CloudMedx’s Algorithms Predict Waitlist Dropout In Patients With Hepatocellular Carcinoma Awaiting Liver Transplantation with 80% Accuracy – Published by UCSF
October 1, 2019
Neil Mehta, Allison J. Kwong, Francis Yao, Bilal Hameed from Department of Medicine, Division of Gastroenterology/Hepatology, University of California and CloudMedx
Title: Machine Learning To Predict Waitlist Dropout In Patients With Hepatocellular Carcinoma Awaiting Liver Transplantation
In patients with HCC listed for liver transplantation (LT), models to predict waitlist dropout exist but often rely on liver disease variables present at listing instead of accounting for real-time clinical changes, such as with response to local-regional treatments. Machine learning algorithms can integrate available liver disease and tumor-related factors into an accurate waitlist dropout prediction calculator.
This machine learning algorithm (CloudMedX) was developed using the United Network for Organ Sharing (UNOS) database and extracted all available waitlist data on HCC patients listed for LT with MELD exception since 2002 (n=24,814) to model time to waitlist dropout. Clinical lab values and tumor burden were recorded at 3-month intervals for each patient with linear regression slope calculated for each of the 1181 distinct clinic and tumor-related variables. We trained our model using 80% of the cohort to predict the probability of dropout within 3, 6, or 12 months while accounting for the competing risk of LT, assessed for accuracy and area under the receiver operating curve (AUC). The model was validated in the remaining 20% of the cohort. To optimize our feature set for the machine learning algorithm, we used both the Spearman correlation and random forest feature importance approach.
The following list of six variables associated with waitlist dropout was ordered by Spearman’s Rank-Order: current INR, current bilirubin, bilirubin slope, presence of ascites, total viable tumor diameter, and current AFP. We trained three models at various time points from listing with these six variables that produced the best accuracy rates. Model accuracy for waitlist dropout prediction at 0-3 months from listing was 80%, at 0-6 months was 79%, and at 0-12 months was 78% with an AUC of 0.86 for this 0-12 month model. We also created a clinical calculator to predict the probability of dropout within 3, 6 and 12 months from listing accounting for the competing event of LT (Figure).
Using a machine learning algorithm, we developed and validated a comprehensive and highly accurate waitlist dropout calculator that accounts for real-time fluctuations in both liver disease (INR, bilirubin, and ascites) and tumor-related (AFP, total tumor diameter) characteristics. This score can be easily applied in the clinical setting to better inform providers and patients of the urgency of LT, and potentially whether extended criteria or live donor LT should be pursued to decrease waiting time.