Temporal shift and predictive performance of machine learning for heart transplant outcomes

Area of Science:

• Cardiology
• Medical Informatics
• Biostatistics

Background:

• Predicting heart transplant outcomes is crucial for patient care and organ offer decisions.
• Machine learning (ML) offers an efficient approach to analyze complex transplant data.
• The United Network of Organ Sharing (UNOS) database provides extensive data for such analyses.

Purpose of the Study:

• To train and test ML and statistical algorithms for predicting cardiac transplant outcomes.
• To evaluate the performance of various algorithms in predicting patient mortality post-transplant.
• To identify factors influencing prediction accuracy, including temporal data shifts.

Main Methods:

• Utilized the UNOS database, including 59,590 adult and 8,349 pediatric heart transplant patients (1994-2016).
• Evaluated three classification and three survival methods using shuffled and rolling 10-fold cross-validation (CV).
• Assessed predictive performance for 1-year and 90-day all-cause mortality via Area Under the Receiver-Operating Characteristic Curve (AUC).

Main Results:

• 12.4% of patients (8,394) died within one year post-transplant.
• Random Forest achieved the highest AUC (0.893) for 1-year survival prediction using shuffled CV.
• Rolling CV showed more modest, comparable performance among models, with XGBoost and logistic regression performing best.

Conclusions:

• ML and statistical models show potential for predicting post-transplant mortality.
• Prediction performance is constrained by temporal variations in patient and donor selection, as indicated by rolling CV results.
• Further research may be needed to address temporal shifts for improved long-term prediction accuracy.