Automated Epilepsy Diagnosis Using EEG With Test Set Evaluation.

Abstract

Several electroencephalogram (EEG)-based predictive models for automated epilepsy diagnosis have been proposed over more than a decade. However, to the best of our knowledge, none have been evaluated on a holdout/test set. A vast majority of these studies have reported accuracies above 95% on a benchmark EEG dataset, but the dataset has been shown here to have certain limitations when used for building classifiers for epilepsy diagnosis. We implemented two previously reported classifiers trained on the benchmark dataset whose accuracies were observed to drop sharply when evaluated on a test set. We propose a feature, engineered specifically, for epilepsy diagnosis that attempts to characterize the neuronal synchronization using scalp EEG by extending the concept of the impulse response of linear time-invariant systems to matrices. This feature was tested on the EEG of 50 epileptics and 50 healthy subjects and yielded an area under the curve (AUC) of 0.87. It outperforms the existing models implemented by us that gave the AUC of 0.80 when trained and tested on scalp EEG data, thereby, setting the new benchmark for automated epilepsy diagnosis on test set evaluation. The feature has also been shown to have statistical consistency across time and vigilance states with robustness against EEG artifacts.