Unsupervised detection of sub-sequence anomalies in epilepsy EEG

Area of Science:

• Biomedical Signal Processing
• Machine Learning
• Neurology

Background:

• Seizures in electroencephalogram (EEG) data are challenging sub-sequence anomalies due to irregular patterns.
• Current seizure detection primarily uses patient-specific supervised models.
• Unsupervised anomaly detection offers potential benefits like generalizability and online application but is less explored.

Purpose of the Study:

• To investigate the effectiveness of state-of-the-art unsupervised anomaly detectors for sub-sequence anomalies in EEG data.
• To compare unsupervised detectors against supervised models for seizure detection.
• To evaluate the efficiency and effectiveness of various unsupervised algorithms.

Main Methods:

• Utilized state-of-the-art unsupervised anomaly detection algorithms.
• Applied algorithms to electroencephalogram (EEG) datasets for seizure sequence detection.
• Conducted extensive experiments to compare performance using diverse metrics.

Main Results:

• Unsupervised detectors achieved sensitivity up to 99.38%, comparable to supervised models (up to 100%).
• Results were obtained without the need for hyper-parameter fine-tuning.
• Demonstrated the viability of unsupervised methods for robust seizure detection.

Conclusions:

• Unsupervised anomaly detection provides a powerful, generalizable alternative for identifying seizure sequences in EEG.
• These methods show comparable performance to supervised approaches, offering advantages in usability and adaptability.
• Further research into unsupervised algorithms can enhance automated seizure detection systems.