Simulated allocation models (SAMs) are used to evaluate organ allocation policies. An important component of SAMs is a module that decides whether each potential recipient will accept an offered organ. The objective of this study was to develop and test accept-or-decline classifiers based on several machine-learning methods in an effort to improve the SAM for liver allocation. Feature selection and imbalance correction methods were tested and best approaches identified for application to organ transplant data. Then, we used 2011 liver match-run data to compare classifiers based on logistic regression, support vector machines, boosting, classification and regression trees, and Random Forests. Finally, because the accept-or-decline module will be embedded in a simulation model, we also developed an evaluation tool for comparing performance of predictors, which we call sample-path accuracy. The Random Forest method resulted in the smallest overall error rate, and boosting techniques had greater accuracy when both sensitivity and specificity were simultaneously considered important. Our comparisons show that no method dominates all others on all performance measures of interest. A logistic regression-based classifier is easy to implement and allows for pinpointing the contribution of each feature toward the probability of acceptance. Other methods we tested did not have a similar interpretation. The Scientific Registry of Transplant Recipients decided to use the logistic regression-based accept-decline decision module in the next generation of liver SAM.
- Machine learning
- Organ transplantation