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

Figures

• 

  Figure 1. PRISMA Study Selection This figure outlines the study screening and selection process using a PRISMA flow diagram. ML = machine learning.

• 

  Figure 2. Top Predictors of 1-Year Mortality This figure presents the 20 strongest predictors of 1-year mortality derived from SHAP-based feature importance across included machine learning models. (Bars show mean absolute SHAP values from study models, grouped by domain (donor = green, recipient = blue, procedural = purple, postop = red). BMI = body mass index; ECMO = extracorporeal membrane oxygenation; EF = ejection fraction; HLA = human leukocyte antigen; HT = heart transplant; PRA = panel reactive antibody; SGOT = serum glutamic-oxaloacetic transaminase; SGPT = serum glutamic-pyruvic transaminase; SHAP = SHapley Additive exPlanations; VAD = ventricular assist device.

• 

  Figure 3. SHAP Predictors by Transplant Era This figure compares leading SHAP-derived predictors of 1-year mortality stratified by transplant era. The era-based comparison highlights evolving clinical determinants of early outcomes following pediatric heart transplantation. Abbreviations as in Figure 2.

• 

  Figure 4. ROC Curves for ML Models The figure illustrates variability in discrimination across models while emphasizing overall robust predictive capacity. AUC = area under the curve; ROC = receiver operating characteristic curve; other abbreviation as in Figure 1.

• 

  Figure 5. Diagnostic Accuracy of ML Models This figure summarizes pooled sensitivity (A), specificity (B), and diagnostic ORs (C) from random-effects meta-analysis of included machine learning models. Pooled estimates demonstrated strong diagnostic accuracy (sensitivity 0.90, specifically 0.91, diagnostic OR 142).

• 

  Figure 6. Survival Curves by Era This figure presents Kaplan-Meier survival curves stratified by transplant era.

• 

  Figure 7. Study Quality and Bias Assessment (A) NOS domain scores for selection, comparability, and outcome assessment. (B) PROBAST ratings for risk of bias and applicability across participants, predictors, outcomes, and analysis domains. Most studies demonstrated moderate risk, with common limitations related to predictor handling and reporting quality. NOS = Newcastle–Ottawa Scale; PROBAST = Prediction model Risk Of Bias ASsessment Tool.

• 

  Figure 8. Publication Bias Analyses (A) Funnel plot symmetry with regression-based testing. (B) Trim-and-fill adjustments estimating potentially missing studies and their influence on pooled effect sizes.

• 

  Figure 9. Deeks’ Funnel Plot for Bias This figure displays Deeks’ funnel plot asymmetry test for diagnostic accuracy studies. The regression line and confidence bands assess small-study effects and publication bias. Nonsignificant asymmetry suggests limited evidence of bias influencing pooled diagnostic accuracy across included machine learning models.

• 

  Central Illustration. Machine Learning for Predicting 1-Year Mortality After Pediatric Heart Transplantation Central illustration summarizing a systematic review and meta-analysis of machine learning models applied to predict 1-year mortality in 33,286 pediatric heart transplant recipients from UNOS and PHTS registries (1987–2020). The figure highlights key machine learning approaches—including random forests, gradient boosting, SHAP-driven models, and Bayesian-tuned ensembles—alongside major predictors such as ECMO/VAD support, congenital heart defects, and renal dysfunction. Bias assessment and validation strategies included NOS, PROBAST domains, rolling cross-validation, and calibration audits. Pooled model performance showed AUROC of 0.86, sensitivity of 90%, specificity of 91%, and diagnostic odds ratio of 142. Findings support improved survival in the recent era and underscore implications for equitable organ allocation.