Risk-Based EHR Alert Doesn’t Improve HF Care or Outcomes: REVEAL-HF

Adding an alert to the electronic health record (EHR) to draw attention to patient risk does not translate into better clinical outcomes or lead to changes in guideline-directed medical therapy for patients with acute heart failure (HF), according to the results of the REVEAL-HF trial.

These alerts, and others like them, shouldn’t be added willy-nilly to the EHR and should be included only if they actually have an impact on care or clinical outcomes, say researchers.  

“We feel that all these interventions [added] to the electronic health record should be studied in a randomized, controlled fashion, and if they don’t have any impact, then they should be removed,” said lead investigator Tariq Ahmad, MD (Yale University School of Medicine, New Haven, CT), earlier this week at the virtual American Heart Association (AHA) 2021 Scientific Sessions. “They can cause alert fatigue, and in the case of a prior study done [at Yale] alerting physicians about acute kidney injury, there is a concern that they can even cause harm. I think we need to study alerts in a scientific fashion.”

Presenting REVEAL-HF at a late-breaking science session, Ahmad said that it’s simply assumed that accurate risk prognostication will lead to better clinical decision-making in HF, which in turn will reduce rates of adverse outcomes. However, these risk prediction tools have never been shown in randomized clinical trials to improve patient outcomes. As a result, the guidelines do not tie the management of HF patients to the accurate quantification of risk, but rather put the responsibility on physicians to estimate patient risk and to make decisions based on those estimations, said Ahmad.

In REVEAL-HF, investigators wanted to test whether providing information to physicians about a patient’s predicted 1-year risk of death would translate into better treatment decisions and lead to a reduction in rehospitalizations or death. The risk score was developed at the Yale School of Medicine and is based on multiple variables, including the patient’s age, weight, systolic blood pressure, red blood cell distribution width, blood urea nitrogen (BUN), monocyte count, percentage of lymphocytes, BUN/creatinine ratio, NT-proBNP, troponin, mean corpuscular volume, ICU admission, and arterial pH.

In past studies, the risk score “performed incredibly well, in terms of discrimination and calibration in our validation of historic Yale heart failure patients,” said Ahmad.  

Tuning Out Alerts and Algorithm Bias

REVEAL-HF was a pragmatic trial conducted through EHRs at four teaching hospitals. In total, 1,590 patients with acute HF were randomized to an intervention in which the treating physician was alerted to the 1-year risk of death when they entered patient information in the EHR and 1,534 patients to usual care (physician received no alert). With the alert, physicians saw the patient’s predicted risk of all-cause mortality at 1 year and was subsequently informed whether that patient was at very low, low, medium, high, or very high risk for death in next 12 months. They did not receive any prescriptive information based on the risk assessment.

Despite the alert, there was no significant difference in the risk of death or rehospitalization at 1 year between the two groups, nor was there any difference in the risk of in-hospital mortality or readmission at 30 days. There also was no evidence of benefit in any of the prespecified subgroups, including patients stratified by their classification of risk.

With respect to clinical decision-making, there was no significant difference in the use of guideline-directed medical therapy. For example, in patients with reduced EF (HFrEF), use of beta-blockers, ACE inhibitors/ARBs, mineralocorticoid receptor antagonists, and sodium-glucose cotransporter 2 (SGLT2) inhibitors was similar in patients treated by physicians who received alerts and those who received usual care. There was also no evidence the alerts increased referrals to palliative care, even among patients at high or very high risk for death at 1 year.

As to why they observed no difference in outcomes, Ahmad said it’s possible that physicians just don’t pay attention to EHR-based alerts. “There is a well-described alert fatigue among physicians,” he said. “There is also something called ‘algorithm bias’ where a physician will trust their gut over an algorithm even if the algorithm has been shown to be superior.”

Based on the results, Ahmad said doctors might need more prescriptive decision support, and the study calls into question the idea that accurate prognostic information leads to better decisions and improved outcomes.

There is a well-described alert fatigue among physicians. Tariq Ahmad

Harriette Van Spall, MD (McMaster University, Hamilton, Canada), who discussed the findings following the AHA presentation, said there are at least 55 validated risk prediction tools for patients with HF, but “they’re complex, and don’t integrate with workflow, and are therefore underutilized at the point of care.” Van Spall led a trial similar to REVEAL-HF, known as PACT-HF, and also found that titrating patient HF services, including home-based nursing care, based on risk scores didn’t improve clinical outcomes.

As to why the results were neutral, Van Spall noted that nearly three-quarters of patients in REVEAL-HF were low risk, and that clustering of this low-risk population might not have allowed physicians to implement a range of different treatments based on the risk scores. Furthermore, the trial largely included HF patients with preserved EF (HFpEF)—just 27.9% of patients had a LVEF less than 40%—for which there are no known pharmacotherapies to alter all-cause mortality.

“Finally, it’s possible and probable that risk scores in and of themselves do not change treatment at the point of care,” said Van Spall. However, the role of risk scores in improving the patient’s understanding of their disease, engagement in self-care, and adherence to therapy is still unknown, she said.