NASHmap: clinical utility of a machine learning model to identify patients at risk of NASH in real-world settings

The NASHmap model is a non-invasive tool using 14 variables (features) collected in standard clinical practice to classify patients as probable nonalcoholic steatohepatitis (NASH) or non-NASH, and here we have explored its performance and prediction accuracy. The National Institute of Diabetes and Digestive Kidney Diseases (NIDDK) NAFLD Adult Database and the Optum Electronic Health Record (EHR) were used for patient data. Model performance metrics were calculated from correct and incorrect classifications for 281 NIDDK (biopsy-confirmed NASH and non-NASH, with and without stratification by type 2 diabetes status) and 1,016 Optum (biopsy-confirmed NASH) patients. NASHmap sensitivity in NIDDK is 81%, with a slightly higher sensitivity in T2DM patients (86%) than non-T2DM patients (77%). NIDDK patients misclassified by NASHmap had mean feature values distinct from correctly predicted patients, particularly for aspartate transaminase (AST; 75.88 U/L true positive vs 34.94 U/L false negative), and alanine transaminase (ALT; 104.09 U/L vs 47.99 U/L). Sensitivity was slightly lower in Optum at 72%. In an undiagnosed Optum cohort at risk for NASH (n = 2.9 M), NASHmap predicted 31% of patients as NASH. This predicted NASH group had AST and ALT mean levels above normal range of 0-35 U/L, and 87% had HbA1C levels > 5.7%. Overall, NASHmap demonstrates good sensitivity in predicting NASH status in both datasets, and NASH patients misclassified as non-NASH by NASHmap have clinical profiles closer to non-NASH patients.