A predictive tool to forecast the Model for End-Stage Liver Disease score: the “MELDPredict” tool

Background: The Model for End-Stage Liver Disease (MELD) score is an effective score to assign and prioritize waitlisted patients for liver transplantation (LT) based on their liver sickness; however, there is no quantitative tool to predict future MELD progression. This study aims to develop and validate a method to forecast with high accuracy MELDNa [MELD with sodium (Na)—a refined MELD score]—based on patients’ previous observations. A successful method would provide a quantitative tool that can be used by clinicians to forecast future MELD scores for patients in need of a liver transplant; thus, better informing the timing of preparing a patient for a transplant. Methods: This is a retrospective study design using data from end-stage liver disease patients who were 18 years and older and were waitlisted to receive a LT. The final cohort included both patients who underwent LT and who were waitlisted but didn’t undergo LT. Both cohorts were derived from electronic health records (EHRs) from a Midwest institution between 2006 and 2021. Data included demographics, MELD scores, and time stamps. We carried out four experiments, each on a wide two classes of machine learning (ML) models: recurrent neural network (RNN) and ensemble learning (EL). All models are experimented within an autoregressive setup. We evaluate the performance of those models, with the baseline being linear regression (LR) model, to test what would be the best model to forecast future MELD scores based on real-world scenarios. Results: A final cohort of 1,498 patients were included. All experimented state-of-the-art ML models performed significantly better than LR models in all experiments. Extreme Gradient Boosting (XGBoost) and Random Forest Regression (RFR) models, specifically, achieve the best performance [best root mean square error (RMSE) across experiments was 2.576 and 2.582, respectively]. Our findings showed that models’ performance generally deteriorates when the forecasting window increases and the amount of training data declines. Conclusions: Our models demonstrated strong potential to accurately predict future MELD scores for end-stage liver disease patients in need of a liver transplant. They maintained consistent performance even with longer prediction intervals, allowing for a quantitative determination of future MELD scores. This study further integrated these models into an automated tool “MELDPredict”, which can serve as an adjunct quantitative tool for end-users’ clinical decision-making in determining the future course of MELD. Specifically, it has the potential to empower patients with information regarding their expected disease progression.