Predicting DMEK Graft Rejection
By Jean Shaw
Selected by Prem S. Subramanian, MD, PhD
Journal Highlights
Translational Vision Science & Technology
2023;12(2):22
Download PDF
Using endothelial cell images, Joseph et al. investigated the ability of machine learning (ML) to predict future episodes of graft rejection in patients who underwent Descemet membrane endothelial keratoplasty (DMEK). They found that ML classifiers—algorithms that sort data into one or more categories—predicted future graft rejections one to 24 months before the rejections became clinically apparent.
For this study, the researchers obobtained endothelial cell images from 44 patients (44 eyes) who had undergone DMEK. Of these, half experienced graft rejection. The images were acquired via specular microscopy at multiple time points following surgery (range, 1-123 months). ML was used for segmentation of images from the patients’ last and second-to-last imaging visits. A bank of 432 features—which represented cellular and image intensity distribution, texture, and shape parameters—was developed; after highly correlated features were removed, the initial 432 features were reduced by approximately 50%. Random forest and logistic regression models were trained to predict the likelihood of future graft rejection.
The results showed that the novel ML classifiers used in this study outperformed those trained in traditional morphometrics, such as measuring endothelial cell density. Features most indicative of graft rejection were cellgraph spatial arrangement, intensity, and shape.
In their discussion, the authors wrote that this study “introduces the potential benefits of two alterations to the standard practice of care for postkeratoplasty patients: consistent and frequent specular microscopy imaging and machine learning models trained on novel quantitative features” outlined in the study. They added that their ML application could help clinicians “identify patients at risk for graft rejection.”
The original article can be found here.