Machine learning for detection of mild cognitive impairment using VR eye-tracking metrics in older adults: a cross-sectional study

Early detection of mild cognitive impairment (MCI) is critical for timely intervention in older adults, yet screening in community and primary care remains challenging. VR-based eye-tracking metrics may provide objective indicators of cognitive function. This study aimed to develop and internally validate machine learning models using these metrics to detect MCI in older adults. This cross-sectional study enrolled 80 older adults with MCI and 100 cognitively healthy controls (HC) from eight communities in Shaanxi Province, China, between February and June 2025. Participants completed VR-based eye-tracking tasks that assessed five oculomotor metrics: fixation offset number, pursuit offset number, pro-saccade latency, anti-saccade latency, and accuracy. Four classifiers including support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost) and logistic regression (LR) were trained and evaluated using leave-one-subject-out cross-validation (LOSO-CV). Model performance was assessed using AUC, accuracy, sensitivity, specificity, precision, and F1-score. Feature importance was evaluated using permutation importance analysis on the classifier. Among the four classifiers, the SVM achieved the best cross-validated performance, with an AUC of 0.849 (95% CI: 0.787–0.906), accuracy of 0.800 (95% CI: 0.744–0.856), sensitivity of 0.662 (95% CI: 0.562–0.763), specificity of 0.910 (95% CI: 0.850–0.960), precision of 0.855 (95% CI: 0.779–0.932), and F1-score of 0.746 (95% CI: 0.667–0.825) for MCI detection. Permutation importance analysis identified anti-saccade accuracy as the most influential feature (mean AUC decrease: 0.272, 95% CI: 0.168–0.388), followed by fixation offset number (0.074, 95% CI: 0.010–0.147) and pursuit offset number (0.055, 95% CI: 0.013–0.107). Anti-saccade latency contributed minimally (0.025, 95% CI: -0.019-0.075), whereas pro-saccade latency showed no detectable contribution to the model. The SVM classifier based on VR eye-tracking metrics shows potential as a detection tool for MCI among community‑dwelling older adults, with anti-saccade accuracy emerging as the strongest predictor. These findings are preliminary, and further external validation in independent cohorts is required before clinical implementation. Not applicable.