Exploratory machine learning models for computer vision syndrome in occupational health.

Computer vision syndrome (CVS) is a growing occupational health concern linked to extensive digital device use, yet robust predictive models for risk identification remain scarce. This study aimed to evaluate and compare traditional and machine learning-based predictive models for CVS by assessing their accuracy in identifying CVS cases as defined by the CVS-Q© questionnaire, which served as the gold standard. A cross-sectional study was conducted among 90 Portuguese workers regularly exposed to digital display devices. Data were collected via a self-administered questionnaire covering age, sex, refractive errors, daily screen time, and symptoms assessed with the validated CVS-Q©. Participants were classified as symptomatic or non-symptomatic (CVS-Q© score ≥6). Predictive models were developed using logistic regression and four machine learning algorithms: random forest (RF), Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), with Synthetic Minority Over-sampling Technique (SMOTE) augmentation and repeated 5-fold cross-validation. Logistic regression and XGBoost demonstrated the highest accuracy (70.6%) and strong discriminative ability (AUC = 0.773 and 0.758, respectively). All models exhibited high sensitivity (≥72.7%), though specificity varied from 16.7% (random forest) to 50% (SVM, XGBoost). Calibration was acceptable across models (Brier scores 0.21-0.25). Female sex was the only statistically significant predictor in logistic regression (OR = 5.14, p = 0.012), while sex, screen time, and age were consistently the most influential variables in machine learning models. Refractive errors contributed minimally. Our results indicate that predictive modeling shows promise for CVS risk identification in occupational settings, particularly using demographic and behavioral variables. However, the limited and unrepresentative sample warrants caution, and further research with larger cohorts is needed to confirm these preliminary findings and assess their scalability.