Towards Label-Free Identification of Electrographic Seizure Patterns: A Machine Learning Anomaly Detection Approach

The electrographic seizure pattern (ESP) identification problem is usually addressed as a supervised classification problem. Nevertheless, in a clinical setting, access to human-labeled data can be limited due to the tedious and time-consuming process of ESP annotation. In addition, when working on longitudinal data, such annotations represent a practical burden for proper patient monitoring. In this work, we address this problem with a semi-supervised machine learning anomaly detection approach. We used intracranial electroencephalographic (iEEG) recordings from patients with drug-resistant epilepsy treated with Responsive Neurostimulation (RNS). From iEEG recordings without ESP, we extracted time-frequency features to train ensembles of one-class support vector machine (OCSVM) classifiers. Training data comprised no-ESP iEEG recordings from the first period after RNS implantation (baseline). The trained classifiers were then used to identify iEEG recordings with ESP as those classified as outliers. The models were initially assessed using baseline recordings and subsequently evaluated on longitudinal data. We investigated the impact of training data regularity and size on model performance and introduced a metric to assess training data quality for OCSVM training. Overall, we present a complete pipeline for ESP identification based on anomaly detection, opening new avenues for the development of robust, reliable, and fast subject-specific ESP detectors.