Northwestern Medicine Builds Machine-Learning Model To Identify Patients With Heart Failure


Scientists and physicians create a custom tool in Epic to help predict which patients may have advanced heart failure.


Hospitals using Epic’s electronic health record software hold the medical records of more than 280 million patients across the U.S. With so much information in one place, a common question facing today’s healthcare experts is how that data can be used to advance patient care.


For a team of scientists and physicians at Northwestern Medicine Bluhm Cardiovascular Institute, the answer to this question involves a form of artificial intelligence (AI). In the fall of 2020, they began creating a custom augmented intelligence-enabled workflow that embedded the output of a machine-learning model in Epic. The model sorts through millions of patient records to make predictions about whether a patient has heart failure — which currently impacts more than 6 million people in the United States* — and more specifically whether that patient is progressing to advanced heart failure.

Prioritizing Patients

An ensemble machine-learning model uses features from three existing peer-reviewed heart failure risk scores as well as input from the clinical team. By examining factors such as lab results, symptoms and medications, the model identifies and classifies patients in different stages of heart failure. Manually sorting through such an enormous dataset to accomplish the same task would be impossible for clinicians.
The model’s output is embedded into an Epic workflow where patients with classified predictions are surfaced for nurse coordinators to review, along with other elements from patient charts. Once a patient is determined to be an appropriate candidate for referral, care is coordinated with the patient’s primary cardiologist or primary care physician.
As a result of this model, more than 680 patients so far have been screened by a nurse coordinator and referred to a clinic for evaluation. Two patients — who would have otherwise gone untreated — have received life-sustaining ventricular assist devices. (Augmented Intelligence to Identify Patients With Advanced Heart Failure in an Integrated Health System, 2022)

Enhancing Workflows

One of the most integral elements of this model is its built-in transparency. The clinicians using the tool know what data elements are being pulled in to support predictions. The feedback loop incorporates clinician review, with additional data and labels generated to train future models.

The project team hopes to improve the AI model by using additional data, such as clinician notes, and optimize the timing of when the model identifies patients for expedited evaluation by a heart failure specialist for advanced therapies.

This technology demonstrates how AI can be integrated into workflows to facilitate earlier identification of patients at risk for progressing to advanced heart failure. In the future, it may be able to identify patients at risk of developing other cardiovascular conditions, saving more lives and supporting better care.

* ​Centers for Disease Control and Prevention