October 16, 2025Open Access

Artificial intelligence analysis of OCT biomarkers to predict visual outcomes following vitrectomy for epiretinal membrane

Key Points

Preoperative OCT biomarkers predicted visual recovery, with functional outcomes measured via best-corrected visual acuity.
The random forest model achieved an area under the curve of 0.71, identifying key biomarkers influencing outcomes.
Introductory analysis revealed thinner outer nuclear layer thickness relates to worse visual outcomes post-surgery.
Integrating automated biomarker analysis into assessments may enhance preoperative patient risk evaluations.

Abstract

Abstract Background Preoperative optical coherence tomography (OCT) biomarkers may help predict visual outcomes after idiopathic epiretinal membrane (ERM) surgery. Artificial intelligence (AI) enables automated, quantitative analysis of retinal structure, potentially improving prognostication. Methods In this multicenter, retrospective study, patients with idiopathic ERM who underwent pars plana vitrectomy (PPV) were included. Preoperative OCT volume scans were analyzed using an AI-based platform (Discovery OCT Biomarker Detector; RetinAI AG) to quantify retinal layer thicknesses and fluid biomarkers within the central 1 mm Early Treatment Diabetic Retinopathy Study (ETDRS) grid. Extracted features included outer nuclear layer (ONL), combined photoreceptor and retinal pigment epithelium complex (PR + RPE), retinal nerve fiber layer (RNFL) thickness, and intraretinal fluid (IRF) volume. A random forest classifier was used to evaluate the importance of these biomarkers in predicting 12-month best-corrected visual acuity (BCVA), categorizing patients as significant improvers (≥ 0.2 logMAR gain) or minimal/non-responders. Results A total of 71 eyes were analyzed. Mean BCVA improved from 0.51 ± 0.41 to 0.25 ± 0.33 logMAR at 12 months postoperatively ( P < 0.001). Thinner preoperative ONL thickness was strongly associated with worse final BCVA ( r = − 0.54), while thicker RNFL ( r = 0.28) and greater IRF volume ( r = − 0.26) were also linked to poorer outcomes. The random forest model achieved an area under the curve (AUC) of 0.71 for predicting visual improvement, identifying PR + RPE thickness, RNFL thickness, and ONL thickness as the most influential predictors. Conclusions Preoperative AI-derived OCT biomarkers, particularly indicators of outer retinal thinning and inner retinal thickening, are associated with limited visual recovery following ERM surgery. Integration of automated biomarker analysis into preoperative assessment may help identify patients at higher risk of suboptimal postoperative vision, informing surgical decision-making and patient counseling.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper