Explainable Machine Learning Accurately Classifies Eye Gaze From Electrooculography Signals

Random Forest and related tree-based machine learning classifiers achieved strong subject-wise performance in classifying eye gaze from electrooculography signals while remaining physiologically interpretable through explainable artificial intelligence techniques, according to a study by Ayse Nur Ay Gul. Electrooculography represents a low-cost biosignal modality designed for gaze-based human–computer interaction, although establishing reliable validation protocols and achieving clear model interpretability have persisted as significant challenges for machine learning gaze classification systems.

To address these ongoing validation and interpretability challenges, the research proposed an explainable gaze classification framework evaluated on the UNIOVI-EOG dataset. The experimental data comprised continuous horizontal and vertical electrooculography recordings collected from nine participants. The researchers converted these biosignal recordings into non-overlapping temporal windows measuring 100 milliseconds each. To preserve data reliability, the filtering process retained only stable windows exhibiting at least 90 percent label purity, which ultimately produced an accepted dataset containing 81,481 windows distributed across nine distinct gaze classes.

The investigation extracted time-domain window features from these accepted segments and tested them using four distinct supervised learning models: Decision Tree, Random Forest, XGBoost, and CatBoost classifiers. Model evaluation was conducted under a subject-wise five-fold cross-validation scheme. This validation structure ensured that participants placed into the validation folds were never included in the training sets, thereby preventing subject overlap between training and testing. In addition, the researchers implemented simple mean-amplitude baseline tests alongside leakage diagnostic analyses to assess and support the validity of the classification framework.

All four tested machine learning algorithms achieved very high subject-wise performance when identifying the nine gaze categories. Among the evaluated approaches, the Random Forest classifier provided the strongest overall results. The authors reported that this robust performance was consistent with the controlled fixation protocol employed in the experiment, as well as the strong separability observed between the horizontal and vertical electrooculography amplitude patterns across the recorded gaze directions.

The research team assessed model interpretability through SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations, commonly designated as SHAP and LIME. These post-hoc explanations revealed that amplitude-related window features—most notably channel means, medians, quantiles, and absolute mean values—served as the primary contributors to the models’ classification decisions. Rank correlation between SHAP and LIME, along with a permutation-based faithfulness analysis, provided quantitative confirmation for these feature explanations. Overall, the findings suggest that leakage-controlled, window-level, explainable machine learning can deliver accurate and physiologically interpretable electrooculography gaze classification in controlled experimental settings.

Subject of Research: Explainable machine learning for electrooculography gaze classification

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

Article References: Ay Gul, A. N. (2026). Interpretable EOG-based gaze classification using tree-based models: A SHAP and LIME analysis. PLOS One, 21(10), e0361971. https://doi.org/10.1371/journal.pone.0361971

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0361971

Keywords: electrooculography, gaze classification, human-computer interaction, machine learning, tree-based models, Random Forest, CatBoost, XGBoost, Decision Tree, SHAP, LIME, UNIOVI-EOG dataset, subject-wise cross-validation, leakage diagnostics, faithfulness analysis, time-domain features, amplitude patterns