Predicting 1-Year Mortality After Pediatric Heart Transplantation Using Machine Learning

Abstract

BACKGROUND: Accurate prediction of post-heart transplant (HT) mortality is essential for donor selection, perioperative management, and equitable organ allocation in children. Traditional risk models often fail to capture nonlinear and temporally dynamic patterns, whereas machine learning (ML) approaches may improve predictive precision across eras. OBJECTIVES: The purpose of this study was to systematically evaluate ML-based models predicting 1-year mortality after pediatric HT, synthesize pooled diagnostic performance, and identify temporal shifts in major risk predictors. METHODS: PubMed, Scopus, and Embase were searched through December 2024. Five studies (n = 33,286) applying ML methods-including semisupervised forests, ensemble algorithms, SHapley Additive exPlanations-driven models, and Bayesian-tuned approaches-met inclusion criteria. Performance metrics were pooled using random-effects meta-analysis. Risk of bias was assessed with the Newcastle-Ottawa Scale, Prediction model Risk Of Bias ASsessment Tool, and Deeks' funnel plot asymmetry test. Era-stratified (1987-2000, 2001-2010, 2011-2020) Kaplan-Meier analyses examined survival and risk shifts, accounting for overlapping United Network of Organ Sharing cohorts with sensitivity analyses. RESULTS: = 72%-88%), reflecting differences in feature selection and validation. Temporal analyses showed a shift from ischemic time and pulmonary vascular resistance to extracorporeal membrane oxygenation/ventricular assist device dependence, donor-recipient mismatch, and socioeconomic factors. CONCLUSIONS: ML-based models outperform traditional approaches and capture evolving risk profiles across eras, supporting their integration into pediatric HT risk stratification to enhance accuracy and equity in transplant decision-making.

Ähnliche Arbeiten