Machine learning for 90-day mortality of liver transplant waitlist

Document Type

Conference Proceeding

Publication Date


Publication Title



Background: Machine learning (ML) models have been more utilized to predict clinical outcomes in medicine field. Our group recently reported the superiority of the neural network waitlist mortality prediction model compared to the MELD-based model in predicting liver transplant (LT) waitlist mortality. In this study, we explored further improvement in the model performance to predict LT waitlist mortality using variety of ML methods. Methods: This study uses data from the OPTN/UNOS for all 192,686 patients listed for LT between 2/27/2002 and 9/3/2020. Patients were excluded if they were listed for multi-organ transplant with thoracic organ(s), pancreas, and/or intestine. The dataset includes values for clinical and disease related variables at listing and longitudinal laboratory variables. A total of 150,361 patients were evaluated to create dataset/cohorts for 90-day waitlist mortality. The mortality outcomes in patients who had transplant within 90 days were imputed using multiple imputation. The imputed outcome data was used as main data for model development. Transplant was considered as a competing risk. Data before 01/01/2019 were randomly divided into a derivation (training) set and a validation set by a 2:1 ratio. Data after 1/1/2019 were used as an independent testing dataset. We used a strategy of a combination of feature selection and ML model for prediction. Specifically, LASSO penalized logistic regression with fivefold cross validation was used for feature selection. The features included each patients' baseline characteristics, patient-specific growth parameters for lab markers (i.e., slope coefficient[s]) and their interaction with baseline HCC status were entered into ML models to predict 90-day mortality. The selected variables were then fitted using the ML models for the training set. Results: The data was split into training (n=87517), validation (n=43105) and testing (n=19737) dataset. The multiple logistic regression followed by LASSO for variable selection performed that best for both of area under the ROC (AUROC) and area under the PR (AUPR). The prediction performance reached plateau at day 60, which indicates that the patients' data up to 60 days could be used to predict 90-day mortality with similar performance (Figure). The dynamic model has better prediction accuracy compared to MELD score for non-HCC patients. The difference in AUROC and AUPR between our model and MELD score were assessed via bootstrap. For all the patients, the AUROC of our model was 0.89 and 0.88 for validation and testing respectively and the AUPR was 0.68 and 0.61 for validation and testing respectively. Our model outperforms MELD score significantly (p<0.001) (Table). Conclusion: Our ML model outperforms MELD-based model in predicting 90-day LT waitlist mortality. The performance could be further improved by incorporating longitudinal laboratory values, which may lead to more equitable LT allocation.

PubMed ID

Not assigned.





First Page