A reliable and explainable EEG-based eye state classification via XGBoost with missing data resilience and SHAP-LIME explanations

This study presents an interpretable, electrode-domain classification framework for Electroencephalography (EEG) eye-state detection using eXtreme Gradient Boosting (XGBoost), evaluated on the UCI EEG Eye State dataset. The approach operates directly on sensor-level voltage time series, bypassing handcrafted spectral or temporal feature extraction, making it suitable for low-latency and physiologically interpretable analysis. Eye-state discrimination from scalp EEG is challenging due to inter-subject variability, low signal-to-noise ratio (SNR), and spatially overlapping neural correlates of eye-open and eye-closed states. The proposed framework was benchmarked against conventional classifiers using rigorous validation protocols (3-, 5-, 7-, and 10-fold cross-validation and an 80/20 hold-out split), achieving 95.73% accuracy and 99.13% Area Under the Curve (AUC), with statistical significance confirmed via t-test, Wilcoxon, and Friedman tests (p<0.001). A comprehensive error–signal analysis, encompassing decision ambiguity, calibration reliability, and sample-wise hardness, showed that most misclassifications occurred in regions of high intra-class variance and inter-class similarity. Calibration curves demonstrated well-aligned predicted probabilities, supporting confidence-aware deployment. Explainability analysis using SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Gini importance consistently identified frontal (F7), parietal (P7), and occipital (O1) electrodes as key contributors, aligning with known neurophysiological mechanisms related to visuomotor processing and alpha-rhythm modulation during eye closure. Robustness was evaluated under random and targeted electrode dropout (1%–20%). Performance remained above 87.4% accuracy (AUC > 94.6%) with 20% missing channels, demonstrating graceful degradation.