Abstract
Purpose :
Prolonged diabetes leads to vision threatening retinopathy (DR) and macular edema (DME). Levels of inflammatory and angiogenic factors such as IL6, VEGF etc have been reported to be elevated in aqueous humor of DR and DME and may find application as companion diagnostics. Therefore, we investigated the potential for patient stratification based on aqueous humor (AH) biomarkers levels and harnessing the predictive power of such biomarkers in the context of disease progression.
Methods :
AH samples from controls and DR eyes were used to measure levels of 8 biomarkers (IL-10, IL-1β, IL-6, MMP-9, sICAM-1, IL-17A, TNF-α and VEGF-A). Biomarker levels from 191 controls and 175 treatment naïve DR eyes with macular edema were used as input in an artificial intelligence (AI) model based on random forest (RF) classifier. In a cohort of additional eyes with no clinical DR (n=404), subjects with VEGF-A levels higher than the AI detected threshold were grouped as vascular disease suspects (VDS). Further, in 517 DR eyes with and without DME where biomarker levels were available, 24 subjects were and followed-up to identify clinical progression using imaging markers (OCT and FFA).
Results :
The RF classifier achieved an AUC of 0.87 in identifying DR. The model demonstrated sensitivity of 70% and specificity of 91% with sICAM-1, VEGF-A and MMP-9 having the highest weightage. AI detected cut-off levels for these markers were: ICAM-1 > 1792, MMP-9 > 2423 and VEGF-A > 300 pg/ml. On follow-up, 66% of the risk group VDS progressed to mild or moderate NPDR.
About 78% of the naïve DME eyes had progressed to severe disease on follow-up. When these follow-up cases were run the RF AI model, 89% of these cases were correctly predicted as DME.
Conclusions :
DR predictability based on a single analyte was very low. The RF AI model accurately identified complex interactions between the 8 biomarkers with excellent predictability. AI determined biomarker thresholds were able to accurately determine progression in VDS eyes and naïve DME, suggestive of their predictive potential for clinical monitoring applications
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.