August 7, 2025 | Hypertrophic cardiomyopathy (HCM) is one of the most commonly inherited heart diseases and is a leading cause of sudden cardiac death among young people and athletes. When a patient is diagnosed with HCM, it is difficult to determine the likelihood of death from sudden cardiac arrest due to arrhythmia. Unlike heart conditions such as infarctions, in HCM it is more difficult to find the signals before it's too late.
But researchers at Johns Hopkins University have found a solution. The team developed Multimodal Artificial intelligence for ventricular Arrhythmia Risk Stratification (MAARS), a deep learning approach to predict lethal arrhythmia in patients with HCM by analyzing multimodal medical data.
Detecting Scarring in the Heart
Infarctions occur after cardiac tissue dies and scars form over the affected area, which are easily picked up by imaging technology. This scarring causes arrhythmia and forces the organ to work harder against increased pressure, ultimately leading to a heart attack.
“In patients with infarction, there is a clinical criterion to predict who will be at the risk of sudden cardiac death called ejection fraction,” elaborates Natalia Trayanova, Murray B. Sachs professor of biomedical engineering at Johns Hopkins and the senior author of the MAARS study. Ejection fraction is the amount of blood being pumped during one heartbeat, she continues, and if it pumps less than 35%, the patient is considered high risk for cardiac death.
However, ejection fraction does not work as a diagnostic marker on HCM because patients with the condition have functional hearts that contract close to normal, despite having thickening walls and the proliferation of fibrosis. Additionally, the scarring of HCM is more distributed throughout the heart wall, which makes it more difficult to assess on cardiac imaging.
HCM patients are usually evaluated at a hospital, where a number of clinical parameters are considered, including demographics and lifestyle. Next, they undergo imaging tests, such as an echocardiogram or an MRI. Then, a radiologist or an echo specialist will read these scans to determine if there are changes in the heart.
To determine the likelihood of a sudden cardiac arrest, MAARS uses transformer-based neural networks to analyze multimodal medical data from electronic health records (EHRs), echocardiogram and CMR reports, and contrast-enhance CMR images (Nature Cardiovascular Research, DOI: https://doi.org/10.1038/s44161-025-00679-1). At the core, there are three unimodal branch networks that process, combine, and fuse data before generating a result.
The first channel is EHRs. MAARS analyzes multimodal medical data from the patient’s EHR. The second channel is information from imaging reports, such as ultrasounds and MRIs. The third data input channel in the multimodal AI predictor is raw contrast-enhanced MRI images. These images are especially critical to risk prediction because they pick up heart scarring, a mechanism by which arrhythmia occurs. Through deep learning, the algorithm can find patterns in these data and associate them to the risk of cardiac arrest.
The team had developed a different model in 2022, which was designed for patients with infarctions. The algorithm used for the 2022 model was not as broad in terms of what it analyzed, but it demonstrated that multimodal AI is “something that really works.” For MAARS, they expanded on the algorithm, and the results were a “surprise.”
“I didn’t expect it to be that much better than the clinical guidelines,” says Trayanova. “The other surprise is how fair it was among the population and the age groups.” MAARS proved that it significantly outperformed all clinical risk assessment tools in the internal and external cohorts. In the internal cohort, MAARS had a predictive capability of 89%, while in the external cohort it was at 81%. The team expected that, with different cohorts and geographic locations, the performance quality would be reduced, but the algorithm proved them wrong (Nature Cardiovascular Research, DOI: https://doi.org/10.1038/s44161-025-00679-1).
Bringing MAARS to Clinics
“We would love to have it as a tool in the hospital,” says Trayanova of the team’s end goal. MAARS can be connected to EHRs of hospitals and clinics and can run in less than 30 seconds, producing results quickly and efficiently.
MAARS is also gaining international attention, she says, recalling someone from a hospital in Egypt inquiring about using it. This shows great promise for the model and opportunities to take it to a global level.
However, as the project was federally funded, there are concerns surrounding budget cuts and grant denials. The team still retains their NIH grants and has not experienced cuts, but Trayanova acknowledges that there is a “fear” of grants being withdrawn. She mentions the possibility of no financial support in coming years to hire researchers. “One grant was withdrawn because it had international components. Now we cannot have international collaborators on the grants the way we did before,” she added.
This has caused research teams to be more conscientious about spending, hiring people, and expanding research. “I have the money right now, but I don’t know what I’ll have tomorrow.”