Consideration of bias in data sources and digital services to advance health equity

In this editorial, I highlight 5 papers that address expanded data sources and services to understand, contextualize, promote, and predict individual health with careful consideration of bias. Coiera et al¹ promote the idea of family informatics to create a set of digital services to support the family network. Two studies in this issue examine the role of social determinants of health (SDOH) in predictive model performance as a strategy for identifying and mitigating bias.^2,3 Lastly, two papers are about data sharing for a variety of purposes. One explores willingness to share as a potential bias in analytic data sets and algorithms⁴ while the other evaluates a privacy-protecting framework for one type of data.⁵ As a group, these papers provide additional foundation to advance health equity.

In a perspective, Coiera et al¹ highlight the importance of understanding individuals in the context of their family and argue that this may require new classes of digital services (ie, family informatics) to address important chronic health challenges such as obesity, mental health, and substance abuse, and to support acute health challenges, and promote self-management capacity. They conceptualize the family network as a multiagent system with distributed cognition. They propose that digital tools can address family needs in four key areas: (1) sensing and monitoring; (2) communicating and sharing; (3) deciding and acting; and (4) treating and preventing illness.

Juhn et al² applied machine learning models for predicting asthma exacerbation in children with asthma. They measured one SDOH, socioeconomic status (SES), using the HOUsing-based SocioEconomic Status measure (HOUSES) index, to assess its influence on predictive model performance. They also compared incompleteness of EHR information relevant to asthma care by SES. Those with lower SES had a higher proportion of missing information relevant to asthma care (eg, asthma severity). The HOUSES index enables assessment of SES bias in predictive model performance.

Amrollahi et al³ compared the performance of sepsis readmission prediction models with and without inclusion of SDOH. They used data from the All of Us Research Program participants across 35 hospitals (n = 8935 septic index encounters) to develop a multicenter validated sepsis-related unplanned 30-day readmission models with and without SDOH to predict 30-day unplanned readmissions. Incorporation of SDOH factors (eg, economic stability) into the model of clinical and demographic features improved area under the receiver operating characteristic curve significantly (from 0.75 to 0.80; P < .001).

Research participant willingness to share types of data sources can influence the representativeness of samples in analytic datasets. Joseph et al⁴ examined the willingness of participants in the National Institutes of Health All of Us Research Program to share EHR information. In a sample of 25 852 participants (White—66.5%, Black—18.7%, Hispanic—7.7%, female—32.5%), 2.3% declined to share EHR data. Younger age (1.26 [1.19–1.33]), female sex (1.74 [1.42–2.14]), and education >high school (2.44 [1.86–3.21]), but not race or ethnicity, were significantly associated with decline to share EHR data.

Concerns about privacy may limit willingness to share data. Bonomi et al⁵ propose a privacy-protecting method for sharing one type of data, individual-level electrocardiography (ECG) time-series data. Their approach leverages dimensional reduction technique and random sampling to achieve privacy protection against an informed adversarial model while enabling useful aggregate-level analysis while maintaining the usability for data analytics. Their evaluation of the approach on two real-world ECG data sets demonstrated significant reduction in privacy risks while retaining data usability for tasks such as predictive modeling and clustering.

Conflict of interest statement

None declared.