Predicting cardiac arrest: AI outperforms cardiologists in risk assessment

People with hypertrophic cardiomyopathy (HCM), a hereditary heart muscle disease, are particularly affected by sudden cardiac death. Until now, it has been almost impossible for doctors to reliably predict which patients are at risk – - a problem that a new AI model is now set to solve. A research team at Johns Hopkins University has developed the MAARS (Multimodal Artificial intelligence for Arrhythmia Risk Stratification) system to predict the individual risk of sudden cardiac death with significantly greater accuracy than traditional clinical risk calculators. "We are [with MAARS] able to predict with very high accuracy whether a patient has a very high risk of sudden cardiac death or not," says Natalia Trayanova, co-author of the paper "Multimodal AI to forecast arrhythmic death in hypertrophic cardiomyopathy" , which was published open access in the journal Nature Cardiovascular Research.

At its core, MAARS analyzes multimodal patient data: electronic health records, cardiological findings and contrast-enhanced MRI images of the heart. Although the latter in particular have so far shown scar tissue, which is considered a risk marker for arrhythmia, they have hardly been systematically analyzed clinically, according to the researchers. MAARS, on the other hand, uses a deep neural network with transformer architecture to extract previously unused information from this three-dimensional image data.

The benefits are evident: while established guidelines from the American Heart Association or the European Cardiology Society only achieve an accuracy of around 50 percent in studies – barely better than chance –, MAARS achieves an accuracy of up to 0.89 in the internal and 0.81 in the external test data set. According to the research team, the accuracy for the age group between 40 and 60 – those with the highest risk – is even 93 percent.

Fewer false alarms, more targeted therapy

A key promise of AI prediction is that it should not only save lives, but also avoid unnecessary medical interventions. Currently, many HCM patients are prophylactically implanted with defibrillators – even though they will never experience a dangerous cardiac arrest. These devices carry their own risks, such as infection or false shocks. MAARS could help to focus these interventions on patients who are really at risk.

The system also provides comprehensible explanations for its decisions. With the help of so-called Shapley values, the most important risk factors can be identified at an individual level, such as the extent of scarring, certain arrhythmias or functional parameters of the heart. Visual explanations are also possible: heat maps on the MRI images show which regions of the heart structure the AI pays particular attention to.

The researchers promise another quality feature: MAARS did not show any systematic bias towards certain age or gender groups in the tests –, a common problem in AI-based medicine. Nevertheless, limitations remain: The database is limited with a total of just under 840 patients across two centers, and the number of actual sudden cardiac deaths during the study period is low. According to the researchers, this makes model validation statistically challenging.

Clinical application is still a long way off. Although the code is publicly available, integration into existing hospital systems and regulatory approval are still pending. MAARS requires high-quality image data and extensive patient information, neither of which is available everywhere.

(mack)