DBTR-AGF: Sieć Transformer-Rekurencyjna z Dwiema Gałęziami i Adaptacyjnym Łączeniem dla Przewidywania Napadów Epilepsji

Abstract Epilepsy is a chronic neurological disorder marked by unpredictable seizures, creating a pressing need for reliable early-warning systems. Despite advances in deep learning for EEG-based prediction, many models show weak patient-independent generalization, elevated false-positive rates, and sensitivity to limited data. We present DBTR-AGF (Dual-Branch Transformer–Recurrent with Adaptive Gating Fusion), which unifies a temporal branch (Transformer → BiLSTM) that captures global-to-local temporal dependencies with a spectral branch (Transformer → Temporal Convolutional Network) that extracts multi-scale spectral structure efficiently. A learnable adaptive-gating module fuses both representations on a per-sample basis. To improve robustness and reduce spurious alarms, we pair comprehensive regularization (Mixup, label smoothing, L2 weight decay) with Hidden Markov Model–based temporal smoothing at inference. On the CHB-MIT scalp-EEG dataset under a clinically motivated leave-one-subject-out protocol, DBTR-AGF attains 99.02% accuracy, 98.95% sensitivity, 99.91% specificity, and 0.12/h false-positive rate, outperforming recent baselines. Performance degrades by only 4.7% when training data are halved, indicating robustness in low-data regimes. These results suggest DBTR-AGF is a strong candidate for practical, patient-independent seizure forecasting.