Doctor working with modern computer interface.

Harms and Biases Associated with the Social Determinants of Health Technology Movement

By Artair Rogers

Many health systems have begun using new screening technologies to ask patients questions about the factors outside of the clinic and hospital that contribute to an individual or family’s health status, also known as the social determinants of health (SDOH). These technologies are framed as a tool to connect patients to needed community resources. However, they also have the potential to harm patients, depending on how patient data is used. This article addresses key harms and biases associated with the SDOH technology movement, and provides suggestions to address these issues going forward.

The Age of SDOH Technology

SDOH interventions typically involve screening patients within a clinical setting using trusted health care providers (nurses, doctors, community health workers, case managers, social workers, etc.) to assess the essential resource gaps of a patient. These needs are often documented within a screening tool developed by the health system or a technology vendor responsible for the SDOH referral platform.

Following the screening process, if a patient has resource gaps, the patient can opt into receiving referral services. Referrals can simply involve providing the patient with information about a social service agency or community-based organizations that may have resources to address needs identified by the patient. Some interventions will provide patients with navigation assistance to increase the likelihood of a resource connection. Many SDOH referral platforms have the ability to conduct a referral directly to the social service or community-based organization electronically. Follow-ups with patients to ensure resource connection, case notes (qualitative data) and connection metrics like number of outreaches to patients (quantitative data) are captured in the platform. Figure 1 shows the places where data capture can happen and who is typically responsible for data capture in the SDOH intervention process.

 

Fig. 1 Potential Data Extraction Points from SDOH Intervention Data Cycle.
Fig. 1 Potential Data Extraction Points from SDOH Intervention Data Cycle.
CBO is an acronym for community-based organization.
SSO is an acronym for social service organization.

The rationale for SDOH interventions is that the identification of resource gaps coupled with closing those resource gaps will yield better health outcomes. However, the business model for the technology vendor is not predicated on the number or percentage of successful resource connections for patients/clients.The technology vendors instead seem to draw value from the amount of data collected, as SDOH data uses have become valuable to the health system.

For example, social determinants of health data is now being utilized in machine learning and predictive algorithms to understand more about a patient’s susceptibility to a medical condition, like cardiovascular disease. Health systems have also used cost of care and health care utilization (i.e., number of emergency department visits and inpatient admissions) as metrics of evaluation for SDOH interventions; therefore, it is conceivable that SDOH data may be used in algorithms to predict cost of care and health care utilization, as well.

Health-system based SDOH efforts are large and growing. A study of 917 hospitals found that a total of $2.5 billion dollars of health system funds were invested into SDOH efforts from 2017 to 2019. Per the study, SDOH represent nearly half of the cost structure of Medicare and Medicaid. The study also indicates that there has been a rise in the average size of equity funding deals for for-profit SDOH organizations, like these technology vendors, from an average size of 11.4 million in 2019 to $75.7 million in 2021. A significant amount of money has been invested in SDOH technology platforms that are intended to be used as: (1) a database that stores SDOH screening data, (2) a repository of community- based organizations and social service organizations along with their corresponding services. (3) a conduit for communication between health system and social service organizations, and (4) a management tool that seeks to improve care coordination for the patient by a patient’s care team.

With this level of investment, we must ask how do we ensure harm is not imposed on individuals as a result of this technology and data. Further, how can we begin to address harm, if it has already occurred?

Informed Consent and Refusal

To move towards a framework of addressing and limiting harm, health systems and SDOH technology companies must first establish easily understood informed consent and refusal processes for all of the data uses that come from the data collection throughout a SDOH intervention. Essentially, in the current approach to SDOH interventions, a patient is typically consenting to data collection by the health system and maybe even data exchange between health systems and community-based organizations and social service organizations for the sake of connecting the patient to needed resources. The patient is typically unaware that the data may be used for research or evaluation studies— let alone for machine-based learning or predictive algorithms. Therefore, a proper informed consent and refusal process would involve naming all the ways days could be used throughout the intervention and allowing patients to opt in or out of the individual data use cases. For example, patients should have the ability to consent to having their data used to connect to a needed resource while opting out of data being used for evaluation/research purposes or machine-based learning use. Further, patients should be in control of who has access to the data— community based organizations, health systems, insurance payers, technology companies, venture capitalists. It is a moral imperative for patients to have agency in how this data is used. Additionally, there is a responsibility to ensure that any informed consent and refusal language is accessible to participants/users of the intervention. Ultimately, justice would involve community members participating in the financial benefits stakeholders receive from the collection and use of their SDOH data, since technology companies are essentially profiting off of these patients.

Measurement Bias

As metrics like food insecurity, housing insecurity, and transportation access can change over time, one-time data capture can create measurement bias. One-time screening, which often happens, may capture a social needs/social determinants of health snapshot or proxy that is oversimplified. This “moment in time” data capture could be outdated when the evaluation or machine-based learning occurs.

Therefore, a recommendation to address this potential bias could be a longitudinal data collection approach, which involves capturing SDOH data on a patient at multiple time points. Additionally, since scaled survey items or questions are typically used to measure social determinants of health, like food insecurity or transportation access, community verification processes in how social determinants are measured could also help reduce measurement bias that may occur in evaluation and machine-based learning (even in longitudinal data collection).

Further, since SDOH data points possess fluidity, it is advantageous and just to have community members, particularly those from marginalized communities, provide context to the SDOH data collected. Researchers, evaluators, health system stakeholders, and technology stakeholders cannot haphazardly separate data from the environment, policies, norms, human choices, and history that cultivated and produced the data. It is helpful to collect qualitative data from those with lived experience to understand the context behind the survey items or data points being used to understand social determinants of health gaps. Further, it is necessary to have this context to also understand the meaning of these data points across individuals of different backgrounds, cultures, and norms. Additionally, it is important to have those with lived experience validate the constructs of algorithms and machine learning predictions.

Aggregation Bias

Ignoring group-specific context (i.e., race, ethnicity, gender, sex, etc.) could lead to oversimplified data narratives. For example, for SDOH interventions, aggregate data analysis and aggregate data use for modeling purposes could assume consistent or similar conditions for patients within a study population. However, social conditions/complexities of different groups (i.e, racial, ethnic, gender, sexual orientation, etc.)  even within a given geography may be vastly different. Ignoring these conditions could lead to analysis, models, and interventions that do not highlight, understand, and address the root causes of inequities that these subpopulations face.

Therefore, given the importance of community context, lived experience expertise must be integrated in all aspects of SDOH interventions to ensure that group-specific context is not ignored in favor of more generalizable models and analyses. Additionally, it is imperative that these generalizable models and analyses are assessed for their ability/likelihood to perpetuate harms, biases, and stereotypes. Equity and justice would push stakeholders to have community members, particularly those with lived experience, as a compensated part of the governance structure of the evaluation, data narrative, and modeling processes.

Learning Bias

Lived experience is also needed to reduce biases that may occur from data narratives that do not fully capture the complex construct of social determinants of health. From a modeling standpoint, the lived experience perspective is needed to ensure that modeling choices do not promote learning biases, like a model prioritizing accuracy over understanding. An example of this could be a model focused on the naming the most socially complex (individuals with most severe social care/ SDOH gaps) and promoting those results without fully considering and mitigating the cost of perpetuating negative stereotypes of marginalized groups.

Integrating lived experience expertise to address this tension may shift the data health systems and technology vendors collect, but also shift how these SDOH data points are labeled. Equity-oriented and justice-oriented approaches may call for survey items and data labeling to not place the onus of the SDOH gap on the individual—but the system. For example, instead of asking about the individual’s ability to pay for utility bills, a SDOH survey could ask about the difficulty or ease of obtaining a living wage in an individual’s living area. There may even be survey items that measure resiliency and assets of the patient in SDOH screening tools. Taking this approach and utilizing community voice may allow SDOH stakeholders to understand the individual’s SDOH snapshot while avoiding the perpetuation of stereotypes.

Deployment Bias

Currently, these SDOH technology tools are intended to increase the likelihood of a patient connecting with a community resource to close a resource gap. However, the deployment of this technology has yet to prove a causal relationship with increased connection rates. Now, the value of SDOH technology seems to be collecting and storing this unique data. If the deployment of the technology itself can shift its purpose and value, there is reason to be wary of deployment bias regarding potential models and algorithms that utilize SDOH data. For instance, modeling that aims to use SDOH data to predict the likelihood of having a clinical condition for the means of better care could shift to the creation of risk assessments that payers could use to penalize individuals regarding insurance coverage.

Conclusion

Given the great influx of investment in SDOH interventions, mapping, naming, and addressing the biases found in the data collection and generation processes, along with the potential data model building are essential to preventing harm to individuals who believe that a SDOH questionnaire is simply being used to help. If these biases and harms are not addressed, SDOH technology and data could further perpetuate historical biases and harmful stereotypes against marginalized groups. Ultimately, SDOH implementers, evaluators, and stakeholders should:

  • Establish mechanisms for informed consent and refusal
  • Ensure that all patients/participants within a SDOH intervention understand all data use cases
  • Assess SDOH models and analyses for their ability and likelihood to perpetuate harms, biases, and stereotypes.
  • Create infrastructure within SDOH interventions for community governance to ensure the most marginalized communities’ needs and priorities are elevated.

Artair Rogers is a first-year doctoral student in the Department of Social and Behavioral Sciences at the Harvard School of Public Health pursuing a PhD in Population Health Sciences. 

The Petrie-Flom Center Staff

The Petrie-Flom Center staff often posts updates, announcements, and guests posts on behalf of others.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.