Epileptic seizure prediction with deep learning-based fusion methods

Accurate prediction of epileptic seizures is important for patient safety and quality of life. This study aims to provide a fair, protocol-controlled comparison of two fusion strategies for EEG-based seizure prediction and to quantify their practical trade-offs. The decision-level pipeline combines posterior probabilities from two independently trained branches: a raw-EEG TCN → GRU with temporal attention model and an STFT-based 2D-CNN → GRU with temporal attention model. Fusion uses a simple calibrated type-2 rule tuned on validation data, and operating thresholds are set by Youden’s J. The feature-level pipeline uses the same two encoders—raw-EEG TCN → GRU and STFT-based 2D-CNN → GRU with temporal attention—to extract embeddings, which are then merged by a lightweight learnable fusion block before the final classifier. All networks are trained from scratch. Evaluation is conducted on the CHB-MIT dataset with stratified 5-fold cross-validation, reporting class-imbalance–robust metrics (PR-AUC and sensitivity at 5 % false-positive rate) in addition to ROC-AUC. The decision-level model attains accuracy 97.50 %, sensitivity 96.86 %, precision 97.57 %, F1 97.33 %, specificity 97.43 %, and AUC 0.99, with PR-AUC 0.994 and Sens@5%FPR 0.967. The feature-level model achieves accuracy 97.70 %, sensitivity 96.64 %, precision 98.47 %, F1 97.44 %, specificity 98.62 %, and AUC 0.99, with PR-AUC 0.995 and Sens@5%FPR 0.986. Post-hoc temperature scaling improved probability calibration (e.g., NLL from 0.089 → 0.083 at decision-level and 0.077 → 0.067 at feature-level) without affecting discrimination. An ablation with non-linear descriptors (Higuchi fractal dimension and fuzzy entropy) yielded modest average gains with added computational cost. These results delineate the conditions under which late posterior fusion versus early representational fusion is preferable and indicate that calibrated fusion improves robustness under realistic class imbalance.