Interpretable EOG-based gaze classification using tree-based models: A SHAP and LIME analysis

Electrooculography (EOG) is a low-cost biosignal modality for gaze-based human–computer interaction, but reliable validation and model interpretability remain important challenges for machine learning-based gaze classification. This study proposes an explainable EOG gaze classification framework using UNIOVI-EOG dataset. Horizontal and vertical EOG recordings from nine participants were converted into 100 ms non-overlapping windows, and only stable windows with at least 90% label purity were retained. Time-domain window features were extracted and evaluated using Decision Tree, Random Forest, XGBoost, and CatBoost classifiers under subject-wise five-fold cross-validation, ensuring that participants in the validation folds were not included in training. The final dataset contained 81,481 accepted windows across nine gaze classes. All models achieved very high subject-wise performance, with Random Forest providing the strongest overall results. The high performance was consistent with the controlled fixation protocol and the strong separability of horizontal and vertical EOG amplitude patterns, as supported by simple mean-amplitude baseline and leakage diagnostic analyses. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME). The explanations showed that amplitude-related features, especially channel means, medians, quantiles, and absolute mean values, were the main contributors to classification decisions. SHAP–LIME rank correlation and permutation-based faithfulness analysis provided additional quantitative support for the explanations. These findings suggest that leakage-controlled, window-level, explainable machine learning can provide accurate and physiologically interpretable EOG gaze classification in controlled experimental settings.