Study site, data sources
Healthcare in Singapore is largely under the responsibility of the Singapore MOH, which uses a mixed financing system that includes nationalized healthcare insurance schemes and deductions from the compulsory savings plan Central Provident Fund (CPF), for Singapore citizens and permanent residents (22, 23). In 2017, Singapore MOH launched the H2H program with the aim of reducing unnecessary hospital admission & utilizations (9). It involves inpatient care coordination and community care navigation by nurses through follow-up calls and home visits for high-risk patients with complex chronic disease to ensure care continuity during the transitional period after hospital discharge. These nurses are full-time, degree-holders with an average of 5 years’ experience in hospital and/or home care in adults, supervised by masters prepared Advanced Practice Nurses. Hospital readmissions and mortality were obtained at 30- and 90-days after discharge. Duration of intervention typically lasts 6 months.
We used routinely collected clinical data from the H2H program. Data included demographics (age, gender and race), medical and socioeconomic characteristics. Based on the commonly accepted age to define an older person (24), age was dichotomized into 2 groups, ≤65 years and >65 years. We included all H2H enrolled adult patients (21 years of age and above) who are Singaporean residents or permanent residents. The Singapore General Hospital (SGH) Population Health and Integrated Care Office approved the usage of collected data for this study. The Centralized Institutional Review Board (iSHaRe Ref. No. 201707-00005) approved this study for ethics. This study has been published as an abstract for the Society for Academic Primary Care 48th Annual Scientific Meeting conference (25).
Variables used in the Latent Class Analysis
Multiple variables straddling medical and socioeconomic conditions were reviewed for inclusion. Importantly, these variables were routinely collected in the program as part of patient assessment.
Medical characteristics:
Five variables related to disease state, cognition and functional status were utilised - Charlson Comorbidity Index (CCI), Abbreviated Mental Test (AMT), Clinical Frailty Score (CFS), clinical insight (26) and Activities of Daily Living (ADLs) dependency. Clinical insight was assessed by the nurses on the presence or absence of the patients’ understanding of their own medical condition. These variables and their grouped categorical scores have been validated across several countries, including locally as a good discriminative tool for predicting disease status and health outcomes (27-31).
Socioeconomic characteristics:
Four variables - Religion, medicine consolidation issues, quality of family support and employment - were utilized. Religion (or a professed faith) was assessed by the presence or absence of it. Medicine consolidation issues were assessed by the presence or absence of the 5 rights of medication administration: Right patient, drug, dose, route and time (32). Quality of family support was categorised into 4 groups: absent (patient has no kin), dysfunctional (presence of a high degree of conflict, misbehaviour, neglect and/or abuse occurring continuously and regularly), distant (presence of kin but minimal contact) and supportive. Employment was categorised into 3 groups: unemployed, retired or employed.
Latent Class Analysis
LCA is a data-driven method utilizing individual level observable data (indicator variables) to identify underlying latent groups of individuals (classes) (33). Examples of successful LCA utilization in population segmentation has been demonstrated by Low et al. in the Singapore regional health system (10) and Yan et al. in a primary care population respectively (34). In this study, 9 identified medical and socioeconomic variables were selected and described above. MPlus Version 8.2 statistical modelling software was used to perform the LCA (35).
The optimal number of classes is determined by the fit statistics and clinical interpretability. Model fit was evaluated using the Bayesian Information Criterion (BIC) and sample-size adjusted BIC (ABIC) (36). Starting with 1 class, a lower value from BIC or ABIC from each successive model, which has one more class than the prior model, indicates a better fit. Additionally, the estimated probabilities of each indicator variables within each class provide information that describes the classes and determines whether the classes are distinct from one another and clinically interpretable. Separate LCA models were generated successively from 1 through 4 class solutions. From the LCA that corresponded to the optimal number of classes identified, the posterior probability of membership for each class is computed for each subject which is assigned to the class with the maximum posterior probability.
Statistical analysis
Firstly, to examine whether significant differences between demographics and disease patterns exist across the classes, we used Fisher exact test for the categorical variables. Next, we identified potentially confounding factors through a univariate analysis of demographics against class and health outcomes. Lastly, to assess the association of class membership on hospital readmissions and mortality, we used logistic regression with Class 3 as reference. The models were adjusted for age. Analyses were performed using SAS, version 0.4 (SAS Institute, Inc., Cary, NC).