Extreme value theory inspires explainable machine learning approach for seizure detection.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Epileptic seizures are unpredictable neurological events requiring accurate detection for effective treatment.
• Current seizure detection methods often rely on supervised machine learning, which can lack generalization.
• Manual analysis of electroencephalogram (EEG) data for seizure detection is time-consuming and labor-intensive.

Purpose of the Study:

• To explore the application of extreme value theory and unsupervised outlier detection for automated seizure detection.
• To develop and test a novel method for identifying seizure events in EEG recordings.

Main Methods:

• Utilized extreme value theory to identify EEG features exhibiting extreme behavior during seizures.
• Applied a one-class Support Vector Machine (SVM), an unsupervised outlier detection algorithm, to classify seizure events.
• Trained the one-class SVM on individual patient data to account for inter-subject variability.

Main Results:

• The proposed approach achieved 77% sensitivity and 12% precision across 83 patients.
• In 60 patients, the sensitivity reached 100%, indicating high detection rates.
• The method demonstrated stability against between-subject variability due to subject-specific training.

Conclusions:

• The study demonstrates a viable convergence of extreme value theory and machine learning for unsupervised seizure detection.
• This approach offers a promising alternative to supervised methods, potentially improving the efficiency and accuracy of seizure analysis.
• The findings highlight the potential of physical concepts in machine learning for solving critical medical challenges.