Do socioeconomic measures improve prediction of cardiovascular disease hospitalization?

Abstract

Background: Risk prediction models for cardiovascular disease (CVD) hospitalization have improved prediction accuracy when individual-level measures of socioeconomic status (SES) are considered. However, it is unclear how validated, area-level SES measures, available at the population-level in Canada, might improve prediction of CVD hospitalization.

Objectives: The research objectives were to (1) test the incremental predictive value of area-level SES measures and (2) compare the incremental predictive performance of different area-level SES measures on time to hospitalization for CVD.

Methods: A retrospective cohort design used Manitoba administrative health records from 2014 to 2020 and area-level SES measures from 2016 Statistics Canada Census data. Individuals 40+ years as of April 1st, 2016 were followed until an acute myocardial infarction (AMI) or stroke hospitalization or loss to follow-up. Covariates included age, sex, comorbid conditions, prior healthcare use, AMI/stroke-related prescription medications, and one or more area-level SES measures, including the Socioeconomic Factor Index - Version 2 (SEFI-2), Material Deprivation Index, Social Deprivation Index, and the Canadian Index of Multiple Deprivation (CIMD). Cox proportional hazards models were assessed for model accuracy (area under the curve; AUC), discrimination (net reclassification improvement; NRI and integrated discrimination improvement; IDI) and calibration (Brier score). Results: Overall predictive performance of models containing one or more SES measures (fully-adjusted model; AUC = 0.753 – 0.757) was similar to model containing all other covariates (partially-adjusted model; AUC = 0.753). Discrimination performance was poor or statistically non-significant (NRI = -0.158 – 0.019; IDI = < 0.000). Prediction error was low for all models (Brier score = 0.022).

Conclusion: Area-level SES measures did not add predictive value to CVD hospitalization risk models. Risk factors available in administrative health data, like demographics and comorbid conditions, already provide a similar amount of information in terms of predictive ability. Area-level SES measures are useful for characterizing and describing populations but may not have strong predictive value to individual-level outcomes. Further studies are recommended to explore their use in prediction of other health conditions and jurisdictions, and comparisons with individual-level SES measures.