Do healthcare needs-based population segments predict outcomes among the elderly? Findings from a prospective cohort study in an urbanized low-income community

The baseline dataset for this prospective cohort study was obtained from a survey among residents of 12 public rental housing blocks in Singapore between December 2016 and March 2017. Residence in public rental housing is heavily government subsidized and is an indicator of low socio-economic status in Singapore [19] given that only financially needy households will qualify [20]. The community is contained within the SingHealth Regional Health System, one of three health clusters that serve as an organizing focus for population health services.

The Singapore Ministry of Health provided 1817 unique addresses of older adults aged 60 years and above living within the target community. Survey personnel then visited the addresses to determine the presence of eligible individuals and to conduct both the recruitment and survey. Eligible study subjects were Singaporean citizens or permanent residents aged 60 years and above who scored 7 or more on the Abbreviated Mental Test (AMT) or had a suitable proxy interviewee if they scored below 7 on the AMT. Respondents (either the study subject or their proxy) provided written consent.

The survey was administered by non-healthcare trained surveyors in the subject’s home and captured health information (specific chronic disease, physical disability through adapted Barthel Index [21] and Lawton IADL scales [22], health promoting behaviours such as medication adherence and health knowledge, health utility score based on responses to the EQ5D Index [23, 24]) and social characteristics (social network, financial status, social stressors, etc) (see Additional file 2 for survey questionnaire). Patient self-reported data for clinical diagnoses and healthcare utilization have been demonstrated to have moderate to substantial degrees of concordance with administrative claims records [25, 26].

Health service utilization and mortality was assessed over 180 days after the baseline survey data. Data were obtained from the SingHealth electronic health record through the Health Intelligence System (eHints) [27]. eHints is a single enterprise data repository that integrates information from multiple healthcare transaction systems including administration, clinical and ancillary systems. Outcomes included mortality and healthcare utilization (hospital admissions, emergency department attendance, specialist outpatient clinic (SOC) attendances).

This study was approved by the SingHealth Centralized Institutional Review Board (CIRB2016/2242).

Subjects were allocated to segments in each of the 4 schemes using a mapping algorithm that matches the original definition of specific population segments from the various alternative schemes considered (see Additional file 3). Study subjects were assigned into population segments by considering whether individual subjects’ characteristics met the pre-defined entry criteria for a specific segment and if not, they would be considered for the next segment based on an approximate highest to lowest healthcare need sequence. This ensures that subjects are assigned into population segments with the highest healthcare need for which they meet admission criteria.

As noted above, the Singapore scheme involved 6 Global Impression categories and a summary of Complicating Factors into 3 levels, denoted ‘Risk for adverse outcome and Actionability’ (RA) (Table 1). Combining 6 Global Impression Levels and 3 RA levels, the Singapore scheme comprises 18 health and social service need-based categories.

The study population was characterized using descriptive statistics on the various population segmentation schemes. Predictive validity of the various population segmentation schemes for the outcomes of mortality and healthcare utilization were plotted based on specific population segments’ mean outcome and confidence interval. Confidence intervals were computed based on normal standard errors for the outcome of SOC visit, Poisson standard errors for the outcome of ED visits and hospital admission, and binomial standard errors for the outcome of mortality. The average outcome rate for all subjects were plotted as a dashed line on the graph. Finally, in order to test the null hypothesis that the population segments have no relationship with the outcomes of mortality and healthcare utilizations, we perform Pearson Chi-squared tests between population segments and the following outcomes: any mortality, any hospital admission, any ED visit, and any SOC visit within the follow-up period of 180 days. Our alternative hypothesis is that population segments are related to outcomes of mortality and healthcare utilizations. We select a significance level of 0.05.

All analysis in this study was conducted using STATA (StataCorp. 2013. Stata Statistical Software: Release 13. College Station, TX: StataCorp LP). The report is consistent with the relevant items from the STROBE guidelines for reporting cohort studies [28] (see Additional file 4).

Of 1324 residents contacted, 928 subjects were recruited for this study (70% response rate). A proxy responded for 47 subjects (5.1%). Compared to the general Singapore population, individuals in this community tended to be older (median of 70 years old vs 41 years old for Singapore overall [29]) with low socioeconomic indicators (100% in subsidized rental flats vs 4.24% of the total population in Singapore living in one and two-room flats, the majority of which are heavily subsidized [29]). All subjects could be assigned to an exclusive segment for all segmentation schemes. A breakdown of study population segments based on our evaluated schemes is shown in Table 2. Some population segments were not represented by any subjects in this study and for clarity were not included in Table 2.

Health utility as assessed by EQ-5D Index was plotted in approximately ascending order of segment severity (Fig. 1). A total of 881 subjects (95%) had recorded data from which to calculate the EQ-5D Index scores. The mean EQ-5D index score was lower at baseline for subjects in more severe segments for each segmentation scheme. Overall, the mean EQ-5D score for all study subjects was 0.83. The population segment with the lowest mean EQ-5D index was Singapore category 12: Long course of decline with high RA with a mean EQ-5D of 0.249).

Mortality rate is shown by segment in Fig. 2. Mortality rate for the overall population was 2%. For each segmentation scheme, there was a general trend of increasing mortality as population segment healthcare need increases. The population segment with the highest mean mortality risk was Singapore category 14: Limited Reserve with high Risk and Actionability (RA), with a 180-day mean mortality rate of 1 out of 7 individuals.

Overall, the proportion of subjects with hospital admission within 180 days was 12.5% with a mean number of admissions per person in that period of 0.189 (Fig. 3). This rate tended to increase with severity of the health segment. The population segment with the largest mean number of admissions per person was Singapore category 14: Limited Reserve with high RA, with a mean number of visits of 1.14.

The overall proportion of subjects with ED visits over 180 days was 17.9% (Fig. 4). The mean number of ED visits per person in that period was 0.252 while the population segment with the highest mean ED visits Singapore category 14: Limited Reserve with high RA, with a mean number of visits of 1.57.

For all study subjects, 30.9% had SOC visits over 180 days (Fig. 5). The mean number of SOC visits for all study subjects was 0.945 visits while the population segment with the highest mean number of SOC visit was Singapore category 14: Limited Reserve with high RA, with a mean number of visits of 3.14.

Based on the test results on Table 3, we observe that there is a statistically significant relationship between all 4 population segments and the outcomes of ‘any hospital admission’, ‘any ED visit’, and ‘any SOC visit’. Meanwhile for the outcome of mortality, we only observe a statistically significant relationship with the Singapore population segmentation scheme.

While there are other studies of the practicality and predictive validity of various segmentation schemes, this study is unique in its application of the schemes on a community survey administered by non-medically trained personnel. Notably, the Lombardy [12], Delaware [11] and North-west London [13] segmentation schemes were originally developed to be applied to data sourced from electronic medical records. Also, while the Singapore scheme was developed to map groups of individuals to similar sets of health and health related social service needs, the other schemes used needs as a framework for classification, but focused on the prediction of per capita medical spending. None of the schemes, including the one developed in Singapore, has been evaluated with regard to the ability to predict responsiveness to interventions. Such demonstrated responsiveness would be the cardinal feature of an ideal segmentation scheme.

The implications of this work should be considered in the context of the broader effort to develop a practical approach to segmenting a heterogenous population into parsimonious groupings with similar need, for the purpose of meeting needs effectively and efficiently. Existing healthcare needs-based population segmentation schemes can be classified as either expert or data driven. Expert based tools [5, 15‐17, 35] have the advantage of being more clinically intuitive while data driven tools [36‐38] have the advantage of being more objective as the population segments are generated using statistical clustering methods such as latent class analysis and hierarchical clustering. Tools that utilize data collected in EHR have the additional advantage of feasibility to segment large populations without the need for additional data collection. Regardless of how they were developed, both types of tools must be practical (i.e. require feasible data collection and produce a parsimonious number of segments), have demonstrable external validity (i.e. generalizable to populations other than the one used for tool development), and show responsiveness to specific package of health services (i.e. serve as a meaningful guide to efficient resource allocation).

In this context, the 4 segmentation schemes evaluated are practical in terms of ease of data collection using a community-based survey. However, to the extent that some tools are intended to be used with an EHR which include complete inpatient and outpatient data, practical application will be limited in jurisdictions with less than comprehensive EHRs. In addition, all tools evaluated here produce a tractable number of segments, ranging from 6 to 14. Notably, the Johns Hopkins’ ACG, which was not assessed in this study, produces up to 92 segments [10].

In a more preliminary way, responsiveness can be assessed in terms of the degree to which the scheme has the potential to guide service allocation (i.e., actionability). This feature distinguishes segmentation schemes that are needs-based (i.e., designed to indicate prototypical services expected to reduce risk) from schemes that are risk-based (predict individuals likely to have poor health outcomes or high health care utilization) [39]. In this regard, the Singapore scheme is distinctive as it was specifically developed to be indicative of services expected to reduce risk of poor outcomes. Compared to the other population segmentation schemes, the Singapore scheme appears to capture the largest number of actionable health and health-related social needs (see Additional file 5 for comparison table).

This study has several strengths. First, it is pragmatic, having been performed in a community that faces practical challenges for the health system. It includes individual who would not have been assessed by use of EHR because they did not have contact with the health system holding that individuals records, either because they successfully avoided medical attention, or used services in other health venues not part of the EHR ecosystem. Second, the study provides an opportunity to evaluate the Singapore segmentation scheme relative to other schemes using a common data set.

There are several limitations to this study. Firstly, not all population segments could be evaluated fully in our analysis given that the community survey did not have all the required variables and subject types (subjects in this study were low-income elderly individuals aged 60 years and above) for fully characterizing all population segments. Specifically, 10 segments from the Delaware scheme were not evaluated as the inclusion criteria for the survey were subjects at least 60 years old, 2 segments from the Lombardy scheme were not evaluated due to a similar reason as well as the lack of utilization data spanning the previous 3 years, 3 segments from the North-West London scheme were not evaluated as the survey did not capture information on alcohol/drug dependency, learning difficulties, as well as an absence of any subjects with dementia but no physical disabilities, and finally, 4 segments from the Singapore scheme were not evaluated given that no subject with limited reserve but low RA exists in the study and the survey did not capture information on end-of-life medical conditions. Secondly, the external validity of our results may be limited by the fact that our study subjects comprise low income elderly individuals aged 60 and above. Individuals with low income may not readily seek healthcare as they have limited financial resources or may be more prone to overutilize services that are subsidized. Thus, our results may not generalize readily to populations from other age group and income levels. Thirdly, some of the population segments had a relatively small number of subjects which resulted in wide confidence intervals that tend to overlap between population segments. For example, Singapore scheme categories 13 and 14 both contained less than 10 subjects and is noted to have relatively large and overlapping confidence intervals for most of the evaluated outcomes. To evaluate mortality and health care use, a competitive risk analysis would have been ideal and should be pursued in future studies should time-to-event data be available. Finally, the existing models still has areas for improvement in terms of its ability to predict mortality.

One line of future work should aim to clarify issues resulting from the limitations noted above. This would include assuring that in a future prospective community-based study captures all variables needed by the various segmentation schemes. Future studies should oversample individuals likely to fall into relatively uncommon segments in order to establish adequate power to assess differences in outcomes. A second, crucial area, for study is to empirically evaluate responsiveness: how meeting service needs impact transition rates to worse health needs states. As this study has demonstrated the feasibility of utilizing patient reported variables for purposes of segmentation, longitudinal follow-up data gathering exercises may potentially be performed through telephone interviews which are logistically easier to execute compared to in-person interviews. Alternatively, core data needed for population segmentation can be embedded into the EHR and clinical workflow to allow ongoing assessment of transition rates between segments and the determinants of those rates. The responsiveness of segmentation schemes can be further addressed by incorporating them into evaluations of integrated packages of health services designed for specific population segments to allow more precise evaluation of the efficacy of these services for improving health outcomes. A third line of future work is to replicate this study in different countries, including additional population segmentation schemes as appropriate in order to gain insights about the nature of health service needs globally, and how individuals in different countries respond to interventions. Lastly, in order to continually improve the segmentation scheme’s ability to predict variability in relevant healthcare outcomes and responsiveness to types and modes of service, future work can include identification of combinations of predictive variables that provide better guides to service allocation, in the spirit of a “learning health system” [40].