The MAESTRIA demonstrator: AI-driven multimodal clinical decision support for cardiovascular risk prediction

Background: Personalised cardiovascular risk prediction remains limited by fragmented data and lack of clinical interpretability. The MAESTRIA project aims to overcome these limitations by developing an AI-powered, multimodal Clinical Decision Support Demonstrator for stroke and atrial fibrillation (AF) risk stratification. Purpose: To develop and validate an interoperable demonstrator integrating imaging, electrophysiology, clinical and omics data using explainable AI to support personalised cardiovascular care.

Methods: We harmonised multimodal datasets across several European centers, including clinical variables, ECG, echocardiography, cardiac-MRI, CT, wearable-datasets and omics. Predictive models were developed using ensemble machine learning algorithms (e.g., RandomForest, Logistic Regression, CatBoost, XGBoost), deep learning approaches for imaging segmentation, and multi-block data integration frameworks. Model interpretability was addressed through SHAP value analysis and partial dependence plots. The demonstrator was built as a GDPR-compliant, decision support tool with clinician-friendly interfaces.

Results: The demonstrator integrates validated models predicting AF recurrence, stroke risk, and atrial cardiomyopathy. By integrating ECG, echo, MRI, CT, and omics data, it significantly outperforms standard clinical scores, boosting prediction accuracy (AUC increase up to 0.15). External validation across multiple cohorts confirmed generalisability and clinical usability.

Conclusion: The MAESTRIA Demonstrator, built on a robust multimodal AI framework, promises to be a powerful clinical tool enabling clinicians to predict, visualize, and interpret cardiovascular risk, particularly atrial fibrillation, through the integration of clinical, imaging, and biological data.