Contrastive self-supervised learning for epileptic seizure detection from EEG with calibration and robustness evaluation under physiology-aligned perturbations

Abstract

Labelled EEG data for epileptic seizure detection are scarce, and clinical recordings routinely suffer from realistic signal degradation including electrode loss, baseline drift, and sampling jitter. We present a proof-of-concept study applying contrastive self-supervised learning (SSL) to EEG epileptic seizure detection, comparing a SimCLR-style InfoNCE encoder pre-trained on unlabelled EEG against a supervised-only baseline sharing the same backbone, and evaluating both through a framework encompassing discrimination, operating-point behaviour, calibration, and physiology-aligned robustness. Stratified random 60/20/20 splits were applied (the Kaggle CHB-MIT derivative lacks patient identifiers; subject leakage is acknowledged as a limitation). On the held-out test set (11,233 segments, 23 channels, 6.26 s at 256 Hz, 6.8:1 imbalance), the SSL model achieves AUROC 0.990 (95% CI: 0.985–0.994) versus 0.989 (0.983–0.994) for the baseline, with overlapping CIs confirming statistical comparability. At threshold 0.5, SSL achieves higher recall (0.951 [0.925, 0.975] versus 0.924 [0.891, 0.953]) at the cost of lower precision (0.669 [0.626, 0.713] versus 0.839 [0.799, 0.878]); non-overlapping CIs confirm the precision gap; the recall advantage has overlapping CIs; both carry alarm-fatigue implications. The baseline is better calibrated on both ECE (0.027 [0.021, 0.034] versus 0.066 [0.057, 0.075]) and Brier score. In robustness tests, SSL is more robust to sampling jitter (± 20%: 0.968 [0.960, 0.976] versus 0.960 [0.947, 0.970]) while the baseline is more robust to strong drift (amplitude 0.40: 0.970 [0.960, 0.979] versus 0.913 [0.894, 0.931]); the drift gap has non-overlapping CIs, the jitter gap marginally overlapping CIs; both trace to augmentation design. Findings inform SSL augmentation design for future EEG monitoring work.