Differential Effects of an ED-to-Home Care Transitions Intervention in an Older Adult Population: A Latent Class Analysis

Abstract

Background:

Older adults frequently return to the emergency department (ED) within 30 days of a visit. High-risk patients can differentially benefit from transitional care interventions. Latent class analysis (LCA) is a model-based method used to segment the population and test intervention effects by subgroup.

Objectives:

We aimed to identify latent classes within an older adult population from a randomized controlled trial evaluating the effectiveness of an ED-to-home transitional care program and test whether class membership modified the intervention effect.

Research Design:

Participants were randomized to receive the Care Transitions Intervention or usual care. Study staff collected outcomes data through medical record reviews and surveys. We performed LCA and logistic regression to evaluate differential effects of the intervention by class membership.

Subjects:

Participants were ED patients (age≥60 years) discharged to a community residence.

Measures:

Indicator variables for the LCA included clinically available and patient-reported data from the initial ED visit. Our primary outcome was ED revisits within 30-days. Secondary outcomes included ED revisits within 14-days, outpatient follow-up within 7- and 30-days, and self-management behaviors.

Results:

We interpreted 6 latent classes in this study population. Classes 1, 4, 5, and 6 showed a reduction in ED revisit rates with the intervention; classes 2 and 3 showed an increase in ED revisit rates. In Class 5, we found evidence that the intervention increased outpatient follow-up within 7 and 30 days (OR: 1.81, 95%CI: 1.13–2.91; OR: 2.24, 95%CI: 1.25–4.03).

Conclusions:

Class membership modified the intervention effect. Population segmentation is an important step in evaluating a transitional care intervention.

Introduction

Older adults (age ≥65), estimated to comprise 21% of the US population by 2030 (1), are a vulnerable population which requires unique healthcare resources to meet their needs (2, 3). They are medically more complex than their younger counterparts, resulting in increased hospital and emergency department (ED) care utilization (3, 4). After ED visits, older adults, especially those with advanced age and frailty, are at increased risk for adverse events (e.g., delirium, readmission) (5, 6).

Approximately 20% of older adults will have a repeat ED visit in the 30 days following an index visit (7, 8). One target to reduce ED readmissions and improve outcomes for these patients is the transition of care (i.e., moving from one area of care to another, such as the ED to home) (8–10). Sub-optimal care transitions lead to poor continuity of care, fragmentation of services, and increased costs, and are a prime target to improve care for high-risk patients (8, 11, 12). Previously studied ED-based transitional care interventions have shown minimal benefit at reducing subsequent ED use (13, 14). These programs have targeted unselected, heterogeneous populations with uniform interventions, a construct poorly aligned with many effective interventions that improve care in older adults. Targeting a program to best align with populations of greatest potential impact could be a path towards improved ED-based transitional care intervention performance (15, 16).

Accordingly, there are increasing calls to focus care coordination interventions on high-risk patients, which have been shown to differentially benefit from transitional care interventions (15, 17). Correctly segmenting the broader patient population to deliver targeted interventions is particularly important in determining the success or failure of new healthcare delivery interventions to inform subsequent policy and best practices to support precision healthcare (18). Latent class analysis (LCA) is a model-based method for breaking down a heterogeneous population into more homogenous classes with respect to a set of measured variables. LCA has been widely used to segment patient populations, including patients deemed to have a high risk of ED readmission (19–24). Studies looking at high use, high cost, and medically complex patients have found distinct classes that may benefit from different interventions (21, 22, 25).

This study aims to identify latent classes within an older adult population from a randomized controlled trial testing the effectiveness of an ED-to-home transitional care program. We use a combination of both clinically available data (e.g., comorbidities), and patient-reported data (e.g., perceived physical health) collected at the baseline study visit, and test whether the effect of the intervention is moderated by class membership.

Methods

Setting

We conducted a single-blind randomized controlled trial to examine the effectiveness of an ED-to-home care transitions intervention (CTI) among community-dwelling older adults (clinicaltrials.gov registration NCT02520661) at three university-affiliated hospital EDs in Madison, WI, and Rochester, NY. The study was approved by institutional review boards at the University of Wisconsin and University of Rochester with written informed consent. Enrollment and data collection occurred between January 2016 and July 2019.

Participants

Eligible subjects were at least 60 years of age; resided in either Dane County, WI, or Monroe County, NY; had a primary care provider affiliated with either health system; had a working telephone; and were discharged from the ED to a community residence within 24 hours of arrival. Subjects were excluded if they did not speak English, were visually or hearing impaired without correction, did not have a permanent residence, were actively enrolled in hospice or a transitions/care management program, presented with a primary behavioral or psychiatric problem, or had an Emergency Severity Index category of 1. Previous participants were excluded.

Study procedures

Study protocol details have been previously published (26, 27). In brief, research coordinators identified and obtained consent from eligible ED patients. Legally authorized representatives (LARs) could provide consent for patients without decisional capacity. Once consented, participants were randomly assigned to either the control (usual care) or the treatment (intervention) group based on within-site block randomization strategies. Research coordinators were blinded until after consent, when the participant’s group assignment was revealed. Baseline survey instruments were verbally administered to all participants prior to ED discharge, assessing self-reported health status, demographic characteristics, cognitive status, and psychosocial measures. LARs could assist in completing most of these measures. For participants randomized to the intervention, a community paramedic coach delivered transitional care content from Coleman’s Care Transitions Intervention (28) at the participant’s home within 48–72h of discharge from the ED. Follow-up phone calls occurred approximately 4 and 30 days after ED discharge to collect patient-reported outcomes, including healthcare utilization, medication adherence, and recollection of “red flags” (signs of a worsening condition) (26). For participants who received the intervention, we asked questions regarding perceived value (in dollars) and likelihood of choosing a hospital that offered the coaching service. We abstracted data from medical records, using best practices for unbiased chart review (29).

Outcomes

Our primary outcome was ED revisits within 30 days of index visit, which is typically used as a quality indicator (11, 30). We included all unplanned ED use, regardless of reason, during the 30 days following discharge from the ED. We examined ED revisits within 14 days to obtain a more nuanced understanding of short-term ED use patterns in this population (11, 30).

Secondary outcomes were outpatient follow-up and patient self-management behaviors, including red flag knowledge (i.e., signs of a worsening condition) and medication adherence. Outpatient follow-up included office visits with primary or specialty providers, telephone calls, and online patient portal messaging (excluding automated reminders, electronic messages without a patient response, laboratory testing, and previously-scheduled outpatient procedures). We examined outpatient follow-up within 7 and 30 days of discharge, allowing us to evaluate short-term follow-up after discharge, which has been shown to improve patient care. All healthcare utilization variables were dichotomized to represent whether any visits or contacts occurred during the specified timeframe.

Red flag knowledge was captured during patient follow-up calls approximately four days after ED discharge. Red flags are clinical signs and symptoms (e.g., dizziness, skin redness) that should lead a patient to seek immediate medical care. We asked patients to report any red flags provided at discharge during the follow-up call. We classified anyone who could correctly recall at least one specific red flag from their discharge instructions in a dichotomous variable, and only participants with specific red flags listed on their discharge instructions were included in this analysis.

Medication adherence was also collected during the four-day follow-up call. We asked participants to self-report any medication changes (stops, starts, or modifications) they had made since discharge. We constructed a dichotomous variable to represent patients who reported making all medication changes as prescribed by the treating clinician at discharge and only included participants who had medication changes on their discharge instructions in this analysis.

Analysis

We conducted an LCA using categorical and continuous measured variables from the entire patient sample recruited for the main study (N = 1,756). (Note that the term LCA technically refers to a finite mixture model with only categorical measured variables; however, we use the term here for simplicity.) Variable selection for the LCA was conducted prior to model fitting. We conducted basic univariate analyses to check for outliers and assess missingness. If a variable had >20% of values missing, it was dropped from analysis. All remaining missing data were assumed to be missing at random and imputed by predictive mean matching using the mice (31) package in R, unless already imputed for previous analyses based on best practices in the literature. Two variables (“own your home” and “red flags on discharge”) had 6% and 10.2% missingness, respectively; all other variables had less than 4% missingness. For each variable with missing data, predictive mean matching fit a regression model to obtain predicted values, imputing data by selecting a donor value from the observed values for the same variable with the nearest predicted value. The distributions for imputed variables were plotted and compared to the corresponding original distributions and found to be similar.

We conducted the LCA using the depmixS4 package in R (depmixS4, version 1.5–0; R Statistical Software, version 4.1) (32). This package was chosen for its flexibility to incorporate continuous, multinomial, and binary measured variables. Table 1 lists the 32 measured variables used in the model along with their definitions and modeled distributions. Appropriate link functions were selected based on variable type, with identity used for continuous/multinomial and logit used for binary. All variables were derived from data either in the patient’s medical record prior to or patient-reported baseline data collected during the index ED visit. Some continuous and categorical variables were recoded as binary due to model convergence issues. We fitted models with an increasing number of latent classes until the Bayesian information criteria (BIC) stopped decreasing. We performed repeated model fitting at each number of classes using different random starts to reach the global maximum. To select a model, we considered the BIC, class size, and class interpretability, specifically evaluating the clinical interpretability of each option. Classification quality was measured using a normalized entropy index. See table, Supplemental Digital Content 1, for model fit information of all candidate models.

Table 1:

Measured Variables and Distributions

Indicator	Distribution	Definition
Marital Status	Binomial	Dichotomous indicator of married or not married
Education	Binomial	Level of education obtained, dichotomized to College or above and Some college or less
Insurance	Binomial	Type of insurance held, either binary Medicare/Tricare/Medicaid or Private/Workers Comp
Own Home	Binomial	Yes or No to indicate whether the subject owns their own home
Residence	Binomial	Where the patient lives, either Independent house, apartment, or condo or Independent retirement community
Live With	Binomial	Dichotomous indicator of whether the subject lives alone or with friends/family
Receives Help	Binomial	Whether the subject receives help with healthcare needs
Activities of Daily Living1	Multinomial	Number of limitations in Activities of Daily Living
Charlson	Gaussian	Number of Charlson comorbidities listed in medical record
Cognitive Impairment2	Binomial	Dichotomous indicator of cognitive impairment, defined by Blessed Orientation-Memory-Concentration score >10
Health Literacy3	Binomial	Adequate or Inadequate per rating of confidence in filling out medical forms (Extremely/Quite a Bit = Adequate, Somewhat – Not at all = Inadequate)
Anxiety4	Binomial	Dichotomous indicator of anxiety, defined by Generalized Anxiety Disorder-2 score ≥3
Depression5	Binomial	Dichotomous indicator of depression, defined by Patient Health Questionnaire-9 score ≥10
Aggregate Mental Score6	Gaussian	Total score of mental health subscale on Short Form-12
Aggregate Physical Score6	Gaussian	Total score of physical health subscale on Short Form-12
Drinks per Week	Gaussian	Number of alcoholic drinks typically consumed per week
Perceived Health Competence Scale7	Gaussian	Total score on Perceived Health Competence Scale
Red Flags	Gaussian	Number of red flags listed on visit summary instructions at ED discharge
Medication Instructions	Multinomial	Number of medication instructions included on visit summary instructions at ED discharge
Follow-Up Instructions	Multinomial	Number of explicit follow-up instructions included on visit summary instructions at ED discharge
Outpatient In-Person Visits	Binomial	In-person visits to any provider (excluding PCP) in the month prior to index ED visit
PCP In-Person Visits	Binomial	In-person visits to subject’s PCP in the month prior to index ED visit
Outpatient e-Communication	Binomial	Telephone or MyChart communication with any provider (excluding PCP) in the month prior to index ED visit
PCP e-Communication	Binomial	Telephone or MyChart communication with the subject’s PCP in the month prior to index ED visit
ED Visits	Binomial	ED visits in the month prior to the index ED visit
Hospitalizations	Binomial	Hospitalizations in the month prior to index ED visit
Urgent Care	Binomial	Urgent Care visits in the month prior to index ED visit
PCP Time	Binomial	Years subject has been seeing their PCP (<5 or 5+)
Count on Team	Binomial	Dichotomous indicator of whether the subject felt they could count on their healthcare team Hardly at all/Somewhat/Moderately or A lot/Completely
Feel Known	Binomial	Dichotomous indicator of whether the subject felt known by their healthcare team Hardly at all/Somewhat/Moderately or A lot/Completely
Healthcare Organization	Binomial	Dichotomous indicator of whether the subject felt their healthcare was Hardly at all/Somewhat/Moderately or A lot/Completely organized
MyChart	Binomial	Dichotomous indicator of whether the subject’s MyChart was active

Footnote: Scale References

^1.Katz S, Downs TD, Cash HR, Grotz RC. Progress in Development of the Index of ADL. The Gerontologist. 1970;10(1_Part_1):20–30. doi:10.1093/geront/10.1_Part_1.20

^2.Katzman R, Brown T, Fuld P, Peck A, Schechter R, Schimmel H. Validation of a short Orientation-Memory-Concentration Test of cognitive impairment. The American journal of psychiatry. 1983 Jun 1983;140(6)doi:10.1176/ajp.140.6.734

^3.Wynia MK, Osborn CY. Health Literacy and Communication Quality in Health Care Organizations. J Health Commun. 2010;15(2):102–115. doi:https://dx.doi.org/10.1080%2F10810730.2010.499981

^4.Wild B, Eckl A, Wolfgang, Niehoff D, et al. Assessing generalized anxiety disorder in elderly people using the GAD-7 and GAD-2 scales: results of a validation study. The American journal of geriatric psychiatry. 2014 Oct 2014;22(10)doi:10.1016/j.jagp.2013.01.076

^5.Manea L, Gilbody S, McMillan D. Optimal cut-off score for diagnosing depression with the Patient Health Questionnaire (PHQ-9): a meta-analysis. CMAJ : Canadian Medical Association Journal. 02/21/2012 2012;184(3)doi:10.1503/cmaj.110829

^6.Ware J, Kosinski M, Keller S. A 12-Item Short-Form Health Survey: Construction of Scales and Preliminary Tests of Reliability and Validity. Medical Care. 1996;34(3):220–233. doi:10.1097/00005650-199603000-00003

^7.Smith MS, Wallston KA, Smith CA. The development and validation of the Perceived Health Competence Scale. Health Education Research. 1995;10(1):51–64. doi:10.1093/her/10.1.510

Once the number of classes was selected, we assigned each participant to their most likely class. Classes were further characterized by socio-demographic variables not used in the LCA. Within each class, we assessed the effect of the intervention on each outcome. We used the 3-step Bolck-Croon-Hagenaars approach to correct for bias due to misclassification of individuals into classes and calculated standard errors based on 10,000 nonparametric bootstrap resamples (33, 34). All analyses were conducted using R statistical software.

Results

Latent Class Analysis

The 7-class solution had the lowest BIC; however, we found that interpretability did not improve, and clinical relevance diminished beyond the 6-class solution. The 7-class solution did not offer any new insights to the groupings and simply split an already homogenous class in two. Therefore, we decided to use the 6-class model, which had a BIC of 76291.26 and entropy of 0.83. A sensitivity analysis of the starting parameter values yielded a different locally optimal solution with moderately lower BIC (76284.32); classes were largely similar with a Cohen’s kappa of 0.74 and outcome analyses were mostly consistent (see additional detail with corresponding tables and figure in Supplemental Digital Content 2, 3, 4, and 5). Broadly, Classes 1–3 had worse than average self-reported mental and physical health; Class 4 had worse than average self-reported physical health but better than average self-reported mental health; and Classes 5 and 6 had better than average self-reported mental and physical health. Classes were further differentiated by prior healthcare use, satisfaction with their care team, and sociodemographic factors. Table 2 presents variable characteristics by class and Table 3 details demographic information that was not used in the model. Among the classes, we see differences of age and gender; additional details are given below.

Table 2:

Measured Variable Characteristics Used in LCA by Class

	Overall	Class 1	Class 2	Class 3	Class 4	Class 5	Class 6
N	1756	324	117	212	360	442	301
Marital Status = Not Married (%) ***	40.8	57.8	23.3	81.6	2.2	5.0	99.0
Education = Some College or Lower (%) ***	39.5	46.8	37.1	60.4	34.3	27.5	41.9
Insurance = Medicare/Medicaid (%) ***	61.6	67.7	58.6	81.1	62.4	48.3	61.1
Own Your Home = No (%) ***	22.5	33.2	25.0	56.6	6.4	4.5	31.6
Live in = Independent Retirement Community (%) ***	94.7	92.9	93.1	88.7	96.4	99.1	93.0
Live Alone = Yes (%) ***	31.8	41.5	15.5	65.6	0.0	0.9	87.4
Help w/Healthcare Needs = Yes (%) ***	45.4	58.8	53.4	48.1	57.1	36.6	25.2
Activities of Daily Living = 1+ Limitations (%) ***	36.3	60.9	38.8	69.8	48.2	4.1	18.3
Comorbidities (mean (SD)) ***	2.72 (1.67)	3.18 (1.54)	3.53 (1.69)	3.97 (1.82)	2.91 (1.53)	1.62 (1.22)	2.44 (1.40)
Dementia = Yes (%) ***	4.7	7.1	9.5	9.4	5.0	0.5	2.7
Health Literacy = Inadequate (%) ***	12.7	20.3	15.5	26.4	13.9	2.3	7.6
Generalized Anxiety Disorder-2 = Anxiety (%) ***	17.0	37.5	18.1	37.3	8.4	5.4	7.6
Patient Health Questionnaire-9 = Moderate to Severe Depression (%) ***	11.2	32.3	12.9	31.6	2.2	0.0	0.3
Short Form-12: Aggregate Mental Score (mean (SD)) ***	53.63 (9.13)	46.97 (11.59)	50.90 (11.04)	49.01 (10.14)	56.72 (6.66)	56.90 (4.78)	56.64 (5.91)
Short Form-12: Aggregate Physical Score (mean (SD)) ***	42.91 (11.65)	35.77 (11.09)	41.07 (11.43)	32.84 (10.49)	38.81 (10.24)	53.05 (4.30)	48.42 (7.17)
Alcoholic Drinks per Week (mean (SD)) ***	8.58 (15.13)	1.16 (0.70)	51.54 (31.59)	11.37 (8.60)	5.82 (4.66)	5.71 (5.29)	5.61 (5.28)
Perceived Health Competence Scale Sum (mean (SD)) ***	30.08 (5.29)	26.18 (4.64)	27.90 (6.47)	26.36 (4.58)	29.52 (4.44)	33.88 (3.42)	32.80 (3.46)
Red Flags on Discharge Summary (mean (SD)) **	5.39 (4.55)	5.24 (4.10)	4.03 (3.50)	6.37 (5.38)	5.36 (4.30)	5.29 (4.69)	5.53 (4.66)
Medication Instructions on Discharge Summary (mean (SD))	0.94 (1.22)	0.87 (1.22)	1.03 (1.30)	0.94 (1.37)	0.90 (1.18)	0.95 (1.23)	1.00 (1.13)
Follow-up Instructions on Discharge Summary (mean (SD))	1.01 (0.63)	1.01 (0.67)	1.13 (0.70)	0.95 (0.56)	1.05 (0.69)	0.97 (0.56)	1.03 (0.64)
Prior In-person Outpatient Visits = 1+ (%) ***	36.2	42.5	34.5	53.8	50.1	19.2	26.2
Prior In-person PCP Visits = 1+ (%) ***	33.7	39.4	44.0	47.6	39.0	23.0	23.3
Prior Outpatient e-Communication = 1+ (%) ***	29.4	40.9	31.0	46.2	40.1	9.9	20.6
Prior PCP e-Communication = 1+ (%) ***	50.7	58.2	53.4	64.6	52.4	38.8	47.2
Prior ED Visits = 1+ (%) ***	9.4	12.3	7.8	22.2	14.2	1.4	4.0
Prior Hospitalizations = 1+ (%) ***	5.1	6.8	1.7	10.4	11.1	0.0	1.3
Prior Urgent Care Visits = 1+ (%)	8.9	10.2	9.5	9.9	5.8	10.2	8.6
PCP Time = ≤4 years (%) ***	38.4	44.0	39.7	50.5	32.9	35.2	34.6
Count on Care Team = Hardly at all - Moderately (%) ***	14.0	38.2	0.9	9.4	8.6	7.9	11.3
Feel Known by Healthcare Team = Hardly at all - Moderately (%) ***	44.6	59.4	23.3	41.5	38.4	44.9	45.8
Feel Healthcare is Organized = Hardly at all - Moderately (%) ***	20.5	42.2	13.8	36.8	14.2	8.1	14.0
MyChart = Not Active (%) *	58.5	60.3	59.5	68.9	56.8	53.5	58.1

Table 3:

Class Demographic and Intervention Characteristics

	Overall	Class 1	Class 2	Class 3	Class 4	Class 5	Class 6
N	1756	325	116	212	360	443	301
Age (mean (SD)) ***	72.39 (8.58)	73.06 (9.03)	72.84 (9.19)	73.46 (8.88)	72.93 (8.25)	69.52 (6.91)	74.30 (9.22)
Gender = Male (%) ***	46.6	36.9	48.3	33.5	59.3	58.0	33.6
Race (%)
Asian	1.0	1.2	0.9	0.5	0.6	1.6	1.0
Black	4.4	8.4	2.6	6.7	3.4	1.4	4.7
Native	0.2	0.3	0.9	0.0	0.3	0.0	0.3
Other	1.6	1.6	1.7	2.4	1.4	1.8	1.0
White	92.8	88.5	94.0	90.4	94.4	95.2	93.0
Ethnicity = Hispanic (%)	1.6	1.6	0.9	3.9	1.1	1.4	1.0
Intervention Characteristics (Randomized to Intervention Only)
N	863	156	54	107	188	204	154
Received home visit (%) ***	84.1	82.7	70.4	74.8	86.2	89.2	87.7
Received one coaching call (%)	91.4	90.2	85.0	88.1	91.5	93.5	93.6
Coaching service value ($) (mean [SD])	71.33 (164.33)	54.56 (78.76)	59.09 (106.22)	95.27 (316.96)	76.66 (134.94)	63.84 (141.90)	79.45 (170.26)
Likelihood of choosing a hospital with the coaching service (%)
Extremely likely	42.1	43.1	44.8	54.0	39.4	36.4	45.0
Likely	45.9	45.1	44.8	30.2	51.2	51.3	42.2
Unlikely	8.2	10.8	6.9	11.1	4.7	9.1	7.3
Extremely unlikely	3.8	1.0	3.4	4.8	4.7	3.2	5.5

^***p≤0.001

Class 1: These patients on average have poorer self-rated mental and physical health and higher scores on depression and anxiety inventories. They are more likely to have limitations in their Activities of Daily Living (ADLs). On average, they have lower scores on measures of perceived health competency and satisfaction with their healthcare team. These patients also have above average healthcare utilization.

Class 2: These patients on average have higher scores on depression and anxiety inventories in addition to higher-than-average alcohol use and comorbidity counts. They’re likely to live with someone, be married and probably have help with their healthcare needs. On average, they have lower scores on measures of perceived health competency but higher scores on measures of satisfaction with their healthcare team.

Class 3: These patients are less educated than the other groups (probability of having a college degree or more <0.5) and are not likely to own their home or live with anyone. On average, they have lower scores on measures of perceived mental and physical health with higher-than-average comorbidity counts and limitations in ADLs. These patients also have higher than average healthcare use, especially ED visits, and are not likely to engage with their healthcare team through a patient health portal (e.g., MyChart).

Class 4: On average, these patients have better scores on measures of self-rated mental health and low scores on depression and anxiety inventories. They have lower than average self-rated physical health with an average comorbidity count. They are likely to live with someone and be married. They also have higher than average healthcare utilization, including ED visits and hospitalizations, with higher-than-average outpatient visits.

Class 5 and 6 are similar. On average, they have better scores on measures of self-rated mental and physical health, low scores on depression and anxiety inventories, and have fewer than average comorbidities. They are likely to report higher education and health literacy and own their home. They have average healthcare utilization, high satisfaction with their healthcare team, and are not likely to have help with their healthcare needs (probably because they don’t need it). Participants in Class 5 are highly likely to live with someone and be married whereas participants in Class 6 are highly unlikely to live with someone or be married.

Variation of Treatment Effect across Classes

Among those randomized to the CTI, there was a significant difference between classes in those who received the coaching home visit—the highest being almost 90% of participants in Class 5 compared to the lowest (70.4%) in Class 2 (Table 3). Classes 2 and 3, whose ED use increased for those assigned to the CTI, were less likely to receive a home visit, resulting in a lower intensity intervention compared to other classes. There were no significant differences in perceived value or likelihood of choosing a hospital with the coaching services between classes.

There was variability across classes in the point estimates of treatment differences in ED revisits within 14 and 30 days. Classes 1, 4, and 5 showed a reduction in ED visits within 14 days; Classes 2 and 3 showed an increase in ED visits within 14 days; and Class 6 showed negligible difference. At the 30-day follow-up, Class 6 showed a reduction in ED visits compared to controls; differences for other classes persisted. There was much statistical uncertainty in these estimates (Figure 1).

Figure 1: Forest Plots for Repeat ED Visits at 14 and 30 Days.

ORs and 95% confidence intervals for ED revisits within 14 days (a) and 30 days (b) of the index visit

The intervention effect on secondary outcomes also varied across classes, though there was much statistical uncertainty. In Class 5, we found a statistically significant increase in the odds of outpatient follow up within 7 days (OR: 1.81, 95%CI: 1.13–2.91) and 30 days (OR: 2.24, 95%CI: 1.25–4.03) of index ED visit (Table 4). There were no significant differences in outpatient follow-up in any other class. Neither recall of red flags nor medication adherence showed significant differences in any class.

Table 4:

Unadjusted Odds Ratios [95% Confidence Interval] for Treatment Effect by Class Membership on Self-Management Behaviors Following the Index ED Visit

	Outpatient Follow-Up (7d)	Outpatient Follow-Up (30d)	Recalled One or More Red Flags	Medication Adherence
Class 1	0.94 (0.50 – 1.76)	0.70 (0.27 – 1.82)	1.17 (0.60 – 2.26)	1.94 (0.37 – 10.06)
Class 2	0.66 (0.25 – 1.71)	0.73 (0.12 – 4.57)	0.50 (0.14 – 1.74)	0.60 (0.10 – 3.60)
Class 3	0.98 (0.46 – 2.08)	3.71 (0.71 – 19.36)	1.46 (0.57 – 3.73)	0.84 (0.20 – 3.61)
Class 4	1.36 (0.71 – 2.59)	0.82 (0.29 – 2.37)	1.72 (0.95 – 3.12)	1.28 (0.43 – 3.80)
Class 5	1.81 (1.13 – 2.91)*	2.24 (1.25 – 4.03)*	0.83 (0.50 – 1.38)	0.51 (0.20 – 1.30)
Class 6	1.10 (0.62 – 1.95)	1.1 (0.50 – 2.39)	1.58 (0.81 – 3.08)	1.00 (0.38 – 2.65)
Observations	1,756	1,756	1,207	469

^*p<0.05

Discussion

This exploratory analysis uses a data-driven LCA approach to examine differential effects of an ED-based CTI on ED revisits, outpatient follow-up, and self-management behaviors. We identified 6 distinct classes of participants, with the effect of the CTI varying across groups. There was considerable statistical uncertainty in estimates due to low healthcare use during the follow-up period. We found evidence of a reduction in ED revisit rates for some participants and an increase in outpatient clinic follow-up for others. Specifically, participants in Class 5, which was characterized by better than average self-reported physical and mental health, health literacy, and education, showed a major reduction in ED revisits within 14 days compared to their controls. The intervention also improved outpatient clinic follow-up within 30 days for participants in Classes 3 and 5.

This study adds to recent findings in the literature which highlight several groups of patients having differential responses to ED-based transitions programs. In a separate sub-analysis, we found that cognitively impaired participants who received the CTI had a significant reduction in ED revisits (35). A recent analysis of a trial evaluating a very different nurse-delivered ED-to-home transitional care program found a reduction in ED revisits for unmarried individuals and an increase in revisits for married individuals when randomized to receive the program (36). The latent classes in our study varied in marital status but we did not see a clear association between marital status and 30-day ED revisit rates. For example, most participants in Class 5 were married (95%) and few participants in Class 6 were married (1%); however, those randomized to the intervention had a lower 30-day ED revisit rate in both classes. One possible explanation beyond the difference in inclusion criteria could be that marital status needs to be considered in the context of other factors, such as health status or education level. Our analysis also diverges from the previous by considering baseline characteristics jointly rather than individually.

The variability of the effects of the intervention across latent classes supports the notion that a successful care transition might look different for different patients (13). For some patients, optimized care might be a reduction in ED revisit rates, as we found for participants in Class 5. For other patients, optimized care might result in more ED visits stemming from timely outpatient referrals where acute symptoms are identified (8, 37, 38), This is illustrated in the findings for Class 3, which is similar to the population recruited in Seidenfeld et al. (36), characterized by a higher than average number of ADLs (70% with one or more limitations), chronic conditions (average of 4), and previous ED visits (22% had at least one ED visit in the previous month). Patients in Class 3 had increased outpatient follow-up, which could be an appropriate path to meeting their needs, though further detail on the nature of these follow-up visits would be necessary to validate this hypothesis. Taken together, these studies indicate that a simple reduction in acute care utilization might not be the appropriate outcome for this type of patient following an ED visit.

For several latent classes, we did not find strong evidence of the intervention impacting either ED revisits or outpatient follow-up. In the context of Andersen’s behavioral model of health services (39), the intervention primarily targets individual health behaviors, but perhaps for some patients the primary driver of their health outcomes are predisposing and enabling contextual and individual factors. The intervention is not designed to address social determinants of health, such as home ownership and education, which play a large role in individual health outcomes (40–42). Patients in Class 3 had fewer years of education (60% did not have a college education), were more likely to be on Medicare or Medicaid (80%), were less likely to own their home (56%) and were more likely to live alone (66%). These factors in combination could have hindered the effectiveness of the intervention.

Behavioral health factors could also play a role in the effectiveness of the intervention. Patients in Class 2 were differentiated by the higher-than-average number of alcoholic drinks per week (31 drinks per week on average). About one-third of the patients in Class 1 scored poorly on measures of depression and anxiety. The ED-adapted CTI did not address underlying factors like substance abuse and other mental health issues that could have a large impact on ED revisits. In addition to their effect on utilization, these structural factors could help explain the disparity in intervention adherence across classes. Classes 2 and 3 were less likely to have received the at-home coaching visit compared to other classes, which could limit intervention effectiveness. This finding suggests a difficulty for some patients to accommodate an in-home follow-up visit, underscoring the need for different strategies to manage the ED-to-home transition and leverage a precision healthcare-based approach. It is also possible that some patients were already sufficiently performing the health behaviors targeted in the intervention, either independently or with the help of a spouse, such as in Classes 4 (having an average comorbidity count but highly likely to be married and living with someone) and 6 (generally healthy and independent despite living alone). Altogether, these findings offer a potential explanation for the overall lack of effect shown in the both the parent study (13) and the literature in general, underscoring the need for a precision-based approach to meet the needs of a heterogeneous population (18).

Limitations

By only including participants with primary care providers within each healthcare system, we likely excluded patients without access to these systems. Additionally, there was limited diversity in the study population, partially resulting from the settings, which likely excluded important patient groups from this study.

While it helped to have many variables available to identify different groups of patients, this need limits the practicability of this work. Unless sufficient information is available in the patient’s medical record or ascertained by staff during the ED encounter, many variables (e.g., depression, anxiety, and health literacy) may be unavailable for use to identify which patients would benefit from the CTI. However, developing informatics-driven approaches, such as machine-learning algorithms, can overcome this (43).

Conclusion

The LCA approach identified six distinct groups of study participants, each with different characteristics and variable outcomes from receiving the ED-adapted CTI, including some evidence of a reduction in odds of repeat ED visits in 14 days (Class 5) and increased odds of outpatient follow-up (Classes 3 and 5). This further supports the value of population segmentation when evaluating interventions to enable precision healthcare.

Supplementary Material

SDC 5 - Figure 1

Supplemental Digital Content 5, Figure 1: Forest Plots for Repeat ED Visits at 14 and 30 Days Using Alternate Solution

ORs and 95% confidence intervals for ED revisits within 14 days (a) and 30 days (b) of the index visit