Subtyping of hepatocellular adenoma: a machine learning-based approach.

Abstract

Subtyping of hepatocellular adenoma (HCA) is an important task in practice as different subtypes may have different clinical outcomes and management algorithms. Definitive subtyping is currently dependent on immunohistochemical and molecular testing. The association between some morphologic/clinical features and HCA subtypes has been reported; however, the predictive performance of these features has been controversial. In this study, we attempted machine learning based methods to select an efficient and parsimonious set of morphologic/clinical features for differentiating a HCA subtype from the others, and then assessed the performance of the selected features in identifying the correct subtypes. We first examined 50 liver HCA resection specimens collected at the University of Washington and Kobe University/Kings College London, including HNF1α-mutated HCA (H-HCA) (n = 16), inflammatory HCA (I-HCA) (n = 20), beta-catenin activated HCA (β-HCA) (n = 8), and unclassified HCA (U-HCA) (n = 6). Twenty-six morphologic/clinical features were assessed. We used LASSO (least absolute shrinkage and selection operator) to select key features that could differentiate a subtype from the others. We further performed SVM (support vector machine) analysis to assess the performance (sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy) of the selected features in HCA subtyping in an independent cohort of liver resection samples (n = 20) collected at the University of Wisconsin-Madison. With some overlap, different combinations of morphologic/clinical features were selected for each subtype. Based on SVM analysis, the selected features classified HCA into correct subtypes with an overall accuracy of at least 80%. Our findings are useful for initial diagnosis and subtyping of HCA, especially in clinical settings without access to immunohistochemical and molecular assays.