LiverCARD: Opportunistic Cardiovascular Risk assessment in MASLD using a multimodal AI pipeline

Introduction: Metabolic dysfunction–associated steatotic liver disease (MASLD), formerly known as non-alcoholic fatty liver disease (NAFLD), affects approximately 30% of the global population and represents a major public health challenge due to its strong association with cardiovascular disease (CVD), the leading cause of mortality in this group. Accurate cardiovascular risk (CVR) assessment in MASLD remains suboptimal, as traditional risk scores do not incorporate liver-related parameters or imaging biomarkers. Aim: The aim of this study was to develop LiverCARD, an artificial intelligence (AI)–based framework designed to improve CVR prediction in MASLD patients by integrating routinely collected clinical data with imaging features extracted from coronary computed tomography angiography (cCTA) native raw images. Methods: A total of 164 patients from two Italian hospitals were retrospectively evaluated. For each patient, ten pre-contrast cCTA images were analyzed. Steatosis was assessed using liver Hounsfield Units, CAC burden was used as an indicator of cardiovascular involvement, and liver fibrosis was estimated through the Fib-4 index. Patients were stratified into high- and low-risk groups according to CAC thresholds and degree of hepatic steatosis. LiverCARD incorporates clinical variables with high-dimensional visual embeddings obtained from pretrained medical deep learning models. Multiple machine learning algorithms were explored, and the best performance was achieved with XGBoost applied to the combined multimodal dataset. Explainability analysis using SHapley Additive exPlanations (SHAP) was performed to highlight the most important variables. Results and Discussion: The integrated approach achieved strong predictive results, with an accuracy of 0.89 in internal validation and 0.72 in external validation, outperforming models trained on clinical or imaging data alone. Moreover, using multiple CT liver slices per each patient, LiverCARD offers a balanced approach, capturing essential anatomical details while maintaining computational efficiency. Importantly, SHAP confirmed that predictions relied on clinically coherent factors such as age, number of comorbidities, Fib-4 values, Type 2 Diabetes, and male sex, while imaging embeddings captured additional discriminative patterns. Conclusion: By enabling the combination of AI-derived imaging features with standard clinical data, LiverCARD provides an enhanced, interpretable, and scalable method for CVR stratification in MASLD. This framework supports the development of personalized preventive strategies and demonstrates the potential of multimodal AI to address unmet clinical needs in cardiometabolic medicine.