Two-year change in latent classes of comorbidity among high-risk Veterans in primary care: a brief report

We conducted a repeated cross-sectional study comparing patients identified by their high-risk status in 2018 to those similarly identified in 2020. Patients were eligible in 2018 or 2020 if they were: 1) actively assigned to a VA primary care team during some portion of the year, and 2) in the top decile for risk of 1-year hospitalization based on the VA Care Assessment Needs (CAN) risk score during any week in the calendar year. The CAN risk score predicts the probability of the patient experiencing an inpatient admission to a VA hospital in the next 12 months The score is generated weekly for all active VA primary care patients with inputs that include demographic, utilization, lab, diagnoses, and medication data [16]. The top decile cut-off was chosen as it is widely used throughout the VA to determine eligibility for initiatives targeting patients with complex health care needs. Each patient’s study entry date was defined as the first day in the observed year with a top decile CAN score recorded. This work was designated non-research, and requirement for IRB approval waived, since it was carried out as a quality improvement evaluation under the terms of a signed attestation of non-research from the VHA Office of Primary Care (OPC).

Patient sociodemographic data, hospitalization dates, and outpatient utilization were obtained from the VA administrative Corporate Data Warehouse (CDW). Mortality data were obtained from VA Vital Status files, which combine information from the VA, Medicare, and Social Security. Primary care site was classified as hospital or stand-alone outpatient clinic using standard VA definitions.

Inputs for latent class models were indicators for 26 health conditions, chosen based on the U.S. Department of Health and Human Services’ multiple chronic conditions framework [17, 18]. Conditions specifically relevant to Veteran health, such as PTSD, were added, and conditions without enough sample prevalence to be included in latent class models, such as HIV and autism, were excluded. Using a 2-year lookback period from cohort entry, each chronic condition was defined as present if the patient had one outpatient or inpatient care encounter containing an ICD-10 code for that condition.

Comorbidity groups were modeled using latent class analysis (LCA) [19]. LCA is a probabilistic clustering approach that assumes patterns in response variables (e.g., medical conditions) reflect meaningful, discrete latent groups of individuals. Estimates from LCA include (1) the identification of latent classes (groups) based on patterns of comorbidity, and (2) a predicted probability of group membership for each patient per group. The predicted probability provides a measure of uncertainty for patient class assignments.

Models of 2 to 7 classes were tested among patients observed in 2018 with the final model chosen based on a combination of clinical interpretability and model fit statistics including AIC, BIC, adjusted BIC, and entropy. Patients were assigned to a latent group if their predicted probability of membership was ≥50%. If no probability was ≥50%, patients were designated as ‘unassigned,’ allowing us to account for uncertainty in class membership. This process was then repeated for patients observed in 2020.

We chose this repeated cross-sectional design as opposed to formally testing longitudinal class status with latent transition analysis [19] to represent the likely application of these models in practice. Specifically, selecting high-risk patients in 2018 and separately in 2020 allows for patients to move in and out of the high-risk population, and reflects the repeated cross-sectional nature of data from primary care encounters.

To assess the population-level stability of latent groups, we compared the prevalence of chronic conditions by latent group for patients observed in 2018 to those in 2020. To describe individual-level changes over time, we followed the 2018 patients through calendar year 2020. Patients were then categorized by their status in 2020 as 1) maintaining a top decile (e.g. “high-risk”) CAN score, 2) moving to a lower-risk CAN score, 3) death, or 4) no longer active in VA primary care. We also examined whether and how individuals’ latent group assignments changed between the two years.

We completed a supplemental descriptive comparison of patients who were assigned to the same group in both years to those who were assigned to different groups. See Supplemental Digital Content for additional methods.

Latent class analysis models were run in MPLUS version 8.2 (Los Angeles, CA: Muthén & Muthén). All other analyses were run in STATA 14 (College Station, TX: StataCorp LP).

We identified 951,771 high-risk patients in 2018 and 978,771 in 2020. Patient characteristics were similar in the two years (Table 1). Mean age was 66.1 (standard deviation = 13.1) and 66.6 (13.2) in 2018 and 2020 respectively. Most patients (92 and 91%) were male. Most patients identified as non-Hispanic White (64.2 and 63.0%) or non-Hispanic Black (22.7 and 23.4%). Over 25% had indicators of low income in both years.

In latent class models testing 2 to 7 groups, each additional group improved BIC but worsened entropy (see Table, Supplemental Digital Content 1). A 5-group model was chosen as the best balance between model fit and clinical interpretability. Similar results were found for 2020 LCA models. In 2018, 871,613 (91.6%) of patients, and in 2020 903,041 (92.3%), were assigned to a group using the cut-off value of 50% predicted probability. We labeled groups based on the high-prevalence conditions within each, relative to the overall average. Profiles of chronic conditions within groups were similar in both years (see Fig. 1). We observed a “Mental Health” group, with high prevalence of post-traumatic stress disorder, depression, and anxiety, a “Substance Use Disorders” group, with high prevalence of mental health diagnoses as well as alcohol, substance, and nicotine use disorders, a “Cardiometabolic” group, with high prevalence of coronary artery disease, congestive heart failure, and arrhythmia, a “Low Diagnosis” group, with low prevalence of most conditions, and a “High Complexity” group, with high prevalence of conditions across multiple physiologic systems. Patients assigned to the High Complexity group had an average of 11.3 (standard deviation = 2.1) chronic disease diagnoses in 2018, compared to the overall average of 6.9 (2.6). See Table, Supplemental Digital Content 2 for patient characteristics by latent group.

Between 2018 and 2020, no condition changed in its overall cohort prevalence by more than 2.5 percentage points (see Table, Supplemental Digital Content 3 for prevalence). Within groups, condition prevalence was also generally stable (Fig. 1). Exceptions include shifts in mental health, substance use disorder, and cardiovascular conditions, especially within the Cardiometabolic and High Complexity groups. For example, diagnoses of alcohol use disorder increased from 28 to 42% among those assigned to the High Complexity group, while depression diagnoses increased from 18 to 29% among those assigned to the Cardiometabolic group.

Among the 951,771 patients observed in 2018, 563,725 (59.2%) were also observed to have a high-risk CAN score in 2020. The remaining patients either died prior to 2020 (14.2%), had an improved CAN score and so were no longer designated as high risk (25.3%), or did not engage in primary care through the VA and therefore had no CAN score recorded in 2020 (1.3%, see Fig. 2). The High Complexity group was the least likely to be observed with an improved CAN score in 2020 (12.8%). In contrast, nearly one third of patients assigned in 2018 to the Substance, Mental Health, and Low Diagnosis groups were observed with an improved CAN score in 2020 and no longer qualified as high-risk by our definition.

Of the 563,725 patients with a high CAN score in both 2018 and 2020, 40.8% (n = 230,185) had a different group assignment in each year (see Table, Supplemental Digital Content 4). Patients assigned to the Cardiometabolic group in 2018 were the most likely to be observed and assigned to the same group in 2020 (76%). Patients assigned to the Substance Use Disorders group and those unassigned in 2018 were the most likely to move into the High Complexity group in 2020 (16% of each). Most high-CAN patients who did not match with a group (‘unassigned’) in 2018 were matched with a specific group in 2020 (82%).

Our descriptive comparison of patients assigned to the same versus different groups in the two years found few meaningful differences, with similar average age, income level, and number of chronic condition diagnoses (see Table, Supplemental Digital Content 5 for patient characteristics).