Artificial intelligence-based multimodal framework for non-invasive detection of digital eye strain using thermal imaging and behavioral metrics

Abstract

Digital Eye Strain is an emerging occupational health concern with significant implications for digital well-being. To address the need for scalable and objective monitoring, EyeStrainNet is proposed. It is a multimodal and explainable health informatics framework that integrates thermal imaging and behavioral metrics for the non-invasive detection of Digital Eye Strain. Thermal images were captured pre- and post-screen exposure using a FLIR Edge Pro camera, and ocular temperature features were extracted from the inner and outer canthus and central cornea. Behavioral data, such as screen exposure duration and distraction levels, were recorded in parallel. A total of 197 samples (34 with significant strain, 163 without) were analyzed. Feature engineering and statistical analysis revealed strong correlations between ocular temperature changes and behavioral factors. The proposed EyeStrainNet, based on a one-dimensional convolutional neural network, was evaluated using 5-fold cross-validation. It achieved 97.5 % accuracy, 92.5 % precision, 94.3 % recall, 92.7 % F1-score, 99.7 % ROC-AUC, and 98.9 % PR-AUC, demonstrating strong performance with tight confidence intervals. EyeStrainNet outperformed baseline models such as One-Class SVM and XGBoost-SVM by 2–3 % in accuracy and 5–10 % in F1-score. SHAP-based explainability analysis identified temperature variation and distraction as dominant predictive features. This multimodal, explainable, and data-driven framework enables early-stage, non-clinical DES detection, promoting proactive digital wellness.