Deep learning for predicting fibrotic progression risk in diabetic individuals with metabolic dysfunction-associated steatotic liver disease initially free of hepatic fibrosis.

Objective: Metabolic dysfunction-associated steatotic liver disease (MASLD) significantly impacts patients with type 2 diabetes mellitus (T2DM), where current non-invasive assessment methods show limited predictive power for future fibrotic progression. This study aims to develop an enhanced deep learning (DL) model that integrates ultrasound elastography images with clinical data, refining the prediction of fibrotic progression in T2DM patients with MASLD who initially exhibit no signs of hepatic fibrosis. Methods: We enrolled 946 diabetic MASLD patients without advanced fibrosis, confirmed by initial liver stiffness measurements (LSM) below 6.5 kPa. Patients were divided into a training dataset of 671 and a testing dataset of 275. Hepatic shear wave elastography (SWE) images measured liver stiffness, classifying participants based on progression. A DL integrated model (DI-model) combining SWE images and clinical data was trained and its predictive performance compared with individual Image and Tabular models, as well as a logistic regression model on the testing dataset. Results: Fibrotic progression was observed in 18.1 % of patients over three years. During the training phase, the DI-model outperformed other models, achieving the lowest validation loss of 0.161 and highest accuracy of 0.933 through cross-validation. In the testing phase, it demonstrated robust discrimination with AUCs of 0.884 and 0.903 for the receiver operating characteristic and precision-recall curves, respectively, clearly outperforming other models. Shapley analysis identified BMI, LSM, and glycated hemoglobin as critical predictors. Conclusion: The DI-model significantly enhances the prediction of future fibrotic progression in diabetic MASLD patients, demonstrating the benefit of combining clinical and imaging data for early diagnosis and intervention.