AI-driven multimodal fusion of tongue images and clinical indicators for identifying MAFLD patients at risk of coronary artery disease: An exploratory study
Jingjing Zhang, Song Feng, Jian Xue, Chao Zhou, Shuangnan Fu, Pengcheng Liu, Zhaoyun He, Man Gong, Hui Feng, Shuang-Nan Zhou, Ning Zhang
Abstract
Metabolic dysfunction-associated fatty liver disease (MAFLD) is associated with coronary artery disease (CAD), but existing risk assessment tools lack precision and scalability. We developed an AI-driven multimodal framework integrating traditional Chinese medicine (TCM) tongue-based diagnosis with clinical biomarkers to stratify CAD risk in MAFLD. In this cross-sectional study with prospective data collection, which comprised 1073 MAFLD patients stratified by CAD status (MAFLD without CAD, n = 942; MAFLD with CAD, n = 131), baseline characteristics were compared using chi-square tests for categorical variables and Mann-Whitney U tests for continuous variables, with p < 0.05 considered significant. We developed two distinct deep learning models: Model-1 (non-invasive) using ResNet18 combined tongue image features with demographic and comorbidity data, and Model-2 (comprehensive) incorporated blood biomarkers alongside the features used in Model-1. Multimodal feature fusion was achieved through a dedicated cross-attention network. Model performance was rigorously evaluated using 10-fold cross-validation, with area under the curve (AUC), sensitivity, and specificity as primary metrics. Focal loss was employed to address class imbalance. The coexistence of CAD was associated with significantly higher rates of hypertension, diabetes, and familial cardiovascular history in patients with MAFLD ( p < 0.001), along with distinct metabolic profiles: elevated systemic inflammation (neutrophil-lymphocyte ratio), advanced fibrosis (FIB-4), elevated fasting glucose levels, and attenuated hepatic inflammation (ALT and AST). Model-1 achieved an AUC of 0.858 (sensitivity 0.778, specificity 0.908), while Model-2 demonstrated superior discrimination (AUC 0.933), particularly in tertiary care. Our AI-driven dual-model framework addresses the critical unmet need for CAD identification in MAFLD patients, providing a community-scalable screening tool (Model-1) and a precision clinical assessment model (Model-2), while offering empirical support for TCM tongue diagnosis. Pending external validation to confirm its generalizability, this approach may serve as a cost-efficient non-invasive screening strategy for this high-risk population.