Background
Social determinants of health (SDOH), such as access to nutritious food, transportation, employment, and stable housing, are significant drivers of health outcomes [
1,
2]. For health care systems to successfully reform delivery towards value, prevention, and effective population health management, they need to assess and respond to social needs associated with downstream consequences of the SDOH [
3,
4]. To this end, health care systems and payers are increasingly collecting population- and individual-level data on social needs, including food insecurity, unemployment, housing instability, and transportation barriers [
5,
6].
In 2016, the Protocol for Responding to and Assessing Patients’ Assets, Risks, and Experiences (PRAPARE) was developed by the National Association of Community Health Centers and partnering organizations as a screening tool and corresponding clinical workflow to assess and respond to patients’ social needs. PRAPARE has been the most prevalent social needs assessment in the United States, increasingly used by hospitals, health systems, and health plans [
7,
8]. The PRAPARE screening assessment bridges social risk and clinical risk indicators by being embedded into electronic health record (EHR) systems and has facilitated national standards surrounding social risk data capture, reporting, and population health and care management activities [
9].
Despite these features and the prevalence of PRAPARE, there is little evidence on the relationship between data from PRAPARE and clinical outcomes of interest [
10]. A recent systematic review of PRAPARE and similar social needs screening assessments found little evidence to evaluate predictive validity [
11]. Addressing this gap is critical as health care systems consider delivery reforms that embrace population health management and care coordination across the health and social care continuum [
4,
12]. With a better understanding of how social needs screening assessment data predicts clinical risk, health systems and payers can identify patients with health related social needs that are most predictive of poor health, provide care management to promote linkages to appropriate wraparound services and community resources to patients most likely to realize health benefits, measure the impact of interventions, manage patient panels, and inform care team composition [
13]. As health systems and payers increasingly invest in collecting this information [
4,
14,
15], there is a need to evaluate the relationship between patients’ social needs and their clinical risk to design tailored interventions.
This cross-sectional study examined the relationship between responses to PRAPARE and cardiometabolic clinical outcomes among patients in a federally-qualified community health center (FQHC). We utilized two approaches to examine the association between social needs assessment data and the likelihood of the following clinical indicators of cardiometabolic health status: obesity, hypertension, and atherosclerotic cardiovascular disease (ASCVD) 10-year risk. These outcomes are important because cardiometabolic disease is the leading cause of death of people in the United States, and obesity and hypertension are both modifiable risk factors [
16]. Our goals were to (1) better understand the social needs and health status of a defined population and (2) evaluate PRAPARE social needs assessment data’s association with cardiometabolic health status to inform risk prediction, stratification, and population health management. We hypothesized that models using social needs data from PRAPARE would have moderate performance for three cardiometabolic health outcomes.
Discussion
Assessing and responding to social needs is a major priority for health care systems seeking to deliver high-value care and improve population health. Efforts to better integrate these activities into routine clinical encounters and standard of care [
6] include social needs EHR documentation strategies [
33], innovative care models [
34,
35], and cross-sector collaboration [
36]. This study builds on the existing literature by evaluating the relationship between responses to a standardized social needs assessment and accepted measures of cardiometabolic health outcomes. Our intention is to highlight practical analytical tools for leveraging social needs information from PRAPARE and similar screening tools [
37,
38] to better understand the association between social needs and commonly-studied cardiometabolic outcomes. The application of these analytical tools has the potential to enhance value-based care, population health management, panel management [
13], and integrated social-medical care model design and implementation [
39,
40].
We evaluated the relationship between data from PRAPARE and three cardiometabolic outcomes using two predictive analytic approaches. We found that social needs were more prevalent in patients with hypertension and borderline-or-higher ASCVD risk. Interestingly, social needs were less prevalent in obese patients compared to those who were not obese. Lack of housing, high stress, and access to medicine and health care were the only social needs that were selected in models across more than one clinical risk outcome. These social needs may be proxies for additional, interrelated non-medical drivers of health and do not represent causal mechanisms between social needs and clinical outcomes.
The presence of social needs was associated with lower prevalence of obesity. Existing literature indicates that this counter-intuitive finding may be because obesity has a unique and multifactorial relationship to social needs and SDOH that varies by cultural context, race, ethnicity, and gender [
41]. This finding highlights the importance of both understanding that associations between social needs and clinical outcomes depend on how adverse outcomes are defined, as well as the need for cautious interpretation of the directionality of the effect based on the variables retained in the model.
We hypothesized that the predictive analytic approaches would demonstrate moderate performance for all three cardiometabolic outcomes (c-statistic > 0.65); we found support for this hypothesis for ASCVD risk and hypertension, but not for obesity. We noted that model performance for ASCVD risk was very high even without including the clinical parameters or health behaviors used to calculate an ASCVD risk score as predictor variables (blood pressure, total and high density lipoprotein cholesterol, diabetes diagnosis, smoking status, and hypertension treatment). The comparison across nested models further contextualizes this finding. The inclusion of social needs variables for Model 4 resulted in a statistically significant increase in prediction performance compared to Model 3 in the backwards stepwise approach. The small decrease in the LASSO approach performance between the same models may be due to overfitting on a small sample, despite regularization to prevent it, and limitations to increases to prediction performance with increasingly granular data on an individual’s unmet social needs. Nevertheless, this still suggests that social needs assessment data as a whole has useful associations with clinical outcomes in the absence of information on health behaviors and biometric data.
We found that the stepwise logistic and machine learning LASSO regression models demonstrated similar performance, a finding consistent with prior studies assessing the performance of predictive models [
42,
43]. A potential explanation is that the advantages of more advanced, resource-intensive machine learning techniques like LASSO regression or random forest models, compared to stepwise logistic regression models, require more observations and dimensionality to become apparent [
31]. The number and functional form of included variables also can influence results, with similar research demonstrating better performance for machine learning approaches when more variables and continuous variables are used [
44]. This underscores the importance of considerations regarding data transformation, variable functional form, and data missingness when using and selecting a predictive analytical approach.
This study has several limitations. Because PRAPARE was administered only to a subset of patients referred to behavioral health, there was less variation in social need levels to base predictive analytics. The smaller sample size of patients from multiple clinics within a FQHC may have limited prediction performance and generalizability. Generalizability may also be limited by the setting—FQHCs in the southeastern United States. Finally, the prediction modeling approaches used in this analysis do not allow for making conclusions on potential causal inference. Despite these limitations, this study contributes to an emerging evidence base that suggests the formal and pragmatic validity of PRAPARE and provides insights into how social needs data could be used in outpatient settings to predict cardiometabolic health outcomes.
Predictive analytics may have the potential to proactively identify patients at higher risk for poor health outcomes who could benefit from an intervention, even in the absence of data obtained in current screening methods for cardiometabolic outcomes; our results align with previous literature suggesting limits to the utility of SDOH data for risk prediction [
45]. In other words, additional SDOH data do not always lead to statistically significant improvements in prediction performance. This is relevant as payors, including state Medicaid programs [
46], collect social needs data for new enrollees to identify patients at risk for worsening medical complexity based on social needs in order to improve population health. Our findings suggest that performance may depend on how clinical outcomes are defined and that relationships between social needs assessment data and outcomes vary by disease pathway.
Future research directions
Future research should evaluate social needs prevalence and association with additional clinical outcomes, using prospective data to understand how social needs data can be used to predict clinical risk and, ultimately, improve population health. Though this study focused on moderate clinical outcomes that may be useful for proactive intervention, outcomes that correspond to more severe clinical conditions (e.g., A1c > 8.0%, ASCVD > 20%) may have differing and perhaps stronger associations with unmet social needs. Ideally, this would include linking multiple data sources to comprehensively describe patient behaviors and environment in addition to information on social needs to predict other clinical risk and health status indicators including uncontrolled diabetes, co-morbidity burden, and behavioral health outcomes. Moreover, risk prediction around social needs will only add significant value if it is coupled with implementing evidence based responses to social needs that meaningfully address social needs to improve health outcomes in a cost-effective manner. These responses, or social care interventions, will need to be rigorously tested in diverse settings among a study sample of sufficient size to detect its impact on outcomes of interest including medication adherence, utilization, and cardiovascular health status.
Understanding the relationship between clinical outcomes and social needs may have important ramifications for how payers adjust for risk. In addition, future research should also evaluate the relationship between social needs assessment data and the likelihood of requiring costly types of heath care utilization including inpatient and emergency department visits. As social need screening becomes wider spread, there is a need to understand how this data can be used to improve health equity as health systems focus on improving population health. For example, our findings and future research can inform the business case for health systems to implement interventions to address social needs, which has the potential to narrow disparities in care resulting from social and economic inequities [
47]. A critical step will be to design quality measures that complement care guidelines to focus support on medically vulnerable patients with unmet social needs [
48].
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.