As expected, many consumers are unsettled by the unregulated use of personal and commercial information to predict sensitive behaviors or health outcomes [4]. An example of such unregulated use of personal information is Google’s acquisition of large amounts of personal health data, from hospitals and clinics across 21 US states, used to predict health and health care use, undisclosed to patients and other parties [73,74]. Social determinants cover sensitive topics, such as poverty, substance misuse, food insecurity, and homelessness. Individuals may fear stigmatization from health providers in revealing their SBDH information [75]. Similarly, individuals may be concerned about the social, employment, and legal effects of revealing SBDH when their data are not protected [75].

To address such concerns, there needs to be an established discourse leading to a national consensus and clear guidelines regarding the ethical use of patients’ SBDH data in the context of a health care predictive model [76]. Lack of transparency in methods, applications, and data protection results in little accountability to ensure that SBDH risk predictions are not used to achieve profits at the expense of health care quality or access, such as using SBDH data to exclude vulnerable patients from a health intervention to ensure greater health care profits [76,77]. Establishing robust and meaningful national guidelines for using SBDH data will require insights from a variety of clinical, social science, and technical perspectives as well as views of patients, community members, policy makers, and ethicists. In particular, patients should participate and be involved in the research that is developing models to safeguard the ethical and transparent use of patient data [78]. Without the perspectives of patients and community members at the forefront of these discussions, rather than moving to a new level of health care equity and access, SBDH predictive analytics could easily slide into domains that many would consider inappropriate use, especially given a special concern and focus on the highest risk members of our communities [76].

One important ethical and technical challenge of SBDH analytics, mostly in the application of statistical modeling, is ingrained model bias. For instance, vulnerable patients, such as those with more social and behavioral risk factors, may not be adequately represented in the data sources used to build the predictive model, leading to the model’s inaccurate predictions for these individuals. Machine learning models on the other hand can address this issue through over- or undersampling. Therefore, being at risk for bias from the original sample is normally corrected in a standard process [79].

The data sources might also lack information on the key SBDH variables that affect the desired outcomes. An example of this challenge might be a predictive model that focuses on health care utilization as the desired outcome and lacks data on health care access for vulnerable populations. Such a model may indicate that individuals with poor access to health care have a low likelihood of future utilization. A model with such ingrained bias would thus underestimate the actual requirement for the greater amount of health care resources necessary to achieve the same health outcomes once these individuals have access to health care [42]. Recently, this situation was observed in a study by Obermeyer et al [80] who assessed a large, commercial health plan’s predictive algorithm. The model systematically underestimated the health needs of African American patients by assuming that health care costs served as an adequate proxy for health needs. The bias arose because the unequal access to care among African American patients resulted in less money spent caring for those patients compared with White patients.

Although many researchers use health care utilization and costs as outcomes for SBDH research, models with these outcomes, proxied for health needs, are biased in that the data underrepresents those with lower access to health care. In recognition of the ingrained model bias, one approach might be to develop guidelines that recommend stratifying the population for key SBDH risk factors. Therefore, separate models would assess health care utilization for each stratum, taking into account unmeasured SBDH risk factors impacting health care utilization (eg, socioeconomic status, which defines insurance type and access to health care).

The future of SBDH-centric predictive modeling faces several challenges related to data sources and model design. One big data–related challenge is that most social and behavioral data found within providers’ EHRs are unstructured, free-text clinical notes and are not standardly interoperable. Although ubiquitous, this information is captured inconsistently and depends on the use of NLP to render the data useful in analytics [81,82]. When NLP is utilized, the SBDH language in the health record may not describe the level of SBDH precisely enough to accurately determine social risk as social determinants such as neighborhood disadvantage may need to reach a threshold to have a significant impact on health-related outcomes [83].

Another important challenge is related to the use of population-level SBDH variables and whether such variables are interpreted as proxies for individual-level factors that cannot be measured, such as low household income, or represent population-level spatial elements, such as a high concentration of low household income in a neighborhood [84]. Proxies are based on assumptions to confer population-level characteristics to an individual. In contrast, geospatial models investigate population-level elements based on the principle of spatial autocorrelation, meaning that data located close together are interrelated by nature [85]. Addressing this challenge is critical to the interpretation of models and requires sufficiently transparent models that allow the proper distinction between the two implications of the population-level SBDH variables.

There are also several technical challenges related to the analytic approach, spanning the choice of analytic model, data sources, discriminatory power, and SBDH temporality. Statistical models, spatial analysis, and machine learning have all been used alone and in combination with various SBDH predictive models. Most often, health care predictive analytics uses regression models for their simplicity and acceptability [86]. However, machine learning models may be useful for finding new dimensions that can accurately classify outcomes according to their predictive characteristics in nonlinear data [86]. However, not all machine learning techniques, which range from transparent decision tree algorithms to unsupervised neural networks, are appropriate for use with SBDH predictive models. Highly autonomous machine learning models may select characteristics that are not clinically relevant for the outcome (eg, family meetings as a predictive characteristic for hospital mortality) when researchers do not remove these characteristics [86]. Models should instead reflect appropriate domain expertise as well as appropriate machine learning techniques. Moreover, for techniques that depend on unsupervised neural networks, there are long-standing controversies regarding the disadvantages of nontransparent, one-of-a-kind models versus more readily explainable logistic regression models [7,86].

There are also challenges related to using SBDH data at the geographic level in predictive modeling, which are often needed to identify SBDH on a population level and for community-level interventions [26]. Geospatial analysts need to choose the appropriate granularity for a model, which may be associated with a model’s discriminatory power to help distinguish those at high- versus low-risk levels [87]. Furthermore, analyzing SBDH data at different geographic levels (eg, census block group, census tract, county, and state) is methodologically complex.

The discriminatory power to distinguish patients with and without social needs also poses a challenge in nongeospatial modeling with the potential to introduce higher-than-desirable false positives and/or negatives [74]. For instance, a study of food security among Medicare patients using clinical data and a needs assessment survey could not accurately predict which patients would benefit from a referral to community resources [88]. Similarly, a predictive model that uses random forest decision methods applied to socioeconomic data did not improve referral rates to community services once at-risk patients were identified [28]. When SBDH data are operationalized in a poorly functioning algorithm, these false positives and negatives indicate that a health system spends unnecessary resources evaluating several patients not at high risk, whereas groups of patients needing social services remain unidentified [74,89]. To address this phenomenon, algorithms may need to be tested with new data as predictive analytics methods that use SBDH risk data have evidenced limited generalizability outside of the original sample data where the model was developed [26,46].

Within a model’s discriminatory power is the challenge of temporality in analytic models. Specifically, further research and development are necessary to determine how to capture changing social risk factors related to changing life circumstances throughout a person’s life or epoch [90]. For example, by structural design, a model may overlook an individual’s loss of income through unemployment or community changes not reflected in neighborhood data [74]. Thus, time-oriented models will be better able to elucidate the persistence or amelioration of disparities.

Further guidance on analytic challenges, such as optimizing the appropriate separation of high- and low-risk cases, will be crucial as part of future, wide-scale dissemination of SBDH-focused predictive modeling tools. To advance predictive analytics and increase generalizability across the United States, there should also be open-source SBDH resources for methods and databases that leverage previous SBDH research and development [91,92]. Globally, the Research Data Alliance could create a working group to spearhead the creation of open-source SBDH data sources and facilitate work toward interoperability [93].

Once a single health care system renders SBDH data useful through advanced data science, they must find ways to disseminate these advances. The lack of standardization of SBDH data and collection processes prevents the interoperability and integration of modeling into diverse platforms [91,92] and impacts the creation of SBDH products for EHRs [94]. For greater interoperability, we need a standard, practical coding system for SBDH factors that goes beyond vendor-specific coding [91,92]. Such an endeavor is presently being pioneered by the Social Interventions Research and Evaluation Network through the HL7 Gravity Project [95].

In recognition of the emerging field of SBDH predictive analytics, steps toward developing consensus and further evaluative work are needed to produce best practice guidelines for the use of SBDH data in predictive modeling [91]. There is wide variability in the choice of data sources, risk factors, targeted outcomes, geographic levels, and analytic approaches in the SBDH predictive models. Each of these model components can impact a tool’s accuracy and appropriateness for use in a particular setting or context. At present, there is a very limited understanding of the impact of these parameters on the effectiveness of the SBDH predictive model. Although endpoints such as health care cost and utilization may seem similar, the choice of health outcome in a model can obscure the path from social risk to health. Best practice guidelines should include transparency of model validation methods for various outcomes to ensure that modeling methods can be replicated in other populations [91]. The use of SBDH variables in predictive modeling is relatively new. Developing consensus might be premature in such circumstances and evaluative work must occur beforehand. However, to form guidelines, it is critical to consider standardization in SBDH predictive analytics and to organize the discourse early on. Such discourse would facilitate data sharing, create open-source tools and algorithms, and set expectations.

Although the methods and analyses addressing SBDH have matured substantially over the past decades, an expanded data infrastructure and more research are necessary to gain a full understanding of how SBDH manifests throughout a person’s life [96]. Present health analytics platforms are generally not built to advance our knowledge base in this area. Rather, they are often intended to give health systems or insurers a leg-up over their competition in achieving financial or pay-for-performance targets. There should be a national agenda to develop and share technology and human resources and strategies to support efficient data extraction, evidence-based development, and effective analytics and reporting within and across institutions in the United States [92]. For-profit entities also have a vested interest to create better predictive models. Such shared desire would be an incentive for them to participate in the development of a shared evidence base, resulting in the creation of better predictive models.