Developing a compact machine learning-based predictor for detecting above-mild ocular surface disease index scores

Mohd Gharib, Nur Shazwani and Che Azemin, Mohd Zulfaezal and Badarudin, Noor Ezailina and Abdul Rahim, Muhammad Afzam Shah and Shanthini, Sharon Viola and Aljarousha, Mohammed (2026) Developing a compact machine learning-based predictor for detecting above-mild ocular surface disease index scores. Medical hypothesis, discovery & innovation in optometry, 7 (1). pp. 41-49. ISSN 2693-8391

PDF - Published Version Restricted to Registered users only Download (985kB) | Request a copy

Abstract

Background: The 12-item Ocular Surface Disease Index (OSDI) is widely used to assess dry eye disease (DED) severity; however, its length may reduce patient compliance and clinical efficiency. Given the expanding role of machine learning in ophthalmology, we aimed to develop and validate a compact OSDI predictor using supervised machine learning techniques to improve the efficiency and accuracy of DED assessment. Methods: This retrospective study analyzed a dataset of complete 12-item OSDI questionnaires obtained from adult residents of the Gaza Strip. OSDI scores were recalculated using the standard scoring formula, with “not applicable†responses treated as missing and excluded from the denominator. Participants were categorized as having moderate-to-severe dry eye (OSDI > 22) or not. Three supervised machine learning models—decision tree, support vector machine, and logistic regression—were developed using Python. Binary feature-importance analysis was initially performed using the full 12-item OSDI, after which each model was retrained using only questionnaire items with a binary feature-importance value of 1. Model performance was evaluated using accuracy, sensitivity, specificity, and precision. Results: Among 452 participants (mean [standard deviation] age, 32.0 [11.8] years; 52.9% male), 252 (55.8%) were classified as having moderate-to-severe dry eye, 200 (44.2%) were not. In the reduced-feature testing models, support vector machine model demonstrated the best overall performance, achieving 94.5% accuracy, 98.0% sensitivity, 90.0% specificity, and 92.6% precision. Logistic regression also showed strong performance, with 93.4% accuracy, 98.0% sensitivity, 87.5% specificity, and 90.9% precision. The decision tree model yielded lower testing accuracy (78.0%) and sensitivity (70.6%) but maintained relatively high specificity (87.5%) and precision (87.8%). Feature-importance analysis identified sensitivity to light, gritty sensation, computer or bank machine use, windy conditions, low-humidity environments, and air-conditioned places as informative predictors in the decision tree model. Both support vector machine and logistic regression models identified gritty sensation, painful or sore eyes, blurred vision, reading, watching television, and air-conditioned places as key predictors. Conclusions: Supervised machine learning models, particularly support vector machine and logistic regression models, effectively classified moderate-to-severe dry eye using recalculated standard OSDI scores and reduced feature sets. The identified predictors underscore the importance of ocular discomfort, visual disturbance, sustained visual activities, and environmental triggers in DED symptom severity. These findings support the potential utility of machine learning-assisted tools for symptom-based DED screening and severity assessment. Further validation in independent clinical populations and integration with objective diagnostic measures are warranted

Item Type:	Article (Journal)
Uncontrolled Keywords:	dry eye disease, questionnaire, ocular surface disease index, OSDI-6, machine learning, AI (artificial intelligence), learning, deep, learning, machine, vision, sciences, optometrist
Subjects:	R Medicine > RE Ophthalmology
Kulliyyahs/Centres/Divisions/Institutes (Can select more than one option. Press CONTROL button):	Kulliyyah of Allied Health Sciences Kulliyyah of Allied Health Sciences > Department of Optometry and Visual Science
Depositing User:	Dr. Mohd Zulfaezal Che Azemin
Date Deposited:	03 Jun 2026 11:59
Last Modified:	03 Jun 2026 11:59
Queue Number:	2026-05-Q3572
URI:	http://irep.iium.edu.my/id/eprint/129155

Actions (login required)

View Item