Abstract
Purpose :
Dry eye disease (DED) has a large disease burden and presents in the clinic with variations in clinical presentation and treatment dilemma. Therefore, it is important to accurately classify patients for treatment and monitoring to improve surgical outcomes. Since tear biomarkers have been shown to be associated with DED, we used AI based classifier strategy to define patients at risk of DED or associated ocular surface inflammation.
Methods :
8 biomarkers (IL-10, IL-1β, IL-6, MMP-9, ICAM-1, IL-17A, TNF-α and VEGF-A) were measured in tears collected in Schirmer’s strips from 640 subjects using multiplex ELISA and divided into clinical groups of controls and DED based on Schirmer’s test, TBUT and OSDI scores. An Artificial Intelligence (AI) model based on decision tree classifier (DTC) used the biomarker levels. Since MMP-9 emerged as the primary discriminator, the same cohort was divided into 4 groups based on MMP-9 statistical quartile ranges and AI determined threshold as Controls (from AI threshold, n = 318); SC-1 (Subclinical inflammation-1: low to moderate MMP-9, n = 72); SC-2 (Subclinical inflammation-2: moderate to high MMP-9, n = 101) and DED (extremely high MMP-9, n = 149). Random Forest (RF) based AI model was used to classify the cohort (MMP-9 was excluded from the AI model) with and without clinical parameters.
Results :
The DCT AI model had an AUC of 0.76 and correctly predicted 68% of Group-1 eyes and 71% of Group-2 eyes. In the second strategy, Controls and DED eyes significantly (p < 0.001) correlated with the clinical parameters (TBUT, Schirmer’s and OSDI). However, this correlation wasn’t significant in the sub-clinical groups (p>0.05) indicating the need for reclassifying them. The RF AI model (MMP9-based grouping) had an AUC of 0.79 with 94.8% sensitivity and 88.6% specificity. When clinical parameters were included in the model, the accuracy slightly improved to 0.81, and specificity towards DED eyes increased to 78%.
Conclusions :
MMP-9 cutoff level for controls was determined by AI model to subdivide apparently healthy eyes into sub-clinical groups at risk of DED. Biomarker-based subgrouping correlated well with clinical parameters-based subgrouping of the DED cohort which is useful for stratification of subjects in a clinical setting.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.