Artificial intelligence in predicting macular hole surgery outcomes: A focus on optical coherence tomography parameters

Purpose To evaluate the predictive performance of optical coherence tomography (OCT)-based indices and artificial intelligence (AI) using a Generative Pre-Trained Transformer (GPT) model and compare them with traditional logistic regression in forecasting anatomical success following macular hole (MH) surgery. Methods This retrospective observational study included 51 eyes of 51 patients who underwent pars plana vitrectomy for idiopathic MH. Preoperative OCT measurements of macular hole index (MHI), traction hole index (THI), hole form factor (HFF), basal hole diameter (BHD), and minimum hole diameter (MHD) were recorded. GPT-based AI predictions were generated using masked input data. A logistic regression model was developed with the same variables. Predictive performance was assessed using accuracy, area under the curve (AUC), positive predictive value (POPV), negative predictive value (NPV), and Kappa statistics. Results Anatomical success was achieved in 72.5% of cases. MHI, THI, and HFF were significantly higher in the successful group (p<0.0001). GPT achieved an accuracy of 77.0% and AUC of 0.770, with perfect POPV (1.000) but low NPV (0.452). Logistic regression outperformed GPT, achieving an accuracy of 84.3%, an AUC of 0.759, a higher NPV (0.800), and better agreement (Kappa 0.568 vs. 0.392). BHD and MHD showed poor predictive power (AUC 0.291). Conclusion OCT-derived indices, especially MHI, THI, and HFF, effectively predict MH surgery outcomes. Logistic regression based on actual patient data demonstrated superior predictive performance compared to GPT. AI models hold potential but require further development, integration of multimodal data, and validation before clinical application.