Home/Work/EEG Seizure Detection
EEG Seizure Detection
Multi-Architecture Benchmark on Pediatric EEG
Problem
Most published seizure detection models report numbers on a single architecture or subject, leaving open which model class actually generalizes across patients in real conditions.
Approach
Built an EDF-to-tensor preprocessing pipeline with MNE, sanitized corrupted PhysioNet timestamps, and used a weighted dataloader with BCEWithLogitsLoss for class imbalance.
Benchmarked 15+ architectures across LSTM variants, Mamba, Mixture-of-Experts, CNNs, Transformers, and classical ML under patient-disjoint splits.
Containerized the full PyTorch training pipeline for reproducible cross-environment experiments.
Results
15+ architectures. AUROC 0.740. 916 hours
- Best AUROC 0.740 under subject-independent evaluation on CHB-MIT
- 916 hours of pediatric EEG across 24 patients compared fairly
- Mamba and small transformers proved most robust under strict generalization
Stack
Learnings
Subject-independent evaluation is brutal but honest. Architectures that look strong on subject-dependent splits often collapse when patients are held out.