CoChat
Sign Up
Cited by 0Machine Learning in HealthcareArtificial Intelligence in Healthcare and EducationHealth disparities and outcomes

Artificial Intelligence and Machine Learning for Identifying Social Determinants of Health in Low‐Income Populations Within United States Health Systems: A Scoping Review

Open access
Razzano, Anthony Vincent·2026·Health Science Reports
DOI 10.1002/hsr2.73083PMC PMC13508128
Open this paper in CoChat — free accountOpen PDF

Abstract

Background and Aims: Low‐income populations in the United States experience disproportionate exposure to adverse social conditions. This scoping review examined artificial intelligence (AI) and machine learning (ML) approaches used in United States health systems to identify financial hardship, housing instability, food insecurity, neighborhood deprivation, and other social determinants of health (SDOH), and characterized interventions implemented in response. Methods: Following the Arksey and O'Malley framework and PRISMA‐ScR guidance, MEDLINE, CINAHL Plus with Full Text, PsycINFO, and Academic Search Ultimate were searched for peer‐reviewed studies published from January 2010 through January 2026. The primary reviewer screened 209 unique records, and two additional reviewers reviewed screening decisions. Disagreements were resolved through discussion and majority vote. Seventeen studies met the inclusion criteria. Results: Studies were conducted in academic medical centers, integrated delivery networks, safety‐net hospitals, Veterans Health Administration facilities, specialty clinics, and Medicaid administrative environments. Methods included predictive ML modeling, natural language processing, and unsupervised clustering, with electronic health records used in 16 of 17 studies. Only one study evaluated a structured intervention: a proactive financial assistance program for ambulatory oncology patients at risk of financial hardship and unmet social needs. The remaining studies focused primarily on risk identification or model development. External validation was uncommon, and several studies reported differential performance across socioeconomic or racial and ethnic groups. Conclusion: AI and ML may support SDOH risk identification in low‐income populations, but evidence of generalizability, fairness, and translation into effective interventions remains limited. Future research should prioritize external validation, equity‐focused evaluation, and rigorous testing of interventions triggered by identified social risks.

AI snapshot

Written from the paper’s full text.

TL;DR

This scoping review maps 17 studies (2018–2025) using AI/ML to identify social determinants of health in low-income populations within U.S. health systems, finding predictive ML and NLP dominate but external validation and intervention evaluation are scarce. Only one included study evaluated a structured intervention, and algorithmic bias against low-income and racial/ethnic minority groups was repeatedly documented.

Key findings

  • Across 17 included studies, 10 used predictive ML, 5 used NLP, and 2 used unsupervised clustering; EHR data were used in 16 of 17 studies.
  • Predictive performance varied widely: AUROC ranged from 0.59 to 0.88 among the 10 predictive ML studies; Patel et al. reported the highest AUROC (0.88) and F1 (0.69) in 14.2 million Medicaid recipients, outperforming non-predictive outreach by 32.5 percentage points.
  • Bias and data gaps were common: Juhn et al. found a balanced error rate ratio of 1.35 for lowest versus other SES quartiles, with 41% of lower-SES children missing relevant EHR records versus 24% of higher-SES peers; Gregory et al. found computable social phenotypes performed poorly for non-White and Hispanic patients.
  • Only Philipp et al. evaluated a structured intervention—a proactive financial assistance program for ambulatory oncology patients—while the remaining studies focused on risk identification or model development; only 2 of 17 studies achieved low overall PROBAST risk of bias.

Methods

  • Scoping review following Arksey and O’Malley and PRISMA-ScR, prospectively registered on OSF, with eligibility requiring an applied AI/ML method, substantive SDOH focus, explicitly low-income/Medicaid/economically disadvantaged population, and U.S. health system setting.
  • Searched MEDLINE, CINAHL Plus with Full Text, PsycINFO, and Academic Search Ultimate via EBSCO for peer-reviewed English articles from January 2010 through January 2026; 209 unique records were screened and 17 studies met inclusion criteria.
  • Data were extracted with a standardized form and appraised using MMAT and PROBAST; findings were synthesized descriptively, with no meta-analysis or pooled effect estimation.

Limitations

  • The paper states that the search was restricted to four EBSCO databases, excluding IEEE Xplore, ACM Digital Library, Scopus, and Embase; grey literature, conference proceedings, and non-English publications were excluded, and studies after January 2026 were not captured.
  • The paper states that requiring explicit low-income, Medicaid-enrolled, or economically disadvantaged populations may have excluded general-population studies with unpublished low-income subgroup analyses; no quantitative synthesis was conducted, and quality appraisal was not used for exclusion.
  • Evident from the text: screening was performed by a primary reviewer with two additional reviewers only reviewing decisions, without fully independent duplicate screening or a chance-corrected inter-rater reliability statistic; data extraction was completed by the primary reviewer and not independently duplicated.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1002/hsr2.73083

Cites

35 of 35 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

100 of 124 cited sources checked · 69 supported · 7 flagged · accuracy not yet measured

  • Findingcitation

    Cross-domain data linkage substantially improved overdose risk prediction versus claims alone.

    “Cross‐domain data linkage substantially improved overdose risk prediction versus claims alone; incarceration history and social service utilization captured key SDOH signals”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [32] W. H. Lo‐Ciganic, J. M. Donohue, E. G. Hulsey, et al., “ Integrating Human Services and Criminal Justice Data With Claims Data to Predict Risk of Opioid Overdose Among Medicaid Beneficiaries: A Machine‐Learning Approach,” PLoS One 16, no. 3 (2021): e0248360, 10.1371/journal.pone.0248360. 33735222 PMC7971495 source

      The Discussion says integrating human services and criminal justice data produced only “slight improvements” over Medicaid claims alone and explicitly says the improvement “was not large.” This contradicts the claim that data linkage substantially improved prediction. (LLM, conf=0.92)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Findingcitation

    SDOH-enriched models improve social service referral prediction.

    “one of the earliest demonstrations that SDOH‐enriched models improve social service referral prediction”
    • Also stated in: § Discussion · plain
      “showed that community‐level SDOH data added predictive value for social service referral need beyond clinical variables alone”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [34] S. N. Kasthurirathne, J. R. Vest, N. Menachemi, P. K. Halverson, and S. J. Grannis, “ Assessing the Capacity of Social Determinants of Health Data to Augment Predictive Models Identifying Patients in Need of Wraparound Social Services,” Journal of the American Medical Informatics Association 25, no. 1 (2018): 47– 53, 10.1093/jamia/ocx130. 29177457 PMC7647142 source

      The Discussion says that adding socioeconomic and public health data did not produce a statistically significant performance improvement in any predictive model. The study demonstrated that social service referral needs could be predicted, but not that SDOH enrichment improved prediction over clinical data alone. (LLM, conf=0.92)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Findingcitation

    Integrating human services and criminal justice data with Medicaid claims substantially improves opioid overdose risk prediction over claims data alone.

    “showed that cross‐sector data linkage substantially improves predictive performance over claims data alone”
    • Also stated in: § Discussion · plain
      “showed that integrating human services and criminal justice data with Medicaid claims substantially improved opioid overdose risk prediction”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [32] W. H. Lo‐Ciganic, J. M. Donohue, E. G. Hulsey, et al., “ Integrating Human Services and Criminal Justice Data With Claims Data to Predict Risk of Opioid Overdose Among Medicaid Beneficiaries: A Machine‐Learning Approach,” PLoS One 16, no. 3 (2021): e0248360, 10.1371/journal.pone.0248360. 33735222 PMC7971495 source

      The Discussion says the integrated models improved prediction only slightly over Medicaid claims alone and explicitly states that the improvement “was not large.” The Conclusions likewise describe small improvements, contrary to the claim of a substantial gain. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Findingnone given

    Yao et al. [29] and Sato et al. [35] relied solely on clinical note text.

    “Yao et al. [29] and Sato et al. [35] relied solely on clinical note text”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [29] L. Yao, J. Tsai, W. Liu, et al., “ Automated Identification of Eviction Status From Electronic Health Record Notes,” Journal of the American Medical Informatics Association 30, no. 8 (2023): 1429– 1437, 10.1093/jamia/ocad081. 37203429 PMC10354775 source
    • cites [35] M. Sato, E. D. Grussing, R. Patel, et al., “ Natural Language Processing for Identification of Hospitalized People Who Use Drugs: Cohort Study,” JMIR Medical Informatics 4 (2025): e63147, 10.2196/63147. PMC12294639 40680182 source

      Sato et al. did not rely solely on clinical note text. The Methods (Abstract > Methods) list ICD-10 codes, toxicology results, medication prescriptions, and NLP-detected text as identification criteria; the Discussion explicitly describes integrating NLP with urine toxicology and medication records. This verdict addresses Sato et al. only, not Yao et al. (LLM, conf=0.92)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Premisecitation

    ML models incorporating socioeconomic variables consistently outperform traditional clinical markers in predicting risk.

    “ML models incorporating socioeconomic variables consistently outperform traditional clinical markers in predicting risk.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [15] Y. Huang, X. Yang, Q. Wang, A. Abula, Y. Dong, and W. Li, “ Development and Validation of Deep Learning‐ and Ensemble Learning‐Based Biological Ages in the NHANES Study,” Frontiers in Aging Neuroscience 17 (2025): 1532884, 10.3389/fnagi.2025.1532884. 40741049 PMC12307447 source

      The Results (§3.2) report that adding behavioral and socioeconomic variables improved mortality-prediction AUC for both ML models compared with their biometric-only versions. But the Discussion (§4) reports that the full-feature DBA model performed slightly worse than PhenoAge V2, so the paper does not support 'consistently outperform' across traditional comparators. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [16] E. Ahmadi and R. Maihami, “ Leveraging Artificial Intelligence for Pandemic Management: Case of COVID‐19 in the United States,” Big Data Research 41 (2025): 100529, 10.1016/j.bdr.2025.100529. source
  • May combine two claims. 27
  • Classified as probably not a substantial claim. 67
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingcitation

    Methods included gradient boosting, random forest, ensemble models, XGBoost, and logistic regression-based ML pipelines with fairness optimization.

    “Methods included gradient boosting [32], random forest [26, 34], ensemble models [26], XGBoost [27], and logistic regression‐based ML pipelines with fairness optimization [21].”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [32] W. H. Lo‐Ciganic, J. M. Donohue, E. G. Hulsey, et al., “ Integrating Human Services and Criminal Justice Data With Claims Data to Predict Risk of Opioid Overdose Among Medicaid Beneficiaries: A Machine‐Learning Approach,” PLoS One 16, no. 3 (2021): e0248360, 10.1371/journal.pone.0248360. 33735222 PMC7971495 source

      The Materials and methods section says the authors developed and tested gradient boosting machine models, not a logistic regression-based pipeline. The Discussion says bias assessment and approaches to improve fairness were needed before implementation; it does not describe fairness optimization as a method used in the study. (LLM, conf=0.88)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [26] S. Choi, R. Fayyad, D. Nolte, C. Riedy, and S.‐L. Normand, “ Predicting Emergency Department Visits for Non‐Traumatic Dental‐Related Conditions Among Medicaid Beneficiaries,” Scientific Reports 15 (2025): 19872, 10.1038/s41598-025-04764-0. 40481164 PMC12144079 source

      The Statistical analysis section describes logistic regression with elastic-net regularization as part of a machine-learning ensemble and a separate logistic-regression comparator. It does not describe fairness optimization, so the qualifier attached to citation [21] is unsupported. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [34] S. N. Kasthurirathne, J. R. Vest, N. Menachemi, P. K. Halverson, and S. J. Grannis, “ Assessing the Capacity of Social Determinants of Health Data to Augment Predictive Models Identifying Patients in Need of Wraparound Social Services,” Journal of the American Medical Informatics Association 25, no. 1 (2018): 47– 53, 10.1093/jamia/ocx130. 29177457 PMC7647142 source

  • Classified as probably not a substantial claim.

    Findingcitation

    Community-level SDOH added meaningful predictive value beyond clinical variables alone.

    “Random forest predicted social service referral needs with considerable accuracy; community‐level SDOH added meaningful predictive value beyond clinical variables alone; among earliest demonstrations of SDOH‐augmented prediction in clinical settings”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [34] S. N. Kasthurirathne, J. R. Vest, N. Menachemi, P. K. Halverson, and S. J. Grannis, “ Assessing the Capacity of Social Determinants of Health Data to Augment Predictive Models Identifying Patients in Need of Wraparound Social Services,” Journal of the American Medical Informatics Association 25, no. 1 (2018): 47– 53, 10.1093/jamia/ocx130. 29177457 PMC7647142 source

      The Discussion states that adding socioeconomic and public health data did not produce a statistically significant performance improvement in any predictive model. The Abstract likewise says that SDH did not result in significant performance improvements, contrary to the claim that community-level SDOH added meaningful predictive value beyond clinical variables alone. (LLM, conf=0.94)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingcitation

    The top 5% risk score had a 28.1% hospitalization rate, approximately 13 times the bottom decile.

    “C‐statistic 0.71; top 5% risk score had 28.1% hospitalization rate (~13x bottom decile); fairness optimization reduced bias without major accuracy loss”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [21] J. Huang, J. Guo, W. T. Donahoo, et al., “ A Fair Individualized Polysocial Risk Score for Identifying Increased Social Risk in Type 2 Diabetes,” Nature Communications 15 (2024): 8653, 10.1038/s41467-024-52960-9. PMC11455957 39369018 source

      The Results section (page unavailable) reports a 27.1% one-year hospitalization rate for the top 10% of iPsRS scores, approximately 21 times the bottom-decile rate. It does not report the claimed 28.1% rate for the top 5% or the approximately 13-fold comparison. (LLM, conf=0.88)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Smartphone systems detect lung disease with high accuracy.

    “AI screening tools designed for resource‐limited settings include smartphone systems that detect lung disease with high accuracy, offering meaningful potential for health equity in underserved regions [9].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [9] Y. Gong, C. Xu, C. Mo, et al., “ AI‐Driven Smartphone Screening for Acute COPD Exacerbations: Enhancing Health Equity in Developing Regions,” npj Digital Medicine 8 (2025): 715, 10.1038/s41746-025-02086-z. 41272082 PMC12638963 source

      The Results report 91.5% pooled accuracy for a smartphone-audio system, and the Discussion describes its potential to improve COPD care where resources are scarce. The system detects acute COPD exacerbations, however, not lung disease generally; that broader wording exceeds what the study establishes. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Medical AI publications grew rapidly after 2017.

    “Medical AI publications grew rapidly after 2017, with research output tracking national economic and innovation capacity even as high‐income and low‐income nations retain unequal access to these tools [3].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [3] D. Klingelhöfer, M. Braun, J. Dröge, D. A. Groneberg, and D. Brüggmann, “ Research on Artificial Intelligence, Machine and Deep Learning in Medicine: Global Characteristics, Readiness, and Equity,” Globalization and Health 21, no. 1 (2025): 36, 10.1186/s12992-025-01128-1. 40484942 PMC12147299 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Integrating big data analytics with genomic sequencing and SDOH data improves personalized cancer care and patient safety.

    “Integrating big data analytics with genomic sequencing and SDOH data improves personalized cancer care and patient safety [7].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [7] M. K. Ahmed, E. Rozario, S. C. Mohonta, et al., “ Leveraging Big Data Analytics for Personalized Cancer Treatment: An Overview of Current Approaches and Future Directions,” Journal of Epidemiology (2025): 1– 15, 10.1155/je/9928467. source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductcitation

    The scoping review was conducted following the Arksey and O'Malley methodological framework and reported in accordance with PRISMA-ScR.

    “A scoping review was conducted following the Arksey and O'Malley methodological framework [18] and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses Extension for Scoping Reviews (PRISMA‐ScR) [19].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [18] H. Arksey and L. O'Malley, “ Scoping Studies: Towards a Methodological Framework,” International Journal of Social Research Methodology 8, no. 1 (2005): 19– 32, 10.1080/1364557032000119616. source
    • cites [19] A. C. Tricco, E. Lillie, W. Zarin, et al., “ PRISMA Extension for Scoping Reviews (PRISMA‐ScR): Checklist and Explanation,” Annals of Internal Medicine 169, no. 7 (2018): 467– 473, 10.7326/M18-0850. 30178033 source

  • May combine two claims.

    Premisecitation

    Studies in low-income primary care settings confirm that household income and neighborhood position each independently relate to disease burden.

    “Studies in low‐income primary care settings confirm that household income and neighborhood position each independently relate to disease burden [17].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [17] A. Kaganovski, A. Bhuiyan, A. Kasi, et al., “ Elevated Prevalence of Age‐Related Macular Degeneration in a Low‐Income Urban Primary Care Setting,” Eye and Vision 13 (2026): 1– 12, 10.1186/s12982-026-01672-0. 41841116 PMC12989009 source

      The Results report that lower income was independently associated with higher prevalence of referable AMD in an urban primary care cohort. However, the Strengths and limitations section says income was approximated using ZIP code–linked census data; the study did not establish separate, independent associations for household income and neighborhood position. Its finding also concerns AMD, not disease burden generally. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Governance frameworks for digital healthcare integrate big data analytics with value-based care principles to protect low-income populations during health crises.

    “Governance frameworks for digital healthcare integrate big data analytics with value‐based care principles to protect low‐income populations during health crises [13].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [13] S. Bagheri, K. Kourtit, and P. Nijkamp, “ Governance of Transformative Value‐Based Healthcare Services in the Digital Era — A Policy and Research Agenda,” Transylvanian Review of Administrative Sciences (2025): 5– 21, Special Issue. 10.24193/tras.SI2025.1. source

      The Abstract proposes a value-based healthcare framework using big data analytics in response to crises such as COVID-19. The Introduction says the pandemic disproportionately affected vulnerable and underserved communities and calls for more equitable, information-driven care. Neither passage specifically identifies low-income populations or establishes that the proposed framework protects them during crises. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Experts have identified eight challenges for equitable deployment in opioid use disorder treatment, including algorithmic transparency, clinical validation, and digital exclusion among low-income individuals.

    “Experts have identified eight challenges for equitable deployment in opioid use disorder treatment, including algorithmic transparency, clinical validation, and digital exclusion among low‐income individuals [12].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [12] M. Amer, R. Gittins, A. M. Millana, et al., “ Are Treatment Services Ready for the Use of Big Data Analytics and AI in Managing Opioid Use Disorder?,” Journal of Medical Internet Research 27 (2025): e58723, 10.2196/58723. 40294410 PMC12070021 source

      The Challenges to Deploying AI Applications section says the authors focus on eight challenges for ethical, effective, and equitable use in OUD treatment. The Implementation Challenges sections identify algorithmic transparency, clinical validation, and digital exclusion, but describe the latter as affecting people who use drugs and other marginalized groups; they do not specifically identify low-income individuals. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Responsible data use principles, including safety, transparency, fairness, and accountability, form a blueprint for interdisciplinary AI collaboration.

    “Responsible data use principles, including safety, transparency, fairness, and accountability, form a blueprint for interdisciplinary AI collaboration [14].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [14] K. P. Chun, T. Octavianti, N. Dogulu, et al., “ Transforming Disaster Risk Reduction With AI and Big Data: Legal and Interdisciplinary Perspectives,” Wiley Interdiscip Rev Data Min Knowl Discov 15, no. 3 (2025): e70011, 10.1002/widm.70011. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Governance frameworks identify big data analytics as essential for improving outcomes in resource-constrained environments.

    “Governance frameworks identify big data analytics as essential for improving outcomes in resource‐constrained environments [13].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [13] S. Bagheri, K. Kourtit, and P. Nijkamp, “ Governance of Transformative Value‐Based Healthcare Services in the Digital Era — A Policy and Research Agenda,” Transylvanian Review of Administrative Sciences (2025): 5– 21, Special Issue. 10.24193/tras.SI2025.1. source

      Section 6 (Lessons and guidelines for policy; page unavailable) calls a framework incorporating analytics of vast healthcare datasets essential to improving healthcare decision-making and delivery. Section 4.2 (Research challenges; page unavailable) links data analytics to better resource use and clinical outcomes. Neither passage specifically addresses resource-constrained environments, so that qualification is unsupported. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Interpretationcitation

    EHR data gaps drive algorithmic bias.

    “Balanced error rate ratio 1.35 for lowest versus other SES quartiles; 41% of lower‐SES children had missing EHR records versus 24% of higher‐SES peers; EHR data gaps drive algorithmic bias”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [23] Y. J. Juhn, E. Ryu, C.‐I. Wi, and K. S. King, “ Assessing Socioeconomic Bias in Machine Learning Algorithms in Health Care: A Case Study of the HOUSES Index,” Journal of the American Medical Informatics Association 29, no. 7 (2022): 1142– 1151, 10.1093/jamia/ocac052. 35396996 PMC9196683 source

      The Discussion reports more unavailable and inaccurate EHR data among lower-SES children with asthma and proposes that differential EHR completeness might contribute to algorithmic bias. It does not establish that EHR data gaps drive the bias; that causal wording is stronger than the paper’s stated conclusion. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Poverty was significantly associated with higher odds of suicidal ideation pre-pandemic among Hispanic mental health care patients.

    “NLP extracted poverty‐related SDOH and suicidal ideation from unstructured notes; poverty significantly associated with higher odds of suicidal ideation pre‐pandemic”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [28] A. Goldstein, E. V. Bailey, and F. A. Wilson, “ Poverty and Suicidal Ideation Among Hispanic Mental Health Care Patients Leading Up to the COVID‐19 Pandemic,” Families, Systems & Health 22, no. 1 (2024): 6– 10, 10.1177/15404153231181110. PMC10271853 37312509 source

  • May combine two claims.

    Findingcitation

    Fairness optimization reduced bias without major accuracy loss.

    “C‐statistic 0.71; top 5% risk score had 28.1% hospitalization rate (~13x bottom decile); fairness optimization reduced bias without major accuracy loss”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [21] J. Huang, J. Guo, W. T. Donahoo, et al., “ A Fair Individualized Polysocial Risk Score for Identifying Increased Social Risk in Type 2 Diabetes,” Nature Communications 15 (2024): 8653, 10.1038/s41467-024-52960-9. PMC11455957 39369018 source

  • Classified as probably not a substantial claim.

    Interpretationcitation

    Social factors are integral to hypertension heterogeneity.

    “Five hypertension subphenotypes identified; one predominantly younger Black women with higher poverty and Medicaid enrollment; social factors integral to hypertension heterogeneity”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [20] S. L. Hall, J. Xu, M. G. Walsh, et al., “ Unsupervised Learning Using Electronic Health Record and Census Data to Identify Distinct Subphenotypes of Newly Diagnosed Hypertension Patients,” PLoS One 20, no. 7 (2025): e0326776, 10.1371/journal.pone.0326776. 40632742 PMC12240299 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    KIRESH-Prompt achieved MCC 0.747 for eviction period prediction.

    “KIRESH‐Prompt achieved MCC 0.747 for eviction period prediction; outperformed BioBERT and BioClinicalBERT; planned deployment as VHA eviction surveillance system”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [29] L. Yao, J. Tsai, W. Liu, et al., “ Automated Identification of Eviction Status From Electronic Health Record Notes,” Journal of the American Medical Informatics Association 30, no. 8 (2023): 1429– 1437, 10.1093/jamia/ocad081. 37203429 PMC10354775 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Five BMI trajectories were identified.

    “Five BMI trajectories identified; children in higher‐risk groups significantly more likely in high‐poverty, high‐unemployment neighborhoods; neighborhood SDOH shapes childhood obesity trajectories”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [24] R. S. Gupta, T.‐L. T. Phan, F. Lê‐Scherban, D. Eckrich, H. T. Bunnell, and R. Beheshti, “ Associations of Longitudinal Body Mass Index‐Percentile Classification Patterns in Early Childhood With Neighborhood‐Level Social Determinants of Health,” Childhood Obesity 21, no. 1 (2025): 65– 75, 10.1089/chi.2023.0157. 39187268 PMC11807911 source

  • May combine two claims.

    Findingcitation

    41% of lower-SES children had missing EHR records versus 24% of higher-SES peers.

    “Balanced error rate ratio 1.35 for lowest versus other SES quartiles; 41% of lower‐SES children had missing EHR records versus 24% of higher‐SES peers; EHR data gaps drive algorithmic bias”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [23] Y. J. Juhn, E. Ryu, C.‐I. Wi, and K. S. King, “ Assessing Socioeconomic Bias in Machine Learning Algorithms in Health Care: A Case Study of the HOUSES Index,” Journal of the American Medical Informatics Association 29, no. 7 (2022): 1142– 1151, 10.1093/jamia/ocac052. 35396996 PMC9196683 source

      The Results section says lower-SES children had more unavailable EHR data for asthma-related variables than higher-SES children. However, the passages from our copy do not give the claimed 41% and 24% figures, so the numerical comparison is not verified. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Children in higher-risk BMI trajectory groups were significantly more likely to be in high-poverty, high-unemployment neighborhoods.

    “Five BMI trajectories identified; children in higher‐risk groups significantly more likely in high‐poverty, high‐unemployment neighborhoods; neighborhood SDOH shapes childhood obesity trajectories”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [24] R. S. Gupta, T.‐L. T. Phan, F. Lê‐Scherban, D. Eckrich, H. T. Bunnell, and R. Beheshti, “ Associations of Longitudinal Body Mass Index‐Percentile Classification Patterns in Early Childhood With Neighborhood‐Level Social Determinants of Health,” Childhood Obesity 21, no. 1 (2025): 65– 75, 10.1089/chi.2023.0157. 39187268 PMC11807911 source

  • Classified as probably not a substantial claim.

    Findingcitation

    NLP outperformed biomarkers, prescriptions, and diagnosis codes alone.

    “NLP positive predictive value 54%; outperformed biomarkers, prescriptions, and diagnosis codes alone; minoritized and lower SVI patients systematically missed by traditional methods; NLP improves identification equity”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [35] M. Sato, E. D. Grussing, R. Patel, et al., “ Natural Language Processing for Identification of Hospitalized People Who Use Drugs: Cohort Study,” JMIR Medical Informatics 4 (2025): e63147, 10.2196/63147. PMC12294639 40680182 source

      The Results and Table 3 report a 54% positive predictive value for NLP alone versus 43% for diagnosis codes alone. The nonabstract Results and Discussion do not provide corresponding performance comparisons with biomarkers or prescriptions, so those parts of the claim are not established by the eligible passages. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Five hypertension subphenotypes were identified using unsupervised learning with EHR and census data.

    “Five hypertension subphenotypes identified; one predominantly younger Black women with higher poverty and Medicaid enrollment; social factors integral to hypertension heterogeneity”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] S. L. Hall, J. Xu, M. G. Walsh, et al., “ Unsupervised Learning Using Electronic Health Record and Census Data to Identify Distinct Subphenotypes of Newly Diagnosed Hypertension Patients,” PLoS One 20, no. 7 (2025): e0326776, 10.1371/journal.pone.0326776. 40632742 PMC12240299 source

  • Classified as probably not a substantial claim.

    Interpretationcitation

    Neighborhood SDOH shapes childhood obesity trajectories.

    “Five BMI trajectories identified; children in higher‐risk groups significantly more likely in high‐poverty, high‐unemployment neighborhoods; neighborhood SDOH shapes childhood obesity trajectories”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [24] R. S. Gupta, T.‐L. T. Phan, F. Lê‐Scherban, D. Eckrich, H. T. Bunnell, and R. Beheshti, “ Associations of Longitudinal Body Mass Index‐Percentile Classification Patterns in Early Childhood With Neighborhood‐Level Social Determinants of Health,” Childhood Obesity 21, no. 1 (2025): 65– 75, 10.1089/chi.2023.0157. 39187268 PMC11807911 source

      The Conclusion reports five distinct early-childhood BMI-percentile subgroups, and the Discussion says all three higher-risk groups were more likely to live in neighborhoods with higher poverty and unemployment. The paper studies associations, however, so the statement that neighborhood SDOH “shapes” trajectories goes beyond its evidence if read as a causal claim. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Findingcitation

    NLP extracted poverty-related SDOH and suicidal ideation from unstructured clinical notes.

    “NLP extracted poverty‐related SDOH and suicidal ideation from unstructured notes; poverty significantly associated with higher odds of suicidal ideation pre‐pandemic”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [28] A. Goldstein, E. V. Bailey, and F. A. Wilson, “ Poverty and Suicidal Ideation Among Hispanic Mental Health Care Patients Leading Up to the COVID‐19 Pandemic,” Families, Systems & Health 22, no. 1 (2024): 6– 10, 10.1177/15404153231181110. PMC10271853 37312509 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    ML consistently outperformed logistic regression.

    “C‐statistics 0.70–0.73; accuracy 0.90–0.95; specificity 0.93–0.97; ML consistently outperformed logistic regression; sociodemographic risk factors identified”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [26] S. Choi, R. Fayyad, D. Nolte, C. Riedy, and S.‐L. Normand, “ Predicting Emergency Department Visits for Non‐Traumatic Dental‐Related Conditions Among Medicaid Beneficiaries,” Scientific Reports 15 (2025): 19872, 10.1038/s41598-025-04764-0. 40481164 PMC12144079 source

      The Discussion says the machine-learning ensemble outperformed traditional logistic regression in discrimination and calibration, particularly among high-risk patients. But “consistently outperformed” is too broad without specifying the metrics: the Results table shows logistic regression had higher sensitivity in all four states and a higher F1 score in two. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    The balanced error rate ratio was 1.35 for the lowest versus other SES quartiles.

    “Balanced error rate ratio 1.35 for lowest versus other SES quartiles; 41% of lower‐SES children had missing EHR records versus 24% of higher‐SES peers; EHR data gaps drive algorithmic bias”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [23] Y. J. Juhn, E. Ryu, C.‐I. Wi, and K. S. King, “ Assessing Socioeconomic Bias in Machine Learning Algorithms in Health Care: A Case Study of the HOUSES Index,” Journal of the American Medical Informatics Association 29, no. 7 (2022): 1142– 1151, 10.1093/jamia/ocac052. 35396996 PMC9196683 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    KIRESH-Prompt outperformed BioBERT and BioClinicalBERT.

    “KIRESH‐Prompt achieved MCC 0.747 for eviction period prediction; outperformed BioBERT and BioClinicalBERT; planned deployment as VHA eviction surveillance system”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [29] L. Yao, J. Tsai, W. Liu, et al., “ Automated Identification of Eviction Status From Electronic Health Record Notes,” Journal of the American Medical Informatics Association 30, no. 8 (2023): 1429– 1437, 10.1093/jamia/ocad081. 37203429 PMC10354775 source

  • May combine two claims.

    Findingcitation

    Accuracy was 0.90-0.95 and specificity 0.93-0.97.

    “C‐statistics 0.70–0.73; accuracy 0.90–0.95; specificity 0.93–0.97; ML consistently outperformed logistic regression; sociodemographic risk factors identified”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [26] S. Choi, R. Fayyad, D. Nolte, C. Riedy, and S.‐L. Normand, “ Predicting Emergency Department Visits for Non‐Traumatic Dental‐Related Conditions Among Medicaid Beneficiaries,” Scientific Reports 15 (2025): 19872, 10.1038/s41598-025-04764-0. 40481164 PMC12144079 source

  • May combine two claims.

    Findingreasoning

    The two studies achieving low PROBAST risk of bias, Patel et al. [22] and Gregory et al. [27], both incorporated multi-site or national data and applied external validation procedures.

    “The two studies achieving low PROBAST risk of bias, Patel et al. [22] and Gregory et al. [27], both incorporated multi‐site or national data and applied external validation procedures.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [22] S. Y. Patel, M. L. Barnett, and S. Basu, “ Predicting Quality Measure Completion Among 14 Million Low‐Income Patients Enrolled in Medicaid,” npj Digital Medicine 8 (2025): 393, 10.1038/s41746-025-01797-7. 40604098 PMC12223227 source

      Methods > Data Sources and Study Oversight reports data from 25 states and Washington, D.C., supporting the multi-site component. Methods > Model Development and Comparison describes a single training/validation/test split, with the test set reserved for final evaluation, rather than external validation. The low-PROBAST rating is the citing authors’ assessment, not a rating established by this source. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [27] M. E. Gregory, S. N. Kasthurirathne, T. Magoc, C. McNamee, C. A. Harle, and J. R. Vest, “ Development and Validation of Computable Social Phenotypes for Health‐Related Social Needs,” JAMIA Open 8, no. 1 (2025): ooae150, 10.1093/jamiaopen/ooae150. 39776620 PMC11706536 source

      The Materials and methods (Setting and sample; page unavailable) confirm that Gregory et al. recruited patients from multiple sites across three health systems in two states. The ML classifier phenotype development section describes a random 80% training/20% testing split and five-fold cross-validation, but does not report external validation in a separate setting. The low PROBAST rating is the citing authors’ assessment, not a finding attributed to Gregory et al. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Rigdon et al. [31] supplemented EHR data with American Community Survey neighborhood data at the census block level.

    “Rigdon et al. [31] supplemented EHR data with American Community Survey neighborhood data at the census block level.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [31] J. Rigdon, K. Montez, D. Palakshappa, et al., “ Predicting Food Insecurity in a Pediatric Population Using the Electronic Health Record,” Clinical and Translational Science 8, no. 1 (2024): e195, 10.1017/cts.2024.645. PMC11626594 39655033 source

  • May combine two claims. The quote only partly states this claim.

    Findingcitation

    ML-guided outreach improved quality gap closure efficiency in Medicaid populations, outperforming non-predictive outreach strategies by 32.5 percentage points.

    “outperforming non‐predictive outreach strategies by 32.5 percentage points”
    • Also stated in: § Interventions in Response to Identified Risks · hedged
      “Patel et al. [22] showed that ML‐guided outreach could improve quality gap closure efficiency in Medicaid populations”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [22] S. Y. Patel, M. L. Barnett, and S. Basu, “ Predicting Quality Measure Completion Among 14 Million Low‐Income Patients Enrolled in Medicaid,” npj Digital Medicine 8 (2025): 393, 10.1038/s41746-025-01797-7. 40604098 PMC12223227 source

      The Results (Machine Learning Model Performance; page unavailable) report that the clinical-only model improved prediction accuracy by 32.5 percentage points over a baseline simulating non-predictive outreach. That figure is not a measured improvement in quality-gap closure or outreach efficiency: the paper reports those outcomes separately as simulations and says actual intervention effects require validation. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    39 records proceeded to full-text review.

    “Proceeding to full‐text review | 39”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    5 studies were excluded at full-text review because SDOH used only as a covariate to predict a clinical outcome.

    “After full‐text assessment, 22 additional studies were excluded: no AI or ML method applied (n = 9); SDOH used only as a covariate to predict a clinical outcome (n = 5); population not identified as low‐income or Medicaid‐enrolled (n = 3); duplicate publication or preprint superseded by an included published version (n = 3); SDOH not a primary focus (n = 1); and ineligible study type (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    1 record was excluded at title/abstract screening because no substantive SDOH focus.

    “The primary reasons for exclusion were setting not within a United States clinical health system (n = 66), population not explicitly identified as low‐income or Medicaid‐enrolled (n = 53), no AI or ML application (n = 50), and no substantive SDOH focus (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    100 duplicates were removed.

    “Duplicates removed | 100”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    170 records were excluded at title/abstract screening.

    “Excluded at title/abstract screening | 170”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    66 records were excluded at title/abstract screening because setting was not within a United States clinical health system.

    “The primary reasons for exclusion were setting not within a United States clinical health system (n = 66), population not explicitly identified as low‐income or Medicaid‐enrolled (n = 53), no AI or ML application (n = 50), and no substantive SDOH focus (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Total records retrieved pre-deduplication were 309.

    “Total records retrieved (pre‐deduplication) | 309”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    209 unique records were screened.

    “Unique records for screening | 209”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    1 study was excluded at full-text review because ineligible study type.

    “After full‐text assessment, 22 additional studies were excluded: no AI or ML method applied (n = 9); SDOH used only as a covariate to predict a clinical outcome (n = 5); population not identified as low‐income or Medicaid‐enrolled (n = 3); duplicate publication or preprint superseded by an included published version (n = 3); SDOH not a primary focus (n = 1); and ineligible study type (n = 1).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Settings included academic medical centers, integrated delivery networks, safety-net hospitals, Veterans Health Administration facilities, and multi-state Medicaid administrative data environments.

    “Settings included academic medical centers, integrated delivery networks, safety‐net hospitals, Veterans Health Administration facilities, and multi‐state Medicaid administrative data environments.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    3 studies were excluded at full-text review because duplicate publication or preprint superseded by an included published version.

    “After full‐text assessment, 22 additional studies were excluded: no AI or ML method applied (n = 9); SDOH used only as a covariate to predict a clinical outcome (n = 5); population not identified as low‐income or Medicaid‐enrolled (n = 3); duplicate publication or preprint superseded by an included published version (n = 3); SDOH not a primary focus (n = 1); and ineligible study type (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    50 records were excluded at title/abstract screening because no AI or ML application.

    “The primary reasons for exclusion were setting not within a United States clinical health system (n = 66), population not explicitly identified as low‐income or Medicaid‐enrolled (n = 53), no AI or ML application (n = 50), and no substantive SDOH focus (n = 1).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    17 studies were included for synthesis.

    “INCLUDED for synthesis | 17”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    9 studies were excluded at full-text review because no AI or ML method applied.

    “After full‐text assessment, 22 additional studies were excluded: no AI or ML method applied (n = 9); SDOH used only as a covariate to predict a clinical outcome (n = 5); population not identified as low‐income or Medicaid‐enrolled (n = 3); duplicate publication or preprint superseded by an included published version (n = 3); SDOH not a primary focus (n = 1); and ineligible study type (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    1 study was excluded at full-text review because SDOH not a primary focus.

    “After full‐text assessment, 22 additional studies were excluded: no AI or ML method applied (n = 9); SDOH used only as a covariate to predict a clinical outcome (n = 5); population not identified as low‐income or Medicaid‐enrolled (n = 3); duplicate publication or preprint superseded by an included published version (n = 3); SDOH not a primary focus (n = 1); and ineligible study type (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    3 studies were excluded at full-text review because population not identified as low-income or Medicaid-enrolled.

    “After full‐text assessment, 22 additional studies were excluded: no AI or ML method applied (n = 9); SDOH used only as a covariate to predict a clinical outcome (n = 5); population not identified as low‐income or Medicaid‐enrolled (n = 3); duplicate publication or preprint superseded by an included published version (n = 3); SDOH not a primary focus (n = 1); and ineligible study type (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    53 records were excluded at title/abstract screening because population was not explicitly identified as low-income or Medicaid-enrolled.

    “The primary reasons for exclusion were setting not within a United States clinical health system (n = 66), population not explicitly identified as low‐income or Medicaid‐enrolled (n = 53), no AI or ML application (n = 50), and no substantive SDOH focus (n = 1).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    22 studies were excluded at full-text review.

    “Excluded at full‐text review | 22”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingcitation

    Non-White and low-income insurance patients were disproportionately associated with housing and transportation terms in clinical notes.

    “Non‐White and low‐income insurance patients disproportionately associated with housing and transportation terms; White/privately insured patients referenced physical activity; clinical notes contain detectable social context signals”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] E. Derton, M. Guevara, S. Chen, et al., “ Natural Language Processing Methods to Empirically Explore Social Contexts and Needs in Cancer Patient Notes,” JCO Clinical Cancer Informatics 7 (2023): e2200196, 10.1200/CCI.22.00196. 37235847 source

  • Classified as probably not a substantial claim.

    Findingcitation

    White/privately insured patients referenced physical activity.

    “Non‐White and low‐income insurance patients disproportionately associated with housing and transportation terms; White/privately insured patients referenced physical activity; clinical notes contain detectable social context signals”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] E. Derton, M. Guevara, S. Chen, et al., “ Natural Language Processing Methods to Empirically Explore Social Contexts and Needs in Cancer Patient Notes,” JCO Clinical Cancer Informatics 7 (2023): e2200196, 10.1200/CCI.22.00196. 37235847 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Model performance remained robust across racial and ethnic strata.

    “EHR‐based model predicted food insecurity, financial distress, or housing instability with an AUC of 0.68; performance remained robust across racial and ethnic strata; external validation was conducted in a separate cohort.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25] R. W. Grant, J. K. McCloskey, C. S. Uratsu, et al., “ Predicting Self‐Reported Social Risk in Medically Complex Adults Using Electronic Health Data,” Medical Care 62, no. 9 (2024): 590– 598, 10.1097/MLR.0000000000002021. 38833715 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    95% of patients were in high-SVI counties.

    “10% Black, 39% Hispanic patients; 95% in high‐SVI counties; proactive data‐driven program reached patients at greatest HRSN and financial distress risk; supports equitable cancer care”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] J. Philipp, A. E. Velozo, A. M. Scott, P. Falconer, and J. Pizzolato, “ Impact of a Proactive Patient Assistance Program for Diverse Ambulatory Oncology Patients: A Retrospective Cohort Study,” Journal of Oncology Navigation & Survivorship (2025), https://www.jons-online.com/issues/2025/september-2025-vol-16-no-9/.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    The proactive patient assistance program served 10% Black and 39% Hispanic patients.

    “10% Black, 39% Hispanic patients; 95% in high‐SVI counties; proactive data‐driven program reached patients at greatest HRSN and financial distress risk; supports equitable cancer care”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] J. Philipp, A. E. Velozo, A. M. Scott, P. Falconer, and J. Pizzolato, “ Impact of a Proactive Patient Assistance Program for Diverse Ambulatory Oncology Patients: A Retrospective Cohort Study,” Journal of Oncology Navigation & Survivorship (2025), https://www.jons-online.com/issues/2025/september-2025-vol-16-no-9/.

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Studies linking EHR data with census-derived neighborhood indicators, Medicaid claims, or cross-sector administrative data achieved stronger predictive performance and captured SDOH dimensions that clinical documentation alone misses.

    “Studies linking EHR data with census‐derived neighborhood indicators, Medicaid claims, or cross‐sector administrative data achieved stronger predictive performance and captured SDOH dimensions that clinical documentation alone misses.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Domain 1 (Participants) and Domain 2 (Predictors) received low bias ratings in 17 and 16 studies, respectively.

    “Domain 1 (Participants) and Domain 2 (Predictors) received low bias ratings in 17 and 16 studies, respectively”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Moderate-rated studies reflected limited confounder adjustment, unclear outcome definitions, or reduced methodological transparency.

    “Moderate‐rated studies reflected limited confounder adjustment, unclear outcome definitions, or reduced methodological transparency.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    MMAT quality ratings classified nine studies as high quality, six as moderate-high, and two as moderate.

    “MMAT quality ratings classified nine studies as high quality, six as moderate‐high, and two as moderate.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingcitation

    The proactive data-driven program reached patients at greatest HRSN and financial distress risk.

    “10% Black, 39% Hispanic patients; 95% in high‐SVI counties; proactive data‐driven program reached patients at greatest HRSN and financial distress risk; supports equitable cancer care”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [36] J. Philipp, A. E. Velozo, A. M. Scott, P. Falconer, and J. Pizzolato, “ Impact of a Proactive Patient Assistance Program for Diverse Ambulatory Oncology Patients: A Retrospective Cohort Study,” Journal of Oncology Navigation & Survivorship (2025), https://www.jons-online.com/issues/2025/september-2025-vol-16-no-9/.

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    External validation, intervention evaluation, and programmatic implementation remain substantially underdeveloped.

    “External validation, intervention evaluation, and programmatic implementation remain substantially underdeveloped.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Data extraction was completed by the primary reviewer using a standardized form.

    “Data extraction was completed by the primary reviewer using a standardized form.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    The additional reviewers did not independently duplicate the complete data-extraction process.

    “The additional reviewers did not independently duplicate the complete data‐extraction process.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    All 209 unique records were screened at the title and abstract level by the primary reviewer and reviewed by two additional reviewers.

    “All 209 unique records were screened at the title and abstract level by the primary reviewer and reviewed by two additional reviewers, both healthcare data and analytics professionals.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The search was restricted to four EBSCO databases to maximize reproducibility and transparency across a single standardized platform.

    “The search was restricted to four EBSCO databases to maximize reproducibility and transparency across a single standardized platform.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The two additional reviewers examined the completed extraction for completeness and consistency.

    “The two additional reviewers examined the completed extraction for completeness and consistency, and any questions identified during verification were resolved through discussion.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The scoping review was prospectively registered on the Open Science Framework Registries before data extraction.

    “This scoping review was prospectively registered on the Open Science Framework (OSF) Registries before data extraction.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Quality appraisal was conducted using the Mixed Methods Appraisal Tool (MMAT) and the Prediction Model Risk of Bias Assessment Tool (PROBAST).

    “Quality appraisal was conducted using the Mixed Methods Appraisal Tool (MMAT) and the Prediction Model Risk of Bias Assessment Tool (PROBAST).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    Grey literature, conference proceedings, and non-English language publications were outside the scope of this search.

    “Grey literature, conference proceedings, and non‐English language publications were outside the scope of this search.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Inconsistent screening decisions were discussed among the three reviewers in an effort to reach consensus; when consensus could not be reached, the final decision was determined by majority vote.

    “Inconsistent screening decisions were discussed among the three reviewers in an effort to reach consensus; when consensus could not be reached, the final decision was determined by majority vote.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    Appraisal results were used to contextualize the evidence and were not used as criteria for study exclusion.

    “Consistent with scoping‐review methodology, appraisal results were used to contextualize the evidence and were not used as criteria for study exclusion.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Evidence of generalizability, fairness, and translation into effective interventions remains limited.

    “AI and ML may support SDOH risk identification in low‐income populations, but evidence of generalizability, fairness, and translation into effective interventions remains limited.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    External validation was conducted in a separate cohort.

    “EHR‐based model predicted food insecurity, financial distress, or housing instability with an AUC of 0.68; performance remained robust across racial and ethnic strata; external validation was conducted in a separate cohort.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25] R. W. Grant, J. K. McCloskey, C. S. Uratsu, et al., “ Predicting Self‐Reported Social Risk in Medically Complex Adults Using Electronic Health Data,” Medical Care 62, no. 9 (2024): 590– 598, 10.1097/MLR.0000000000002021. 38833715 source

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Underrepresentation of employment, childcare, domestic violence, and civic participation is likely because they are documented less consistently in clinical records.

    “likely because they are documented less consistently in clinical records”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    ML-based risk identification creates the technical precondition for targeted intervention, but identification alone does not reduce social need.

    “ML‐based risk identification creates the technical precondition for targeted intervention, but identification alone does not reduce social need.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Expanding SDOH capture in both structured screening and unstructured clinical documentation would broaden the domain coverage achievable through AI-based identification.

    “Expanding SDOH capture in both structured screening and unstructured clinical documentation would broaden the domain coverage achievable through AI‐based identification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Consistent reliance on EHR data reflects the infrastructure available to health systems and the operational feasibility of integrating clinical records with predictive models.

    “Consistent reliance on EHR data reflects the infrastructure available to health systems and the operational feasibility of integrating clinical records with predictive models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The search was restricted to four databases available through the EBSCO platform and did not include IEEE Xplore, ACM Digital Library, Scopus, or Embase.

    “The search was restricted to four databases available through the EBSCO platform and did not include IEEE Xplore, ACM Digital Library, Scopus, or Embase”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Grey literature, conference proceedings, and non-English publications were excluded.

    “Grey literature, conference proceedings, and non‐English publications were excluded.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Unstructured documentation has high information density.

    “underscoring the information density in unstructured documentation”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationcitation

    The proactive patient assistance program supports equitable cancer care.

    “10% Black, 39% Hispanic patients; 95% in high‐SVI counties; proactive data‐driven program reached patients at greatest HRSN and financial distress risk; supports equitable cancer care”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [36] J. Philipp, A. E. Velozo, A. M. Scott, P. Falconer, and J. Pizzolato, “ Impact of a Proactive Patient Assistance Program for Diverse Ambulatory Oncology Patients: A Retrospective Cohort Study,” Journal of Oncology Navigation & Survivorship (2025), https://www.jons-online.com/issues/2025/september-2025-vol-16-no-9/.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    NLP offers both identification capacity and a lens for documenting disparities that structured data cannot reveal.

    “NLP offers both identification capacity and a lens for documenting disparities that structured data cannot reveal.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Unstructured clinical notes contain detailed social histories, patient-reported circumstances, and provider observations that structured EHR fields fail to capture.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Quality appraisal was not used as an exclusion criterion, consistent with scoping review methodology.

    “Quality appraisal was not used as an exclusion criterion, consistent with scoping review methodology”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    AI and ML approaches show meaningful capacity to identify SDOH risks among low-income populations within United States health systems, with predictive ML modeling and NLP providing the strongest and most consistent evidence.

    “AI and ML approaches show meaningful capacity to identify SDOH risks among low‐income populations within United States health systems, with predictive ML modeling and NLP providing the strongest and most consistent evidence.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Because this is a scoping review, no formal quantitative synthesis or meta-analysis was conducted.

    “Because this is a scoping review, no formal quantitative synthesis or meta‐analysis was conducted”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Studies published after January 2026 are not captured in this review.

    “studies published after January 2026 are not captured in this review”
    Quote in the paperNot a substantial claim
85 of 85

  • The paper relies on this without support.

    Low-income populations in the United States experience disproportionate exposure to adverse social conditions.

    “Low‐income populations in the United States experience disproportionate exposure to adverse social conditions.”

  • The paper relies on this without support.

    Electronic health records provide the foundational data infrastructure for AI-driven SDOH identification.

    “Electronic health records (EHRs) provide the foundational data infrastructure for AI‐driven SDOH identification.”

  • The paper relies on this without support.

    Integrating AI and SDOH data advances value-based healthcare.

    “Integrating AI and SDOH data advances value‐based healthcare.”

  • The paper relies on this without support.

    Few scoping reviews systematically characterize how AI and ML identify SDOH risks for low-income individuals in United States health systems or examine interventions responding to identified risks.

    “Few scoping reviews systematically characterize how AI and ML identify SDOH risks for low‐income individuals in United States health systems or examine interventions responding to identified risks.”

  • The paper relies on this without support.

    Existing research frequently separates methodology from SDOH content or overlooks economic vulnerability in safety-net settings.

    “Existing research frequently separates methodology from SDOH content or overlooks economic vulnerability in safety‐net settings.”

  • The paper relies on this without support.

    Health disparities among low-income populations in the United States reflect persistent structural failures in the modern healthcare system.

    “Health disparities among low‐income populations in the United States reflect persistent structural failures in the modern healthcare system.”

  • The paper relies on this without support.

    AI and ML technologies identify and quantify SDOH-driven risks.

    “AI and ML technologies identify and quantify SDOH‐driven risks.”

  • The paper relies on this without support.

    Unstructured clinical notes contain detailed social histories, patient-reported circumstances, and provider observations that structured EHR fields fail to capture.

    “Unstructured clinical notes contain detailed social histories, patient‐reported circumstances, and provider observations that structured EHR fields fail to capture.”

  • The paper relies on this without support.

    ML model performance frequently degrades when models are applied to new settings.

    “performance frequently degrades when models are applied to new settings”

  • The paper relies on this without support.

    Financial hardship, housing instability, and food insecurity are the most frequently screened SDOH in health system settings.

    “Financial hardship, housing instability, and food insecurity are the most frequently screened SDOH in health system settings”
10 of 10

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Choi study described as both four-state and single-state

    Minor

    The study summary says Choi et al. used Medicaid beneficiaries from four states, while the risk-of-bias table describes its Medicaid claims as single-state.

  • Conflicting reference standards for Gregory study

    Minor

    The study summary says performance was assessed against patient-reported reference standards, but the risk-of-bias table calls chart review the gold standard.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Philipp J., 2025, Impact of a Proactive Patient Assistance Program for Diverse Ambulatory Oncology Patients: A Retrospective Cohort Study, Journal of Oncology Navigation & Survivorship

Authors
Razzano, A. V.
Affiliations
University of Massachusetts Lowell (US)
Corresponding author
Razzano, Anthony Vincent
ORCID
Razzano, Anthony Vincent
Venue
Health Science Reports
Publisher
Wiley
Published
2026-08-26
Volume
9
Issue
9
Pages
e73083-e73083
ISSN
2398-8835 (ISSN-L)
DOI
10.1002/hsr2.73083
PMC
PMC13508128
OpenAlex
W7153037082
Cited by
0
References
35
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Health Science Reports (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (acceptedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • Open Science Framework (public-domain)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Social Sciences › Health › Health disparities and outcomes
Keywords
social determinants of health, low-income populations, artificial intelligence, machine learning, financial hardship, housing instability, food insecurity, neighborhood disadvantage, social needs, predictive modeling, natural language processing, unsupervised clustering, electronic health records, health equity, algorithmic fairness, safety-net hospitals, Medicaid, financial assistance programs, cancer patients

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-27.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-27.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review