Artificial Intelligence
in Liver transplant waitlist
How New AI Predictive Technology Could Cut Liver Transplant Waitlist Mortality
Liver transplantation (LT) is currently the single best way to extend the lives of those suffering from end-stage liver disease. The number of people in need of liver transplants continues to grow while the number of available livers does not.
It has been the goal of the OPTN (Organ Procurement and Transplant Network) to allocate organs to candidates according to medical urgency, while also minimizing wait time and excluding patients for whom transplantation is least likely to be successful.
Since 2002, the Model for End-Stage Liver Disease (MELD), changed to Meld-Na (MELD-Sodium) in 2016, has been used to allocate donor-livers to patients with the highest medical need. However, there are concerns as to whether the current MELD and MELD-Na driven designation and allocation process properly and equitably identify patients who are deemed to be at the highest risk for mortality while on the waitlist.
In this study, we explored whether there are more efficacious predictive models for identifying the risk of mortality on the waitlist for a liver transplant.
We leveraged Neural Networks to construct models to predict the risk of waitlist mortality at 30, 90, 180, and 365 days since a patient record is updated to UNOS (United Network for Organ Sharing).
We also compared the MeLD score against the 90-day mortality prediction model constructed in this study.
The data was acquired from UNOS which contained pre-transplant and post-transplant data on over 194K patients registered with OPTN.
Only patients who died or survived on the waitlist were included in this study. A total of 44 pre-transplant variables/features were extracted and processed to prepare data for training.
Four models were constructed using Neural Networks for predicting the risk of mortality on the waitlist at 30, 90, 180, and 365 days. The base network architecture comprised of an input layer, hidden layers, dropout layers, and an output layer. The network configuration and parameters for each outcome were determined through hyperparameter tunning
Model performance was evaluated for each outcome on multiple subsets of data based on Ethnicity, Gender, Age, Diagnosis, Region, and Year. The 90-day mortality prediction model was compared against MeLD score to assess the difference in performance.
Models training was performed using scalable cloud infrastructure on Google Cloud Platform (GCP) such as cloud storage and AI platform. Python and TensorFlow were used to preprocess data and construct Neural Networks.
30-Day NN Mortality Prediction
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
90-Day NN Mortality Prediction
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
180-Day NN Mortality Prediction
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
365-Day NN Mortality Prediction
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
In this study, we explored the ability of Machine Learning to recognize mortality risk for patients awaiting a liver transplant. The models constructed in this study were observed to have high performance and the 90-day mortality model significantly outperformed MeLD score.
These models can be used to recognize patients with a high risk of mortality on the waitlist for varying time intervals and can be used as an alternate to MeLD score. The methods used in this study to develop predictive models can be applied to post-transplant outcomes also.
This isn’t the only way that our research into AI in healthcare is improving patient outcomes. Check out our guide to 6 exciting applications of AI in healthcare.