AI tool can predict heart failure from genetic and health record data

One of the real promises of artificial intelligence and machine learning in health care is disease prediction: leveraging computing power to help patients understand their health.

A University of Michigan study demonstrates that combining two health related data sources–genetic information and clinical data–for an individual patient can help predict heart failure, the world’s leading cause of hospitalization, a decade before it’s diagnosed.

Led by Xu Shi, Ph.D., of the Department of Biostatistics in the U-M School of Public Health, Cristen Willer, Ph.D., formerly of the Department of Human Genetics and Kuan-Han H. Wu, Ph.D., formerly of the Department of Computational Medicine and Bioinformatics at the U-M Medical School, the study team was intrigued by the idea of developing something akin to a polygenic risk score, which uses genetic data to assess a patient’s risk of a future disease or condition, using clinical data found in an electronic health record. 

“Our main question was can we predict who will develop heart failure long before they ever show any sign of it using the data we already have in an electronic health record linked to a biobank of genetic data,” Shi said.

They began by training their AI/ML model on vast amounts of genetic and medical data.

The team used biobank sets from the Global Biobank Meta-Analysis Initiative to develop a genome-wide association score for heart failure.

Clinical data was pulled from the EHRs of de-identified cohorts of Michigan Medicine patients to develop the clinical risk score.

Their model, Shi explained, treats the medical diagnosis codes (known as ICD codes) found in electronic medical records as “words” in human language.

Using natural language processing (the technology behind AI models like ChatGPT), it analyzes around 30,000 codes in the data to generate one score associated with the development of heart failure.

With the trained model, they were able to predict who would develop heart failure eight years prior to diagnosis.

Combining both scores together enabled prediction of heart failure 10 years before diagnosis.

“This study shows that by combining a person's genetic data with detailed information from their medical records, doctors can predict who will develop heart failure much earlier - in some cases, years before it actually happens. This breakthrough could allow for earlier counseling and prevention, potentially helping people avoid or better manage heart failure in the future,” said co-author Michael Mathis, M.D., of the Department of Anesthesiology.

To that end, Shi hopes to improve the model so that it can make predictions based on hypothetical changes in health behaviors.

“The prediction model is just telling you one possibility, but there are so many possibilities according to your behavior after the prediction,” said Shi.

For example, a patient who is told they may develop heart failure in 10 years might start to exercise or change their diet and ultimately prove the model incorrect.

A precision predictive model would offer outputs that account for these changes, says Shi.

The widespread use of these models has enormous potential for improving health and healthcare cost savings by intervening with actionable information before signs and symptoms have set in.

Additional authors: Brooke N. Wolford, Jiacong Du, Xianshi Yu, Nicholas J. Douville, Michael R. Mathis, Sarah E. Graham, Ida Surakka, Whitney E. Hornsby, Jiang Bian, Lili Zhao

Funding/disclosures: K.H.W., S.E.G. and C.J.W. work at Regeneron Pharmaceuticals. N.J.D. received funding from the Foundation for Anesthesia Education and Research (Mentored Research Training Grant).

Paper cited: “Integrating large scale genetic and clinical information to predict cases of heart failure,” Communications Medicine. DOI: 10.1038/s43856-025-01198