CoChat
Sign Up
Cited by 0Machine Learning in HealthcareArtificial Intelligence in HealthcareDiabetes Management and Research

Machine learning-driven risk prediction for post-hospitalization diabetes case management: Integrating clinical and social determinants of health

Open access
Lee, Seung-Yup, Saleem, Mohammad, Land, Andrew M., DeLaney, Erin W., Garretson, Alison R., Patel, Mahee, Hall, Allyson G., Westrick, Salisa +3 more·2026·International Journal of Medical Informatics
DOI 10.1016/j.ijmedinf.2026.106604PMC PMC13525843
Open this paper in CoChat — free accountOpen PDF

Abstract

Objective Patients hospitalized for diabetes-related conditions face elevated risks of emergency department (ED) visits post-discharge, driven by both clinical factors and social determinants of health (SDoH). This study aimed to develop and validate predictive models integrating clinical and SDoH data to identify high-risk patients for post-hospitalization diabetes case management. Methods We conducted a retrospective cohort study using electronic health record data from the University of Alabama at Birmingham Medical Center, including 162,063 inpatient encounters (January 2020-June 2024) for training and testing and 16,164 encounters (January-May 2025) for temporal validation. Patients were identified by diabetes-related ICD-10 codes or HbA1c ≥ 6.5 %, reflecting the scope of the institution's diabetes case management program. Predictors included demographics, diabetes-related comorbidities, surgical procedures, laboratory values, medications, and both area-level and individual-level SDoH. Logistic regression, decision trees, and XGBoost models were developed to predict diabetes-related ED visits within 3 months post-hospitalization. Hyperparameters for decision tree and XGBoost models were tuned via 10-fold cross-validation, and calibration was assessed using Brier scores and calibration plots. Results Among 162,063 hospitalizations, 6.2 % resulted in a diabetes-related ED visit. XGBoost achieved the best performance (area under the curve [AUC] 0.846, precision 0.420, sensitivity 0.296, specificity 0.972), maintained on temporal validation (AUC 0.842). Key predictors included past ED visit frequency, insulin prescriptions, age, and area-level SDoH indices. Individual-level SDoH factors, including home safety issues and work disability, also contributed to prediction. Targeting the top 20 % of predicted risk captured 64.1 % of all ED visits. Model discrimination was consistent across racial subgroups (AUC range: 0.843-0.851). Calibration was clinically acceptable across datasets. Conclusions Integration of clinical and SDoH data achieved effective prediction of post-hospitalization ED visits. XGBoost provided excellent discrimination with temporal stability. Decision trees offered greater interpretability. A pilot implementation delivering daily risk-stratified patient lists to the diabetes case manager is underway, demonstrating a practical pathway from model development to clinical decision support.

AI snapshot

Written from the paper’s full text.

TL;DR

This retrospective cohort study developed and temporally validated machine learning models integrating clinical and social determinants of health data to predict 3-month post-hospitalization diabetes-related ED visits. XGBoost achieved AUC 0.846 and maintained discrimination on temporal validation (AUC 0.842), with the top 20% of predicted risk capturing 64.1% of ED visits.

Key findings

  • Among 162,063 hospitalizations, 6.2% resulted in a diabetes-related ED visit within 3 months; XGBoost achieved AUC 0.846, precision 0.420, sensitivity 0.296, and specificity 0.972, outperforming logistic regression (AUC 0.816) and decision trees (AUC 0.810).
  • Temporal validation showed consistent XGBoost performance (AUC 0.842; sensitivity 0.294; precision 0.419; specificity 0.969), with Brier scores of 0.051 (test) and 0.055 (validation) and clinically acceptable calibration.
  • Key predictors included past 3-month ED visit frequency, insulin prescriptions (glargine and aspart), age, and area-level SDoH indices (SVI, ADI); individual-level SDoH such as home safety concerns, work disability, household size, and Medicaid insurance appeared among the top 30 predictors.
  • Targeting the top 20% of predicted risk captured 64.1% of all ED visits, compared with 29.6% using the traditional 0.5 probability threshold; model discrimination was consistent across racial subgroups (AUC range 0.843–0.851).

Methods

  • Retrospective cohort using electronic health record data from the University of Alabama at Birmingham Medical Center: 162,063 inpatient encounters (January 2020–June 2024) for training/testing (85%/15% split) and 16,164 encounters (January–May 2025) for temporal validation. Patients were identified by diabetes-related ICD-10 codes or HbA1c ≥6.5%.
  • Predictors included demographics, diabetes-related comorbidities, surgical procedures, laboratory values, medications, past 3-month ED visit frequency, area-level SDoH (ADI, SVI, LILA food desert indicator), and individual-level SDoH from insurance status and PRAPARE survey. The outcome was a diabetes-related ED visit within 3 months post-hospitalization.
  • Logistic regression, CART decision trees, and XGBoost were compared; CART and XGBoost hyperparameters were tuned via 10-fold cross-validation. Performance was assessed with AUC, sensitivity, precision, specificity, Brier scores, calibration plots, XGBoost feature importance, and race-stratified AUC.

Limitations

  • The paper states: high missingness in individual-level self-reported SDoH variables limited their direct contribution to the model; the single-institution southern US setting may limit generalizability; the operationally broad cohort definition may introduce heterogeneity; the outcome restricted to primary-diagnosis codes may undercount ED visits where diabetes was contributing but not primary; and patients may have sought ED care outside UAB.
  • Evident from the text: the unit of analysis was the individual hospitalization encounter, so patients with multiple hospitalizations contributed multiple observations, which may not account for within-patient correlation.

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

Integrity signals

Registry

Crossref · 10.1016/j.ijmedinf.2026.106604

Cites

90 of 90 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

38 of 60 cited sources checked · 7 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Social determinants of health influence outcomes among patients with diabetes, particularly in vulnerable populations.

    “At the same time, there is growing recognition that social determinants of health (SDoH)—including socioeconomic status, food access, housing stability, and transportation—influence outcomes among patients with diabetes, particularly in vulnerable populations [31–34].”
    Quote in the paperSubstantial claim4 of 4 sources checked
    • cites [31] Berkowitz SA, Orr CJ. Three Lessons About Diabetes and the Social Determinants of Health. Diabetes Care. 2023: dci230045. PMC10465981 37354315 source
    • cites [32] Walker RJ, Williams JS, Egede LE. Influence of race, ethnicity and social determinants of health on diabetes outcomes. The American journal of the medical sciences. 2016; 351(4): 366– 373. 27079342 PMC4834895 source

      The Social Determinants of Health and Diabetes section (page unavailable) links social determinants, including socioeconomic circumstances, to diabetes outcomes; the Neighborhood Factors and Diabetes section links food insecurity to poor glycemic control. The paper also discusses disparities affecting vulnerable groups and lists housing as a material factor, but the passages do not establish housing stability or transportation as influences on diabetes outcomes. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [33] Hill-Briggs F, Adler NE, Berkowitz SA, Social determinants of health and diabetes: a scientific review. Diabetes care. 2020; 44(1): 258. 33139407 PMC7783927 source
    • cites [34] Abbott EE, Apakama D, Richardson LD, Chan L, Nadkarni GN. Leveraging artificial intelligence and data science for integration of social determinants of health in emergency medicine: scoping review. JMIR Medical Informatics. 2024; 12: e57124. 39475815 PMC11539921 source

      The Introduction (page not provided) says SDoH can affect every patient’s health trajectory and clinical outcomes. The Results, under “SDOH Domains” (page not provided), include socioeconomic status, food insecurity, and housing insecurity. The paper addresses these issues in emergency medicine, but the passages do not establish their effects on diabetes outcomes specifically or address transportation as a determinant. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    These methods can capture complex, non-linear interactions among multiple factors that traditional models may miss.

    “These methods can capture complex, non-linear interactions among multiple factors that traditional models may miss [29,30].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [29] Carvalho DV, Pereira EM, Cardoso JS. Machine learning interpretability: A survey on methods and metrics. Electronics. 2019; 8(8): 832. source
    • cites [30] Gilpin LH, Bau D, Yuan BZ, Bajwa A, Specter M, Kagal L. Explaining explanations: An overview of interpretability of machine learning. IEEE; 2018: 80– 89. source

      The discussion of linear proxy models (A. Explanations of Deep Network Processing, p. 3) and local approximations (A. Processing, p. 8) concerns how to explain complex networks. Searches for non-linear interactions, input-feature relationships, and comparisons with traditional models found no passage saying these methods capture interactions that traditional models may miss. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
  • Premisecitation

    Individual-level SDoH data collection remains evolving and is not yet consistently available across large populations.

    “individual-level SDoH data collection remains evolving and not yet consistently available across large populations [95–98]”
    Quote in the paperNot a substantial claim3 of 4 sources checked
    • cites [95] Ashe JJ, Baker MC, Alvarado CS, Alberti PM. Screening for health-related social needs and collaboration with external partners among US hospitals. JAMA Network Open. 2023; 6(8): e2330228– e2330228. 37610754 PMC10448297 source

      The Discussion’s Limitations section (page unavailable) says the national hospital survey identifies which hospitals screen for social needs but cannot show whether screening reaches all patients; it calls for more granular data. That supports a gap in patient-level screening data, but the paper does not establish that individual-level SDoH data are inconsistently available across large populations generally. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [96] Brewster AL, Rodriguez HP, Murray GF, Lewis VA, Schifferdecker KE, Fisher ES. Trends in screening for social risk in US physician practices. JAMA network open. 2025; 8(1): e2453117– e2453117. 39752156 PMC11699528 source

      The Results report that systematic screening for all five social risks rose from 15% of physician practices in 2017 to 27% in 2022, supporting the claim that collection practices are evolving and uneven. But the Methods measure whether practices have screening systems, not whether individual-level SDoH data are consistently available across large populations; the Discussion’s Limitations also say the survey cannot show how screening is implemented. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [97] Gold R, Cottrell E, Bunce A, Developing electronic health record (EHR) strategies related to health center patients’ social determinants of health. The Journal of the American Board of Family Medicine. 2017; 30(4): 428– 447. 28720625 PMC5618800 source

      The Background (page not available) says some patient-reported SDoH domains are not regularly collected by federally funded clinics and that CHCs have historically addressed patients’ needs through manual, ad-hoc approaches. This supports the claim that collection is still developing, but the paper does not establish availability across large populations generally; its evidence is focused on CHCs. (LLM, conf=0.79)

      Experimental — accuracy not yet measured
    • cites [98] Magnan S Social determinants of health 101 for health care: five plus five. NAM perspectives. 2017; PMC8406598 34532697

      Not checked in the free run.

  • Findingown data

    Among 162,063 hospitalizations between January 2020 and June 2024 involving patients with diabetes-related conditions, 6.2% returned to the ED for diabetes-related reasons within 3 months.

    “Among 162,063 hospitalizations between January 2020 and June 2024 involving patients with diabetes-related conditions, 6.2% returned to the ED for diabetes-related reasons within 3 months.”
    • Also stated in: § Results · plain
      “Among 162,063 hospitalizations, 6.2% resulted in a diabetes-related ED visit.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Patients who returned to the ED within 3 months were significantly more likely to be Black than those who did not (56.7% vs. 38.9%, p<.001).

    “Compared with patients who did not return to the ED within 3 months, patients who returned were significantly more likely to be Black (56.7% vs. 38.9%, p<.001), have Medicaid insurance (37.9% vs. 31.9%, p<.001), and live in areas with higher deprivation (ADI: 76.4 vs. 71.4, p<.001) and social vulnerability (SVI: 0.64 vs. 0.57, p<.001).”
    Quote in the paperNot a substantial claim
  • May combine two claims. 37
  • Classified as probably not a substantial claim. 33

  • May combine two claims.

    Interpretationcitation

    Past ED visit frequency emerged as the strongest predictor, consistent with literature on recurrent acute care utilization and likely reflecting underlying health instability, insufficient outpatient follow-up, and barriers to primary care access.

    “Past ED visit frequency emerged as the strongest predictor, consistent with literature on recurrent acute care utilization and likely reflecting underlying health instability, insufficient outpatient follow-up, and barriers to primary care access [2,9].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [2] Hudon C, Courteau J, Krieg C, Vanasse A. Factors associated with chronic frequent emergency department utilization in a population with diabetes living in metropolitan areas: a population-based retrospective cohort study. BMC Health Services Research. 2017; 17: 1– 9. 28778161 PMC5544976 source

      Split into 2 atoms — 0 confirmed, 1 partially supported, 1 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [9] Rubin DJ, Shah AA. Predicting and preventing acute care re-utilization by patients with diabetes. Current Diabetes Reports. 2021; 21: 1– 13. 34480653 PMC8418292 source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    Models integrating both clinical and SDoH data may more effectively predict adverse events such as ED visits.

    “Models integrating both clinical and SDoH data may more effectively predict adverse events such as ED visits [35–37].”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [35] Lee S-Y, Hayes LW, Ozaydin B, Integrating Social Determinants of Health in Machine Learning–Driven Decision Support for Diabetes Case Management: Protocol for a Sequential Mixed Methods Study. JMIR Research Protocols. 2024; 13(1): e56049. 39321449 PMC11464948 source
    • cites [36] Basu S, Narayanaswamy R. A prediction model for uncontrolled type 2 diabetes mellitus incorporating area-level social determinants of health. Medical care. 2019; 57(8): 592– 600. 31268954 source
    • cites [37] Chi W, Andreyeva E, Zhang Y, Kaushal R, Haynes K. Neighborhood-level social determinants of health improve prediction of preventable hospitalization and emergency department visits beyond claims history. Population Health Management. 2021; 24(6): 701– 709. 34010058 source

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

    Study design and conductcitation

    Due to the high frequency of missing data in self-reported SDoH, we coded missing values as 'not measured' rather than imputing, as missingness may reflect lower use of safety-net services and thus warrant review.

    “Due to the high frequency of missing data in self-reported SDoH, we coded missing values as “not measured” rather than imputing [46,47], as missingness may reflect lower use of safety-net services and thus warrant review [70,71].”
    Quote in the paperSubstantial claim3 of 4 sources checked
    • cites [46] Lin W-C, Tsai C-F. Missing value imputation: a review and analysis of the literature (2006–2017). Artificial Intelligence Review. 2020; 53: 1487– 1509. source
    • cites [47] Mansoor H, Ali S, Alam S, Khan MA, Hassan UU, Khan I. Impact of missing data imputation on the fairness and accuracy of graph node classifiers. IEEE; 2022: 5988– 5997. source

      The Results and Discussions section reports that imputation techniques affect the fairness and accuracy of graph node classifiers, which offers a general reason for caution about imputing. It does not address self-reported SDoH, establish a high frequency of missingness there, or recommend coding missing values as “not measured.” The safety-net-services explanation is attached to different citations in the citing sentence. (LLM, conf=0.79)

      Experimental — accuracy not yet measured
    • cites [70] Henry AJ, Hevelone ND, Lipsitz S, Nguyen LL. Comparative methods for handling missing data in large databases. Journal of vascular surgery. 2013; 58(5): 1353– 1359. e6. 23830314 source

      The Analytic methods section (page unavailable) describes assigning a separate indicator to records with missing categorical data, which supports the coding approach. But the paper studies missing race data in hospital records, not self-reported SDoH, and its Abstract warns that the missing-indicator method produced the most biased coefficients; it does not recommend that approach over imputation for this setting. The safety-net-services clause has separate citations and is not assessed here. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [71] Hughes RA, Heron J, Sterne JA, Tilling K. Accounting for missing data in statistical analyses: multiple imputation is not always the answer. International journal of epidemiology. 2019; 48(4): 1294– 1304. 30879056 PMC6693809 source

      The Introduction (page unavailable) says that multiple imputation is not appropriate in every situation and that the choice of method depends on the amount, patterns, and reasons for missing data. That supports considering a non-imputation approach, but the paper does not support coding missing self-reported SDoH values as a “not measured” category on the basis of their high frequency. The safety-net-services clause is attributed to different citations, not this paper. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The post-hospitalization period is critical for diabetes care: gaps in follow-up, medication mismanagement, or worsening conditions can precipitate acute complications and costly ED visits.

    “The post-hospitalization period is critical for diabetes care: gaps in follow-up, medication mismanagement, or worsening conditions can precipitate acute complications and costly ED visits [7,10–12].”
    Quote in the paperSubstantial claim1 of 4 sources checked
    • cites [7] McCants KM, Reid KB, Williams I, Miller DE, Rubin R, Dutton S. The impact of case management on reducing readmission for patients diagnosed with heart failure and diabetes. Professional case management. 2019; 24(4): 177– 193. 31145236 source
    • cites [10] Black RL, Duval C. Diabetes discharge planning and transitions of care: a focused review. Current diabetes reviews. 2019; 15(2): 111– 117. 29992890 source
    • cites [11] Garnica P Transition of care for patients with diabetes. Current diabetes reviews. 2017; 13(3): 263– 279. 27881054 source
    • cites [12] Demidowich AP, Batty K, Zilbermint M. Instituting a successful discharge plan for patients with type 2 diabetes: challenges and solutions. Diabetes Spectrum. 2022; 35(4): 440– 451. 36561646 PMC9668725 source

      The paper’s opening section (page unavailable) emphasizes discharge medications and follow-up, identifies worsening conditions as a reason to adjust treatment, and links diabetes-related readmissions to financial implications. Its outpatient education section also calls continued care after discharge essential. These passages support the importance of the post-hospitalization period, but they do not specifically say that the listed gaps precipitate acute complications and costly ED visits. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The 3-month timeframe is clinically and operationally significant as the hospital's diabetes case manager follows up on discharged patients during this period, frequently identified as the high-risk period in the literature.

    “The 3-month timeframe is clinically and operationally significant as the hospital’s diabetes case manager follows up on discharged patients during this period, frequently identified as the high-risk period in the literature [39–41].”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [39] Saxena FE, Bierman AS, Glazier RH, Association of early physician follow-up with readmission among patients hospitalized for acute myocardial infarction, congestive heart failure, or chronic obstructive pulmonary disease. JAMA Network Open. 2022; 5(7): e2222056– e2222056. 35819782 PMC9277500 source

      The Introduction (page unavailable) calls the days immediately after discharge a period of increased risk. The Methods follow patients for 90 days, but the paper does not identify the entire three-month period as high risk; it studies patients hospitalized for AMI, CHF, or COPD, not diabetes patients. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [40] Davison BA, Metra M, Senger S, Patient journey after admission for acute heart failure: length of stay, 30-day readmission and 90-day mortality. European journal of heart failure. 2016; 18(8): 1041– 1050. 27114058 source

      The Results sections report readmissions within 30 days and deaths within 90 days after discharge among patients hospitalized for acute heart failure. This supports the relevance of early post-discharge risk, but the paper does not identify the full three-month period as high-risk for discharged diabetes patients. The hospital’s case-manager follow-up is the citing authors’ own context, not a finding attributed to this source. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [41] Culler SD, Kugelmass AD, Cohen DJ, Understanding Readmissions in Medicare Beneficiaries during the 90-Day Follow-up Period of an Acute Myocardial Infarction Admission. Journal of the American Heart Association. 2019; 8(21): e013513. 31663436 PMC6898831 source

      The Discussion (page not provided) reports readmissions within 90 days after acute myocardial infarction and identifies both type 1 and type 2 diabetes as predictors of readmission. That supports concern about risk during a three-month follow-up period, but the study concerns Medicare beneficiaries discharged after a heart attack; it does not establish the entire three months as a high-risk period for discharged diabetes patients generally. The hospital’s case-manager practice is context supplied by the citing authors, not a finding of this source. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Patients with diabetes are at higher risk of hospital admissions and frequent ED visits, driven by comorbidities such as heart failure, chronic kidney disease, and poor glycemic control.

    “Patients with diabetes are at higher risk of hospital admissions and frequent emergency department (ED) visits [2], driven by comorbidities such as heart failure, chronic kidney disease, and poor glycemic control [3].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [2] Hudon C, Courteau J, Krieg C, Vanasse A. Factors associated with chronic frequent emergency department utilization in a population with diabetes living in metropolitan areas: a population-based retrospective cohort study. BMC Health Services Research. 2017; 17: 1– 9. 28778161 PMC5544976 source

      The Background describes patients with diabetes as high health care users, and the Results report frequent ED use within the diabetes cohort. The paper does not establish that patients with diabetes have a higher risk of hospital admission or frequent ED visits than patients without diabetes. The comorbidities clause is attached to a different citation and is not assessed here. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [3] Agvall B, Jonasson JM, Galozy A, Halling A. Factors influencing hospitalization or emergency department visits and mortality in type 2 diabetes following the onset of new cardiovascular diagnoses in a population-based study. Cardiovascular Diabetology. 2024; 23(1): 124. 38600574 PMC11007935 source

      The Discussion (page unavailable) reports a higher hospitalization risk for people with both type 2 diabetes and cardiovascular disease, and higher hospitalization and ED-visit risk for the subgroup whose cardiovascular disease was diagnosed in the ED. It does not establish the broader claim that patients with diabetes generally have more hospital admissions or frequent ED visits. The comorbidity explanation is attached to citation [3], not this source. (LLM, conf=0.83)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Interpretationreasoning

    These findings suggest that predictive modeling benefits from capturing both scales of SDoH, with area-level indices providing broad contextual signals and individual-level measures offering more actionable insights for tailoring case management interventions.

    “These findings suggest that predictive modeling benefits from capturing both scales of SDoH, with area-level indices providing broad contextual signals and individual-level measures offering more actionable insights for tailoring case management interventions [33,36,37].”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [33] Hill-Briggs F, Adler NE, Berkowitz SA, Social determinants of health and diabetes: a scientific review. Diabetes care. 2020; 44(1): 258. 33139407 PMC7783927 source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [36] Basu S, Narayanaswamy R. A prediction model for uncontrolled type 2 diabetes mellitus incorporating area-level social determinants of health. Medical care. 2019; 57(8): 592– 600. 31268954 source
    • cites [37] Chi W, Andreyeva E, Zhang Y, Kaushal R, Haynes K. Neighborhood-level social determinants of health improve prediction of preventable hospitalization and emergency department visits beyond claims history. Population Health Management. 2021; 24(6): 701– 709. 34010058 source

  • Classified as probably not a substantial claim.

    Premisecitation

    Machine learning and artificial intelligence approaches have demonstrated growing promise in diabetes risk prediction, complication detection and healthcare utilization forecasting.

    “More recently, machine learning and artificial intelligence approaches have demonstrated growing promise in diabetes risk prediction, complication detection and healthcare utilization forecasting [21–28].”
    Quote in the paperSubstantial claim7 of 8 sources checked
    • cites [21] Shin J, Kim J, Lee C, Development of various diabetes prediction models using machine learning techniques. Diabetes & Metabolism Journal. 2022; 46(4): 650– 657. 35272434 PMC9353566 source

      The Discussion reports good performance for the paper’s 1- and 2-year diabetes prediction models, supporting the risk-prediction part of the claim. The paper does not report machine-learning detection of diabetes complications or forecasting of healthcare utilization. Our copy provides section labels but no page numbers. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [22] Chiu YM, Courteau J, Dufour I, Vanasse A, Hudon C. Machine learning to improve frequent emergency department use prediction: a retrospective cohort study. Scientific Reports. 2023; 13(1): 1981. 36737625 PMC9898278 source

      The Introduction and Discussion support machine-learning prediction of frequent emergency-department use, a form of healthcare utilization forecasting. The Discussion > Model performance also reports a separate study predicting uncontrolled diabetes, but that is not prediction of who will develop diabetes. Our copy does not support the complication-detection part of the claim, and its Conclusions say the models showed no clear superiority over simpler models for ED-use prediction. The indexed passages provide no page numbers. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [23] Guan Z, Li H, Liu R, Artificial intelligence in diabetes management: advancements, opportunities, and challenges. Cell Reports Medicine. 2023; 4(10) PMC10591058 37788667 source

      The sections on predicting diabetes onset and screening diabetic complications describe promising AI applications, including risk models and detection of diabetic retinopathy. The passages found do not describe forecasting healthcare utilization; using hospitalization records to predict complications is not the same thing. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [24] Xiong K, Cao G, Jin M, Ye B. A Multi-modal Deep Learning Approach for Predicting Type 2 Diabetes Complications: Early Warning System Design and Implementation. Journal of Theory and Practice of Engineering Science. 2025; 5(1): 33– 45. source
    • cites [25] Wang SC, Nickel G, Venkatesh KP, Raza MM, Kvedar JC. AI-based diabetes care: risk prediction models and implementation concerns. NPJ digital medicine. 2024; 7(1): 36. 38361152 PMC10869708 source

      The Risk prediction section describes AI models for predicting type 2 diabetes, and the Introduction reports an AI system that detects diabetic retinopathy in screening images. The paper does not report healthcare utilization forecasting, so it supports only part of the attributed claim. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [26] Alizadeh JM, Patel JS, Tajeu G, Predicting Emergency Department Visits for Patients with Type II Diabetes. arXiv preprint arXiv:241208984. 2024; source
    • cites [27] Wändell P, Wierzbicka M, Sigurdsson K, Development and evaluation of a machine learning prediction model for short-term mortality in patients with diabetes or hyperglycemia at emergency department admission. Cardiovascular Diabetology. 2025; 24(1): 383. 41044571 PMC12492943 source

      The Introduction says ML has shown promise in emergency-department outcome prediction, including complications and length of stay. The Abstract Conclusions report accurate prediction of 30-day mortality among patients with diabetes. The paper does not demonstrate prediction of diabetes risk or healthcare utilization in this population; it discusses length-of-stay prediction as prior work in broader emergency settings. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [28] Tan JK, Quan L, Salim NNM, Machine learning–based prediction for high health care utilizers by using a multi-institutional diabetes registry: model training and evaluation. JMIR AI. 2024; 3(1): e58463. 39418089 PMC11528163 source

  • May combine two claims.

    Interpretationcitation

    Statin prescriptions also appeared among important predictors, consistent with their indication for patients with diabetes and cardiovascular risk factors whose greater clinical complexity increases the likelihood of experiencing acute events requiring emergency care.

    “Statin prescriptions also appeared among important predictors, consistent with their indication for patients with diabetes and cardiovascular risk factors whose greater clinical complexity [86,87] increases the likelihood of experiencing acute events requiring emergency care [88–90].”
    Quote in the paperNot a substantial claim1 of 5 sources checked
    • cites [86] Stern MP. Diabetes and cardiovascular disease: the “common soil” hypothesis. Diabetes. 1995; 44(4): 369– 374. 7698502 source
    • cites [87] Grundy SM, Benjamin IJ, Burke GL, Diabetes and cardiovascular disease: a statement for healthcare professionals from the American Heart Association. Circulation. 1999; 100(10): 1134– 1146. 10477542 source
    • cites [88] Committee ADAPP. 10. Cardiovascular Disease and Risk Management: Standards of Care in Diabetes—2024. Diabetes Care. 2023; 47(Supplement_1): S179– S218. doi: 10.2337/dc24-S010 PMC10725811 38078592 source
    • cites [89] Mansi IA, Chansard M, Lingvay I, Zhang S, Halm EA, Alvarez CA. Association of statin therapy initiation with diabetes progression: a retrospective matched-cohort study. JAMA internal medicine. 2021; 181(12): 1562– 1574. 34605849 PMC8491130 source
    • cites [90] Grundy SM, Stone NJ, Blumenthal RS, High-intensity statins benefit high-risk patients: why and how to do better. Elsevier; 2021: 2660– 2670. 34531060 source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Race was included not as a biological variable but as a structural equity proxy, reflecting disparities in healthcare access, treatment patterns, and social conditions that are associated with diabetes outcomes and health decline.

    “Race was included not as a biological variable but as a structural equity proxy, reflecting disparities in healthcare access, treatment patterns, and social conditions that are associated with diabetes outcomes and health decline [32,42].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [32] Walker RJ, Williams JS, Egede LE. Influence of race, ethnicity and social determinants of health on diabetes outcomes. The American journal of the medical sciences. 2016; 351(4): 366– 373. 27079342 PMC4834895 source

      The sections “Racial Differences in Diabetes Outcomes” and “Social Determinants of Health and Diabetes” describe inequities in diabetes care and link access to care and social conditions with outcomes (page numbers unavailable). The paper does not say that race was included as a structural equity proxy rather than a biological variable; that is the citing authors’ characterization, not a finding established by this source. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [42] Nicklett EJ. Socioeconomic status and race/ethnicity independently predict health decline among older diabetics. BMC Public Health. 2011; 11: 1– 8. 21888645 PMC3175469 source

      The Background (page unavailable) links racial and ethnic health disparities to differences in medical decision-making, quality and access to care, discrimination, and socioeconomic conditions. The Discussion (page unavailable) finds race/ethnicity predicts health decline among older adults with diabetes. The paper does not explicitly say race was included as a structural equity proxy rather than a biological variable, and its Conclusions describe the proposed mechanisms as needing further study. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

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

    Premisecitation

    SDoH variables reflect structural inequities, and model outputs using them could inadvertently reinforce disparities if misapplied.

    “Because SDoH variables reflect structural inequities, there is a risk that model outputs could inadvertently reinforce disparities if misapplied [91,92].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [91] Rajkomar A, Hardt M, Howell MD, Corrado G, Chin MH. Ensuring fairness in machine learning to advance health equity. Annals of internal medicine. 2018; 169(12): 866– 872. 30508424 PMC6594166 source
    • cites [92] Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019; 366(6464): 447– 453. 31649194 source

      The Discussion (p. 8) says prediction labels can reflect structural inequalities and that apparently effective proxies can produce large biases. The paper supports the risk of reinforcing disparities, but it does not specifically establish that SDoH variables, rather than the labels and cost proxy it examines, reflect those inequities. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Marital status was included as a proxy for social support, which has been associated with diabetes management and health outcomes.

    “Marital status was included as a proxy for social support, which has been associated with diabetes management and health outcomes [43].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [43] Ramezankhani A, Azizi F, Hadaegh F. Associations of marital status with diabetes, hypertension, cardiovascular disease and all-cause mortality: a long term follow-up study. PloS one. 2019; 14(4): e0215593. 31009512 PMC6476533 source

      The Methods (Exposure) identify marital status as the study’s exposure, and the Discussion links marriage to health-related behaviors and marital status to type 2 diabetes. The paper does not describe marital status as a proxy for social support or examine diabetes management; its diabetes outcome is occurrence of type 2 diabetes, not management among people with diabetes. Page numbers are not available in our copy. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

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

    Interpretationcitation

    This finding may also reflect medication management challenges including nonadherence related to cost, storage, or dosing complexity.

    “This finding may also reflect medication management challenges including nonadherence related to cost, storage, or dosing complexity [83–85].”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [83] Dankers M, Van den Berk-Bulsink M, van Dalfsen-Slingerland M, Nelissen-Vrancken H, Mantel-Teeuwisse A, van Dijk L. Non-adherence to guideline recommendations for insulins: a qualitative study amongst primary care practitioners. BMC primary care. 2022; 23(1): 150. 35698052 PMC9189803 source

      The Results section (no page supplied) describes patients who cannot resuspend NPH insulin or use certain devices, supporting a medication-management difficulty. But the paper studies practitioners’ non-adherence to prescribing guidelines, not patients’ medication nonadherence; its discussion of higher costs concerns insulin selection, and the cited passages do not link nonadherence to storage or dosing complexity. (LLM, conf=0.81)

      Experimental — accuracy not yet measured
    • cites [84] Haycox A Insulin aspart: an evidence-based medicine review. Clinical drug investigation. 2004; 24: 695– 717. 17523734 source
    • cites [85] Evans M, Wilkinson M, Giannpolou A. Fast-acting insulin aspart: the rationale for a new mealtime insulin. Diabetes Therapy. 2019; 10: 1793– 1800. 31485918 PMC6778592 source

      In “The Role of Faster Aspart in Clinical Practice” (page unavailable), the paper says patients struggle with premeal insulin timing and reports that 24% of people with type 2 diabetes fail to follow dosing-timing guidance. This supports a medication-management and nonadherence concern, but the paper does not link nonadherence to cost, storage, or dosing complexity. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Temporal validation with calibration assessment is increasingly recommended but not yet routine in diabetes prediction modeling.

    “an approach increasingly recommended but not yet routine in diabetes prediction modeling [94]”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [94] Steyerberg EW, Harrell FE Jr. Prediction models need appropriate internal, internal-external, and external validation. Journal of clinical epidemiology. 2015; 69: 245. 25981519 PMC5578404 source

      In the main text (page not provided), the paper describes temporal external validation and recommends validation approaches that can split data by calendar time. It also discusses assessing calibration in external validation. It does not address diabetes prediction modeling or establish that temporal validation with calibration assessment is not yet routine in that field. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation

    Ongoing health system efforts to incorporate SDoH into routine care continue to advance.

    “ongoing health system efforts to incorporate SDoH into routine care continue to advance [99,100]”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [99] Gottlieb LM, Alderwick H. Integrating social and medical care: could it worsen health and increase inequity? The Annals of Family Medicine. 2019; 17(1): 77– 81. 30670400 PMC6342587 source
    • cites [100] Figueroa JF, Frakt AB, Jha AK. Addressing social determinants of health: time for a polysocial risk score. Jama. 2020; 323(16): 1553– 1554. 32242887 source

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

    Findingown data

    Most patients were male (51.6%).

    “Most patients were White (54.8%) and male (51.6%), with the majority insured by commercial plans (34.5%), followed by Medicaid (32.2%) and Medicare (24.8%).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Most patients were White (54.8%).

    “Most patients were White (54.8%) and male (51.6%), with the majority insured by commercial plans (34.5%), followed by Medicaid (32.2%) and Medicare (24.8%).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The majority were insured by commercial plans (34.5%), followed by Medicaid (32.2%) and Medicare (24.8%).

    “Most patients were White (54.8%) and male (51.6%), with the majority insured by commercial plans (34.5%), followed by Medicaid (32.2%) and Medicare (24.8%).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The average patient age was 56.1 years (SD = 17.9).

    “The average patient age was 56.1 years (SD = 17.9).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Patients with 3-month ED visits were significantly more likely to have been prescribed all insulin types (all p<.001), with the largest differences for aspart (90.1% vs. 55.8%) and glargine (67.5% vs. 29.9%).

    “Patients with 3-month ED visits were significantly more likely to have been prescribed all insulin types (all p<.001), with the largest differences for aspart (90.1% vs. 55.8%) and glargine (67.5% vs. 29.9%).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The decision tree provided interpretable risk pathways, with primary splits on insulin status, prior ED visits, and age.

    “The decision tree provided interpretable risk pathways (Supplemental Figure S1), with primary splits on insulin status, prior ED visits, and age.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Four individual-level SDoH-home safety concerns, work disability, household size, and Medicaid insurance-appeared among the top 30 predictors, alongside clinical variables such as high blood glucose, diabetic-range HbA1c, and chronic kidney disease.

    “Notably, four individual-level SDoH—home safety concerns, work disability, household size, and Medicaid insurance—appeared among the top 30 predictors, alongside clinical variables such as high blood glucose, diabetic-range HbA1c, and chronic kidney disease.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Calibration plots demonstrated close agreement between predicted and observed outcomes, with only mild overprediction at the highest risk levels.

    “Calibration plots (Figure 4) demonstrated close agreement between predicted and observed outcomes, with only mild overprediction at the highest risk levels.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Feature importance analysis from the XGBoost model identified past 3-month ED visit frequency as the strongest predictor by Gain, followed by insulin prescriptions (glargine and aspart), age, and area-level SDoH (SVI, ADI).

    “Feature importance analysis from the XGBoost model (Figure 2) identified past 3-month ED visit frequency as the strongest predictor by Gain, followed by insulin prescriptions (glargine and aspart), age, and area-level SDoH (SVI, ADI).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    XGBoost outperformed the other models with an AUC of 0.846, compared with 0.816 for logistic regression and 0.810 for decision trees (p<.001).

    “XGBoost outperformed the other models with an AUC of 0.846, compared with 0.816 for logistic regression and 0.810 for decision trees (p<.001) (Figure 1).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    XGBoost achieved the highest precision (0.420 vs. 0.331-0.335) and sensitivity (0.296 vs. 0.240-0.243).

    “XGBoost also achieved the highest precision (0.420 vs. 0.331–0.335) and sensitivity (0.296 vs. 0.240–0.243).”
    Quote in the paperSubstantial claim

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

    Findingown data

    Pregnancy-related diagnoses (O00-O9A) were less common in the ED visit group (0.9% vs. 4.2%, p<.001).

    “Pregnancy-related diagnoses (O00–O9A) were less common in the ED visit group (0.9% vs. 4.2%, p<.001).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Calibration intercepts were slightly negative and slopes below 1.

    “Calibration intercepts were slightly negative and slopes below 1 (Table 5), indicating mild overprediction at highest risk levels that remained clinically acceptable.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Circulatory system diagnoses (I00-I99) were most prevalent (72.1%) and more common among patients with 3-month ED visits (82.5% vs. 71.4%, p<.001).

    “Circulatory system diagnoses (I00–I99) were most prevalent (72.1%) and more common among patients with 3-month ED visits (82.5% vs. 71.4%, p<.001), followed by mental/behavioral disorders (F00–F99; 54.1% vs. 47.8%, p<.001) and genitourinary diagnoses (N00–N99; 47.8% vs. 39.8%, p<.001).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Self-reported SDoH variables showed significant differences between groups, though these variables were frequently missing (e.g., housing instability: 82.5% not measured; stress: 87.6% not measured).

    “Self-reported SDoH variables showed significant differences between groups, though these variables were frequently missing (e.g., housing instability: 82.5% not measured; stress: 87.6% not measured).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Several surgical procedures-including maternity care and delivery (0.54% vs. 3.4%), nervous system (6.2% vs. 10.9%), and respiratory system (4.8% vs. 9.4%)-were less prevalent among patients with ED visits.

    “Several surgical procedures—including maternity care and delivery (0.54% vs. 3.4%), nervous system (6.2% vs. 10.9%), and respiratory system (4.8% vs. 9.4%)—were less prevalent among patients with ED visits.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Aspart was the most commonly prescribed insulin (57.9%), followed by glargine (32.3%) and regular insulin (16.9%).

    “Aspart was the most commonly prescribed insulin (57.9%), followed by glargine (32.3%) and regular insulin (16.9%).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    Decision trees offered greater interpretability for clinical workflows.

    “decision trees offered greater interpretability for clinical workflows”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    XGBoost provided excellent, temporally stable discrimination with acceptable calibration.

    “XGBoost provided excellent, temporally stable discrimination with acceptable calibration”
    Quote in the paperSubstantial claim

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

    Findingnone given

    The study cohort represents underserved, diverse populations with high disease burden.

    “it represents underserved, diverse populations with high disease burden”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The model's discrimination was comparable across racial subgroups, with AUCs ranging from 0.843 to 0.851 and overlapping confidence intervals.

    “Our stratified analyses demonstrated comparable discrimination across racial subgroups (AUC range: 0.843–0.851), with overlapping confidence intervals indicating consistent performance.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Temporal validation was performed on the January 2025-May 2025 cohort, six-months after the last testing observation.

    “Temporal validation was performed on the January 2025–May 2025 cohort, six-months after the last testing observation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    No resampling techniques (e.g., SMOTE) were applied despite the low prevalence (6.2%), as sufficient positive cases (n ≈ 9,000), models' probability outputs, and XGBoost's scale_pos_weight parameter adequately addressed class imbalance.

    “No resampling techniques (e.g., SMOTE) were applied despite the low prevalence (6.2%), as sufficient positive cases (n ≈ 9,000), models’ probability outputs, and XGBoost’s scale_pos_weight parameter adequately addressed class imbalance.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    A pilot implementation using the XGBoost model is currently underway: each morning, an automated list of high-risk patients hospitalized for ≥48 hours is delivered to the diabetes case manager, enabling same-day engagement and structured 90-day post-discharge follow-up.

    “A pilot implementation using the XGBoost model is currently underway: each morning, an automated list of high-risk patients hospitalized for ≥48 hours is delivered to the diabetes case manager, enabling same-day engagement and structured 90-day post-discharge follow-up.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    The six month-gap was intentional to ensure clear temporal distinction between model development and validation and to reduce temporal autocorrelation between datasets.

    “The six month-gap was intentional to ensure clear temporal distinction between model development and validation and to reduce temporal autocorrelation between datasets.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    AUC values above 0.7 are considered acceptable and above 0.8 excellent.

    “Model performance was evaluated using area under the curve (AUC), sensitivity, precision, and specificity, with AUC values above 0.7 considered acceptable and above 0.8 excellent [72].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [72] Hosmer DW, Lemeshow S, Sturdivant RX, ProQuest. Applied logistic regression. Third edition ed. Wiley series in probability and statistics. Wiley; 2013. source

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

    Study design and conductnone given

    CART and XGBoost hyperparameters were tuned via 10-fold cross-validation.

    “CART and XGBoost hyperparameters were tuned via 10-fold cross-validation (details in Supplementary Appendix A).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Because the PRAPARE survey was administered primarily in community health clinics and EDs, patients with survey data were generally more socioeconomically vulnerable, while missingness may indicate less frequent use of safety-net services.

    “Given that the PRAPARE survey was administered in community health clinics and ED settings, patients with PRAPARE entries were generally more socioeconomically vulnerable, while missing data may suggest less frequent use of safety-net services due to better access to care.”
    • Also stated in: § Principal Findings · hedged
      “Because the PRAPARE survey was administered primarily in community health clinics and EDs, patients with survey data were generally more socioeconomically vulnerable, while missingness may indicate less frequent use of safety-net services.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    XGBoost achieved excellent discrimination (AUC 0.846) maintained on temporal validation, with clinically acceptable calibration, supporting its potential for clinical decision support.

    “XGBoost achieved excellent discrimination (AUC 0.846) maintained on temporal validation, with clinically acceptable calibration, supporting its potential for clinical decision support.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Area-level SDoH indices (SVI, ADI, LILA) were among the strongest predictors, underscoring how neighborhood-level disadvantage and food access barriers contribute to ED utilization risk.

    “Area-level SDoH indices (SVI, ADI, LILA) were among the strongest predictors, underscoring how neighborhood-level disadvantage and food access barriers contribute to ED utilization risk.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    Insulin prescriptions-particularly glargine (long-acting) and aspart (rapid-acting)-were the next strongest clinical predictors, suggesting that patients requiring both basal and rapid-acting insulin have more complex or less stable disease.

    “Insulin prescriptions—particularly glargine (long-acting) and aspart (rapid-acting)—were the next strongest clinical predictors, suggesting that patients requiring both basal and rapid-acting insulin have more complex or less stable disease [81,82].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [81] Bolli GB, Owens DR. Insulin glargine. The Lancet. 2000; 356(9228): 443– 445. 10981882 source
    • cites [82] Mudaliar SR, Lindberg FA, Joyce M, Insulin aspart (B28 asp-insulin): a fast-acting analog of human insulin: absorption kinetics and action profile compared with regular human insulin in healthy nondiabetic subjects. Diabetes care. 1999; 22(9): 1501– 1506. 10480516 source

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    In the case management context, the model serves as a prioritization tool rather than a diagnostic test.

    “In the case management context, the model serves as a prioritization tool rather than a diagnostic test.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Several surgical procedures-including maternity care, nervous system, and cardiovascular procedures-were less prevalent among patients with ED visits (Table 3). These associations likely reflect the composition of distinct patient subgroups within the operationally defined cohort.

    “Several surgical procedures—including maternity care, nervous system, and cardiovascular procedures—were less prevalent among patients with ED visits (Table 3). These associations likely reflect the composition of distinct patient subgroups within the operationally defined cohort.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Calibration was assessed using Brier scores, calibration intercepts and slopes from logistic calibration models, and calibration plots with locally estimated scatterplot smoothing (LOESS) and 95% confidence intervals, to assess the agreement between predicted risks and observed outcomes.

    “Calibration was assessed using Brier scores, calibration intercepts and slopes from logistic calibration models, and calibration plots with locally estimated scatterplot smoothing (LOESS) and 95% confidence intervals, to assess the agreement between predicted risks and observed outcomes.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    To assess model fairness, we examined XGBoost discrimination (AUC with 95% confidence intervals) stratified by racial subgroup (White, Black, Other) in the test dataset.

    “To assess model fairness, we examined XGBoost discrimination (AUC with 95% confidence intervals) stratified by racial subgroup (White, Black, Other) in the test dataset.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Targeting the top 20% of predicted risk captured 64.1% of all ED visits (Table 4), offering a practical operating point for programs with limited capacity.

    “Targeting the top 20% of predicted risk captured 64.1% of all ED visits (Table 4), offering a practical operating point for programs with limited capacity.”
    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

    The 'not measured' category was used only for descriptive analyses, and its binary indicator was excluded from prediction modeling.

    “The “not measured” category was used only for descriptive analyses, and its binary indicator was excluded from prediction modeling.”
    Quote in the paperNot a substantial claim

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

    Interpretationcitation

    The study's cohort scope enhances the likelihood of successful implementation.

    “enhancing the likelihood of successful implementation [93]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [93] Bates DW, Saria S, Ohno-Machado L, Shah A, Escobar G. Big data in health care: using analytics to identify and manage high-risk and high-cost patients. Health affairs. 2014; 33(7): 1123– 1131. 25006137 source

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

    Interpretationnone given

    The single-institution setting in the southern United States may limit the generalizability of the findings.

    “the single-institution setting in the southern United States may limit generalizability”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The operationally broad cohort definition may introduce heterogeneity.

    “the operationally broad cohort definition may introduce heterogeneity”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationnone given

    Focusing on recently hospitalized patients and their returns within the same system provides a perspective that directly supports case management efforts.

    “the focus on recently hospitalized patients and their returns within the same system provides a perspective that directly supports case management efforts”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The study's broad cohort definition reflects real-world case management practice.

    “our cohort reflects real-world case management scope”
    • Also stated in: § Limitations · plain
      “aligns with real-world case management practice”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Restricting the outcome to primary-diagnosis codes may undercount ED visits where diabetes was contributing but not primary.

    “Fourth, the outcome restricted to primary-diagnosis codes may undercount ED visits where diabetes was contributing but not primary.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    Patients may have sought ED care outside UAB.

    “patients may have sought ED care outside UAB”
    Quote in the paperNot a substantial claim
59 of 59

  • The paper relies on this without support.

    The ICD-10 codes used to define diabetes-related ED visits accurately identify diabetes-related ED visits.

    “A diabetes-related ED visit was counted if any of the ICD-10 codes listed above appeared as the primary diagnosis.”

  • The paper relies on this without support.

    The temporal validation cohort (January–May 2025) is comparable to the training/testing cohort except for time, so performance differences reflect temporal stability.

    “Temporal validation was performed on the January 2025–May 2025 cohort, six-months after the last testing observation.”

  • The paper relies on this without support.

    The encounter-level unit of analysis reflects the operational use case of the case management team.

    “This encounter-level approach reflects the operational use case, in which the case management team assesses risk at each hospitalization.”
3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Alizadeh JM, Patel JS, Tajeu G, et al. Predicting Emergency Department Visits for Patients with Type II Diabetes. arXiv preprint arXiv:241208984. 2024.

Cited by

Still being checked.

Related

Authors
Lee, S.-Y., Saleem, M., Land, A. M., DeLaney, E. W., Garretson, A. R., Patel, M., Hall, A. G., Westrick, S., Ovalle, F., Cherrington, A. L., & Banaszak-Holl, J. C.
Affiliations
  • University of Alabama at Birmingham (US)
  • UAB Medicine (US)
  • Auburn University (US)
Corresponding author
Lee, Seung-Yup
ORCID
Venue
International Journal of Medical Informatics
Publisher
Elsevier BV
Published
2026-11
Volume
220
Pages
106604
ISSN
1386-5056 (ISSN-L)
DOI
10.1016/j.ijmedinf.2026.106604
PMC
PMC13525843
OpenAlex
W7168023483
Semantic Scholar
e9815a6d99cab27a4d4d2503eba2903ce3c50cdd
Cited by
0
References
90
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • International Journal of Medical Informatics (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion)
Funders
  • University of Alabama
  • National Center for Advancing Translational Sciences
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Health Sciences › Medicine › Endocrinology, Diabetes and Metabolism › Diabetes Management and Research
Keywords
diabetes mellitus, social determinants of health, risk prediction, XGBoost, decision tree, home safety, work disability, risk stratification, clinical decision support, temporal validation, hospital readmission risk, insulin prescription, HbA1c, diabetes comorbidity
MeSH terms
Machine Learning, Emergency Room Visits, Boosting Machine Learning Algorithms, Predictive Learning Models, Classification Algorithms, Prediction Algorithms, Aged, Alabama, Diabetes Mellitus, Female, Hospitalization, Humans, Male, Middle Aged, Retrospective Studies, Risk Assessment, Case Management, Electronic Health Records, Social Determinants of Health

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 2026-10-28.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-28.

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