Validation and application of a needs‐based segmentation tool for cross‐country comparisons

Abstract Objective To compare countries' health care needs by segmenting populations into a set of needs‐based health states. Data sources We used seven waves of the Survey of Health, Aging and Retirement in Europe (SHARE) panel survey data. Study design We developed the Cross‐Country Simple Segmentation Tool (CCSST), a validated clinician‐administered instrument for categorizing older individuals by distinct, homogeneous health and related social service needs. Using clinical indicators, self‐reported physician diagnosis of chronic disease, and performance‐based tests conducted during the survey interview, individuals were assigned to 1–5 global impressions (GI) segments and assessed for having any of the four identifiable complicating factors (CFs). We used Cox proportional hazard models to estimate the risk of mortality by segment. First, we show the segmentation cross‐sectionally to assess cross‐country differences in the fraction of individuals with different levels of medical needs. Second, we compare the differences in the rate at which individuals transition between those levels and death. Data collection/extraction methods We segmented 270,208 observations (from Austria, Belgium, Czech Republic, Denmark, France, Germany, Greece, Israel, Italy, the Netherlands, Poland, Spain, Sweden, and Switzerland) from 96,396 individuals into GI and CF categories. Principal findings The CCSST is a valid tool for segmenting populations into needs‐based states, showing Switzerland with the lowest fraction of individuals in high medical needs segments, followed by Denmark and Sweden, and Poland with the highest fraction, followed by Italy and Israel. Comparing hazard ratios of transitioning between health states may help identify country‐specific areas for analysis of ecological and cultural risk factors. Conclusions The CCSST is an innovative tool for aggregate cross‐country comparisons of both health needs and transitions between them. A cross‐country comparison gives policy makers an effective means of comparing national health system performance and provides targeted guidance on how to identify strategies for curbing the rise of high‐need, high‐cost patients.


| INTRODUCTION
Chronic conditions have been called "the healthcare challenge of this century" by the World Health Organization. 1 The prevalence of patients with numerous and complex health care needs related to one or more chronic conditions is expected to increase as life expectancy continues to rise. 2,3 An increasing life expectancy also exposes individuals to a greater number of changes over their life course, particularly to older-age life transitions not captured in their medical diagnoses. Changes over the life course, such as changes in social networks, relationships, living arrangements, or employment, influence the complexity of individuals' health care, including their ability to meet their own basic needs and rely on support from others. 4 Therefore, the health system is facing not only an increase in the number of individuals with specific clinical diagnoses 3 but also a growing number of individuals with complex needs related to a combination of multiple medical conditions and health-related social needs.
Lynn et al. 5 developed a needs-based population segmentation approach for improving a health system's effectiveness and efficiency in meeting the needs of the population it serves. Within a population, this approach identifies segments or clusters of individuals with similar care goals and similar types and intensity of needs. Such an ontology can, for example, measure the extent to which a particular population segment has less than desirable outcomes relative to a given benchmark. Policy makers can use these estimates to assess whether that segment's health and health-related social service needs are met and, if not, to develop targeted interventions. 5 While no clear definition of a "need" exists, a service is defined as "needed" when a typical individual with a set of characteristics that define a segment will likely benefit from receiving that service (in terms of reducing the likelihood of experiencing a more adverse health state). 6 To date, the operationalization of population needs-based segmentation has been restricted to specific subgroups (e.g., the frail elderly or individuals with possible palliative care needs) [7][8][9][10] or to datadriven approaches relying on electronic records to cluster individuals according to the risk of both poor outcomes and high cost. 8 In this paper, we introduce an adapted version of the Simple Segmentation Tool (SST) 11 -developed by Duke-NUS Medical School-that we call the "Cross-Country Simple Segmentation Tool" (CCSST). The CCSST segments the population into clinically significant global impressions (GI) segments (i.e., categories in terms of health status and medical complexity) and complicating factors (CFs) (e.g., experiencing fragmented care) are indicative of health and health-related social services needs, respectively. Operationally, for the needs-based population segmentation approach, a consensus panel of medical and social service experts defined "need" as any actionable service with a high probability of notable benefit. 7 In our framework, health needs are principally services that would need to be provided by a physician or other health provider trained and licensed to diagnose, prescribe, and perform procedures; and health-related social services needs are those that a physician, nurse, social worker, family member, volunteer, or even a specific technology can provide. While the SST was originally developed for clinical settings, it was modified for administration through special-purpose community surveys, 11,12 from which the CCSST is adapted. The CCSST uses data from the Survey of Health, Aging and Retirement in Europe (SHARE) and related international surveys available from the Gateway to Global Aging Data. 13 Using data that are readily available in many countries allows us to compare health care needs across countries.
The objective of this article is to present the CCSST and demonstrate a cross-country benchmarking application, comparing countries according to (1) the fraction of individuals with high medical needs and (2) the relative hazards of transitioning between lower and higher needs segments, and between those segments and death.

| METHODS
The methods section is structured as follows. First, we briefly introduce the survey data. Second, we present and describe the CCSST, a segmentation tool comprising five ordinal GI segments and four binary CF indicators. The GI segment categorizes individuals into one of five ordinal medical complexity categories. The four CFs identify patient characteristics that, if present, would increase the complexity of care for the conditions in the GI designation. Third, using Cox proportional hazard models, we assess the CCSST for predictive validity by modeling the association of the CCSST segments with mortality. Fourth, we describe how we applied the CCSST to cross-country comparisons by

| Description of survey data
The SHARE is a multidisciplinary, cross-national panel database consisting of seven biennial survey waves. It contains microdata on health, socioeconomic status, and the social networks of individuals older than 50 years. 10

| Development of the CCSST: Mapping survey data to GI and CFs
The CCSST is based on both clinical and survey versions of the SST, which has been validated for inter-rater reliability in a clinical setting, 7 and for predictive validity in both clinical and survey settings 11,12 as well as having face validity. 7 In both of these applications, questions were added to the data collection instruments, with the specific goal of assessing a specific set of GI categories and CFs. However, given that the survey data were not created for this purpose, we designed and constructed a mapping algorithm by using the foundational principles of the SST (see Supporting Information Data S1 for the original SST).
We found the data sufficient for the five ordinal GI segments: healthy, chronic asymptomatic, chronic symptomatic, long course of decline, and limited reserve with serious exacerbation-in line with the previous survey version of the SST. 12 Although the original SST included a sixth state (short period of decline before dying), predicting rapid decline with survey data is not feasible. We, therefore, collapsed the sixth state, short period of decline before dying, into the fifth state, limited reserve, and serious exacerbation. 12 The qualifying criteria for each GI segment are defined through clinical indicators, self-reported physician diagnosis of chronic disease, and performance-based tests conducted during the survey interview.
A trade-off to limiting ourselves to data in aging surveys, 13,14 thereby making the tool widely accessible, was that the CCSST was able to capture only four of the eight CFs described in the original SST: functional assessment (dependence on caregiver assistance), social support in case of need, frequent transitions between inpatient and outpatient care, and needing to take five or more prescription medications daily. The other CFs are need for nursing or rehabilitation services, activation in care, disruptive behavioral features, and having multiple health care providers in multiple locations. Even though we were able to capture only four CFs, preliminary results on the effect of the four we identified had been found earlier to have a significant association with adverse health outcomes. 11 We confirmed the consistency of all variables used for both GI segments and CFs across all waves for all countries. The descriptive mapping file for the GI segments and CFs is given in Table 1.
The clinical indicators we used for segmentation include frailty, the global activity limitation index, activities of daily living, the EURO-D depression score, and cognitive impairment. Frailty was determined through the SHARE operationalization 15 of Fried's "Frailty as a phenotype." 16 The global activity limitation index is a global single-item instrument that measures long-standing activity limitations (6 months or more) that general health problems cause and that inhibit activities commonly undertaken by survey participants. 17,18 The EURO-D depression symptom scale is a 12-point index identifying an individual's validated depression state for comparing the prevalence and risk of depression across countries. [19][20][21] Finally, we measure cognitive impairment by individuals' scoring below a cutoff value in a performance-based cognitive function score consisting of the mean normalized score of five cognitive function tests: verbal fluency, immediate recall, delayed recall, orientation, and numeracy. 22 The cutoff value was defined as 1.5 standard deviations below the mean score obtained by the total survey Wave 1 sample.
Chronic disease diagnoses-classified as either potentially lifethreatening or non-life-threatening-were determined through discussions with physicians on the project team and individuals involved in designing previous versions of the SST. 7,11,12 As chronic (50-59 years, 60-69 years, 70-79 years, 80+ years), and the presence of any CF as covariates. 27 Estimates from these log-linear models are exponentiated and interpreted as hazard ratios.
We ensured a reliable model convergence by (a) minimizing the number of groups (high medical needs, no high medical needs, and death) by including CFs as a time-inhomogenous covariate, rather than as part of the state stratification, and (b) restricting our analyses to countries with at least 8 years of observations to ensure that enough state transitions occurred. We evaluated different combinations of GI segments per health state to make certain that the rankings were robust and not the result of modeling Simpson's paradox, in which a trend appears in several different groups of data but disappears or reverses when these groups are combined. 9 We performed the analyses in the R statistical computing environment 10 by using the msm software package. 11 Although we estimate transition probabilities using a continuous-time Markov process, a Markov simulation model is beyond the scope of this paper.

| Predictive validity of the CCSST
In the Cox regression adjusted for age, age-squared, and gender, hazard ratios for mortality associated with high medical needs were significantly greater than 1 for all countries (see results on the predictive validity of GI segments in Table 2). Similarly, having a CF was associated with a higher likelihood of mortality, with 12 of 14 countries having hazard ratios statistically significant at the 1% level (see results on the predictive validity of CF segments in Table 2

| Cross-country ranking of per capita high medical needs
The most recent country-specific per capita proportions of GI segments with and without CFs are shown in Figure 1.

| Cross-country comparison of hazard of transition
We compare disease progression and recovery rates between countries using hazard ratios shown in Figure 2. First, the hazard ratio for transitioning from medically severe to death is generally inversely proportional to the fraction of the population in the high medical needs segment (i.e., the fewer individuals with high needs, the more likely they are to die). Second, with exceptions (e.g., Germany, Israel, Poland, and Switzerland), the relative hazard for moving between high and low medical needs categories is similar (i.e., the hazard of developing high needs is similar to the hazard of moving to low needs).
Third, for the countries with the lowest fraction of high medical needs individuals, not only is the hazard of transition from high needs to death higher, the hazard of transitioning from low to high medical needs is lower (i.e., for countries with the lowest fraction of individuals with high medical needs, the risk of entering the high medical needs category is lower, and the risk of exiting via death is higher).
Our results use 95% confidence intervals for the hazard ratios to determine whether differences or similarities in mean values are due to chance. The non-overlapping confidence intervals in Prevalence No high medical needs Transition from no high medical needs to high medical needs Transition from high medical needs to no high medical needs Transition from high medical needs to death High medical needs F I G U R E 2 Cross-country comparison of (1) prevalence of having high medical needs (i.e., classified in either of the two most severe global impressions (GI) segments) or not, and (2) hazard rates between no high medical needs and high medical needs segments, and between high medical needs and death [Color figure can be viewed at wileyonlinelibrary.com] delivery, policy makers need new performance metrics to help them learn which strategies are successful. As the CCSST allows for international comparison, it has great potential for supporting support policy makers in achieving their goals.
Unlike population segmentation approaches that focus on predicting poor outcomes, utilization, or cost, the CCSST methodology is explicitly linked to typical, actionable needs. Since the GI segments were defined by clinicians as meaningfully different in terms of type and intensity of medical service needs, 7 any transition across GI segments represents a clinically significant change to an individual's health.
Therefore, by extension, any transitions between collapsed categories indicate a clinically significant change in an individual's health.
Analyzing transition rates, therefore, helps policy makers think about and determine how undesirable transitions across health states-high rates of transition from low needs to high needs and high needs to death, or low rates of transition from high needs to low needs-might be handled at a whole-system level. For example, high rates of transition to high needs segments could be the result of unaddressed risk factors (e.g., smoking, excessive alcohol use, untreated diabetes, low rates of either vaccination for preventable conditions or screening for major treatable ones). High rates of mortality among individuals with high medical needs could be attributed to problems with the accessibility and affordability of medical services or lack of attention to CFs, thereby making medical services more difficult and less effective. In short, these metrics provide a guide to better assessing where the health system is less than effective in meeting health and health-related social service needs, in turn directly leading to interventions aimed at efficiently meeting those needs.
While we propose that applying a needs-based segmentation measure such as the CCSST can contribute to positive action, we recognize that, in isolation, neither comparisons between the proportion of the population within each health state nor the differences in proportional hazards of transitions can be directly used for assessing the quality of health system performance. Health outcomes reflect a myriad of causes, some of which may not be within the resources or ambit of the health system. Nevertheless, our needs-based segmentation approach can help identify ecological factors, such as social determinants of health (e.g., income, housing, pollution), thereby allowing health system leadership to provide critical input to broader public policy making.

| Strengths and limitations of the study
This study has several major strengths. First, the CCSST is based on a conceptual framework that explicitly links the features of individuals to needs that, if met, have the potential for improving health. Second, the approach provides an opportunity for evaluating and comparing needs-based population segments across time, across jurisdictions, thanks to the availability of compatible data collection efforts. The CCSST can be applied to most countries that conduct a health and aging survey, thereby allowing a form of benchmarking across countries and health systems in a period when dynamic complexity in health care service provision is rapidly increasing. Third, while the CCSST may lose some of the classification precision of a clinician's assessment, the CCSST has the advantage of not requiring information that depends heavily on health service utilization. It focuses instead on the self-report of health conditions and the impact of those conditions on perceived well-being.
Indeed, segmenting the population without relying on utilization data constitutes an advantage over the original survey version of the SST. While the original survey version of the SST was designed for mapping the clinician-completed version, it uses emergency room hospitalizations in assigning segments. 12 This feature of the CCSST will be incorporated into future survey versions of the SST.
One limitation of this study is that the segmentation is not necessarily equally accurate from country to country, because the compari-

| Implications
Future applications of this work include country-specific projections of the proportion of the future health status by using simulation modeling 12 to examine the potential impact of various policy actions aimed at more effectively meeting population needs. 28 Countryspecific modeling applications could potentially increase their number of states to incorporate CFs into their state stratification, because a larger number of states would reliably converge without the computational complexity of a 14-level country covariate. Other applications include determining the overall health system performance across countries through a utility-based evaluation of the CCSST health states. The first step would be to transform the SHARE quality of life indicator, the CASP-19, 13 into cardinal utilities through, for example, discrete choice experiments. 14 Further research is also needed for comparing the degree of bias present in the segmentation across countries. Future studies should compare the assessments resulting from the application of the CCSST to either (a) the assessments of an international group of clinicians or (b) entirely objective measures such as biomarkers, potentially including analyses of dried blood spot assays. 29 Additional further research includes applying the CCSST to related international surveys, 13 including the US Health and Retirement Study, to accelerate a deeper understanding needed to improve population health globally.

| CONCLUSIONS
The CCSST is an innovative tool for aggregate cross-country comparisons of both health needs and transitions between them. A crosscountry comparison gives policy makers an effective means of benchmarking national health system performance and provides targeted guidance on how to identify strategies for curbing the rise of high-need, high-cost patients and generally promoting coherent efforts at system improvement.