Shap-driven explainable AI with simulated annealing for optimized seizure detection using multichannel EEG signal

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Signal Processing

Background:

• Epilepsy seizure detection remains challenging, requiring accurate and interpretable models.
• Current methods often struggle with patient variability and data imbalance.
• Explainable AI (XAI) offers potential for transparent and reliable diagnostic tools.

Purpose of the Study:

• To develop a novel framework for enhanced seizure detection by integrating Explainable AI (XAI) with advanced optimization techniques.
• To improve both patient-specific and patient-non-specific seizure detection models.
• To introduce a new method, SHAP-RELFR, for effective patient-non-specific feature selection.

Main Methods:

• Utilized discrete wavelet transform (DWT) for multiband EEG feature extraction.
• Employed Simulated Annealing (SA) for Random Forest (RF) hyperparameter optimization.
• Applied SHAP (SHapley Additive exPlanations) values for feature selection, including the novel SHAP-RELFR method.
• Incorporated SMOTE (Synthetic Minority Over-sampling Technique) to address imbalanced datasets.

Main Results:

• The proposed methodology significantly improved seizure detection performance on CHB-MIT and Siena datasets.
• Achieved high average performance in patient-non-specific detection: 96.58% accuracy (CHB-MIT) and 94.81% (Siena).
• Demonstrated the effectiveness of SHAP-RELFR for patient-non-specific feature selection, enhancing model generalizability.

Conclusions:

• Combining XAI, SMOTE, and metaheuristic optimization yields enhanced seizure detection.
• The SHAP-RELFR method offers effective patient-non-specific feature selection, increasing clinical applicability.
• The framework provides interpretable and versatile seizure detection models, supporting clinical decision-making.