Machine learning seizure prediction: one problematic but accepted practice.

Area of Science:

• Neurology
• Machine Learning
• Biomedical Signal Processing

Background:

• Epilepsy is a common neurological disorder impacting quality of life.
• Accurate seizure prediction using electroencephalography (EEG) is crucial for timely intervention.
• Current machine learning approaches for EEG seizure prediction report high accuracy but lack clinical translation.

Purpose of the Study:

• Investigate a potential reason for the discrepancy between reported and clinical seizure prediction accuracy.
• Examine a commonly used data labeling method in EEG seizure prediction research.
• Identify and demonstrate a confound in machine learning-based seizure prediction.

Main Methods:

• Conducted an empirical study on a prevalent EEG data labeling technique for seizure prediction.
• Analyzed a method involving temporal grouping and random selection of EEG windows for validation.
• Trained an artificial neural network on artificially generated 'fake seizures' decoupled from biological signals within real EEG data.

Main Results:

• Identified that non-seizure signals, when labeled using the studied method, can create misleading decision surfaces for machine learning models.
• Demonstrated that this labeling approach can lead to falsely high prediction accuracy on validation datasets.
• Showcased that machine learning systems can inadvertently learn this confound, even with 'fake seizures'.

Conclusions:

• Many existing studies on machine learning for seizure prediction may report inflated accuracy due to this labeling confound.
• Future research must implement stricter requirements to mitigate this confound for clinically relevant results.
• Addressing this issue is critical for advancing the development of effective seizure prediction solutions.