Same pipelines, different answers: how evaluation design shapes EEG seizure prediction

In CHB-MIT scalp EEG seizure prediction, evaluation design dominates architecture: PSD+LDA ranges from AUC 0.815 with random window allocation to 0.572 with intact preictal blocks held out.

• Why it matters: Reported accuracies can look clinically useful, but they may reflect how test windows are separated from training rather than true generalization.
• What they did: Five fixed seizure-prediction pipelines were run on CHB-MIT with different subject- and within-patient partitioning, including a within-patient split that kept preictal blocks intact and a random window allocation; window manifests and partitions were released.
• The result: With whole-subject holdout, all pipelines stayed near chance (mean AUC 0.527–0.553), and an index-based within-patient split leaked raw EEG across train-test boundaries in all ten eligible cases.

AI summary of the abstract · may contain errors · not medical advice · read the paper.