Significant participation bias and suboptimal cardiovascular risk management in learning healthcare systems due to non-consent

Abstract

Abstract Background Computer-aided clinical decision support algorithms require continuous adjustments (learning) to account for changes in patient-characteristics, treatment(s), and outcomes, also referred to as a "Learning Healthcare System" (LHS). A LHS’s accuracy depends on adequate reflection of all patients, and participation bias may distort its effectiveness substantially. Therefore, we studied participation bias in a single center cardiovascular cohort study. Methods The cardiovascular cohort was set up as a LHS aiming for guideline-based treatment according to the individual risk profile for all cardiovascular disease patients in routine clinical care. Patients visiting an outpatient clinic for first time evaluation of a cardiovascular disease or risk factor received information on the cardiovascular cohort prior to their visit. During visit, cardiovascular risk management (CVRM) data was collected in the electronic health record (EHR) and informed consent was asked. We studied differences in characteristics between consenting, non-consenting and non-responding patients, and used multivariable logistic regression to identify determinants of non-consent. Multinomial regression was used for exploratory analyses of non-response determinants. A waiver (19/641) was obtained from the institutional review board for this study. Results In total, 2378 patients were consenting, 1907 non-consenting, and 1445 non-responding. In short, non-consent was related to higher cardiovascular risk (e.g., a cardiovascular disease history (OR 1.43, 95%CI 1.23-1.66)) and lower education (OR 0.76, 95%CI 0.60-0.97) than consent (Figure 1). Non-response, in contrast, was associated with higher physical activity (OR 1.04, 95%CI 1.01-1.08) and education level (OR 1.43, 95%CI 1.04-1.98) compared to consent. Non-consent and non-response patients were, respectively, younger and older. Presence of CVRM indicators in the EHR was 5-30% lower in non-consenting and 37-69% lower in non-responding patients compared to consenting patients. Conclusion A traditional informed consent procedure leads to participation bias in a LHS. Furthermore, it affects structured registration of CVRM indicators in the EHR, which is detrimental for the LHS feedback loop, and potentially leads to suboptimal cardiovascular risk management in patients not willing to participate. This study underlines the importance of reassessing the need of a traditional informed consent procedure for the use of routine care data in learning healthcare systems.