Prediction of Outcomes after Heart Transplantation Using Machine Learning Techniques

Purpose Machine learning (ML) techniques can improve predictive modeling over more traditional methods by identifying higher dimensionality and non-linear relationships between variables. We hypothesized that an AutoML algorithm would be superior to a logistic regression (LR) model for prediction of outcomes in HT. Methods The UNOS database was queried for patients receiving single-organ heart transplantation between 2006-2016. Pre-transplant variables for the donor and recipient were used for the prediction of one-year mortality or re-transplant. Auto machine-learning (AutoML) with stacking of Gradient Boosting Machine (GBM) derived algorithms was used to create a ML predictive meta-model. These results were compared to a traditional LR model using receiving operating characteristic (ROC) values. Results During this time period 18,612 patients with HT were identified. Observed one-year mortality or re-transplant was 11.5%. The AutoML derived model performed modestly (ROC=0.66, Figure 1) but showed improvement in outcome prediction over the LR model (ROC=0.62, Figure 2). Strongest predictive variables in the LR model were recipient bilirubin, creatinine, mechanical ventilation, donor age and ischemic time. The meta-model structure of AutoML precludes direct assessment of individual variable weight. Conclusion Using contemporary input from the UNOS database, AutoML meta-modeling outperformed LR for prediction of one-year outcomes in HT. Automation of predictive modeling using ML in HT is powerful, albeit limited by the “black box” effect on individual variables. AutoML in HT warrants further investigation.