Shifting Determinants of Mortality Risk After Orthotopic Heart Transplantation Identified by Machine Learning.

Area of Science:

• Cardiology
• Transplantation Medicine
• Data Science in Healthcare

Background:

• Orthotopic heart transplantation (OHT) is the standard treatment for end-stage heart failure.
• Accurate prediction of postoperative mortality in OHT patients is crucial but challenging.
• Individualized risk assessment is needed to optimize patient outcomes.

Purpose of the Study:

• To develop and interpret random forest models for predicting 30-day and 1-year mortality after OHT.
• To identify key predictors of mortality and examine temporal shifts in their importance.
• To evaluate the performance of models using preoperative and postoperative data.

Main Methods:

• Analysis of 581 OHT patients (2012-2024) with 9.9% 30-day and 17.6% 1-year mortality.
• Development of random forest models using 87 preoperative and 48 postoperative variables.
• Model validation via five-fold cross-validation and explainability assessment using SHapley Additive exPlanations (SHAP).

Main Results:

• Preoperative models achieved AUCs of 0.62 (30-day) and 0.67 (1-year).
• SHAP analysis revealed distinct predictors: acute stress (hepatic dysfunction, inflammation, hemodynamic instability) for short-term, and renal function, metabolic reserve, frailty for long-term.
• Incorporating postoperative data significantly improved AUCs to 0.98 (30-day) and 0.86 (1-year), driven by organ dysfunction severity and persistence.

Conclusions:

• Random forest models integrating preoperative and postoperative data effectively predict short- and mid-term mortality post-OHT.
• SHAP analysis highlighted dynamic shifts in predictor importance, emphasizing the need for dynamic, data-driven risk assessment.
• These findings support enhanced, individualized patient management strategies in transplant care.