Development and validation of mortality prediction models based on the social determinants of health.

Background: There is no standardised approach to screening adults for social risk factors. The goal of this study was to develop mortality risk prediction models based on the social determinants of health (SDoH) for clinical risk stratification.

Methods: Data were used from REasons for Geographic And Racial Differences in Stroke (REGARDS) study, a national, population-based, longitudinal cohort of black and white Americans aged ≥45 recruited between 2003 and 2007. Analysis was limited to participants with available SDoH and mortality data (n=20 843). All-cause mortality, available through 31 December 2018, was modelled using Cox proportional hazards with baseline individual, area-level and business-level SDoH as predictors. The area-level Social Vulnerability Index (SVI) was included for comparison. All models were adjusted for age, sex and sampling region and underwent internal split-sample validation.

Results: The baseline prediction model including only age, sex and REGARDS sampling region had a c-statistic of 0.699. An individual-level SDoH model (Model 1) had a higher c-statistic than the SVI (0.723 vs 0.708, p<0.001) in the testing set. Sequentially adding area-level SDoH (c-statistic 0.723) and business-level SDoH (c-statistics 0.723) to Model 1 had minimal improvement in model discrimination. Structural racism variables were associated with all-cause mortality for black participants but did not improve model discrimination compared with Model 1 (p=0.175).

Conclusion: In conclusion, SDoH can improve mortality prediction over 10 years relative to a baseline model and have the potential to identify high-risk patients for further evaluation or intervention if validated externally.