Adding Social Determinants of Health to AI Models Boosts HF Risk Prediction in Black Patients

Fran Lowry

July 15, 2022

The addition of social determinants of health (SDOH) to machine-learning risk-prediction models improved forecasts of in-hospital mortality in Black adults hospitalized for heart failure (HF), but didn't show similar ability in non-Black patients, in a study based in part on the American Heart Association–sponsored Get with the Guidelines in Heart Failure (GWTG-HF) registry.

The novel risk-prediction tool bolstered by SDOH at the zip-code level — including household income, number of adults without a high-school degree, poverty, and unemployment rates, and other factors — stratified risk more sharply in Black patients than more standard models, including some based on multivariable logistic regression.

"Traditional risk models that exist for heart failure assign lower risks to Black individuals if everything else is held constant," Ambarish Pandey, MD, MSCS, University of Texas Southwestern Medical Center, Dallas, told | Medscape Cardiology.

"I think that is problematic, because if Black patients are considered lower risk, they may not get appropriate risk-based therapies that are being provided. We wanted to move away from this approach and use a more race-agnostic approach," said Pandey, who is senior author on the study, published July 6 in JAMA Cardiology, with lead author Matthew W. Segar, MD, Texas Heart Institute, Houston.

The training dataset for the prediction model consisted of 123,634 patients hospitalized with HF (mean age, 71 years), of whom 47% were women, enrolled in the GWTG-HF registry from 2010 through 2020.

The machine-learning models showed "excellent performance" when applied to an internal subset cohort of 82,420 patients, with a C statistic of 0.81 for Black patients and 0.82 for non-Black patients, the authors report, and in a real-world cohort of 553,506 patients, with C statistics of 0.74 and 0.75, respectively. The models performed similarly well, they write, in an external validation cohort derived from the ARIC registry, with C statistics of 0.79 and 0.80, respectively.

The machine-learning models' performance surpassed that of the GWTG-HF risk-score model, C statistics 0.69 for both Black and non-Black patients, and other logistic regression models in which race was a covariate, the authors state.

"We also observed significant race-specific differences in the population-attributable risk of in-hospital mortality associated with the SDOH, with a significantly greater contribution of these parameters to the overall in-hospital mortality risk in Black patients vs non-Black patients," they write.

For Black patients, five of the SDOH parameters were among the top 20 covariate predictors of in-hospital mortality: mean income level, vacancy and unemployment rates, proportion of the population without a high-school degree, and proportion older than 65 years. Together they accounted for 11.6% of population-attributable risk for in-hospital death.

Only one SDOH parameter — percentage of population older than 65 years — made the top 20 for non-Black patients, with a population-attributable risk of 0.5%, the group reports.

"I hope our work spurs future investigations to better understand how social determinants contribute to risk and how they can be incorporated in management of these patients," Pandey said.

"I commend the authors for attempting to address SDOH as a potential contributor to some of the differences in outcomes among patients with heart failure," writes Eldrin F. Lewis, MD, MPH, Stanford University School of Medicine, Palo Alto, California, in an accompanying editorial.

"It is imperative that we use these newer techniques to go beyond simply predicting which groups are at heightened risk and leverage the data to create solutions that will reduce those risks for the individual patient," Lewis states.

"We should use these tools to reduce racial and ethnic differences in the operations of healthcare systems, potential bias in management decisions, and inactivity due to the difficulty in getting guideline-directed medical therapy into the hands of people who may have limited resources with minimal out-of-pocket costs," he writes.

The models assessed in the current report "set a new bar for risk prediction: integration of a comprehensive set of demographics, comorbidities, and social determinants with machine learning obviates race and ethnicity in risk prediction," contend JAMA Cardiology deputy editor Clyde W. Yancy, MD, and associate editor Sadiya S. Khan, MD, both from Northwestern University Feinberg School of Medicine, Chicago, in an accompanying editor's note.

"This more careful incorporation of individual-level, neighborhood-level, and hospital-level social factors," they conclude, "is now a candidate template for future risk models."

Pandey discloses grant funding from Applied Therapeutics and Gilead Sciences; consulting for or serving as an advisor to Tricog Health, Eli Lilly, Rivus, and Roche Diagnostics; receiving nonfinancial support from Pfizer and Merck; and research support from the Texas Health Resources Clinical Scholarship, the Gilead Sciences Research Scholar Program, the National Institute on Aging GEMSSTAR Grant, and Applied Therapeutics. Segar discloses receiving nonfinancial support from Pfizer and Merck. Other disclosures are in the report. Lewis reported no disclosures. Yancy and Khan had no relevant disclosures.

JAMA Cardiol. Published online July 6, 2020. Abstract, Editorial, Editor's note

For more from | Medscape Cardiology, follow us on Twitter and Facebook.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.