AI-Based Visual Prognosis in Full-Thickness Macular Hole Surgery Using the ILM Flap Technique

Abstract

Objectives This study evaluated the clinical utility of a multimodal large language model-based artificial intelligence (AI) model in predicting postoperative visual outcomes following full-thickness macular hole (FTMH) surgery using the inverted internal limiting membrane (ILM) flap technique. Methods A retrospective analysis was conducted on 45 patients who underwent pars plana vitrectomy for FTMH at a tertiary eye care center between January 2021 and December 2023. Preoperative optical coherence tomography (OCT) images, demographic data, and clinical parameters were analyzed using the AI model to predict best-corrected visual acuity (BCVA) at four postoperative time points. The predicted BCVA values were then compared with actual clinical outcomes. Results Preoperatively, the mean AI-predicted BCVA was 1.13±0.20 logMAR compared with the actual value of 1.24±0.33 logMAR (p=0.192). At 6 months, the predicted BCVA was 0.65±0.20 logMAR versus the actual value of 0.67±0.25 logMAR (p=0.528), and at 12 months, it was 0.47±0.17 logMAR versus 0.55±0.27 logMAR (p=0.155). However, at 7 days postoperatively, the model significantly overestimated visual impairment, predicting 1.37±0.28 logMAR versus the actual value of 1.07±0.33 logMAR (p<0.001). Spearman correlation analysis showed the strongest association between AI-predicted and actual BCVA at 6 months (rs=0.5885, p<0.001), with a moderate correlation at 12 months (rs=0.4156, p=0.005), a weak correlation preoperatively (rs=0.3029, p=0.043), and no significant correlation at 7 days (rs=0.1949, p=0.199). Conclusion These findings demonstrate the model’s potential as a supportive tool for visual outcome prediction after FTMH surgery. The AI platform showed clinically relevant predictive performance for preoperative and later postoperative visual outcomes after FTMH surgery, particularly at 6 and 12 months, but was less reliable in the early postoperative period.