Explainable AI-Enhanced Ensemble Protocol Using Gradient-Boosted Models for Zero-False-Alarm Seizure Detection from...

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy impacts over 50 million globally, necessitating advanced seizure detection.
• Current automated systems face limitations in sensitivity, false alarms, or interpretability.
• Developing patient-independent, reliable seizure detection is crucial for clinical application.

Purpose of the Study:

• To present a patient-independent EEG-based seizure detection framework.
• To achieve high sensitivity with zero false alarms using interpretable machine learning models.
• To validate the framework's generalizability across different datasets and populations.

Main Methods:

• Extracted 27 time-, frequency-, and nonlinear-domain EEG features from 5-second windows.
• Trained five ensemble classifiers (XGBoost, CatBoost, LightGBM, Extra Trees, Random Forest) using leave-one-subject-out cross-validation.
• Employed SHAP and LIME for biomarker identification and interpretability.

Main Results:

• Achieved 95% sensitivity with zero false alarms in a pediatric cohort (CHB-MIT).
• Demonstrated cross-dataset generalization on an adult cohort (Siena Scalp EEG Database) with 95% event-level sensitivity.
• Identified key EEG biomarkers, including theta-band power and amplitude variability, consistent across populations.

Conclusions:

• Calibrated gradient-boosted ensembles with interpretable EEG features enable clinically safe seizure detection.
• The developed framework shows robust performance and cross-dataset generalizability.
• Findings support the use of interpretable machine learning for reliable epilepsy management.