Abstract
Background/Objectives: This study compared eight machine-learning classifiers for three dataset-defined digital eye strain (DES) categories and examined the robustness and interpretation of their predictions. Methods: A publicly available dataset containing 800 observations and 11 nonconstant predictors was analyzed using training-only preprocessing, repeated stratified cross-validation, randomized hyperparameter optimization, and an 80:20 stratified train–test split. Additional analyses included nested cross-validation of support vector machine (SVM) and logistic regression (LR), feature ablation, probability calibration, and SHAP-based interpretation. Results: SVM achieved the highest mean repeated-CV macro-F1 (0.9861). On the held-out test set, SVM achieved 0.9625 accuracy and 0.9636 macro-F1, while LR achieved 0.9688 and 0.9714, respectively. Nested-CV macro-F1 values were similar for SVM (0.9777) and LR (0.9768). Excluding symptom-related predictors reduced test macro-F1 to 0.5141 (SVM) and 0.4902 (LR); additionally, excluding daily screen time reduced these values to 0.3714 and 0.3724. Multiclass Brier scores were 0.0506 for SVM and 0.0456 for LR. Screen time and ocular symptoms also contributed prominently to SHAP explanations. Conclusions: Classification performance depended strongly on symptom and screen-exposure variables. The reported metrics quantify agreement with dataset-defined labels, not clinical diagnostic accuracy or the ability to predict future DES risk. The label-generation criteria are undocumented, supporting clinical gold-standard evidence is unavailable, and potential overlap between predictors and label-generation criteria cannot be excluded. Without validation against independently established clinical outcomes and in external populations, these findings remain dataset-specific classification results and do not establish real-world clinical risk-prediction capability.