Machine learning identifies predictors of amniotic membrane graft failure in neurotrophic keratitis

Abstract Introduction Neurotrophic keratitis (NK) is a rare, vision-threatening corneal disease characterized by impaired epithelial healing and frequent recurrence of defects. Amniotic membrane transplantation (AMT) is an established therapeutic option, but outcomes remain variable and predictors of failure are poorly defined. We aimed to identify clinical and systemic factors associated with recurrence following AMT using both classical statistics and machine learning approaches. Methods We retrospectively analyzed 66 patients with NK who underwent AMT between 2019 and 2025 at a tertiary referral center. The primary outcome was epithelial defect recurrence following AMT, and the a priori primary model was a multivariable logistic regression complemented by machine learning approaches. A total of 66 patients had complete data for analysis; missing data were minimal and addressed using complete-case analysis. Clinical, demographic, and ophthalmological variables were retrospectively collected. Predictors of recurrence were assessed using bivariate tests, multivariable regression, principal component analysis (PCA), hierarchical clustering, and random forest modeling. Results Epithelial defect recurrence occurred in 46 patients (70%). Multivariable analysis identified systemic immunosuppressive therapy (odds ratio [OR] 19.9; p = 0.023) and number of AMTs (OR 2.73 per graft; p = 0.040) as independent predictors, with prior ocular surgery showing a borderline association (p = 0.060). PCA and clustering revealed three phenotypic subgroups, including a high-risk cluster (72% recurrence) characterized by immunosuppression, multiple AMTs, prior surgery, and inflammatory complications. Random forest classification confirmed the predictive role of these variables, achieving an AUC of 0.82 with balanced sensitivity (74%) and specificity (81%). Conclusion Systemic immunosuppression, repeated AMTs, and prior ocular surgery are key predictors of AMT failure in NK. Combining regression models, clustering, and machine learning provides a robust framework for risk stratification. These findings support the development of personalized monitoring and treatment strategies to improve surgical outcomes in NK.