity improvement of 1-3% is observed on the majority of subjects and a large sensitivity improvement of >12% is observed on three subjects with <1% impact on specificity.