EHR-Derived Risk Estimates for Clinicians Do Not Change Heart Failure Patient Outcomes – The REVeAL-HF Trial

Key Points:

• Previously, the impact of risk prediction on clinical decision-making and outcomes in heart failure had not been tested in a randomized control trial
• The REVeAL-HF trial showed that mortality estimates of admitted heart failure patients during standard EHR-based workflow did not lead to significant reductions in rehospitalization or mortality
• Mortality estimates did not significantly influence use of medical therapies, ICD implantation, HT/VAD or referrals to palliative care
• Bedside clinicians may require more prescriptive decision support though alert-based systems should be studied in a randomized fashion prior to implementation

The medical community continues to seek effective tools to reduce heart failure hospitalizations. One such proposed method has been accurate risk prognostication tools which could aid clinicians in accurately classifying patients’ prognoses to provide appropriate clinical interventions with the hopes of improving heart failure outcomes. No guidelines currently emphasize risk quantification tools as part of clinical decision making. Therefore, the REVeAL-HF (Risk EValuation And its Impact on ClinicAL Decision Making and Outcomes in Heart Failure) trial sought to definitively evaluate the impact of knowledge about prognosis on clinical decision making and patient outcomes.

The REVeAL-HF trial was a pragmatic, unblinded, parallel randomized control trial. It included patients hospitalized for heart failure (defined as N-terminal pro–B-type natriuretic peptide >500 pg/ml and receiving intravenous diuretics within 24 hours of admission) across 4 New Haven hospitals between November 2019 and March 2021. Patients randomized to the intervention had EMR-generated alerts of their 1-year mortality risk score presented to their providers during clinician order entry. Patients randomized to usual care had no alert provided. The hypothesis of the study was that accurate prognostication of heart failure patients would enhance clinical outcomes. The risk score itself was derived based on historic data of Yale University patients and included variables that have been shown to have prognostic value in heart failure (such as age, weight, systolic blood pressures, and several laboratory values). The risk score performed favorably when compared to other published risk scores with ROC of 0.74.

There were no significant differences among the baseline characteristics of 3,124 patients randomized to the alert or usual care groups. Results presented by lead investigator Dr. Tariq Ahmad at AHA’s Late Breaking Science Session on November 14^th showed that “there is no evidence that prognostic information provided to clinicians caring for heart failure patients made any statistically significant differences in the primary outcome.” Specifically, the risk stratification tool did not affect the primary composite outcome of one-year all-cause mortality or 30-day re-hospitalization (p=0.82) or either of its components. Secondary endpoints including length of stay, discharge doses of HF therapies, palliative care referral, referral to electrophysiology, and referral for advanced heart failure therapies were also not affected. Across prespecified subgroups based on age, gender, race, CKD, ejection fraction, NT-proBNP levels, ICU status and risk stratification levels there was also no evidence that prognostic information affected patients in any significant way.

Several hypotheses exist for the neutral results of this trial. Though the alert was dynamic in its ability to adjust based on updated patient data, in the discussion during his AHA Session Dr. Ahmad cited alert fatigue bias as a possible cause for the results of this trial. He stated that clinicians were likely to override automated alerts with their own risk assessment. It is possible the metrics chosen to evaluate the impact of alerts was also insufficient to capture their affects, for example pharmacotherapy was unlikely to change based on risk stratification for the predominately HFpEF population. The discussion also focused on the notion that had clinicians rather than patients been randomized, the treatment effects might have been different. The pragmatic design and clustering of mostly low-risk patients are other possible explanations. Overall, Dr. Ahmad suggests that it is important for these kind of EHR-based risk assessment tools be studied in randomized control trials. He remarks that if the risk tools are not making a positive impact on patient outcomes that such alerts should be promptly removed from EHR systems and that further scientific study is needed.