Explainability Analysis of Black Box SVM models for Hepatic Steatosis Screening

Abstract

Non-Alcoholic Fatty Liver Disease (NAFLD) or HS is one of the major causes of chronic liver diseases worldwide. Identifying the NAFLD condition at an early stage allows for preventative care and potential disease remission.To this end, our research group is addressing this issue by developing a computational model for decision support for Hepatic Steatosis (HS) or NAFLD screening. Our recent work included the development of machine learning models using seven physiological parameters (demographic, lipids, and liver biochemical parameters). Although the developed models show potential for screening, there is a need for further improving the model performance. Considering the complex nature of this condition and its interaction with different physiological parameters, we identified the contribution of the individual parameters in predicting the target (HS). The objective of this paper is to identify how different features contribute to a given model prediction by using an explainable artificial intelligence (XAI) technique called Partial Dependency. Results from partial dependency analysis and plots are summarized in this paper along with insights related to model performance. We identified the top three individual important predictors (ALT, AST, and Glucose levels) for both male and female. The models both for the male and female populations were analyzed separately to incorporate the pathobiological difference in NAFLD morphology in male vs female population.Clinical Relevance—The current study and obtained results do not have immediate clinical implications. However, this work paves the path for a potential computational model, which after required validation and testing, could be used as a decision support system for Hepatic Steatosis screening.