Study of XAI in Metabolome Estimation
Please login to view abstract download link
Early detection of liver diseases such as non-alcoholic fatty liver disease (NAFLD) is important, but conventional diagnostic methods impose substantial financial and physical burdens. Machine learning models based on blood amino acid profiles provide a non -invasive approach to estimating liver fat accumulation. However, the basis of their predictions is often unclear, making interpretation difficult. Therefore, methods to improve model interpretability are required. In this study, a neural network was used to estimate liver fat accumulation from blood amino acid profiles, and explainable artificial intelligence (XAI) techniques were applied to interpret the prediction results. Results obtained using SHAP, Partial Dependence Plots (PDP), and LIME suggested that liver fat accumulation is influenced by multiple amino acids and that interactions among them may exist. Overall, the application of XAI was found to improve the interpretability of metabolome-based prediction models to some extent and to be potentially useful for understanding their prediction behavior.
