Multichamber Myocardial Strain for Cardiovascular Risk Stratification in Liver Transplant Recipients Using Interpretable Machine Learning

BACKGROUND: Cardiovascular events are a leading cause of mortality after liver transplantation (LT), and existing risk scores incompletely capture myocardial vulnerability. OBJECTIVES: The objective of the study was to evaluate whether pre-LT myocardial strain improves machine-learning risk stratification for major adverse cardiovascular events (MACE) following LT. METHODS: We retrospectively studied 309 LT recipients with pretransplant echocardiography. Left ventricular global longitudinal strain, left atrial reservoir strain, and right ventricular free wall strain were measured. The primary endpoint was post-transplant MACE. Penalized Cox regression, random survival forest (RSF), and gradient boosting survival models were developed and benchmarked against coronary artery disease in LT and cardiovascular risk in orthotopic LT scores. Unsupervised clustering identified phenotypes. RESULTS: The mean age was 55.8 ± 11.4 years, and 33.3% were females. Over a median 4.6 years, 65 patients experienced MACE (21.0%), with worse right ventricular free wall strain and left ventricular global longitudinal strain. RSF showed the highest discrimination (C-index 0.636; integrated Brier score 0.138). Adding multichamber strain produced small, statistically nonsignificant changes in discrimination (RSF ΔC-index +0.019; gradient boosting +0.028) and none in penalized Cox regression. The models did not outperform penalized Cox regression or the established LT-specific risk scores. By SHapley Additive exPlanations, strain ranked among the most important contributors. Clustering identified "cardiometabolic" and "hepatic" phenotypes. CONCLUSIONS: Machine-learning models showed only modest discrimination for post-transplant MACE and did not outperform conventional approaches or established LT risk scores; the incremental value of strain was not statistically significant. Nevertheless, multichamber strain consistently ranked among the strongest contributors in feature attribution, suggesting complementary biological information. These findings are hypothesis-generating and support further evaluation of strain within LT risk models in larger, externally validated cohorts.