Enhanced survival prediction using explainable artificial intelligence in heart transplantation.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Transplantation Medicine

Background:

• The scarcity of donor organs limits heart transplantation success.
• Accurate prediction of transplant outcomes can optimize organ utilization and patient survival.
• The accuracy-interpretability trade-off in machine learning models is a key challenge in clinical decision support.

Purpose of the Study:

• To model 1-year mortality in heart transplantation using a self-explaining neural network.
• To benchmark the self-explaining model against a deep learning model.
• To assess the interpretability and predictive performance of machine learning models on tabular clinical data.

Main Methods:

• Development and external validation of a self-explaining neural network and a deep learning model.
• Utilized two independent datasets: UNOS transplants (2017-2018) and Scandinavian transplants (1997-2018).
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and calibration analysis.

Main Results:

• Self-explaining and deep learning models demonstrated comparable predictive performance in both datasets.
• AUROCs for the UNOS dataset were 0.628 and 0.635, respectively.
• The self-explaining model showed good calibration on the Scandinavian dataset (AUROC 0.626 with missing data, 0.634 without).

Conclusions:

• Transparent, self-explaining neural networks can achieve full predictive performance on tabular data.
• Interpretability does not necessitate a compromise in accuracy for clinical prediction models.
• These findings support the use of interpretable machine learning for enhanced clinical decision support in heart transplantation.