Long-Term Clinical Outcomes of nDSAEK and Machine Learning-Based Prediction of Graft Survival in Corneal Endothelial Decompensation.

Purpose: To evaluate long-term outcomes of non-Descemet stripping automated endothelial keratoplasty (nDSAEK) in low-vision patients with corneal endothelial decompensation (CED), and to develop a machine-learning framework for predicting graft failure. Methods: This retrospective study included 114 eyes (Fuchs' and non-Fuchs' etiologies) treated with nDSAEK. Best-corrected visual acuity (BCVA), endothelial cell density (ECD), and complications were assessed over a long-term follow-up. Linear mixed-effects models (LMM) analyzed ECD kinetics. An XGBoost model using 10 clinical features was constructed to predict graft failure, interpreted via SHapley Additive exPlanations (SHAP) analysis. Results were compared against established endothelial keratoplasty benchmarks. Results: The median follow-up was 41.50 months (IQR: 25.65-55.75 months). Mean BCVA improved significantly from 1.80 logMAR baseline to 1.20 logMAR at 6 months, remaining stable thereafter. At 3 years, mean ECD was 1724 ± 279 cells/mm Conclusion: nDSAEK offers stable long-term visual and anatomical outcomes for CED. The integration of AI frameworks offers an exploratory framework for individualized prognostic screening, though further external validation is required before direct clinical integration.