Authors:
Gnaneswari Gnanaguru, S. Silvia Priscila, B. M. Praveen

Addresses:
Department of Information Technology, Institute of Engineering and Technology, Srinivas University, Dakshina Kannada, Karnataka, India. Department of Computer Applications, CMR Institute of Technology, Bengaluru, Karnataka, India. Department of Computer Science, Bharath Institute of Higher Education and Research, Chennai, Tamil Nadu, India. Institute of Engineering and Technology, Srinivas University, Dakshina Kannada, Karnataka, India.

Abstract:

Epilepsy is a recurrent and chronic neurologic condition that leads to a change in the psychological comfort and safety of patients, so elaboration of efficient tools of detection and early prediction is especially important. Conventional methods that rely on handcrafted features, CNNs, and LSTMs have performed well but are hampered by their inability to generalize findings from patient to patient, to integrate multiple inputs and modalities effectively, and to provide clinical interpretability. To address these difficulties, this paper proposes a multimodal framework combining EEG, ECG, and clinical metadata for robust seizure detection and prediction using a Temporal Fusion Transformer (TFT). The model leverages variable selection networks, static covariate encoders, temporal attention, and gated residual networks to capture both short- and long-term dependencies and ensure interpretability through feature importance and temporal heat maps. Experiments with benchmark datasets (CHB-MIT, TUH EEG Seizure Corpus) indicate that the proposed framework significantly surpasses the baseline models (CNN, LSTM, Transformer) in sensitivity (94.1%), specificity (92.3%), F1-score (93.6%), and AUC-ROC (0.96), and has a low false alarm rate essential to successful real-world deployment. These findings demonstrate the usefulness of TFT in consolidating multimodal evidence, making fewer false-positive predictions, and increasing clinical confidence, thereby supporting its application in real-time seizure management systems.

Keywords: Epilepsy Condition; Seizure Detection; Temporal Fusion Transformer (TFT); Seizure Prediction; Multimodal Deep Learning; Clinical Interpretability; Conventional Methods.

Received on: 22/03/2025, Revised on: 25/05/2025, Accepted on: 22/08/2025, Published on: 08/03/2026

DOI: 10.69888/FTSHSL.2026.000594

FMDB Transactions on Sustainable Health Science Letters, 2026 Vol. 4 No. 1, Pages: 59-72

• Views : 39
• Downloads : 15

Download PDF