Trustworthy IoMT: Explainable Deep Learning (XAI) Framework for Automated Seizure Prediction from Multi-Channel EEG

Abstract

Epileptic seizure prediction remains a critical challenge in clinical neurology, particularly for patients with drug-resistant epilepsy. Recent advances in deep learning have improved predictive performance; however, the lack of interpretability and reliability limits their adoption in real-world healthcare settings. This paper proposes a trustworthy Internet of Medical Things (IoMT) framework for automated seizure prediction from multi-channel EEG signals, integrating explainable artificial intelligence techniques with a hybrid deep learning architecture. The proposed approach employs a CNN–BiLSTM model integrated with a channel-wise attention mechanism to enhance EEG preprocessing and feature extraction across various domains. SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) are used to show how the model makes decisions on both a global and a local level. The framework is evaluated using the benchmark CHB-MIT scalp EEG dataset through patient-wise cross-validation to mitigate data leakage. The average accuracy of the experiments was 94.7%, the sensitivity was 95.6%, and the AUC was 98.2%. Also, the calibration analysis shows a very small Expected Calibration Error of 0.018, which means the probability predictions are good. These results show that the new method does a great job of balancing accuracy, clarity, and trustworthiness. This makes it a good choice for helping doctors figure out when someone might have a seizure.