Application of Interpretable Machine Learning Models Based on Ultrasonic Radiomics for Predicting the Risk of Fibrosis Progression in Diabetic Patients with Nonalcoholic Fatty Liver Disease

Abstract

Patients with nonalcoholic fatty liver disease (NAFLD) and type 2 diabetes mellitus (T2DM) face a significant risk of hepatic fibrosis. Liver stiffness measurement (LSM) is commonly used to exclude advanced fibrosis, but its effectiveness in predicting fibrosis progression, especially in initially fibrosis-free patients, remains under-investigated. Although radiomics and machine learning (ML) models show promise in interpreting intricate data and predicting clinical outcomes, their application in assessing the fibrosis progression risk has not been fully explored. This study aimed to address this gap by developing and validating ML-based models to identify patients at risk of fibrosis progression using clinical data and multimodal radiomics features, thereby enhancing NAFLD and T2DM management. Methods: The study involved a retrospective analysis of 618 diabetic patients with NAFLD. These patients were divided into training and external validation cohorts. Based on LSM values, patients were classified into "Low-risk" and "Fibrosis-risk" groups. Radiomics features from multimodal ultrasound imaging were extracted, standardized, and utilized to develop various ML models. The models were internally validated based on these radiomics or clinical data, and the optimal model's feature importance was analyzed using the Shapley Additive Explanations (SHAP) approach, followed by external validation. Results: Of the 618 patients, 18.1% demonstrated an LSM≥6.5kPa, indicating a higher risk of hepatic fibrosis. The study identified 25 significant fibrosis-related radiomics features, with the support vector machine (SVM) model demonstrating superior performance in both internal and external validations. The SHAP analysis identified five key determinants of fibrosis risk, which included three radiomics features from shear wave elastography (SWE) and two from grayscale imaging. Conclusion:This study demonstrates the utility of an SVM model based on radiomics features derived from SWE and grayscale imaging for predicting fibrosis progression in diabetic patients with NAFLD, thereby enabling timely and effective therapeutic interventions.

Machine Learning Model for Non-Alcoholic Steatohepatitis Diagnosis Based on Ultrasound Radiomics

“…In this study, we utilized machine learning (ML) models due to their ability to handle complex nonlinear relationships between variables and outcomes, surpassing traditional linear prediction models (21).The shape value analysis obtained through the ML LightGBM algorithm revealed that body weight and LDL-C play a primary role in predicting NASH, which aligns with the ndings of Vilar-Gomez E's research. They discovered a correlation between weight loss and improvement in histological features of NASH tissue.…”