Predicting Success in Descemet Membrane Endothelial Keratoplasty Using Machine Learning

Area of Science:

• Ophthalmology
• Medical Informatics
• Biostatistics

Background:

• Descemet membrane endothelial keratoplasty (DMEK) is a surgical procedure for corneal diseases.
• Early graft failure (GF) after DMEK can negatively impact visual outcomes.
• Predicting GF is crucial for optimizing patient selection and post-operative care.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting early graft failure (GF) after Descemet membrane endothelial keratoplasty (DMEK).
• To identify donor-related characteristics associated with GF risk.

Main Methods:

• Utilized five ML classification methods: random forest, support vector machine, k-nearest neighbor, RUSBoosted tree, and neural networks.
• Trained models on donor data including age, sex, systemic diseases, medications, ICU stay, death-to-preservation time (DPT), endothelial cell density, and surgical factors.
• Employed holdout validation with 75% data for training and 25% for testing; evaluated predictive accuracy, sensitivity, specificity, f-score, and ROC AUC.

Main Results:

• Achieved a maximum classification accuracy of 96% with ML models.
• Reported precision, recall, and f1-score values of 0.95, 0.81, and 0.90, respectively.
• Identified increased intensive care unit duration and death-to-preservation time (DPT) as significant predictors of GF risk (P < 0.05).

Conclusions:

• Machine learning models demonstrate high efficacy in automatically predicting early graft failure (GF) in DMEK.
• Elevated intensive care unit duration and DPT are strongly associated with increased GF risk.
• These findings support the use of ML for proactive risk assessment in DMEK recipients.