Unsupervised detection of sub-sequence anomalies in epilepsy EEG.
Journal:
Computers in biology and medicine
Published Date:
Apr 24, 2025
Abstract
Seizures in electroencephalogram (EEG) data constitute a special case of sub-sequence anomalies in multivariate data with numerous challenges. These challenges include the irregular patterns exhibited even by the same individual, making seizures difficult to detect. The common approach seizure detection relies on supervised models trained specifically for individual patients. However, unsupervised anomaly detection techniques, despite their potential advantages - such as out-of-the-box usability, easy generalization to diverse individuals, and suitability for online applications - have not been explored as extensively as supervised models. This paper aims to address this gap by investigating the effectiveness of state-of-the-art unsupervised detectors for sub-sequence anomalies in EEG data. We present concrete evidence that unsupervised detectors can yield results on par with supervised models in detecting seizure sequences in EEG recordings. Extensive experiments demonstrate that unsupervised detectors can achieve up to 99.38% sensitivity (compared to up to 100% sensitivity in supervised models) without requiring hyper-parameter fine-tuning. Furthermore, the effectiveness and efficiency of different types of unsupervised algorithms are compared using a diverse set of metrics to ensure a fair and comprehensive evaluation. This paper aims to address this gap by investigating the effectiveness of state-of-the-art unsupervised detectors for sub-sequence anomalies in EEG data. We present concrete evidence that unsupervised detectors can yield results on par with supervised models in detecting seizure sequences in EEG recordings. Extensive experiments demonstrate that unsupervised detectors can achieve up to 99.38% sensitivity (compared to up to 100% sensitivity in supervised models) without requiring hyper-parameter fine-tuning. Furthermore, the effectiveness and efficiency of different types of unsupervised algorithms are compared using a diverse set of metrics to ensure a fair and comprehensive evaluation.