| Popis: |
Liver transplant allocation policies evolve over time. With no universal algorithm to predict the outcome of liver transplantation allocation, donor-recipient matching still relies heavily on the experience of the transplant team. Machine learning models, using data collected on-site could offer more reliable and relevant ranking systems in comparison to traditional prognostic scores. 27 donor and recipient parameters were collected from 656 patients who underwent liver transplantation from March 2013 through December 2018 at the Merkur Clinical Hospital in Zagreb. We designed a machine learning survival ranking system, testing several different models: the regularized Cox regression, the random survival forest, gradient-boosted trees, and the support vector machine (SVM). The models were evaluated using nested cross-validation and interpreted using Shapley additive explanations (SHAP). We compared the performance of our various ranking systems and more traditional scores, namely BAR and ET-DRI. As a measure of comparison, we employed Uno’s concordance index estimator, which was appropriate for the high censorship rates in our data. Our best ranking system, based on gradient boosted trees, achieved an average concordance index of 65% on the data, outperforming traditional scores that achieved a maximum concordance index of only 53%. Random survival forests performed similarly to the gradient-boosted trees, while the regularized Cox and SVM models showed inferior ranking capability. Interpretation of the ranking systems yielded several relevant predictors of survival: donor CRP and sodium levels, liver steatosis, cold ischemia time, and recipient age. With machine learning, we can improve existing metrics by creating an interpretable ranking system that better fits our data. Such a system could potentially provide superior assistance in donor-recipient matching and identify risk factors in the process. |
