Multi-Stage Explainable Deep Learning System for Binary Epilepsy Classification

In this study, an explainable artificial intelligence-based multi-stage deep learning system for epilepsy detection from EEG signals is developed. The proposed method identifies information-rich regions of the signal with the captum integrated gradients approach and analyzes only significant time windows with the attribution-based masking technique. The 1D-CNN architecture underlying the model achieved high accuracy in discriminating healthy and epileptic signals. The baseline model achieved 88 percent accuracy and an F1-Macro value of 0.87. With the implementation of the XAI-guided model, the accuracy increased to 90 percent, the AUROC value increased to 0.96, and the learning stability of the model increased significantly. The models combined with the ensemble strategy showed the highest performance with 97 percent accuracy and 0.98 AUROC. In the calibration phase, the probability estimates were rescaled using the isotonic regression method, reducing the brier score to 0.06 and the expected calibration error to 0.76 percent. These results show that the system achieves not only high accuracy but also strong reliability and calibration success. The proposed captum IG+ ensemble+ calibration-based system provides a new reference framework for detecting explainable epilepsy in EEG signals and improves the effectiveness of interpretable artificial intelligence models in clinical applications.