A machine learning approach to identify predictors of adverse events in posterior lamellar keratoplasty: a nationwide registry study

Purpose: To evaluate the influence of variations in practice patterns on the incidence of graft detachment and graft failure after posterior lamellar keratoplasty (DMEK/DSEK) with a machine learning approach.

Methods: All DMEK and DSEK procedures recorded in the prospective Dutch Cornea Transplant Registry between 2015 and 2018 were included and center-specific practice patterns were acquired using a 70-item survey. All available variables of the donor, graft, patient, practice patterns, and clinical course were collected for unsupervised machine learning analysis. Random oversampling using ROSE-sampling was performed to balance the database for the prediction model. A least absolute shrinkage and selection operator (LASSO) model and classification tree analysis (CTA) was used to select predictor variables and estimate subsequent effect sizes regarding graft detachment requiring rebubbling and graft failure.

Results: A total of 3647 transplants in 16 centers were included for analysis; 996 DMEK (27%) and 2651 (ultrathin-)DSEK (73%). The overall incidence of rebubbling was 17.4% for DMEK and 7.1% for DSEK. The incidence of graft failures was 8.4% and 3.8% respectively. The explained variance regarding clinical outcomes by the two models (LASSO/CTA) was 0.68/0.66 for rebubbling and 0.66/0.77 for graft failure. The classification tree revealed relevant protective variables and predictor variables.

Conclusions: Our analysis identified predictor variables in the multifactorial origin of graft detachment and graft failure following DMEK and DSEK. These outcomes provide insight in the mechanisms which lead to the development of adverse events following DMEK and DSEK, and could be used to as a guidance to reduce post-operative complications.