Abstract
MACHINE LEARNING SCORES ACCURATELY CLASSIFY INDIVIDUALS AT INDETERMINATE RISK OF INCIDENT CIRRHOSIS INTO LOW AND HIGH RISK GROUPS
Background: Risk stratification in non-alcoholic fatty liver disease (NAFLD) using non-invasive scores including Fibrosis-4 (FIB4) and NAFLD fibrosis score (NFS) is recommended by clinical guidelines. However, FIB4 and NFS values are indeterminate in 20-50% of patients with NAFLD. We aimed to develop machine learning models to improve upon FIB4 and NFS, especially in the indeterminate-risk range.
Methods: We included two cohorts, Michigan Medicine (MM) patients with NAFLD, and UK Biobank (UKBB) participants without excess alcohol intake or chronic liver diseases other than NAFLD. We adopted a two-stage approach to train in MM a model to identify incident cirrhosis. First, we used a least absolute shrinkage and selection operator (LASSO) model to select features. Second, we generated using those features a random forest (RF) model based on downsampling. We externally validated this model in UKBB and compared it to FIB4 and NFS based on time-dependent area under the curve (tAUC) models for risk of incident cirrhosis at 10 years. We compared incidence rates based on quintile of predicted risk by the RF model.
Results: The MM and UKBB cohorts included 28,684 and 480,651 participants, respectively. LASSO regression identified six predictors of incident cirrhosis in MM which were included in the RF model: age, AST, ALP, platelets, diabetes, and hypertension. In the UKBB, the RF model had 10-year tAUC of 0.84 (0.82-0.86) which outperformed FIB4 [0.81 (0.79-0.83)] and NFS [0.76 (0.73-0.78)], p<0.05 for both. The RF model identified considerable heterogeneity in risk of incident cirrhosis in patients with indeterminate FIB4 (1.3-2.67) or NFS (-1.455 to 0.675) with >10-fold variation of cirrhosis incidence between the bottom quintiles (1-2) and top quintile (5) of RF model risk (Table). Patients with indeterminate FIB4/NFS and the lowest 40% of RF model risk had incidence of cirrhosis similar to those with low FIB4 (<1.3) or NFS (<-1.455), and those in the top 20% had similar incidence to that of high NFS (>0.675) or FIB4 (>2.67). The RF model reclassified 30-40% of indeterminate-risk patients into high or low risk categories. These findings were consistent in subgroups including diabetes or elevated transaminases.
Conclusion: We trained and externally validated a simple RF model that identified heterogeneity among indeterminate-risk patients. This model accurately risk stratifies 30-40% of indeterminate-risk individuals using commonly-available predictors, thus providing a scalable pathway to optimize tertiary care referrals.