For question 1, we analyzed the full All of Us cohort (N=372,397) and characterized their responses to all the SDoH questions identified by the expert panel (described in the Variables section). For question 2, we analyzed all participants (n=12,913) who had valid responses to the SDoH questions identified in question 1 and used them to identify subtypes and their risks of specific outcomes.

For question 1, the expert panel was asked to review all 1113 questions across 6 All of Us non–COVID-19 health surveys, each of which was attempted once per participant (The Basics, Lifestyle, Overall Health, Personal/Family Health History, Health Care Access and Utilization, and SDoH), and the 2843 Systematized Medical Nomenclature of Medicine (SNOMED) codes related to SDoH [67]. The expert panel arrived at a consensus for the SDoH across the surveys and the SNOMED codes. As the SDoH-related SNOMED codes in the EHRs had very low use (see Multimedia Appendix 2 for a characterization), they were not further analyzed.

For question 2, to identify and analyze the SDoH subtypes, we used the following variables:

To analyze the range and responses to the survey questions, we first characterized all SDoH in All of Us at two levels of granularity—(1) SDoH questions based on the surveys used to collect the data and (2) SDoH factors, which were categories of the SDoH questions to form a coarser-grained classification (see Figure 2, which explains SDoH questions, factors, and subtypes).

To identify the SDoH questions, members of the expert panel independently used their domain knowledge about SDoH to identify and code the SDoH questions and examine their range with respect to the 5 HP-30 domains using the following steps: (1) reviewed all 1113 questions across 6 health surveys (excluding 2 related to COVID-19) and extracting all SDoH questions that were relevant, (2) transformed all positive or value-free questions into negative phrases and abbreviating them for interpretability in the graphs (eg, How often do you have someone help you read health-related materials? was changed to No one to help read health materials), (3) reverse coded and dichotomized the abbreviated SDoH questions (eg, always or often=1 and never, occasionally, or sometimes=0), and (4) categorized the SDoH questions into 1 of the 5 HP-30 SDoH domains (economic stability, education access and quality, health care access and quality, neighborhood and built environment, and social and community context). The expert panel subsequently met and collaboratively resolved any differences among their coding schemes to arrive at a consensus (see Multimedia Appendix 3 for the 110 SDoH questions and their consensus coding by the expert panel).

To characterize the SDoH factors, the expert panel arrived at a consensus to categorize one or more of the aforementioned SDoH questions in All of Us into SDoH factors and examined their range with respect to HP-30 using the following steps: (1) reviewed the subgrouping labels of questions in the All of Us surveys and integrating them to categorize the SDoH into factors, (2) coded a participant as having a “1” for an SDoH factor if they answered one or more of the questions within that factor with a “1,” and (3) categorized the SDoH factors into 1 of the 5 HP-30 SDoH domains (economic stability, education access and quality, health care access and quality, neighborhood and built environment, and social and community context; see Multimedia Appendix 3 for the 110 SDoH questions, their consensus coding into 19 SDoH factors, and mapping to the 5 SDoH domains from HP-30).

The aforementioned knowledge-based classification of SDoH questions and SDoH factors was analyzed to examine their range (with respect to the 5 HP-30 domains) and their responses (across all participants in All of Us) using the following four methods: (1) bar graph displaying the number of participants who had valid answers (all responses other than “skip” or “choose not to answer”) to each of the SDoH questions, sorted by survey based on mean response, and then sorted by raw response within each survey. Finally, to analyze their range, each bar was colored to denote one of the 5 SDoH domains defined by HP-30; (2) Venn diagram showing how many participants had cross-sectionally valid responses to all identified SDoH questions or factors; (3) table describing the number and proportion of race, ethnicity, sex, gender, and age between those who answered the SDoH questions or factors and those who did not have valid responses; and (4) frequency distribution of the number of SDoH questions or factors across participants who had valid responses for all the SDoH questions. The aforementioned plots are shown in the Results section.

The original data collection by the All of Us program was approved by an institutional review board as described on the web [76].

The secondary analysis of the All of Us data conducted in this work did not receive approval or exemption from an institutional review board. Such an approval or exemption is not required as described on the web [77].

Therefore, the authors had permission to conduct a secondary analysis of the data.

The expert panel identified 110 questions from 4 surveys (The Basics, Overall Health, Healthcare Access and Utilization, and SDoH). Of these 110 questions, 110 (100%) were abbreviated, and 48 (43.6%) were negatively worded and coded (Multimedia Appendix 3). The 110 SDoH questions were further categorized into 19 SDoH factors (one of these was Delayed medical care, which was used as an outcome).

As shown in Figure 4A, the number of valid responses to each of the 110 SDoH questions was largely dictated by the surveys in which the responses were solicited. SDoH from 2 surveys (The Basics and Overall Health) had the most valid responses (mean 349,434, SD 23,556), followed by Healthcare Access and Utilization (mean 149,898, SD 6146) and, finally, the SDoH survey (mean 55,960, SD 1083). This pattern of responses matched how answers to each of the surveys were solicited—at enrollment, all participants are required to do The Basics and Overall Health surveys, and then, on a rolling basis, responses to the other surveys are solicited. The SDoH survey was the last survey that was solicited, which explained it having the lowest number of responses. As shown in Figure 4B, this pattern of missingness held for the responses at the SDoH factor level, which was not unexpected as the SDoH factors were aggregations of the SDoH questions. However, as shown in Figures 4A and 4B by the uneven number of valid responses within each survey block, there were several SDoH questions that had invalid responses (“skip” or “choose not to answer”) at both levels of granularity: The Basics (339,254/5,655,412, 5.99%), Healthcare Access and Utilization (341,516/5,587,957, 6.11%), Overall Health (32,669/744,126, 4.39%), and SDoH (83,699/3,206,035, 2.61%). Furthermore, the proportion of valid to invalid responses among them was significantly different for the SDoH questions (N=365,237, χ²₂=57.5; P<.001) and for the SDoH factors (N=372,063, χ²₂=75.6; P<.001).

As shown by the colored bars in Figure 4, the surveys spanned the full range of the 5 SDoH HP-30 domains. The SDoH questions in The Basics and Overall Health surveys were predominantly related to economic stability (blue) and social and community context (purple), those in the Healthcare Access and Utilization survey were all related to that topic (green), whereas those from the SDoH survey were a mix of all 4 domains. Overall, the 4 surveys contained 110 SDoH questions that together had 100% coverage of the 5 HP-30 domains (social and community context: n=38, 34.5%; neighborhood and built environment: n=19, 17.3%; economic stability: n=10, 9.1%; education access and quality: n=2, 1.8%; health care access and quality: n=42, 38.2%). This characterization suggests that, while the SDoH in All of Us have broad domain coverage across the surveys, analysis of them requires access to all 4 surveys, each of which has different levels of completion and valid responses.

Given the large degree and systematic nature of missingness in 2 of the 4 surveys, we could not use multiple imputation to estimate the values. Therefore, we had to find a subset of participants who had valid responses to all the SDoH questions. An examination revealed that 2 SDoH questions had <10% of responses (English verbal frequency: 6193/371,942, 1.67%; Neighborhood has no recreation spaces: 31,152/371,942, 8.38%), accounting for the largest loss in cohort size with valid responses. Therefore, these questions were dropped from further analysis. Furthermore, one question required a branched response (Living situation branching to Did not live in a house), and these responses were merged. Finally, as we used Delayed medical care as an outcome, 8.2% (9/110) of the questions related to that topic were removed, resulting in a total of 98 SDoH questions. As shown in Figure 5, a Venn diagram of the overlap among the valid responses across the surveys revealed that 3.47% (12,913/372,397) of the participants had valid responses to all 98 SDoH questions.

As shown in Figure 6, participants had a median of 15 SDoH question co-occurrences and a median of 9 SDoH factor co-occurrences. Furthermore, participants of racial and ethnic minority groups who had valid responses to the 110 SDoH questions had a significantly higher median number of co-occurring SDoH compared to the equivalent White population (median 20 for participants of racial and ethnic minority groups; median 14 for White participants; P<.001). These results show the high co-occurrences of SDoH at both levels of granularity, with a significant difference in median co-occurrences between the White participants and the participants of racial and ethnic minority populations with valid responses.

As the cohort size dropped to 3.47% (12,913/372,397), we analyzed how that impacted the demographic distribution compared with the overall All of Us dataset. As shown in Table 1, there were statistically significant differences in race (N=372,397, χ²₅=2073.1; P<.001) and ethnicity (N=372,397, χ²₉=6292.2; P<.001) between the 2 cohorts after multiple testing correction, with a higher proportion of White participants having valid answers than participants of racial or ethnic minority groups. Furthermore, there was a statistically significant difference in age between the participants who had valid answers and those who did not (H1=148.08; P<.001). These results show the demographic differences between the cohort with complete and valid answers to the SDoH questions and the full All of Us dataset, necessitating the use of IPW–generated weights to address those imbalances, as discussed in question 2.

The cohort used to identify the subtypes consisted of 12,913 participants, of whom 12,886 (99.79%) had valid IPW–generated weights. The latter cohort was split randomly into the training and replication datasets, each with complete data for 18 SDoH factors (identified in question 1) in addition to the 3 outcomes (depression, delayed medical care, and ER visits in the last year) and covariates (demographics). The results are organized based on the 3 parts of the HIT framework described in Figure 3.

The subtypes were identified by using a bipartite network where the edges were weighted using the IPW-generated weights to account for the imbalance in demographics between our cohort and the full All of Us dataset. The weighted bipartite network of the training dataset (n=6492) and the 18 SDoH factors revealed 4 biclusters with statistically significant bicluster modularity (Q=0.13; random-Q=0.11; z=7.5; P<.001). As shown in Figure 7, there were 4 clusters with participant subgroups and their most frequently co-occurring SDoH factors (Cluster 1 [pink]: low educational attainment, low literacy, low income, not employed, food insecurity, and housing insecurity; Cluster 2 [green]: difficulty affording medical care, discriminatory experiences in everyday life, discriminatory experiences in medical settings, and poor interactions with providers; Cluster 3 [blue]: poor neighborhood cohesion and poor relationships with others; and Cluster 4 [gray]: language barrier, lack of health care coverage, mismatched provider characteristics, disadvantaged neighborhood characteristics, disadvantaged demographics, and low supportive relationships). These co-occurrences of SDoH factors were significantly replicated in the replication dataset (real RI=0.88; random RI=0.62; P<.001). As shown in Figure 8, while the 18 SDoH factors have a hierarchical relationship with the 5 knowledge-driven HP-30 domains (shown on the left), those same SDoH factors have a more complex relationship with the 4 data-driven biclusters (shown on the right).

Table 2 shows the association of each subtype with the 3 outcomes. As shown by the italicized row, Cluster 1 (low educational attainment, low literacy, low income, not employed, food insecurity, and housing insecurity) had a significantly higher OR for each of the 3 outcomes than Cluster 4 (mismatched provider characteristics, disadvantaged neighborhood characteristics, lack of health care coverage, disadvantaged demographics, low supportive relationships, and language barriers). Furthermore, within the Depression outcome, each of the clusters had a significantly higher OR than one other cluster, forming a ranking of risk among all the 4 clusters (1>3>2>4). In contrast, Delayed medical care had 2 other significant associations (2>1 and 3>4), with ER visits in the last year having only 1 significant pairwise association that fit into the overall trend.

As shown in Table 3, this trend continued in the enrichment analysis of association with living in a state with No Medicaid expansion. As shown, Cluster 1 had a significantly higher OR than Cluster 4 in addition to having a significantly OR than the other clusters. The overall results suggest that Cluster 1 and Cluster 4 form “book ends” representing the high and low ends of the risk spectrum among the clusters.

The expert panel examined the co-occurrences of SDoH factors within each bicluster shown in the network visualization (Figure 7) and integrated them with the quantitative ORs in Tables 2 and 3. The consistent “book ends” result where Cluster 1 had significantly higher ORs than Cluster 4 across all 4 variables was of strong interest and interpreted as follows: (1) Cluster 1 was labeled Socioeconomic barriers as it comprised multiple high-risk SDoH. These co-occurring SDoH could have resulted from cascades over time, such as low educational attainment, potentially leading to lower rates of employment and lower income with higher rates of food and housing insecurity. Such cascading factors can be perceived as being relatively unmodifiable, leading to a higher risk of chronic stress and depression. Furthermore, the strong association of this subtype with the outcomes Delayed medical care and ER visits in the last tear and the fact that participants in this subtype were more likely to be from a US state with No Medicaid expansion provided a more comprehensive understanding of this high-risk SDoH subtype; (2) Cluster 4 was labeled Sociocultural barriers as it contained a combination of SDoH related to disadvantaged neighborhood characteristics and low supportive relations in addition to language barriers and mismatched provider interactions. In contrast to the socioeconomic barriers in Cluster 1, many of the sociocultural barriers could be perceived as potentially modifiable, resulting in a lower risk of depression, delayed medical care, and ER visits. Participants who matched this profile could be screened for language and communication barriers, useful for providing culturally competent care, identifying providers who better match the profile of the individuals, and providing resources to facilitate contact with matching nationality or cultural groups in the vicinity or on the web.

While Clusters 1 and 4 formed the “book ends” of risk across the 3 outcomes, potentially caused by relative differences in the unmodifiability of their frequently co-occurring SDoH, Cluster 2 was flagged as critical and labeled Lived experience barriers. The SDoH in this cluster included discriminatory experiences in everyday life and in medical settings in addition to poor interactions with providers and difficulty affording medical care. These frequently co-occurring SDoH could explain why this subtype had a significantly higher OR for Delayed medical care than Cluster 1. Finally, Cluster 3 was labeled Social context barriers as the SDoH were related to poor neighborhood cohesion and relationships with others. While not as critical as Clusters 1 and 2, this cluster still had a significantly higher OR for Depression than Cluster 4. Together, the 4 clusters could explain how different degrees of unmodifiability in frequently co-occurring SDoH might impact health outcomes.

The expert panel and the ethicist concluded that clinicians treating patients who match each subtype profile could be alerted of specific risks and, consequently, motivate a discussion about mental health and consequences of delayed medical care with the goal of collaboratively exploring options and solutions with the patients. The results could also be useful for resource planning in hospitals to ensure that there is adequate staff to address the needs of the populations they serve, and for proposing health care policies to address the critical connection between specific combinations of SDoH and their impact on public health. For example, many health equity policies categorize Americans based on sociodemographic variables like race and income, which are proxies of need rather than the needs themselves. Instead, such policies could categorize Americans based on SDoH subtypes and their risks to more precisely allocate resources based on combinations of real-world needs.

Furthermore, the subtypes did not have a one-to-one mapping to the 5 SDoH domains defined by HP-30. As shown in Figure 8, these data-driven clusters have a complex relationship with the SDoH domains and factors. While one subtype belonged to a single domain (the Social context subtype belonged to the Social and community context domain), 3 of the 4 subtypes belonged to ≥2 domains (eg, the Socioeconomic barriers subtype belonged to the Economic stability and Education access and quality domains). Such interdomain relationships reflect how SDoH co-occur in the real world, reflecting the complex cross-domain interactions described in the Dahlgren-Whitehead model (Figure 1). These relationships could be useful for refining conceptual models to explain the complex association between SDoH and adverse health outcomes and build more accurate SDoH models for predicting adverse health outcomes.

The analysis revealed 3 types of missingness. The first was rollout missingness. This type of missingness was largely dictated by how the surveys were rolled out to participants. As all participants at enrollment are required to do The Basics and Overall Health surveys, these surveys had the highest number of responses, followed by the later solicited Healthcare Access and Utilization and SDoH surveys rolled out more recently in 2022. This order of rollout was the main source of missingness, resulting in a precipitous reduction in cohort size for those who had answers to all the SDoH questions. The second type was valid answer missingness. As participants can choose not to answer any survey questions, the data contained “skip” and “choose not to answer” responses. However, these responses accounted for a much smaller reduction in cohort size for complete data than rollout missingness. The third type was low use missingness. Although there were 259 SDoH SNOMED codes, only 93 (35.9%) had such information for >20 participants that are allowed to be reported. This could be because most clinicians currently do not screen for SDoH as it is typically done by a social worker. Furthermore, we also attempted to use 3-digit zip codes to determine which subtypes had a significant association with living in a state that did not offer Medicaid expansion. However, 13.07% (1688/12,913) of the participants did not have zip code information (which was adjusted by using IPW).

Together, the aforementioned 3 types of missingness impacted the size of the resulting cohort that had valid answers in the following 2 ways, which is the minimum number of participants that are allowed to be reported by All of Us. First, there was a drastic reduction in cohort size by 93.5%. However, because of the size of the overall dataset (N=372,397), we were still left with a large cohort (n=12,886), which, to the best of our knowledge, is the largest set of individuals to be analyzed for such a wide range of SDoH. Second, there were significant differences in the proportion of race, ethnicity, and age in the aforementioned cohort when compared to the overall All of Us population. Specifically, the cohort with valid answers had significantly more White, non-Hispanic, or older participants when compared to the overall cohort. This could potentially be because, once a participant has been enrolled, there is a 90-day delay in sending subsequent solicitations to complete surveys, a policy that is currently being reassessed due to its impact on missingness. Therefore, we had to correct this imbalance in demographic proportions by using IPW with the goal of identifying subtypes that were representative of the overall All of Us cohort.

Because our goal was to use machine learning methods to identify SDoH subtypes, we encountered uneven granularity in the SDoH questions. Some questions were fine grained and highly correlated and, therefore, would cluster more strongly because of the nature of the granularity of the questions, not because of the SDoH mechanisms. To address this uneven granularity and make the results more interpretable, we used SDoH factors that had a coarser but more consistent level of granularity. We chose this approach because SDoH factors had already been defined; were understood by the expert panel, enabling high domain fidelity; and appeared to be at the right level of abstraction useful for clinical applications, such as referring a patient to the appropriate social services. However, because the use of coarse-grained variables loses information, future research could explore aggregating only those SDoH questions that are highly correlated while preserving the rest at the finer level of granularity and explore computational methods to merge SDoH questions into SDoH factors.

We used 3 types of code to conduct the analysis for both research questions. The first was automatically generated code to extract the cohort, produced by All of Us once a cohort was selected using the point-and-click interface. This code was adequately scalable and generalizable and so will not be discussed further. The second was customized code to extract specific parts of the data. For example, the analysis of co-occurrences required customized code in R (R Foundation for Statistical Computing) to plot the diagrams in Figures 4–6. As expected, these tasks required strong programming skills, but fortunately, we did not encounter any coding or execution problems using the R or Python (Python Software Foundation) programming languages. However, there were significant server issues that hampered our analysis. Although the workbench instructions stated that code running on the workbench for >2 weeks would be terminated and all intermediate results would be deleted, we frequently encountered our work disappearing at shorter intervals. These disruptions resulted in a higher consumption of the free server time credits resulting in fewer analyses that could be conducted. The third type of code was machine learning code that we had previously developed and disseminated on Comprehensive R Archive Network [78-80] to conduct the bipartite network analysis and the significance testing and visualize the network. As this code was designed to be generalizable and scalable, we did not encounter any issues in the execution of our code (in addition to the same server issues mentioned previously). Finally, the visualization of our networks worked as expected, and we used them to help interpret the patterns in the data.

Our code for the first 2 steps of the HIT framework is in Project Jupyter notebooks and has been used to analyze other cohorts that were filtered for age and previous conditions. For example, we extracted a cohort (n=4090) of participants with diabetes aged ≥65 years with complete data on 18 SDoH variables selected through consensus by 2 experienced health services researchers and guided by the Andersen behavioral model. The analysis [81,82] revealed 7 SDoH subtypes with statistically significant modularity compared with 100 random permutations of the data (All of Us=0.51; random mean 0.38 SD 0.0065; z=20; P<.001) and that were not only clinically meaningful but also significant in different degrees for the outcome. Our subsequent attempt at increasing the number of SDoH variables from 18 to 110 for participants with diabetes who had valid answers led to an extremely small cohort size (n=926; Multimedia Appendix 5) due to the missingness that we described previously. While this reduction resulted in our current strategy of analyzing all participants regardless of condition or age, these experiments demonstrate that our approach is generalizable to other subsets of the data.

The HIT model is designed to be extensible to include other methods. For example, the model could use other biclustering (eg, nonnegative matrix factorization [83]) and causal modeling methods and different types of classification (eg, deep learning [84]) and prediction methods (eg, subgroup-specific modeling [40]) to build the decision support system in the translational step (Figure 2). Furthermore, the model can integrate a wide range of datatypes to enable analysis of how each subtype is associated with them, resulting in a layered interpretation of the SDoH subtypes, as we have demonstrated. For example, as the percentage of participants who have genomic information increases (currently, >25% of our cohort have missing genomic information), our pipeline will be able to integrate such information into our analysis. Finally, the integration of different datatypes required a diverse team consisting of experts in machine learning, biostatistics, programming, clinical care, health services research, gerontology, and ethics to enable a 360° analysis and interpretation of the subtypes that was, therefore, aligned with the human-centered artificial intelligence approach [64-66]. Furthermore, the use of the workbench to share results through visualizations operationalized team-centered informatics [85] designed to facilitate multidisciplinary translational teams [86] working more effectively across disciplinary boundaries with the goal of analyzing subtypes and designing targeted interventions.

This set of notebooks generates four plots that can be used by other researchers on All of Us to characterize any cohort: (1) valid response plot to show how many participants have data with valid responses and colored by SDoH domain, (2) Venn diagram showing how many participants have valid responses for all questions within each survey, and (3) frequency distribution plot showing co-occurrence of SDoH across the selected cohort. This set of tools should enable researchers to characterize SDoH across different cohorts to help determine methods that are appropriate to adjust for missingness in those cohorts.

This set of notebooks can be used to conduct the following analyses: (1) bicluster modularity of a cohort with the 18 SDoH factors to identify the number and members of biclusters and the measure Q representing the quality of the biclustering, (2) visualization of the bipartite network, and (3) significance of the network with respect to null models.