Reducing Readmissions in the Safety Net Through AI and Automation

ABSTRACT

Objectives: To implement a technology-based, systemwide readmission reduction initiative in a safety-net health system and evaluate clinical, care equity, and financial outcomes.

Study Design: Retrospective interrupted time series analysis between October 2015 and January 2023.

Methods: The readmission reduction initiative standardized inpatient care for patients through a novel, electronic health record–integrated, digitally automated point-of-care decision-support tool. A predictive artificial intelligence algorithm was utilized to identify patients at the highest risk of readmission in both the inpatient and outpatient settings, allowing a population health team to perform proactive outpatient management in medical and social domains to avoid readmission.

Results: Readmission rates declined from 27.9% in the preimplementation period to 23.9% in the postimplementation period (P < .004) by the end of 2023. A significant gap in readmission rates between Black/African American patients and the general population was eliminated over the course of the evaluation period. Survival analysis demonstrated a reduction in all-cause mortality in the postimplementation period (HR, 0.82; 95% CI, 0.68-0.99; P = .037). Improvement in readmission rates allowed the health system to retain $7.2 million of at-risk pay-for-performance funding.

Conclusions: This technology-based readmission reduction initiative demonstrated efficacy in reducing readmission rates, closing equity gaps, improving survival, and leading to a positive financial impact in a safety-net health system. This approach could be an effective model of technology-based, value-based care for other resource-limited health systems to meet pay-for-performance metrics and retain at-risk funding while improving clinical and equity outcomes.

Am J Manag Care. 2025;31(3):In Press

Takeaway Points

A safety-net health system implemented a technology-based readmission reduction initiative, leading to cost savings and improved clinical and equity outcomes.

• Through a Lean-based, data-driven analysis, the health system identified a chronic disease driving readmission rates and focused the readmission initiative on that population.
• Electronic health record–based automation tools standardized a safe discharge workflow for hospitalized patients at the point of care.
• Predictive artificial intelligence identified patients with the highest readmission risk, allowing a population health team to perform proactive outpatient management.
• This technology-based methodology offers a framework for other safety-net systems to meet performance-based funding metrics while improving patient care.

Safety-net health systems are defined as a health care delivery institution that, by legal obligation or mission, provides health care for individuals regardless of their insurance status or ability to pay. Amid the shift toward performance-based funding in US health care, these health systems face the dual challenge of adhering to performance metrics to retain at-risk funding without compromising patient outcomes or equity. Hospital readmission reduction metrics, a core component of pay-for-performance programs, directly financially penalize health systems with higher-than-expected 30-day readmission rates, creating a significant challenge for safety-net providers. Since CMS launched the Hospital Readmissions Reduction Program (HRRP),¹ readmission reduction has been widely adopted as a federal and state metric.² Although readmissions appear to have declined nationally,³ safety-net hospitals—serving a high proportion of vulnerable and underserved patients—have struggled. These health systems are often disproportionately penalized despite risk-adjusted metrics,^4,5 leading to reduced funding for health systems that are most in need. Moreover, programs to reduce readmissions have sometimes inadvertently increased mortality and exacerbated health care equity gaps.^6,7

Despite the substantial at-risk funding associated with readmission metrics, few health system–wide initiatives have successfully improved readmission metrics in safety-net health systems. A meta-analysis of 57 system-level interventions reported zero successful interventions to reduce readmissions in the safety-net setting.⁸ Indeed, few interventions are tested within the safety net, but those that are focus primarily on improving medical factors and often neglect interventions in the social domain, despite research demonstrating that medical and social risk factors are drivers of poor readmission outcomes in safety-net settings.^9-11

Health system–level performance improvement initiatives to reduce readmissions in safety-net health systems must be designed and implemented to address several unique challenges. First and foremost, readmission reduction programs must successfully reduce readmission rates. However, reductions in readmission rates must be accompanied by stability or improvement in overall mortality rates and any equity gaps in readmission. It would be a failure for a safety-net health system, already caring for vulnerable and underserved patients, to meet readmission reduction metrics at the expense of worsened clinical or equity outcomes. Second, readmission reduction efforts must lead to overall cost savings, given the significant financial resource constraints of these institutions. Finally, due to the annual repetition of readmission reduction metrics, the readmission reduction program must lead to durable and sustained improvement.

Zuckerberg San Francisco General Hospital (ZSFG) is an urban, academic, safety-net hospital within the San Francisco Health Network that, before 2017, experienced elevated readmission rates as well as significant disparities in the rates of readmission between Black/African American (B/AA) patients and the general patient population. Failure to meet readmission reduction metrics imperiled $1.2 million per year of funding. Here, we present the methodology and outcomes of a multiyear, systemwide readmission reduction program designed and implemented at ZSFG that used predictive artificial intelligence (AI) and electronic health record (EHR)–based automation tools to navigate these challenges effectively.

METHODS

Understanding the Problem and Formulating an Approach

Utilizing Lean methodology, ZSFG performed a data-driven analysis to understand the system-based challenges contributing to elevated all-cause 30-day readmission rates. A pivotal finding was that heart failure (HF) accounted for more than 40% of unplanned readmission events, causing an outsized effect on readmission performance. Preliminary modeling indicated that successful reductions in all-cause 30-day HF readmission rates would result in the health system meeting overall readmission reduction metrics. Therefore, an approach targeting HF readmission reduction would have the greatest likelihood of success. It would focus resources and efforts in a specific clinical area, making countermeasures easier to implement and iterate. An analysis was then performed to understand the critical drivers of 30-day unplanned readmission in HF patients, including the following:

• Patients’ social determinants of health (SDOH) had an outsized negative effect on readmission rates.¹¹ Patients with both HF and methamphetamine use were at the highest risk, for example.
• There was no standardized approach to HF care, contributing to substantial care variation sometimes driven by underlying treatment biases.
• Clinical teams had difficulty identifying patients at the highest risk of readmission.

Piloting Interventions

Based on these findings, pilot countermeasures were designed and implemented for a limited 6-month trial period on a single inpatient service. These pilot countermeasures included the following:

• Use of an evidence-based inpatient checklist that standardized care for all patients with HF¹²: This checklist ensured that all patients admitted with HF underwent full diuresis before discharge, received social risk–informed HF medical therapy, and received expedited 7-day follow-up in primary care and cardiology clinics.
• Creating a specialized team to care for the highest-risk patients with HF: This Heart Team combined previously siloed health care providers into a multidisciplinary group of HF experts; primary care providers; and specialists in addiction medicine,¹³ palliative care, and social medicine.¹⁴

Adapting Successful Pilots Into a Health System–Wide Improvement Program

Although these pilot countermeasures yielded promising results, specific barriers to usage and deployment were identified: (1) The inpatient checklist was paper-based and lacked direct integration with clinical workflows and the EHR, and (2) the Heart Team struggled to identify high-risk patients with HF and relied on word-of-mouth referrals.

The team determined that the only practical implementation strategy was to deploy the pilot interventions as digital tools within the EHR and create a digital platform for HF readmission care coordination. The team determined this could be accomplished by innovating existing EHR technology and utilizing predictive AI.

Converting the Paper-Based Discharge Checklist Into an EHR-Integrated, Adaptive, Point-of-Care Automated Decision Support Tool

We determined that the checklist could be converted into a digital decision-support tool if 3 criteria were met: (1) It was integrated into the EHR to prevent providers from having to leave their digital workspace to interact with the tool, (2) it adapted treatment recommendations to individual patient medical and social needs based on provider inputs and EHR patient data, and (3) it was automated, meaning it streamlined provider workflows, improved clinical efficiency, and decreased provider cognitive burden by collecting and processing health data to inform guideline-directed care.

Using EHR technology, we adapted the discharge checklist into a logic-based, point-of-care decision-support tool housed within a custom-built user interface (Figure 1). This tool provided guideline-based recommendations about HF care to inpatient providers.

Utilization of Predictive AI to Identify Patients at High Risk of Readmission

We localized an AI model predicting readmission risks to the ZSFG population to provide a framework for readmission risk stratification. This model was a proprietary random forest ensemble learning model from Epic Systems (Risk of Unplanned Readmission Version 2) utilizing 70 features. The area under the curve of the model was 0.72, the true positive rate was 79.7%, and the positive predictive value was 21.3%.

Predictions were surfaced in the decision-support tool and guided providers to place high-priority follow-up referrals (Figure 1). Linking a predictive output to a specific provider action is critical to successfully implementing predictive AI to improve clinical care. Prior research has demonstrated that surfacing a prediction without actionable guidance does not improve outcomes.¹⁵

Development of Population-Level HF Dashboards for Anticipatory Population Health Management

We created an HF dashboard within the EHR that displays predictive AI-derived readmission risk outputs for all patients with HF in real time and identifies key clinical metrics (Figure 2). Through this dashboard, the Heart Team transitioned to an anticipatory population health management approach in which, instead of identifying high-risk patients based on prior frequency of admissions, they focused on patients with a high risk of predicted unplanned readmission. Specifically, the Heart Team met every month and developed multidisciplinary management plans addressing both the medical and social needs of the patients at the highest predicted risk of readmission. The Heart Team then coordinated and executed this management plan with the patient’s care team.

Analyzing Outcomes

We performed a retrospective time series analysis of readmission and mortality rates for patients with HF between October 2015 and January 2023. We utilized internal historical control data and data from 5 comparator safety-net hospitals in California to assess this program’s efficacy and ensure the external validity of our results. All data were retrospectively collected and reported by the third-party health care service Vizient¹⁶ in March 2023. California state death registry records were matched to ZSFG data to identify mortality events. The institutional review boards at the University of California, San Francisco; ZSFG; and Vizient approved this research.

All-cause 30-day readmission after an HF admission was determined on a rolling basis by the month of discharge at the patient level. Readmission was defined by the Quality Incentive Pool Program definition provided by the California Department of Health Care Services in conjunction with CMS to align the readmission definition with performance metrics.¹⁷ The statistical efficacy of cumulative interventions over time compared with comparator hospitals was interrogated with a multivariate mixed-effects logistic regression–interrupted time series analysis. Models were corrected for age, sex, and Charlson Comorbidity Index (CCI) score,¹⁸ similar to the covariates used to adjust readmission rates by the HRRP.¹ Mixed-effects logistic regression was performed to examine the impact of the readmission reduction program on census-based racial and ethnic subpopulations. Cox proportional hazards models adjusted for age, sex, CCI score,¹⁸ and Social Deprivation Index were used to interrogate the program’s effect on postdischarge survival. Mortality data were available only for patients from ZSFG; therefore, we could not compare survival trends with those of other hospitals.

All statistical analysis was performed using R 4.2.2 (R Foundation for Statistical Computing). The significance cutoff for all analyses was set at a P value less than .05.

RESULTS

Effect on All-Cause 30-Day Readmission Rates

At ZSFG, HF readmission rates declined from 27.9% in the preimplementation period to 23.9% in the postimplementation period (P < .004) by the end of 2023. Compared with peer hospitals, the odds of 30-day readmission were significantly higher at ZSFG in the preimplementation period (OR, 1.58; 95% CI, 1.21-2.06; P < .001), and readmission odds trended upward over time before implementation (OR, 1.06 per year; 95% CI, 1.00-1.13; P = .065). The decline in readmission odds following program implementation was significantly higher at ZSFG compared with other hospitals (OR, 0.91 per year; 95% CI, 0.84-0.98; P = .015) (Figure 3). At baseline, ZSFG had the highest 30-day unplanned readmission rate among peer safety-net hospitals in California. As of January 2023, ZSFG exhibited one of the lowest readmission rates among safety-net hospitals in California.

Effect on Readmission Equity

In 2018, ZSFG exhibited a significant gap in readmission rates between B/AA patients with HF and the general HF population (Figure 4), and adjusted odds of readmission were significantly higher for B/AA patients (OR, 1.49; 95% CI, 1.13-1.98; P = .005) than for other census-based racial (Asian/Pacific Islander, White) or ethnic (Hispanic, non-Hispanic) subpopulations, which did not significantly differ from one another (all P ≥ .5). By 2022, there was no difference in readmission rates (Figure 4), and there was a significant decline in readmission risk in the postimplementation period among B/AA patients with HF (OR, 0.87 per year; 95% CI, 0.80-0.94; P < .001). Compared with peer safety-net hospitals, readmission risk declined at a significantly higher rate for B/AA patients with HF at ZSFG in the postimplementation period (OR, 0.85; 95% CI, 0.78-0.94; P < .001).

Effect on Mortality

Cox proportional hazards models adjusted for age, sex, CCI score, and Social Deprivation Index revealed a significant reduction in risk of mortality in patients with HF in the postimplementation period compared with patients with HF in the 3 years before implementation (HR, 0.82; 95% CI, 0.68-0.99; P = .037) (Figure 5).

Effect on Pay-for-Performance Readmission Metrics and Financial Impact on the Health System

As a county safety-net hospital, ZSFG was enrolled in federal and state pay-for-performance programs.^1,2 At-risk pay per year was $1.2 million, and ZSFG had yet to meet readmission metrics before the implementation of this program in 2018. After implementation, the health system met both metrics annually from 2018 to 2023. This resulted in the retention of $7.2 million of at-risk funding compared with a development cost of $1 million for the decision-support tool. Thus, the overall return on investment for the health system has been over $7 to $1.

DISCUSSION

This health system–wide performance improvement initiative in a safety-net health system demonstrated the feasibility of utilizing technology to meet readmission reduction metrics while simultaneously closing readmission equity gaps and improving mortality. Programs implemented in other health systems to improve readmissions have resulted in variable success. Some programs have not successfully reduced readmission rates,¹⁹ whereas readmission reduction has come at the cost of increased mortality for others.²⁰ To our knowledge, this is the first readmission reduction program in a safety-net health system to report improvement in readmission rate, survival, and a known equity gap.

We postulate that 3 key factors were responsible for the success of this program, none of which would have been possible without the use of technology: standardization of medical therapy to limit variation and potential biases in care, addressing individual patient social needs, and identifying high-risk patients. Automation tools our team developed in the EHR allowed for standardizing HF care across all patients admitted with HF from admission to discharge. Standardization ensures that patients are provided with contemporary, guideline-directed therapy and appropriate social interventions, regardless of social circumstances that may lead to unintended biases in care. This is particularly important in safety-net settings, where standardized medical therapy alone is inadequate to improve poor outcomes compounded by substantial social risk factors.^4,11 A negative relationship between social risk factors and care provision is well known. For example, patients who are experiencing homelessness and those deemed by health systems to be actively using illicit substances are less likely to receive guideline-directed HF care or access ambulatory care systems.^21,22 Interventions described in this improvement program addressed a limited subset of social needs. Still, by identifying the most significant social risk factors for poor outcomes in our population, we could direct limited resources toward social interventions with the most significant population-level impact. In our case, the majority of the social interventions were directed toward substance use, particularly methamphetamine use.

Standardization and social interventions would not have been as far-reaching or impactful without predictive AI. An AI model that predicted readmission risk allowed our team to focus limited resources on our highest-risk patients and continuously monitor the entirety of the HF population for temporal changes in risk. Before using these models, there was no method to identify high-risk patients. Predictive AI algorithms require a relatively small number of trained staff with expertise, with low marginal cost or significant resource burden to the health care system, making it especially well-suited for safety-net health systems with an integrated EHR.

Despite these successes, we encountered several hurdles during the implementation of this program. First, general interaction rates with EHR-based decision-support aid across our health system were meager, sometimes less than 1% in the case of best-practice advisories. Digital tools are only effective when their intended users use them. To address this, we conducted workshops with providers using the digital tool to integrate their design feedback and encourage buy-in to facilitate the tool’s success. Post implementation, we conduct in-person monthly orientations to the tool and, through these orientations, solicit further feedback for design improvement. Current metrics show that providers interact with the tool for decision support in 56% to 75% of inpatients with HF. Second, although the predictive AI model provided a framework for readmission risk stratification, it had a high specificity but low positive predictive value. As such, some patients were inaccurately labeled as high risk for readmission and received priority referrals. However, the overall number of referrals did not increase because all admitted patients with HF received a discharge referral to the HF clinic.

In August 2024, we transitioned to a gradient-boosted tree model that we developed internally with an equivalent area under the curve but improved positive predictive value. This model utilizes 49 features, some of which incorporate social determinants of health, which were absent from the initial model and are critical for readmission prediction performance and bias mitigation given the outsized effect of social determinants of health on patient outcomes in our population.¹¹

Limitations

There are several limitations to our outcome analysis. We cannot distinguish the effects of each independent system-level intervention from the others and, therefore, we cannot quantify the effect of the predictive AI algorithm vs the standardization of care and rigorous dissemination of evidence-based treatment guidelines on the outcomes of this initiative. Although we suspect that combining these interventions resulted in cumulative improvements over time, we also believe that health systems without access to predictive AI tools can still successfully improve patient outcomes through automated decision-support tools. We believe these tools should be the first digital solutions implemented in health systems to leverage technology for performance improvement, given their relative simplicity compared with AI and high degree of impact. Many system-level interventions carry the risk of systematically benefiting a specific subpopulation, which may improve or worsen existing care gaps. Through statistical comparison of patient populations at 5 additional safety-net hospital systems in California over the same period, we attempted to evaluate whether these changes resulted from population-level changes occurring irrespective of interventions at our hospital. Although this method increases our confidence that these results are likely due to system-level changes, without an intrahospital control group, we cannot entirely exclude the possibility that population changes are confounding our results or regression to the mean. Finally, the reported financial impact of this initiative is not a complete cost-benefit analysis. It is limited to comparing the cost of developing the technological tools with the retention of at-risk pay-for-performance funding.

CONCLUSIONS

Implementing population-level performance improvement initiatives is often difficult for resource-limited health care systems due to financial and logistical constraints. However, these programs are vital for meeting performance metrics and advancing health outcomes and equity. We believe technology has significant potential to overcome inherent resource limitations affecting successful improvement initiatives in safety-net health systems. Our methodology and technology-based improvement initiative may be a framework for other safety-net systems to avoid pay-for-performance penalization and increase available financial resources while delivering value-based patient care. 

Author Affiliations: School of Medicine, University of California, San Francisco (DJB), San Francisco, CA; Department of Epidemiology and Biostatistics, University of California, San Francisco (JF, AK), San Francisco, CA; Division of Hospital Medicine (SG), Division of Cardiology (MSD, JD, LSZ), and Department of Anesthesia (JM), Zuckerberg San Francisco General Hospital (SE), San Francisco, CA; San Francisco Department of Public Health (SG, SE), San Francisco, CA; Department of Family and Community Medicine, University of California, San Francisco (LMG), San Francisco, CA; Department of Medicine, University of California, San Francisco (MSD, SE, JD, LSZ), San Francisco, CA; Division of Clinical Informatics and Digital Transformation, University of California, San Francisco (LSZ), San Francisco, CA.

Source of Funding: Zuckerberg Priscilla Chan Qualitative Improvement Fund via the San Francisco General Hospital Foundation.

Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Authorship Information: Concept and design (DJB, JF, LMG, MSD, SE, JD, LSZ); acquisition of data (DJB, LSZ); analysis and interpretation of data (DJB, JF, SG, AK, MSD, LSZ); drafting of the manuscript (DJB, SG, LMG, SE, LSZ); critical revision of the manuscript for important intellectual content (DJB, JF, SG, LMG, MSD, JM, SE, JD, LSZ); statistical analysis (DJB, JF, AK, LSZ); provision of patients or study materials (JD); obtaining funding (JM, LSZ); administrative, technical, or logistic support (AK, JM, LSZ); and supervision (JF, JM, SE, LSZ).

Address Correspondence to: Lucas S. Zier, MD, MS, Zuckerberg San Francisco General Hospital, 1001 Potrero Ave, Room 5G2, San Francisco, CA 94110. Email: lucas.zier@ucsf.edu.

REFERENCES

1. Hospital Readmissions Reduction Program (HRRP). CMS. Accessed July 31, 2024. https://www.cms.gov/medicare/quality/value-based-programs/hospital-readmissions

2. McIlvennan CK, Eapen ZJ, Allen LA. Hospital Readmissions Reduction Program. Circulation. 2015;131(20):1796-1803. doi:10.1161/CIRCULATIONAHA.114.010270

3. Wadhera RK, Yeh RW, Joynt Maddox KE. The Hospital Readmissions Reduction Program — time for a reboot. N Engl J Med. 2019;380(24):2289-2291. doi:10.1056/NEJMp1901225

4. Figueroa JF, Joynt KE, Zhou X, Orav EJ, Jha AK. Safety-net hospitals face more barriers yet use fewer strategies to reduce readmissions. Med Care. 2017;55(3):229-235. doi:10.1097/MLR.0000000000000687

5. Joynt KE, Jha AK. Characteristics of hospitals receiving penalties under the Hospital Readmissions Reduction Program. JAMA. 2013;309(4):342-343. doi:10.1001/jama.2012.94856

6. Wadhera RK, Joynt Maddox KE, Wasfy JH, Haneuse S, Shen C, Yeh RW. Association of the Hospital Readmissions Reduction Program with mortality among Medicare beneficiaries hospitalized for heart failure, acute myocardial infarction, and pneumonia. JAMA. 2018;320(24):2542-2552. doi:10.1001/jama.2018.19232

7. Huckfeldt P, Escarce J, Sood N, Yang Z, Popescu I, Nuckols T. Thirty-day postdischarge mortality among Black and White patients 65 years and older in the Medicare Hospital Readmissions Reduction Program. JAMA Netw Open. 2019;2(3):e190634. doi:10.1001/jamanetworkopen.2019.0634

8. Ruppar TM, Cooper PS, Mehr DR, Delgado JM, Dunbar-Jacob JM. Medication adherence interventions improve heart failure mortality and readmission rates: systematic review and meta-analysis of controlled trials. J Am Heart Assoc. 2016;5(6):e002606. doi:10.1161/JAHA.115.002606

9. Lindenauer PK, Lagu T, Rothberg MB, et al. Income inequality and 30 day outcomes after acute myocardial infarction, heart failure, and pneumonia: retrospective cohort study. BMJ. 2013;346:f521. doi:10.1136/bmj.f521

10. Struthers AD, Anderson G, Donnan PT, MacDonald T. Social deprivation increases cardiac hospitalisations in chronic heart failure independent of disease severity and diuretic non-adherence. Heart. 2000;83(1):12-16. doi:10.1136/heart.83.1.12

11. Steverson AB, Marano PJ, Chen C, et al. Predictors of all-cause 30-day readmissions in patients with heart failure at an urban safety net hospital: the importance of social determinants of health and mental health. Am J Med Open. 2023;10:100060. doi:10.1016/j.ajmo.2023.100060

12. Marano PJ, Steverson AB, Chen C, et al. Effect of a novel, evidence-based, standardized discharge checklist on 30-day all-cause readmissions in patients hospitalized for heart failure in an urban safety net hospital. Am J Cardiol. 2022;182:40-45. doi:10.1016/j.amjcard.2022.06.058

13. Martin M, Snyder HR, Coffa D, et al. Time to ACT: launching an Addiction Care Team (ACT) in an urban safety-net health system. BMJ Open Qual. 2021;10(1):e001111. doi:10.1136/bmjoq-2020-001111

14. Chase J, Bilinski J, Kanzaria HK. Caring for emergency department patients with complex medical, behavioral health, and social needs. JAMA. 2020;324(24):2550-2551. doi:10.1001/jama.2020.17017

15. Ahmad T, Desai NR, Yamamoto Y, et al. Alerting clinicians to 1-year mortality risk in patients hospitalized with heart failure: the REVEAL-HF randomized clinical trial. JAMA Cardiol. 2022;7(9):905-912. doi:10.1001/jamacardio.2022.2496

16. Vizient Inc. Accessed July 31, 2024. https://www.vizientinc.com/

17. Directed Payments Quality Incentive Pool program. California Department of Health Care Services. Accessed October 20, 2024. https://www.dhcs.ca.gov/services/Pages/DP-DPH-QIP.aspx

18. NCI Comorbidity Index Overview. National Cancer Institute. Updated April 19, 2024. Accessed July 31, 2024. https://healthcaredelivery.cancer.gov/seermedicare/considerations/comorbidity.html

19. Gwadry-Sridhar FH, Flintoft V, Lee DS, Lee H, Guyatt GH. A systematic review and meta-analysis of studies comparing readmission rates and mortality rates in patients with heart failure. Arch Intern Med. 2004;164(21):2315-2320. doi:10.1001/archinte.164.21.2315

20. Gorodeski EZ, Starling RC, Blackstone EH. Are all readmissions bad readmissions? N Engl J Med. 2010;363(3):297-298. doi:10.1056/NEJMc1001882

21. Vlasnik JJ, Aliotta SL, DeLor B. Medication adherence: factors influencing compliance with prescribed medication plans. Case Manager. 2005;16(2):47-51. doi:10.1016/j.casemgr.2005.01.009

22. Kushel MB, Vittinghoff E, Haas JS. Factors associated with the health care utilization of homeless persons. JAMA. 2001;285(2):200-206. doi:10.1001/jama.285.2.200