E-POSTER DETAIL | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Validation of a machine learning model to predict graft survival post keratoplasty in an Indian cohort

Authors
Tanmay Gokhale, Poornima Tandra, Kavya Chandran, Pravin Krishna, Srinivas Rana

Presenting
Tanmay Gokhale

PURPOSE:
Keratoplasty is one of the most widely performed organ transplantation procedure globally. Graft survival post keratoplasty is dependent on a number of pre operative and intraoperative factors which can be estimated prior to surgery. The purpose of this study was to prospectively validate machine learning models to predict the outcome of keratoplasty

METHODS:
We developed time-to-event models to predict graft survival following corneal transplant surgery using right-censored data. Six survival models — including Cox PH, Weibull AFT, Random Survival Forest, and XGBoost Survival Embeddings — were trained with Optuna optimization across overall and surgery-type-specific variants. Models were first filtered by C-index, then the Random Survival Forest (RSF) was selected based on the minimum Integrated Brier Score over a 5-year horizon. Pre operative and post operative regression models to predict BCVA (logMAR) following corneal transplant surgeries used demographics, diagnosis, and visual acuity and postoperative models additionally incorporated intraoperative and early postoperative features. Nine algorithms were trained with Optuna hyperparameter optimization.

METHODS:

We developed time-to-event models to predict graft survival following corneal transplant surgery using right-censored data. Six survival models — including Cox PH, Weibull AFT, Random Survival Forest, and XGBoost Survival Embeddings — were trained with Optuna optimization across overall and surgery-type-specific variants. Models were first filtered by C-index, then the Random Survival Forest (RSF) was selected based on the minimum Integrated Brier Score over a 5-year horizon. Pre operative and post operative regression models to predict BCVA (logMAR) following corneal transplant surgeries used demographics, diagnosis, and visual acuity and postoperative models additionally incorporated intraoperative and early postoperative features. Nine algorithms were trained with Optuna hyperparameter optimization.

RESULTS:
On the test set, the Cox PH model achieved a C-index of 0.684 (minimum time-to-event) and 0.654 (maximum time-to-event), indicating moderate discriminative ability in ranking patients by graft survival risk. On the test set, the best preoperative model (XGBoost) achieved MAE 0.465 logMAR and R-squared 0.413 (N=1,222), while the best postoperative model (XGBoost) achieved MAE 0.397 logMAR and R-squared 0.575 (N=718), outperforming statistical baselines.

RESULTS:

On the test set, the Cox PH model achieved a C-index of 0.684 (minimum time-to-event) and 0.654 (maximum time-to-event), indicating moderate discriminative ability in ranking patients by graft survival risk. On the test set, the best preoperative model (XGBoost) achieved MAE 0.465 logMAR and R-squared 0.413 (N=1,222), while the best postoperative model (XGBoost) achieved MAE 0.397 logMAR and R-squared 0.575 (N=718), outperforming statistical baselines.

CONCLUSIONS:
The models are capable of estimating survival of a corneal transplant and the expected visual potential. This tool will help clinicians to counsel patients regarding expected outcomes post keratoplasty