Interpretable machine learning models based on CMR radiomics for predicting left ventricular diastolic dysfunction in patients with metabolic-associated steatotic liver disease and type 2 diabetes mellitus.

BACKGROUND: Obesity-induced left ventricular diastolic dysfunction (LVDD), associated with ectopic fat and dysfunctional epicardial adipose tissue (EAT), is emerging as a key research area due to its increasing prevalence and links to metabolic-associated steatotic liver disease (MASLD) and type 2 diabetes mellitus (T2DM). This underscores the importance of early risk assessment and intervention to prevent the progression of LVDD. We developed an interpretable machine learning (ML) model combining cardiac magnetic resonance (CMR) radiomics and clinical data to assess LVDD risk in MASLD/T2DM patients, enabling proactive treatment customization.

METHODS: We prospectively analyzed 175 MASLD/T2DM patients, splitting them into training and external validation groups. After categorizing them as LVDD+ or LVDD-, we collected clinical data and extracted standardized CMR radiomics features to develop ML models. The optimal model was internally validated, interpreted using Shapley Additive Explanations (SHAP), and externally validated.

RESULTS: LVDD prevalence was similar in both cohorts (45.5% vs 46.2%, χ2 = 0.108, P = 0.743) among 175 MASLD/T2DM patients. The extreme gradient boosting (XGBoost) model, combining CMR radiomics and clinical data, outperformed in both internal and external validations. SHAP analysis revealed five critical determinants of LVDD: three radiomics features from CMR images and two clinical variables.

CONCLUSION: The XGBoost model, incorporating radiomics from CMR images and clinical data, outperformed other ML models in predicting LVDD risk in patients with T2DM and MASLD, enhancing risk assessment accuracy. This improvement allows for timely treatment adjustments, potentially preventing LVDD progression more effectively.