USE OF NEURAL NETWORK MODELS TO PREDICT MORTALITY/ SURVIVAL AMONG PATIENTS ON THE LIVER TRANSPLANT WAITLIST
Nagai S, Nallabasannagari AR, Moonka D, Reddiboina M, Nanna M, Chau LC, Yeddula S, Kitajima T, Bajjoka-Francis I, and Abouljoud MS. USE OF NEURAL NETWORK MODELS TO PREDICT MORTALITY/ SURVIVAL AMONG PATIENTS ON THE LIVER TRANSPLANT WAITLIST. Hepatology 2020; 72:2A-3A.
Background: While use of the MELD-Na scores have shown success in predicting waitlist mortality in liver transplant (LT), questions remain whether there are more efficacious models. The objective of this study is to develop Neural Network (NN) models that more accurately predict waitlist mortality. NNs are a type of Machine Learning algorithm. The fundamental building blocks of a NN are layers which are composed of units of calculation called neurons. NN performs calculations on input data and extracts meaningful patterns for the problem.
Methods: This study used data from the OPTN/UNOS registry, which includes data for 194,299 patients listed for LT between Feb 27, 2002 and Dec 31, 2018. Subsets of the data were used for the creation of 4 separate NN models. These models were constructed to predict mortality at different timeframes at 30, 90, 180, and 365 days. We excluded patients who received LTs before the outcome timeline, patients with liver cancer, patients who received MELD exceptions, and patients who were listed for combined organ transplants other than kidney. The Liver Data and the Liver Wait List History files in the OPTN/UNOS registry were combined and a total of 44 variables were selected, including recipient characteristics, trend of liver and kidney function during waiting time, UNOS regions, and registration year. Age, ethnicity, and gender were not included in the NN model to avoid assigning waitlist priority based on these factors. For each model, the data were split using random sampling into training, validation, and test dataset in a 60:20:20 ratio. The performance of the models was assessed using Area Under Receiver Operating Curve (AUC-ROC) and Area Under Precision-Recall curve (PR- AUC).
Results: According to NN prediction models, the AUC- ROC for 30-Day, 90-Day, 180-Day, and 365-Day Mortality was 0.949, 0.928, 0.915, and 0.899 and the PR-AUC was 0.689, 0.730, 0.769, and 0.823, respectively. The 90-Day Mortality NN model outperformed MELD score for both AUC-ROC and PR-AUC. It also outperformed MELD score for Recall (Sensitivity), Negative Predictive Value (NPV), and F-1 score. The 90-Day Mortality model specifically identified more waitlist deaths with a higher Recall (Sensitivity) of 0.833 vs 0.308 (P<0.001). MELD score performed better for Specificity and Precision. (Figure) The performance metrics were compared by breaking the test dataset into multiple subsets based on Ethnicity, Gender, Region, Age, Diagnosis Group, and Year of listing. The 90-Day Mortality NN model significantly outperformed MELD scores across all subsets of the data for predicting waitlist mortality.
Conclusion: Prediction models using NN more accurately identified waitlist mortality which outperformed MELD score. Using NN will improve predictive ability for waitlist mortality and lead to developing a more accurate and equitable allocation system with the ultimate goal of reducing LT waitlist mortality.