An Explainable AI Framework Integrating Variational Sparse Autoencoder and Random Forest for EEG based Epilepsy Detection

Epilepsy is a medical condition characterized by sudden and frequent sensory disruptions which is commonly detected by electroencephalogram (EEG) analysis. However, analyzing these signals is challenging for traditional classifiers due to their non-stationary nature and high dimensionality. Deep learning (DL) techniques offer significant potential for fast and accurate medical decisions, especially when addressing imbalanced medical datasets. Therefore, this research proposes a novel artificial neural network architecture called the Variational Sparse Autoencoder (VSAE), which combines the strengths of a Sparse Autoencoder (SAE) and a Variational Autoencoder (VAE). The VSAE produces compact, sparse, and informative features for Random Forest (RF) classification, while Explainable AI techniques (XAI) methods like SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) enhance interpretability and transparency. LIME provides local interpretability whereas SHAP offers global interpretability by identifying the most influential EEG features contributing towards seizure detection. Additionally, 10-fold cross-validation (CV) is used to validate the proposed model. Compared to other conventional linear and non linear models, the proposed VSAE model demonstrates accuracy of 96.81%, precision 94.03%, recall 89.74% and F1 score 91.83% respectively.