This study examined data from 497 adults referred to Durham Homeless Care Transitions (DHCT), a transitional care and medical respite program for PEH, from July 1, 2016 through June 30, 2020. DHCT inclusion criteria were adults who were (a) experiencing homelessness; (b) able to participate in and maintain a safe and harm-free environment; (c) willing to follow rules of housing setting, which may include abstinence from substances and alcohol; (d) willing to participate in case management visits and treatment plans; (e) competent in activities of daily living (self-toileting, simple meal preparation); (f) psychiatrically stable (e.g., no active threats of harm to self or others); and (g) cleared by physical therapy for home discharge when applicable. Data were collected and analysis was conducted between March 12, 2020 and May 27, 2021. The Duke Health Institutional Review Board approved all study procedures (Pro00103699).

We extracted electronic health record (EHR) data from January 1, 2014, through June 30, 2020. We retrieved demographic data (i.e., sex, age, race, referral source, insurance), medical diagnoses (e.g., cardiovascular, chronic pulmonary, other end organ disease, diabetes), substance use diagnoses (e.g., alcohol use disorder, opioid use disorder, stimulant use disorder), and mental illness diagnoses (e.g., psychotic disorder, mood disorder, anxiety/trauma related disorders) according to ICD, Tenth Revision (ICD-10) codes (19–21) (see Supplementary 1). We chose these 10 diagnoses (risk factors) based on the high prevalence of such disorders amongst PEH (22, 23) and consensus of three individuals with expertise in providing medical and psychiatric care for PEH. We obtained encounter information (i.e., admissions, length of stay, ED visits, and outpatient visits within 12 months after program referral) and clinical outcomes (i.e., mortality and drug overdose at 12 months after program referral).

LCA statistical testing of model fit and class membership is probabilistic, with latent group membership probabilities computed for each subject from the set of candidate risk factors. For this LCA, the presence or absence of 10 risk factors at the referral date was used to define the latent groups: (a) substance use disorders (i.e., alcohol, opioid, and stimulants) (b) mental illness (i.e., psychotic disorder, anxiety/trauma, mood disorder), and (c) medical conditions (i.e., cardiovascular, diabetes, end organ, pulmonary). Following standard practice, we began with a one class model and then added classes in successive models (24). We used the Bayesian Information Criterion (BIC) to determine the optimal number of classes, as previous literature has shown BIC to be superior to other criteria at correctly identifying the number of groups in LCA processing, particularly when using categorical outcomes (24). In addition, we evaluated the resulting models for interpretability using clinical judgment (25). Each participant was assigned to their modal class based on the computed membership probabilities.

After defining the latent classes, our next analytical objective was to determine how the set of classifying variables (presence/absence of 10 risk factors) and demographic variables varied by latent class group using logistic regression or ordinary least square regression, as appropriate. To prevent overtesting, the utilization outcomes were evaluated using a referent latent group that was selected and compared to each of the other latent groups using Poisson regression.

The LCA was performed using the PROC LCA procedure in SAS v9.4 (26), which provides likelihood-based information indices to aid in assessing the number of latent classes needed to fit the data. To validate class selection, using simple random sampling, the 497 subjects were separated into two independent groups and the PROC LCA was conducted separately on each group. LCA for each sample resulted in an optimal 5-class solution. Zero-inflated Poisson (ZIP) regression was used to model the association of LCA group membership with the outcomes of medical utilization. Number of hospital admissions, emergency department encounters, and outpatient encounters are counts in nature, and the length of stay of admissions fit the Poisson distribution. The abundance of zero counts necessitated the use of the ZIP method. The occurrence of mortality and drug overdose were modeled using logistic regression. For both sets of analyses, the LCA classes were treated as nominal with “Low morbidity” as the referent group. The analyses were conducted using SAS version 9.4 (SAS Institute, Inc., Cary, NC).

A 5-group model, depicted in Figure 1, was chosen based on the optimal number of latent classes by comparing the BIC of the candidate models (25). The BIC of the five-class solution was 784.6 and the entropy was 0.87 (Supplementary 2). Figure 1 indicates the percent of subjects in each group with each of the 10 pre-specified medical, psychiatric and substance use conditions. The x-axis represents the 10 medical, psychiatric and substance use conditions and the y-axis represents the percent of subjects in each group with that condition. In Group 1, prevalence of medical, psychiatric and substance use conditions was relatively low compared to the four other groups. Group 2 had the highest prevalence of chronic pulmonary disease and anxiety disorders, and high prevalence of cardiovascular, end organ disease, diabetes, and mood disorders despite relatively lower prevalence of substance use disorders. Group 3 had the highest prevalence of end organ disease, opioid and stimulant use disorders, and mood disorders as well as a relatively high prevalence of all other conditions. Group 4 had the highest prevalence of alcohol use and psychotic disorders, and Group 5 had the highest prevalence of cardiovascular disease and diabetes.

Based on these findings, we named Group 1 [the referent group; n = 100 (20.1%)] “Low morbidity,” Group 2 [n = 102 (20.5%)] “high comorbidity,” Group 3 [n = 127 (25.6%)] “high tri-morbidity,” Group 4 [n = 56 (11.3%)] “high alcohol use,” and Group 5 [n = 112 (22.5%)] “high medical illness.”

Of the 497 patients included in this study, the mean age was 50.5 (±11.0) years, 68.6% were male, and 61.1% were Black. A total of 36.8% of patients were uninsured, 40.6% had public insurance, and 2.4% had private insurance. For the entire cohort, the prevalence of (a) cardiovascular disease was 75.9%, (b) chronic pulmonary disease was 36.2%, (c) other end organ disease was 36.4%, and (d) diabetes was 32.4%. Alcohol use disorder was present in 38.8% of the cohort, and opioid use disorder and stimulant use disorder were present in 17.7 and 37.6% of the cohort, respectively. Finally, 13.9% of the cohort had a psychotic disorder, 50.9% had a mood disorder, and 43.5% had an anxiety disorder. Socio-economic and clinical characteristics varied across LCA groups. See Table 1 for complete demographic and clinical details. Supplementary 3 includes odds ratios and confidence intervals for all risk factors across each class.

Groups 2–5 had more admissions, longer duration of admissions, and more ED visits in the 12 months following referral compared to the “low morbidity” group. All groups except the “high alcohol use” group had more outpatient visits in the 12 months after referral compared to the “low morbidity” group. In comparison with the “low morbidity” group, the “high tri-morbidity” group had the largest magnitude increase in admissions (RR, 1.93; 95% CI, 1.32–2.81; p = 0.0007) and in ED visits (RR, 3.58; 95% CI, 2.79–4.58; p < 0.0001), and the “high comorbidity” group had the largest magnitude increase in mean length of admissions (RR, 2.11; 95% CI, 1.86–2.39; p < 0.0001) and in number of outpatient visits (RR, 1.39; 95% CI, 1.26–1.52, p < 0.0001). Complete health care utilization data is presented in Table 2.

Only the group with “high medical illness” had a greater mortality 12 months after referral compared to the “low morbidity” group (OR, 2.53, 1.03–6.16; 95% CI, 1.03–6.16; p = 0.04). Both the “high comorbidity” group (OR, 5.23; 95% CI, 1.57–17.39; p = 0.007) and “high tri-morbidity” group (OR, 4.20; 95% CI, 1.26–14.01; p = 0.02) had a greater 12-month drug overdose risk after referral compared to the referent group. Complete clinical outcome data are present in Table 3.