Non-Invasive Dry Eye Disease Detection Using Infrared Thermography Images: A Proof-of-Concept Study

<b>Background/Objectives:</b> Dry Eye Disease (DED) significantly impacts quality of life due to the instability of the tear film and reduced tear production. The limited availability of eye care professionals, combined with traditional diagnostic methods that are invasive, non-portable, and time-consuming, results in delayed detection and hindered treatment. This proof-of-concept study aims to explore the feasibility of using smartphone-based infrared thermography (IRT) as a non-invasive, portable screening method for DED. <b>Methods:</b> This study included infrared thermography (IRT) images of 40 subjects (22 normal and 58 DED). Ocular surface temperature changes at three regions of interest (ROIs): nasal cornea, center cornea, and temporal cornea, were compared with Tear Film Break-up Time (TBUT) and Ocular Surface Disease Index (OSDI) scores. Statistical correlations and independent <i>t</i>-tests were performed, while machine learning (ML) models classified normal vs. DED eyes. <b>Results:</b> In these preliminary results, DED eyes exhibited a significantly faster cooling rate (<i>p</i> < 0.001). TBUT showed a negative correlation with OSDI (r = -0.802, <i>p</i> < 0.001) and positive correlations with cooling rates in the nasal cornea (r = 0.717, <i>p</i> < 0.001), center cornea (r = 0.764, <i>p</i> < 0.001), and temporal cornea (r = 0.669, <i>p</i> < 0.001) regions. Independent <i>t</i>-tests confirmed significant differences between normal and DED eyes across all parameters (<i>p</i> < 0.001). The Quadratic Support Vector Machine (SVM) achieved the highest accuracy among SVM models (90.54%), while the k-Nearest Neighbours (k-NN) model using Euclidean distance (k = 3) outperformed overall with 91.89% accuracy, demonstrating strong potential for DED classification. <b>Conclusions:</b> This study provides initial evidence supporting the use of smartphone-based infrared thermography (IRT) as a screening tool for DED. The promising classification performance highlights the potential of this approach, though further validation on larger and more diverse datasets is necessary to advance toward clinical application.