Eeg-based detection of epileptic seizures in patients with disabilities using a novel attention-driven deep learning framework with SHAP interpretability

Epileptic seizures are neurological events caused by abnormal electrical activity in the brain, frequently resulting in loss of consciousness, involuntary movements, or cognitive deficits. Electroencephalograms (EEGs) are essential for diagnosing epilepsy, but conventional detection techniques depend on manual analysis, which can be labor-intensive and susceptible to inaccuracies. Recent developments in artificial intelligence (AI) and deep learning have facilitated the automation of seizure detection from EEG signals with improved accuracy. Nevertheless, current models frequently face challenges related to feature selection, interpretability, and computational demands. In this research, we introduce a cutting-edge deep learning methodology for the automated prediction of epilepsy, incorporating a Novel Attention Module (NAM) into a new Convolutional Neural Network (CNN) to improve the extraction of features from EEG signals. The proposed system employs Fourier Transform for feature extraction, utilizes Principal Component Analysis (PCA) for reducing dimensionality, and applies an optimized stochastic gradient descent approach with the Adam optimizer to enhance the learning process. We articulate the mathematical characteristics of feature selection driven by NAM, delineate the convergence attributes of the loss function, and present measures of explainability through Shapley Additive Explanations (SHAP). The model underwent training, validation, and testing with three publicly accessible EEG datasets sourced from PhysioNet and ResearchGate, thereby ensuring strong generalization across various datasets. A series of experiments were carried out to assess the effectiveness of the model by utilizing key performance metrics such as accuracy, sensitivity, specificity, and F1-score. The proposed methodology attained an accuracy of 99.3 %, an F1-score of 99.5 %, and both sensitivity and specificity at 99 %, showcasing its superior performance compared to existing models. Additionally, the computational complexity of the proposed framework was evaluated in terms of floating-point operations per second (FLOPs) and the total number of parameters, ensuring its efficiency for real-time biomedical applications. The incorporation of explainability techniques, including Shapley Additive Explanations (SHAP), enhances model transparency, which is beneficial for clinical decision-making. These findings suggest that the proposed attention-enhanced CNN framework serves as a reliable and interpretable tool for the early detection of epilepsy and patient monitoring.