Machine Learning Prediction Model of Waitlist Outcomes in Patients with Primary Sclerosing Cholangitis
Xun Zhao1, Maryam Naghibzadeh1, Yingji Sun1
1Ajmera Transplant Program, University Health Network, Toronto, ON, Canada.
Transplantation Direct
|April 1, 2025
Summary
A new machine learning model accurately predicts liver transplant waitlist outcomes for primary sclerosing cholangitis (PSC) patients. This Random Survival Forest model outperforms existing MELD scores, improving predictions for this specific liver disease.
Area of Science:
- Hepatology
- Transplant Surgery
- Machine Learning in Medicine
Background:
- Primary Sclerosing Cholangitis (PSC) patients face disparities in liver transplantation access.
- The Model for End-Stage Liver Disease (MELD) score may inaccurately predict waitlist mortality for PSC patients due to unique comorbidities.
- Accurate prediction of waitlist outcomes is crucial for managing PSC patients awaiting liver transplants.
Purpose of the Study:
- To develop a more accurate predictive model for liver transplant waitlist outcomes in PSC patients.
- To incorporate complex clinical and disease-specific variables into a predictive model.
- To improve upon the predictive accuracy of existing MELD scores for PSC.
Main Methods:
- Developed three machine learning architectures using data from 4,666 PSC patients (Scientific Registry of Transplant Recipients - SRTR).
- Validated models on an institutional dataset of 144 PSC patients (University Health Network - UHN).
- Compared the time-dependent concordance index (C-index) of models against MELD-sodium and MELD 3.0 for mortality prediction.
Main Results:
- The Random Survival Forest (RSF) model demonstrated superior performance over MELD-sodium and MELD 3.0 in both SRTR and UHN datasets.
- RSF achieved C-indices of 0.868 (SRTR) and 0.771 (UHN), outperforming MELD scores.
- An RSF model trained with PSC-specific variables on UHN data achieved a C-index of 0.91.
- Key predictive features included high MELD score, white blood cell count, waitlist time, platelet count, autoimmune hepatitis-PSC overlap, AST, sex, age, stricture dilation history, and body weight extremes.
Conclusions:
- The developed RSF model provides more accurate waitlist outcome prediction for PSC patients.
- Integrating PSC-specific variables significantly enhances predictive performance, highlighting the need for disease-specific models.
- This approach can better predict clinical trajectories for distinct disease presentations like PSC.
