Health risk assessment and stratification in an integrated care scenario

Authors: Ivan Dueñas-Espín*, Emili Vela*, Steffen Pauws, Cristina Bescos, Isaac Cano, Montserrat Cleries, Joan Carlos Contel, Esteban de Manuel Keenoy, Judith GarciaAymerich, David Gomez-Cabrero, Rachelle Kaye, Maarten Lahr, Magí Lluch-Ariet, Montserrat Moharra, David Monterde Joana Mora, Marco Nalin, Andrea Pavlikova, Jordi Piera, Sara Ponce,Sebastià Santaeugenia, Helen Schonenberg, Stefan Störk,Jesper Tegner, Filip Velickovski, Christoph Westerteicher, and Josep Roca

Large scale deployment and adoption of integrated care services in Europe is seeking health efficiencies with simultaneous reduction of outcome variability within and among regions [1][2][3][4]. It is well recognized that health risk assessment is relevant for regional adoption of integrated care [5,6]. Therefore, not surprisingly, the on-going deployment processes of integrated care are generating novel requirements, and exciting opportunities, in both population-based health risk assessment and clinical risk prediction with potential impact on the design of healthcare services and clinical management of chronic patients [6], respectively.
The current study aimed to meet the challenges generated by large-scale deployment of integrated care services in the area of health risk prediction. That is, to fulfil the need for comprehensive risk assessment, both at the population level and the clinical scenario. In the latter, health risk assessment is needed to support adaptive case management strategies [7,8] aiming to cover the evolving requirements of chronic patients over time.
The research was carried out within the frame of the Advancing Care Coordination and TeleHealth program (http://www.act-programme.eu/) [8] involving five leading EU regions in terms of scaling up integrated care: Basque Country (ES), Scotland (UK), Lombardy (I),

Groningen (NL) and Catalonia (ES).
In the healthcare services domain, population-based risk predictive modelling facilitates the elaboration of stratification maps characterizing risk strata distribution of the entire population in a given geographic location. It allows identification of subsets of citizens with similar healthcare requirements facilitating both case finding and screening. The former, case finding, identifies highly vulnerable patients, allocated at the tip of the risk pyramid who are prone to major deleterious health events such as unplanned hospital admissions/readmissions, fast functional decline and/or death [9]. It is acknowledged that case finding fosters cost-effective preventive interventions despite compromised strength of prediction.
Likewise, screening looks for discovery of cases with non-manifest illnesses that may benefit from early diagnosis and cost-effective preventive interventions [6].
In the clinical management domain, risk prediction of well-defined medical problems (i.e. prediction of survival in acute exacerbations of COPD) [10] can support health professionals in the decision making process. Moreover, clinical risk prediction may contribute to patient classification in the optimal healthcare tier, helping to define shared care arrangements between primary care and specialists. However, it is acknowledged that modelling tools addressing specific clinical issues with a high predictive power may present limitations for their general application outside the source population [11].
In the current research, it is hypothesized that clinical prediction for any given specific medical purpose could be significantly improved by including the allocation of the individual in the population risk pyramid into the modelling approach.
The study addressed two specific aims. Firstly, to analyse population-based health risk assessment strategies, including assessment tools and health indicators, in the five European regions in order to identify current barriers and to elaborate recommendations for large scale deployment of integrated care at European level. The second purpose was to propose strategies toward enhanced health risk predictive modelling in the clinical scenario.

Method
The general characteristics of the population and healthcare organization of these five regions have been reported in detail [8]. The analysis of population-based health risk assessment and stratification strategies in place in the five ACT regions was done focusing on two specific components: i) analysis of health risk predictive modelling tools; and, ii) comparison of reported health indicators.

Population-based health risk predictive modelling
We performed a two-phase survey approximately eight months apart (Summer 2014 and Spring 2015) addressing the person(s) responsible for the development/maintenance of the risk predictive modelling tools at regional level.
In the first survey, systematic responses to a standardized questionnaire elaborated for this purpose by Opimec® [12] were collected. Participants answered the questionnaire by mail and subsequently underwent an interview. We captured information on several key dimensions characterizing the risk predictive modelling tools, namely: (i) modelling approach, (ii) source sample, (iii) main and summary statistics, (iv) outcome (dependent) variables and covariates, (v) update periodicity, (vi) target population; and, (vii) maturity of clinical implementation. This facilitated the elaboration of an initial map of regional practices.
The second survey had a twofold purpose: (i) to fill specific information gaps; and, (ii) to ask additional questions inquiring on existing plans for evolving the risk predictive modelling tool in place. Also, we assessed the potential for transferability across regions at EU level. To this end, four main items were analyzed: (i) openness of algorithms; ii) flexibility for adjustments to other populations; (iii) licence costs associated with the use of the case finding tool; and, (iv) licence agreements binding its applicability to specific territories.
The comparative analysis of health risk assessment tools among ACT regions was carried out taking into account a clear distinction regarding the characteristics of the source population. That is, health risk assessment tools generated from modelling the entire population of a given region (or geographical area) with a holistic approach were considered to follow a population health approach, as proposed by Kindig D et al. in 2003 [13]. On the other hand, health risk assessment derived from modelling patient populations were regarded as following a population medicine approach [14].
Because of our interest on case finding analysis, the current study focused on healthcare forecasting [15] that implies predicting an individual's healthcare utilization for interventional purposes with either preventive or therapeutic goals. Comprehensive descriptions of the characteristics of health risk predictive modelling and the logistics required for deployment are reported elsewhere [16][17][18]. It is of note that other analyses like risk adjustment [19][20][21] or actuarial approaches [22] were not considered in the current research.

Health indicators
A semi-structured questionnaire including indicators to evaluate health status at population level was sent via email to the ACT coordinator in each of the five regions. The indicators were shared by the five ACT regions and had been previously defined and agreed within the consortium in order to facilitate comparability of the effects of integrated care interventions over time within and across regions.
The study aimed to identify a common set of indicators useful for evaluating the impact of health interventions at population level in order to facilitate comparability of the effects of integrated care services within and across regions.
Elaboration of a proposal for enhanced clinical risk assessment We analysed the potential of population-based health risk assessment to contribute to enhance clinical risk predictive modelling. To this end, we assessed the flexibility of the different health risk predictive modelling tools to contribute to clinical risk estimation.

Population-based health risk assessment tools
The main characteristics of the health risk prediction modelling tools in place in four out the five ACT regions are depicted in Table 1. Groningen (NL) is not represented because the site does not use any population-based health risk prediction modelling for the two integrated care programs currently deployed [8]. Instead, Groningen prioritized individual health risk characterization based on information collected in the electronic health records.
The four regions ( Table 1) perform periodic updates of their respective population-based stratification. The table indicates that a population health approach [13] is currently only adopted in the Basque Country [23][24][25][26] and in Catalonia. Since 2010, Scotland [27] is clearly evolving in this direction. The source population of the current health risk predictive modelling tool already covers 63 per cent of the entire Scottish population. Strategically, it is moving from a strong focus on use of hospital-related resources (e.g., emergency department consultations, unplanned hospital admissions and/or early re-admissions) toward integration of needs for frail patients, including social support and long-term care.
In contrast, Lombardy [28] has a population medicine approach consisting of a classification system based on stratification by health costs. It serves the coordinated care program for chronic patients, especially those with conditions such as chronic obstructive pulmonary disease (COPD), cardiovascular disorders and diabetes mellitus types I and II.
The analysis of the risk-strata distributions resulting from the different regions showed poor comparability ( Table 2). This is explained by differences among risk predictive modelling tools, and by the diverse classification criteria used to define risk groups.
We identified significant constraints for transferability across regions due to three main factors, namely: i) lack of openness of algorithms, ii) rigidities due to inclusion of expertbased criteria in the morbidity groupers; and, iii) license bindings constraining applicability of health risk assessment tools to other EU regions. It is of note that only Catalonia and Scotland ( Table 2) have white-box tools owned by the regions, which, in principle, implies high potential to properly deal with the limitations for transferability described above.
The four regions indicated in Table 1 provide information on case finding for primary care doctors. We identified a consensus on the need for transferring information on high-risk patients to practicing clinicians in order to trigger preventive interventions and to support clinical decision-making processes. However, we observed different degrees of maturity in the interactions with clinicians, from only providing a list of high-risk candidates for interventions to the display of simple clinical decision support systems in the clinical workstation of primary care physicians.
The two surveys carried out during the project lifetime did not indicate relevant conceptual differences among regions in terms of the basic aspects that should be covered in an ideal health-risk assessment strategy. Accordingly, a high acceptability of the population health approach [13] for elaboration of health risk predictive models was confirmed. Table 3 indicates the characteristics recommended for an ideal population-health risk assessment tool showing transferability among regions and potential to generate synergies with clinical risk predictive models. The practicalities for deployment of health risk assessment tools at regional level are summarized in Table 3S (on-line supplementary material).
Because the survey showed agreement among all regions on the need for using predictive models, showing statistics indicating sensitivity/specificity of the predictions was selected as a recommendation for good practice of population-based health risk assessment ( Table 3).
Both Basque Country and Scotland risk predictive modelling tools provide information on sensitivity and specificity; by contrast, Catalonia and Lombardy classify individuals into specific percentiles of the risk-strata pyramid thereby neglecting the metrics assessing robustness of predictions.

Health indicators
The list of indicators identified by the consortium is included in Table 4S. We found that despite availability of most of the data at the regional level, two main limiting factors precluded comparisons among the regions, namely: (i) insufficient data harmonization (e.g. different versions of International Classification of Diseases (ICD) coding) [29]; and, (ii) differences on data reporting (i.e. different levels of data aggregation and/or differences in calculation of complex indices).

Enhanced clinical risk predictive modelling
We propose to incorporate the classification of the individual in the risk stratification pyramid as one of the covariates of the clinical predictive models. Among the different populationbased risk assessment tools evaluated in ACT (Table 1) Overall, the GMA approach shows flexibility and transferability, as demonstrated by its recent adoption by thirteen out of the seventeen regional healthcare systems in Spain, covering 92% of the overall Spanish population. Figure 1 illustrates the contribution of the morbidity grouper GMA to explain the variance of four relevant outcomes, namely: mortality, hospital admissions, emergency department admissions and total healthcare expenses.
Statistical refinement of the computational modelling of the current GMA [32], in order to generate an enhanced GMA fulfilling all the requirements indicated in Table 3, is recommended as the first milestone to enhance clinical risk predictive modelling. In a second step, we propose to prospectively assess the benefit of using the predictions from the enhanced GMA as an additional covariate into clinical risk predictive modelling. Both conceptual grounds and statistical feasibility support the proposal. However, its implementation will require further work beyond the scope of the current research.

Toward personalized medicine
The study also explored a systems approach [33] considering all dimensions influencing patient health as potential covariates to be taken into account for elaboration of a roadmap toward personalized medicine for chronic patients.
Three categories of covariates have been identified to show potential for inclusion into clinical risk predictive modelling, as displayed in have the potential to articulate the three categories of variables potentially allowing for dynamic assessment of health risk both for population-based purposes, but also for specific clinical problems. Nowadays, a Digital Health Framework, as depicted in Figure 3, is only a conceptual formulation, but it contains the seeds to foster the concept of the "exposome", as defined by Coughlin SS [33], which provides the basis for personalized medicine for chronic cases. There is no doubt that the implementation of the Digital Health Framework constitutes an ambitious endeavour requiring an stepwise approach to effectively overcome major challenges involved in the transitional process to make it operational.

Summary of main findings
The results of the two surveys indicated a high degree of conceptual agreement among the five ACT regions on the relevant role of population-based health risk assessment for regional deployment of integrated care. Its usefulness for service commission, case finding and screening was shared by the entire ACT consortium. There was also consensus on the use of a population health approach [13] as the optimal strategy for population-based risk assessment.
However, the health risk predictive modelling tools in place showed marked heterogeneities that precluded comparability of the risk pyramid distributions across regions. Moreover, we identified a clear need for evolving toward risk predictive modelling tools allowing proper quantification of the estimations ( Table 3) [35]. Likewise, different well-identified problems mostly associated to data reporting precluded appropriate comparisons of the recommended health indicators described in Table 4S.
The current study identified transferability across regions and potential for evolving, that is flexibility, as two key requirements for any population-based health risk assessment tool.
Factors such as: i) license binding constraints, ii) insufficient public availability; iii) lack of availability for inspection; and/or, iv) rigidity of some computational algorithms (i.e. due to inclusion of expert-based criteria in some morbidity groupers) are currently limiting transferability. These factors might also preclude adaptation of the current risk prediction tools toward evolving requirements such as: i) integration between healthcare and social services; and, ii) implementation of synergies between population-based and clinically oriented risk predictive modelling, as described in the study.
We acknowledge some intrinsic limitations of population-based predictive modelling in terms of strength of estimations. However, their potential for allocation of individuals into the risk stratification pyramid facilitates both design and implementation of preventive strategies that have shown high potential to generate healthcare value. For example, for those individuals in the tip of the pyramid (~5%) accounting for high use of healthcare resources (~ 36% total healthcare costs, as assessed for Catalonia in 2014, Figure 2S).
The study reports on the conceptual steps required for development of innovative strategies for clinical risk predictive modelling with potential to enhance its supporting role for decision making in the clinical scenario. We acknowledge, however, that further studies evaluating feasibility, benefits and applicability of the proposals for enhanced clinical risk prediction are needed.

Toward an European strategy for population-based health risk prediction
While the ACT project has confirmed the role initially ascribed to population-based health risk assessment in regional deployment and adoption of integrated care services, the core lesson learnt from the current study is that a common European strategy is needed and it constitutes a priority for any region planning adoption of integrated care.
Two basic pillars for a future European strategy should be: i) the characteristics of the risk predictive modelling tools, as displayed in Table 3; and, ii) the ability to report on the list of basic indicators depicted in Table 4S. The current heterogeneities among regions clearly indicate that adjustment of the current settings to the recommended good practice will require site-specific transitional strategies whose common goals and basic principles are described in the current study. Key operational steps needed for practical implementation of a regional strategy for population-based health risk predictive modelling are summarized in Table 3S.
There is a lively debate regarding management modalities associated with generation and exploitation of population-based health risk predictive modelling. Should model generation and maintenance be publicly funded (i.e. Department of Health) or should there simply be policies promoting open market in terms of private suppliers of risk predictive tools [16,36]?.
The current study only stresses the need for openness, flexibility and transferability of risk predictive modelling in order to fulfil their core purposes. However, as stated below, we acknowledge the complexities of the issue, also involving ethical aspects. Doubtless, this issue will require proper regulation irrespective of the finally adopted business adoption.

Clinical health-risk assessment
The authors acknowledge that only a small proportion of the huge potential of risk predictive modelling is currently applied for health forecasting purposes in the clinical arena. A detailed description of the bottlenecks constraining the developments recommended for enhanced clinical risk predictive modelling, as proposed in the current study, are reported in [37]. Under the subheading, we are highlighting only a few key aspects that shall be addressed to accomplish successfully the roadmap proposed in the current study.
One key requirement of enhanced clinical risk predictive modelling is to set-up the concept of Digital Health Framework (DHF) (

Part I -Population-based health risk assessment in Catalonia
CatSalut is the Catalan public agency acting as unique payer of regional healthcare services covering the entire population of approximately 7.5 million inhabitants. The Agency is commissioned by the Department of Health of the Catalan Government to generate a regional population health strategy for health risk assessment and stratification.
Until very recently (early 2015), the risk predictive modeling in place was based on Clinical Risk Groups (CRG) [1]. However, CatSalut has developed its own system, the GMA (Adjusted Morbidity Groups), refined during the last years and fully implemented into the primary care clinicians workstation by May 2015. The reasons for moving from CRG to GMA were twofold: (i) to decrease costs, and, (ii) to increase flexibility of the risk predictive modeling tool allowing its adaptation to the evolving needs such as integration of social support. There has also been an active policy to foster transferability to other regions. As described in the main text, the GMA is being successfully implemented in thirteen out of the seventeen regional healthcare systems in Spain, which represents coverage of 92% of the Spanish population.

Regional source datasets
The current catalan population-based risk assessment tool is updated every 6 months using the dataset depicted in Figure 1S that includes information from Primary Care, Hospital-related events, Pharmacy, Mental Health, and Socio-sanitary services.
Analyses of use of healthcare resources, pharmacy consumption, prevalence of key disorders and calculation of adjusted morbidity groups, using the GMA morbidity grouper, constitute the basis for periodic updates of the regional health-risk strata pyramid. The multiple regression use as covariates: (i) age, (ii) sex, (iii) ZIP code location (as a proxy of socio-economic status using adjusted territorial income and health services accessibility), (iv) GMA morbidity grouper; and (v) use of healthcare resources.
Interestingly, the GMA morbidity grouper is one of the components providing marked flexibility/transferability to the catalan health-risk assessment tool because it is not built on fixed expert knowledge, but it relies on population-based statistical information.
Additional key features are algorithm openness and flexibility regarding licensing agreements. Table 1S indicates main advantages of the GMA compared with the CRG previously used in Catalonia. Risk classification using GMA -The GMA grouper is a new tool for assessing multimorbidity, which classifies individuals into unique and mutually exclusive groups taking into account: (i) type of disease, (ii) occurrence of multi-morbidity; and, (iii) case complexity. Briefly, the risk classification criteria combines two dimensions: i) Morbidity, including a total of seven morbidity groups, and, ii) Case Complexity, as depicted in Table 2S. Patients with a chronic disease in 4 or more systems 1 2 3 4 5 Patients with a chronic disease in 2 or 3 systems 1 2 3 4 5 Patients with a chronic disease in 1 system 1 2 3 4 5 Patients with an acute diseases 1 2 3 4 5 Pregnancy and delivery 1 2 3 4 5 Healthy population 1 The GMA classification (Table 2S)  The main requirement to elaborate the GMA grouper is availability of all health diagnosis, events and use of pharmacy obtained from the registry of insured people, as displayed in Figure 1S. The core information is obtained from Primary Care datasets.
Additional information from other healthcare tiers is useful to refine the GMA grouper but it is not strictly necessary.
The use of the GMA grouper provides allocation of each citizen into the risk stratification pyramid. A summary representation of the update carried out by the end of 2014 grouping the results in four main risk strata is depicted in Figure 2S. The four main strata are identified according to the criteria indicated below:  GMA-1 or low risk stratum: it corresponds to 50% of the population, with a lower complexity level.
 GMA-2 or moderate risk stratum: it corresponds to 30% of the population, which has higher complexity than the previous risk stratum.
 GMA-3 or high risk stratum: it corresponds to 15% of the population, which has greater complexity than the risk stratum.
 GMA-4 or very high-risk stratum: it corresponds to 5% of the population, which has the highest complexity level. Figure 2S -Stratification of the Catalan population (2014) using the GMA. The third and fourth columns depict rates of mortality and hospital admissions, respectively. The fifth column indicates the cost per inhabitant per year expressed in € and the last column refers the percentage of total healthcare expenditure by risk strata. It is of note that the closer the patient is to the tip of the pyramid, the higher are: mortality, risk of hospital admission and healthcare expenses. Green color (bottom) indicates healthy status whereas red (tip) corresponds to maximum risk of admissions and highest mortality risk.

GMA evaluation protocol
The GMA morbidity grouper was evaluated using two different approaches: i) Statistical evaluation using a comparative analysis of the contribution of different covariates to prediction a specific healthcare outcomes, namely: mortality, unplanned admissions, emergency department consultations and healthcare expenditure, as displayed in Figure 1 (main text); and, ii) Clinical evaluation carried out by general practitioners.

Statistical evaluation
A set of models based on multiple linear regression analysis including different covariates were used to assess the performance of the GMA grouper for health risk assessment (Figure 1). The population of Catalonia in 2014 was taken as a reference for the analysis and the four healthcare outcomes indicated in Figure 1 were the dependent variables. Two main statistics were used for comparison among the models obtained using different covariates: i) Akaike's Information Criterion (AIC), as a measure of the relative quality of statistical models for a given set of data; and, ii) Rsquare that should be interpreted as the proportion of uncertainty in the relevant outcome that has been explained by the model [4].  Figure 3S: Goodness of the classifications generated by the two morbidity groupers: CRG (grey) and GMA (orange) by level of complexity assigned by the general practitioner. The last column provides a summary analysis.

Clinical workstation in Primary Care
The outcome from the GMA for a given citizen/patient It includes a brief systematic description of main recommendations for implementation and evaluation of a health risk assessment strategy at regional level (Table 3S), as well as the list of main domains and specific indicators for regional population-based risk assessment (Table 4S).
C O N F I D E N T I A L Table 3S. Recommended operational steps toward implementation of a regional strategy for health risk assessment Recommended operational steps 1. Health risk predictive modelling implementation -Use a population health risk assessment tool fulfilling the requirements indicated in Table 3 (main text), either by fostering the evolution of your own risk assessment tool or by adopting an existing risk assessment tool that fits your local needs, that can be used without any license bindings and supports an open market of suppliers. Screen your population on a regular and repeated basis. Be aware of the logistics required at regional level to develop operational health risk prediction strategies: i) identify and overcome the practical local hurdles and barriers for accessing and linking routine administrative and clinical data and, ii) estimate the cost of running a tool, software platform, data integration, as well as labor for operations.
2. Define and activate specific functionalities -Use population-health risk stratification to understand the needs and risks of your population to target and prioritize effective integrated care. Make the outcome to be predicted operational (  The number of patients using ≥15 drugs (per 1000), last 12 months, due to any cause (The number of patients using ≥15 drugs in the last 12 months, due to any cause/total population)*1000 No.
a The indicators are expressed over the population in a given year; some indicators could be specified for being applied to the population with specific diseases (Xi disease). b In this domain, the deprivation index is calculated based on the next regional indicators domains: Barriers