INTERACTION

Cardiovascular diseases are the leading cause of mortality in developed countries, with sudden cardiac death (SCD) often resulting from the occurrence of ventricular arrhythmia (i.e.: ventricular tachycardia [VT] or ventricular fibrillation [VF]. Current risk stratifiers, such as left ventricular ejection fraction (LVEF), have limited sensitivity and specificity. Other approaches using imaging biomarkers, ECG analysis, and genetic risk scores have been proposed, but they rely on single modalities and provide only partial observations of the cardiovascular system.

The INTERACTION project aims to develop a multimodal AI generative model of the cardiovascular system to better predict the risk of VT / VF in high-risk patients, particularly in post-infarct patients with a low LVEF.

This project addresses the limitations of current unimodal risk assessment methods and responds to the need for multiparametric risk scores combining structural and electrical risk markers.

The INTERACTION project will build on preliminary results from project partners, including knowledge-based digital twins and a pilot genome-wide association study (GWAS) that revealed significant genetic variants associated with SCD risk in patients with severe cardiomyopathy. The project will also leverage data from the SMART-DEF clinical trial, which aims to assess the predictive power of imaging-based biomarkers in patients with remote myocardial infarction and low LVEF.

The project is divided into five work packages:

1) Clinical trial management: Obtaining ethics approval for modifications to the SMART-DEF trial and additional task management for the ancillary study.

2) Identification of actionable common variants to stratify the risk of SCD in patients with ischaemic cardiomyopathy: Expanding the GWAS study and perform subgroup analyses to uncover genetic markers associated with various biomarkers. This will involve implementing a Polygenic Risk Score (PRS) for VT / VF risk stratification.

3) Multimodal Representation Learning: Developing a fusion method for information from multiple modalities, including clinical data, genetic data, ECG time-series, and medical images. This will involve using advanced AI architectures like MAMBA models and training on large-scale databases such as UK Biobank.

4) Phenotyping of the SMART-DEF database: Applying the multimodal representation to better characterize the SMART-DEF population, including fine-tuning the model and using clustering approaches to extract subgroups indicative of long-term outcomes.

5) Digital twins and explainability: Linking data-driven mechanisms with knowledge-based cardiac digital twins using a graph-based methodology to investigate VT / VF risk stratification in post-infarct patients with a low LVEF.

The main outcome of the INTERACTION project will be a data-driven multimodal AI digital twin of the cardiovascular system that fuses information routinely acquired in clinical practice. This model could be useful for generating missing modalities in retrospective clinical trials and developing methodologies for better introspection of data-driven models.

The project involves a diverse team of experts, including cardiologists (Prof de Chillou), clinical trial managers, geneticists (Dr Barc), and researchers specializing in signal processing, image processing, and machine learning (Dr Oster, Dr Karfoul, Dr Kachenoura, Dr Yochum) and digital twins (Dr. Le Rolle et Dr. Hernandez).

The INTERACTION project aligns with the objectives of the PEPR Digital Health program, focusing on cardiovascular health and multimodal analysis techniques. It complements other funded projects by proposing a fully data-driven approach to uncover a latent representation of the cardiovascular system, that could lead to finer phenotyping of populations and potentially reveal unknown physiological mechanisms. By combining large-scale genetic studies, advanced imaging techniques, and innovative AI methodologies, the INTERACTION project aims to significantly improve risk stratification for sudden cardiac death and advance our understanding of cardiovascular disease mechanisms.