URMC – Dreisbach

Social Determinants of Health Factors as Upstream Predictors of Postpartum Hemorrhage

Sponsor:

Caitlin Dreisbach

Coauthors:

Team 4, ​​DSCC 383W Data Science Capstone

¹Goergen Institute for Data Science and Artificial Intelligence, University of Rochester, Rochester, NY, USA

²School of Nursing, University of Rochester, Rochester, NY, USA

1. Introduction

Our project builds on this infrastructure by analyzing whether specific SDOH constructs are associated with an increased risk of postpartum hemorrhage. Using All of Us survey data and linked EHR records, we aim to identify which upstream social factors—such as income, educational attainment, or perceived discrimination—are predictive of PPH in a pregnancy cohort. By focusing on the broader social context in which pregnancy and childbirth occur, our goal is to highlight modifiable population-level risk factors and inform interventions that advance maternal health equity.

2. Methods

2.1 Data Source: All of Us Research Program

This study utilizes data from the Controlled Tier Dataset v7 of the All of Us Research Program, a large-scale research initiative by the National Institutes of Health (NIH). All data access and analysis were conducted through the secure All of Us Researcher Workbench platform. Data include structured electronic health records (EHR), curated survey responses, and demographic metadata for participants who are historically underrepresented in biomedical research. This project specifically draws from pregnancy-related OMOP data and responses to the Social Determinants of Health (SDOH) survey module.

2.2 Cohort Builder

Using the Cohort Builder interface within the Workbench, we constructed a cohort of participants who meet the following inclusion criteria:

• Assigned female at birth (AFAB)
• At least one recorded pregnancy-related event in the EHR

To ensure comparability across participants, we limited inclusion to only each individual’s first pregnancy episode, as identified by the HIPPS (Hierarchical Pregnancy Phenotyping System) algorithm-based concept sets. This was done using the Full_HIPPS_Episodes concept, which classifies complete pregnancy episodes based on structured visit and procedure data. Participants who had 16 or more pregnancy episodes were excluded.

2.3 Concept Sets

We constructed and applied multiple Concept Sets to extract relevant data from both the EHR and SDOH domains:

1. Pregnancy Episodes: Using associated OMOP concepts, we extracted each participant’s first recorded pregnancy.
2. Postpartum Hemorrhage (PPH): PPH cases were identified using a delivery-anchored PPH indicator variable embedded in the HIPPS system.
3. SDOH Constructs: SDOH variables were obtained from the All of Us survey module and included validated constructs such as:
   • Perceived discrimination
   • Neighborhood cohesion
   • Housing instability
   • Food insecurity
   • Emotional support
   • Instrumental support

Survey variables were already pre-processed and linked by person_id. Participants with missing values in any of these key constructs were excluded from analysis.

2.4 Participant Flow and Cohort Derivation

The analytic cohort was derived from 74,451 All of Us participants assigned female at birth (AFAB) who had documented evidence of pregnancy. This initial dataset included 44,369 pregnancy episodes representing 24,094 unique individuals. To support downstream analyses, the dataset was enriched with key variables including socio-demographic characteristics, postpartum hemorrhage (PPH) status, mode of delivery, and Social Determinants of Health (SDOH).

To construct the final analytic cohort, the following exclusion criteria were applied:
(1) individuals with incomplete SDOH data were removed,
(2) those with 16 or more pregnancy episodes were excluded as outliers (n = 4), and
(3) only the first recorded pregnancy episode per participant was retained to ensure uniformity in the unit of analysis.

After applying all criteria, the final cohort consisted of 5,246 unique individuals with complete clinical and contextual data.

Table 1:  Cohort Derivation and Participant Flow

Participant Selection Step (number of individuals)	Reason for Exclusion (n)
All AFAB* participants in dataset (n = 172,401)	–
Identified first pregnancy episode (n = 24,094)	No pregnancy-related records (n = 50357)
Completed SDOH survey (n = 5,250)	Missing SDOH survey data (n = 18844);
Had 15 or less pregnancy episodes (n = 5246)	Had 16 or more pregnancy episodes (n = 4)
Had PPH  and completed SDOH survey (n = 106)	Did not have PPH (n = 4735);Incomplete SDOH construct data (n = 409)
Final analytical cohort (n = 5,246)	–

*AFAB= Assigned female sex at birth

Participants were excluded if they: (1) were under 16 years of age at the time of pregnancy, (2) had missing delivery-related information, or (3) lacked complete data in any of the SDOH domains. A flow diagram summarizing the cohort refinement process is provided in Table 1.

Subgroup analyses were performed within this cohort. A total of 515 participants experienced postpartum hemorrhage (PPH) in their first pregnancy, and 106 had both PPH and complete SDOH data, allowing for detailed stratified and predictive modeling.

2.5 Statistical Analysis

We conducted several statistical procedures to investigate the association between social determinants of health and postpartum hemorrhage:

• Descriptive statistics and group comparisons were used to summarize the cohort and compare characteristics between participants with and without PPH. Chi-square or Fisher’s exact tests were applied for categorical variables.
• Correlation analyses were performed among SDOH constructs within the full analytic pregnancy cohort (n = 5,250) to explore interrelationships among upstream social risk factors. This was visualized using a heatmap of pairwise correlations.

All analyses were conducted using harmonized electronic health record (EHR) and survey data accessed through the All of Us Researcher Workbench.

3. Results

3.1 Group Differences in SDOH Scores

We compared Social Determinants of Health (SDOH) scores between participants who experienced postpartum hemorrhage (PPH) during their first pregnancy and those who did not. On average, individuals in the PPH group exhibited lower levels of emotional support, instrumental support, and spiritual well-being, while reporting higher stress and perceived discrimination scores. These findings suggest that participants who experienced PPH may have faced more psychosocial challenges during pregnancy. The results are presented in Figure 2.

3.2 Variability in SDOH Domains

Beyond mean differences, we assessed variability within SDOH constructs. As shown in Figure 3, across most measures, the standard deviations are similar, indicating consistent variability within both groups. The standard deviation of emotional support and spiritual well-being was notably higher in the PPH group compared to the non-PPH group. This suggests a broader heterogeneity in psychosocial environments among individuals who experienced PPH, possibly reflecting diverse lived experiences and risk exposures.

Figure 2:  Mean SDOH Score Differences: First Pregnancy vs. With PPH

Figure 3: Standard Deviation of SDOH Scores: First Pregnancy vs. With PPH

3.3 Correlation Structure of SDOH Constructs

Using correlation heatmaps, we explored the structure of inter-variable relationships across three populations: all AFAB participants, those in their first pregnancy, and the PPH subgroup.

• Among all AFAB participants, we observed tight clustering within support-related scores (emotional, instrumental, and general), strong correlations within healthcare discrimination metrics (hcd_count, hcd_sum, hcd_mean), and a distinct cluster of psychosocial stress variables (loneliness and stress_sum). Neighborhood variables such as spa and nei appeared more independent, capturing distinct environmental dimensions (see Figure 4).
• In the first-pregnancy group, similar clustering patterns held, reinforcing the robustness of these constructs (see Figure 5).
• However, in the PPH subgroup, these relationships weakened considerably. Support-related scores were less tightly correlated, and stress/discrimination clusters were more fragmented. This disruption may indicate a more unstable or inconsistent relationship among upstream determinants in individuals with adverse maternal outcomes (see Figure 6).

Figure 4: Correlation Map of SDOH Domains in the AFAB Population

Figure 5: Correlation Map of SDOH Domains in the First Pregnancy Group

Figure 6: Correlation Map of SDOH Domains in the First Pregnancies With PPH

3.4 Study Population Overview

The analytic sample comprises 44,369 pregnancy episodes, representing 24,094 participants, and offers a diverse representation across various demographic and socioeconomic dimensions.

Figure 7: Distribution of Race and Ethnicity

3.4.1 Demographic Characteristics

• Race and Ethnicity: Among participants with recorded race, 41.2% identified as White, 17.7% as Black or African American, 3.0% as Asian, and 4.1% as more than one race. A substantial portion (31.1%) did not report race (see Figure 7).
• English Proficiency: Among non-native speakers, 2.9% report speaking English “very well,” 0.8% “well,” 0.4% “not well,” and 0.1% “not at all.”(see Figure 8)

Figure 8: Distribution of English Level

• Marital Status: 50.3% of the participants reported being married, while 11.6% were living with a partner. 21.1% of participants had never married, and 8.6% were divorced (see Figure 9).

Figure 9: Distribution of Marital Status

• Education:
  • 21.1% of participants have attained a college degree, and 19.4% hold an advanced degree.
  • Conversely, 21.2% have a high school diploma or equivalent, and 7.2% have education levels between ninth and eleventh grade (see Figure 10).

Figure 10: Distribution of Educational Level

Figure 11: Distribution of Annual Income

• Income:
  • 13.2% report annual incomes less than $10,000, and 11.0% between $10,000 and $25,000.
  • A significant portion (approximately 30%) chose not to disclose their income, indicating potential sensitivity around this information (see Figure 11).

3.4.2 Social Determinants of Health (SDOH)

• Housing Quality: 4.7% of participants report poor housing conditions, such as issues with ventilation, mold, or structural problems.
• Food Insecurity: 3.6% indicate experiencing food insecurity, reflecting challenges in accessing sufficient and nutritious food.
• Healthcare Discrimination: 15.3% report having experienced discrimination within healthcare settings, highlighting systemic issues that may impact health outcomes.
• Religious Participation: Religious engagement varies, with 6.3% never or almost never attending services, and 4.7% identifying as not religious. Regular attendance (monthly or more) is less common.
• Neighborhood Density: 13.6% reside in low-density areas, while 7.0% live in high-density environments, suggesting a range of urban and rural living situations that may affect access to resources.

3.5 Model Evaluation and Threshold Tuning

To evaluate the predictive utility of Social Determinants of Health (SDOH) variables for postpartum hemorrhage (PPH), we assessed model performance under two decision thresholds: the default value of 0.5 and an optimized threshold of 0.8348 derived by maximizing the F1-score along the precision-recall tradeoff curve.

At the default threshold of 0.5, the model achieved:

• Accuracy: 64%
• Recall for PPH cases: 63%
• Precision for PPH cases: 0.02
• F1-score: 0.03

This indicates that the model was sensitive to identifying PPH cases (high recall) but also generated an excessive number of false positives, significantly limiting its practical use in clinical screening (see Figure 12a).

To improve balance, we computed an optimal threshold by identifying the point on the precision-recall curve where the F1-score peaked, yielding a new cutoff at 0.8348. At this threshold:

• Accuracy: increased to 98%
• Precision for PPH: rose to 0.06
• Recall for PPH: dropped to 0.08
• F1-score: improved slightly to 0.07

As shown in the confusion matrix at this threshold (see Figure 12b), the model successfully minimized false positives but missed the majority of true PPH cases—correctly identifying only 10 out of 128. This reflects a classic tradeoff in imbalanced classification: higher thresholds improve specificity but reduce sensitivity for the minority class.

 (a) Default Threshold (0.5)                     (b) Optimized Threshold (0.83)

Figure 12: Confusion Matrix at Two Thresholds: (a) Default Threshold, (b) Optimized Threshold

Figure 13: Precision and Recall Across Thresholds with Optimal Cutoff

The precision-recall curve (Figure 13) visually highlights this tradeoff. As the decision threshold increases, precision improves marginally while recall declines steeply. This pattern is a direct consequence of extreme class imbalance in our dataset (only 128 PPH cases among 13,311 participants), which biases standard classification metrics such as accuracy.

These results underscore that while SDOH features contain some predictive signal, their standalone use may not support reliable prediction of rare and high-stakes outcomes like PPH. Future modeling efforts may require alternative strategies such as resampling techniques, ensemble models, or the integration of richer clinical variables.

4. Discussion

4.1 Implications for Postpartum Hemorrhage. The intersection of low income and limited education can exacerbate the risk of PPH through several mechanisms:

• Delayed or Inadequate Prenatal Care: Financial constraints and limited health literacy may lead to delayed initiation of prenatal care or fewer prenatal visits, increasing the risk of undetected complications^{10, 13}.
• Limited Access to Quality Healthcare: Lower-income individuals may have restricted access to high-quality obstetric services, including emergency care necessary to manage PPH effectively¹¹.
• Increased Stress and Comorbidities: Socioeconomic stressors are associated with higher levels of chronic stress and comorbid conditions, which can complicate pregnancy and delivery¹⁴.

These factors underscore the importance of addressing socioeconomic disparities to improve maternal health outcomes.

4.2 Marital Status and Social Context

Half of the participants in our cohort reported being married, while an additional 11.6% were living with a partner. Together, these groups may benefit from household-level social support, which previous studies have associated with improved maternal outcomes and reduced risk of complications like postpartum hemorrhage. Conversely, 21.1% of participants had never married, and 8.6% were divorced, with smaller portions reporting separation or widowhood. These groups may face varying degrees of social isolation or economic vulnerability, which can contribute to stress and reduced access to healthcare resources, indirectly impacting maternal outcomes^{15, 16}.

4.3 Racial and Ethnic Diversity

The racial breakdown in this sample provides an opportunity to explore whether similar disparities are observed in our PPH subgroup and to what extent social determinants of health (SDOH) mediate these differences. For example, previous studies suggest that racial differences in PPH may be partially explained by differences in income, insurance status, neighborhood disadvantage, and perceived discrimination¹⁹.

5. Conclusion

This study examined the relationship between upstream Social Determinants of Health (SDOH) and postpartum hemorrhage (PPH) among participants in the All of Us Research Program. Three consistent patterns emerged across our analyses. First, individuals who experienced PPH exhibited higher levels of psychosocial strain, as evidenced by elevated stress and perceived discrimination scores. Second, correlation structures among SDOH constructs—such as support and neighborhood cohesion—appeared notably weaker in the PPH subgroup, suggesting environmental and social instability may accompany adverse maternal outcomes. Third, although predictive modeling based solely on SDOH features yielded limited performance due to class imbalance, the findings point to the potential utility of social context variables in informing risk stratification when combined with clinical data.

Future studies should address these gaps by leveraging larger, longitudinal datasets and integrating clinical features alongside SDOH. Logistic regression models can be applied within a 1:1 matched sample, pairing participants with and without PPH on key confounders such as age and parity. This approach could help isolate independent associations between specific SDOH factors (e.g., housing instability, food insecurity) and PPH risk. Additionally, as discussed with our faculty advisor (see Figure X), future work could investigate survey timing by comparing the timestamp of SDOH survey completion to the estimated start date of pregnancy. This would clarify whether upstream exposures preceded pregnancy and thus support stronger causal inferences. Ultimately, more granular, time-sensitive modeling of SDOH data could improve our understanding of how social adversity contributes to maternal health disparities—and inform upstream interventions to mitigate them.

Acknowledgements

We extend our deepest appreciation to Dr. Caitlin Dreisbach, Assistant Professor at the University of Rochester School of Nursing and affiliated faculty at the Goergen Institute for Data Science, for her continuous mentorship, expert feedback, and domain insight throughout the duration of this project.

We also thank the University of Rochester Data Science Program, particularly the DSCC 383W Capstone faculty team, including Prof. Ajay Anand and Prof. Cantay Caliskan, for creating a learning environment that enabled this research to take shape.

Finally, we thank the All of Us Research Program for providing secure access to an exceptionally diverse and rich dataset, without which this analysis would not have been possible.

References

[1] Committee on Practice Bulletins—Obstetrics. Practice Bulletin No. 183: Postpartum Hemorrhage. Obstet Gynecol. 2017;130(4):e168–e186. https://doi.org/10.1097/AOG.0000000000002351

[2] World Health Organization. Postpartum haemorrhage. 2023. Available at: https://www.who.int/teams/sexual-and-reproductive-health-and-research-(srh)/areas-of-work/maternal-and-perinatal-health/postpartum-haemorrhage

[3] Arab H. Coming Back from A Severe Postpartum Hemorrhage (PPH). J Gynecol Res Rev Rep. 2025;7(214). https://doi.org/10.47363/JGRRR/2025(7)214

[4] Sentilhes L, Gromez A, Clavier E, Resch B, Descamps P, Marpeau L. Long-term psychological impact of severe postpartum hemorrhage. Acta Obstet Gynecol Scand. 2011;90(6):615–620. https://doi.org/10.1111/j.1600-0412.2011.01119.x

[5] Centers for Disease Control and Prevention. Racial/Ethnic Disparities in Pregnancy-Related Deaths — United States, 2007–2016. MMWR Morb Mortal Wkly Rep. 2019;68(35):762–765. https://www.cdc.gov/mmwr/volumes/68/wr/mm6835a3.htm

[6] Kaiser Family Foundation. Racial Disparities in Maternal and Infant Health: Current Status and Efforts to Address Them. 2023. https://www.kff.org/racial-equity-and-health-policy/issue-brief/racial-disparities-in-maternal-and-infant-health-current-status-and-efforts-to-address-them/

[7] Koleck TA, Dreisbach C, Zhang C, et al. User guide for Social Determinants of Health Survey data in the All of Us Research Program. J Am Med Inform Assoc. 2024;31(12):3032–3041. https://doi.org/10.1093/jamia/ocae214​

[8] Healthy People 2030. Social Determinants of Health. U.S. Department of Health and Human Services, Office of Disease Prevention and Health Promotion. https://health.gov/healthypeople/objectives-and-data/social-determinants-health​odphp.health.gov

[9] Sentilhes L, Gromez A, Clavier E, et al. Long-term psychological impact of severe postpartum hemorrhage. Acta Obstet Gynecol Scand. 2011;90(6):615–620. https://doi.org/10.1111/j.1600-0412.2011.01119.x

[10] Haelterman E, Qvist R, Barlow P, Alexander S. Social determinants of pregnancy-related mortality and morbidity in the United States: a systematic review. Obstetrics & Gynecology. 2020;135(4):896–915.

[11] Kim Y, et al. Income-based disparities in preeclampsia and postpartum hemorrhage: a population-based study. BMC Pregnancy and Childbirth. 2016;16:262.

[12] World Health Organization. Social determinants of health. Accessed May 4, 2025.

[13] Centers for Disease Control and Prevention. Pregnancy-related deaths. Accessed May 4, 2025.

[14] Braveman P, et al. The role of socioeconomic factors in Black–White disparities in preterm birth. American Journal of Public Health. 2015;105(4):694–702.

[15] Shah PS, et al. Social support and maternal health: a systematic review and meta-analysis. Maternal and Child Health Journal. 2011;15(7):1159–1170.

[16] Hyman I, et al. The impact of social isolation and relationship instability on maternal health outcomes. Journal of Women’s Health. 2019;28(8):1125–1134.

[17] Petersen EE, et al. Racial/ethnic disparities in pregnancy-related deaths — United States, 2007–2016. MMWR Morb Mortal Wkly Rep. 2019;68(35):762–765. https://www.cdc.gov/mmwr/volumes/68/wr/mm6835a3.htm

[18] Howell EA. Reducing disparities in severe maternal morbidity and mortality. Clinical Obstetrics and Gynecology. 2018;61(2):387–399.

[19] Admon LK, et al. Disparities in postpartum hemorrhage by race and ethnicity. Obstetrics & Gynecology. 2020;136(1):89–96.