University of Michigan researchers were recently awarded a $500,000, two-year grant through the American Heart Association (AHA) for a project focused on predicting the presence of hypertrophic cardiomyopathy (HCM) in patients who were previously undiagnosed from common tests available in the electronic health record like electrocardiograms and echocardiograms. HCM can be a serious heart condition in which the heart muscle becomes thickened, most often due to genetic abnormalities that leads to poor cardiac function and rhythm abnormalities.

The project lead is Sardar Ansari, PhD, Research Assistant Professor of Emergency Medicine. Co-investigators include Precision Health Co-Director Brahmajee Nallamothu, MD, Stevo Julius Research Professor of Cardiovascular Medicine, Adam Helms, MD, Assistant Professor of Internal Medicine, and John Donnelly, MSPH, PhD, Research Assistant Professor, Learning Health Sciences, all of University of Michigan Medical School. They will be working alongside multiple other sites including John Hopkins, UT Southwestern and Stanford University. The AHA is working on identifying additional community hospitals to participate.

At the University of Michigan, the two-year grant will be approached as an on-going partnership between Precision Health and The Max Harry Weil Institute for Critical Care Research and Innovation. Dr. Ansari and colleagues will be working with AHA to train & validate a model to predict HCM using a unique machine learning approach called ‘federated learning’.  AHA is acting as the central coordinating entity in this case, and will stay involved, providing on-going support and collaboration with all sites.

In the conventional machine learning approaches, there can be serious concerns around data-sharing and ownership, issues associated with data security, and potential risks to patient privacy; this often makes institutions hesitant to participate in multi-center studies involving large scale patient datasets because of these potential risks. In contrast, using a federated learning approach allows each institution/site to retain their PHI data, perform model training locally, and patient data is not transferred out of the health system. Validation can happen locally as well. These steps will comprise the first phase of the project.

Once trusted models are derived, there is still a need to study their application in real world settings to demonstrates a positive impact on patient outcomes. “HCM may be tough to diagnose due to its variable signs and symptoms,” says Dr. Nallamothu. He adds, “As a result, HCM often goes unrecognized for years and early detection with a machine learning tool like this could help clinicians initiate treatment earlier in at-risk patients.”

Challenges to adoption of machine learning at the bedside include understanding and preparing for their impact on system resources, including clinical workflows, referral, and treatment patterns as well as a governance framework to support their safe implementation. Phase 2 of the project will entail simulating implementation of the now-developed model across all sites, determining impact on clinical practices, assessing how much the model may improve clinical outcomes, and how much additional cost/workload would be incurred if adopted, and the feasibility of adoption at local and national levels.

Dr. Ansari says of the project, “My hope is that this project can become a blueprint of how to do federated learning with the involvement of organizations/funders that have interest in a particular disease, and to see how models can be implemented at several institutions. This project is much more likely to reach the bedside and improve patient care with the on-going support we’ll received from AHA, as well as the other institutions participating.”