Development and Clinical Validation of the DMEK Risk and Outcome Prediction (DROP) Score: A Dynamic Temporal Machine Learning Framework

TL;DR

The DROP Score is a new machine learning model that predicts outcomes for DMEK surgery by combining patient, donor, surgical, and center data.

Contribution

The DROP Score introduces a dynamic, multi-domain risk prediction framework validated with clinical and imaging data for DMEK outcomes.

Findings

High-risk patients had worse visual acuity and higher poor prognosis rates at 12 months.

Diabetes mellitus was the strongest prognostic factor for DMEK outcomes.

Tissue classification accuracy using IVCM imaging reached 96.2%.

Abstract

Background/Objectives: To develop and validate the DMEK Risk and Outcome Prediction (DROP) Score—a benchmarking model integrating patient, donor, surgical, and center-specific parameters for individualized risk assessment following DMEK. Methods: The DROP Score comprises four subscores, namely the Patient Risk Profile (PRP), Donor Tissue Quality (DTQ), Surgical Complexity Index (SCI), and Center Performance Factor (CPF), with literature-derived weights (α = 0.40, β = 0.25, γ = 0.20, δ = 0.15) validated by sensitivity analysis (K = 0.82–0.91). Clinical validation included 76 DMEK eyes and 89 controls (2019–2023). Machine learning models utilized EfficientNetV2B3 transfer learning with Random Forest classifiers and patient-level data partitioning. IVCM imaging comprised 6200 images. Results: The mean DROP Score was 39.35 ± 7.61 (Moderate: 92.1%; High: 7.9%). High-risk patients showed worse 12-month BCVA (0.50 vs. 0.31 logMAR) and higher poor prognosis rates (50.0% vs. 34.3%). The DROP Score showed significant correlations with BCVA (r = 0.305, p = 0.007) and ECD (r = −0.352, p = 0.002). Tissue classification accuracy reached 96.2%. Diabetes mellitus emerged as the strongest prognostic factor (OR: 4.34, p = 0.012), followed by hypertension (OR: 2.65, p = 0.078). Conclusions: The DROP Score provides transparent, individualized DMEK risk assessment. Diabetes mellitus and hypertension emerged as dominant systemic prognostic factors, while rebubbling showed no adverse impact on long-term outcomes. Complete four-domain validation requires ongoing prospective data collection.

Figures