Should health insurers target prevention of cardiovascular disease?: a cost-effectiveness analysis of an individualised programme in Germany based on routine data

KardioPro is an individualised prevention programme, which includes screening for CHD, risk factor assessment, early detection and secondary prevention of CHD. KardioPro was first introduced in Munich and later in Coburg, Berlin, Karlsruhe, Erlangen and North Rhine-Westphalia. To enrol in KardioPro, participants needed to be at least 45 years of age. Furthermore, CHD prevalent subjects of any age were eligible for inclusion. All subjects who met the enrolment criteria received information in writing about KardioPro. Individuals in KardioPro were divided into different pathways: patients known to have CHD, suspected to have CHD and without CHD. Each pathway had a different setting depending on the individual risk or the severity of the disease. For risk assessment, the original Prospective Cardiovascular Münster (PROCAM) score was used [10]. This score defined the 10-year risk of sudden cardiac death or a definite fatal or nonfatal myocardial infarction and stratified diagnostic and therapeutic strategies. Furthermore, all KardioPro participants were categorised as either CHD prevalent subjects or subjects with high, medium or low risk, according to the PROCAM score.

The management plan for each individual patient was provided by the participating cardiologists whose task was to prescribe the optimal medical treatment for the patients and to manage the risk factors according to European guideline recommendations. Furthermore, KardioPro included recent techniques such as CT for calcium scoring (Agatston score) and/or CTA for diagnosis. In the case of necessary coronary stenting advanced types of drug-eluting stents were allowed. In addition, an electronic health record was created offering easy access for patients and facilitating communication between specialists and institutes participating in KardioPro. Patients were also sent reminders about follow-up visits.

For the evaluation of KardioPro, we included participants who enrolled 2007, 2008 and 2009 and ended the observation at 31 December 2010. As the treatment paths in KardioPro differed according to risk classification, the cost-effectiveness of KardioPro was intended to be determined by risk group. However, the risk classification that was applied to KardioPro participants was not suitable for analysis, as risk classification was based on information that has only been observed for participants. The applied PROCAM score requires information about smoking behaviour, blood pressure and cholesterol level, which is not included in German sickness funds routine data [10]. Applying propensity score matching to each of the KardioPro risk groups would most likely have yielded biased control groups (selection bias), as the outcome of KardioPro depends on CHD risk. To achieve unbiased results, risk group classification needed to be consistent between KardioPro participants and controls. Furthermore, to develop efficient strategies for targeting selected risk groups (i.e. offering or advertising KardioPro to selected individuals), the risk groups needed to be defined independently from participation. Thus, for economic evaluation, new groups of patients at risk of CHD were defined based on health insurance company routine data. However, our definition of risk groups was inspired by the risk groups originally defined in KardioPro.

Firstly, we identified CHD prevalent subjects based on administrative morbidity classification. This information is included in routine data, as the presence of CHD has to be reported for each insured person within the German morbidity-based risk structure compensation scheme. The remaining subjects were divided into subjects with low, medium or high CHD risk. CHD risk scores (such as the PROCAM or the Framingham score) could not be calculated directly based on routine data. However, the PROCAM score [10], which measures the probability of CHD incidence within the next 10 years, was available for KardioPro participants. Thus, we built a regression model to predict the PROCAM score based on routine data. The routine data from CHD non-prevalent participants with a known PROCAM score at baseline of KardioPro were used as input data. As the PROCAM score is a probability and thus has a valid range from 0 to 1, we chose a quasi-beta regression model for prediction [11]. The explanatory variables used for prediction were age, sex, hypertension, obesity and diabetes, and referred to the year prior to KardioPro participation. These variables were selected as they best corresponded to prediction variables in the original PROCAM score [10]. As data for three accounting years have been combined within the regression model, we also included a random intercept for the calendar year (2007, 2008 and 2009). The concordance of the prediction model and the original PROCAM score was quantified via the Spearman’s rank correlation coefficient. The threshold values to define the low, medium and high CHD risk groups were set in such a way that the same proportions were achieved, as observed for the originally defined KardioPro risk groups (i.e. 9.72% high risk, 11.16% medium risk and 79.12% low risk among non-CHD KardioPro participants). All risk groups that we used for analysis were built based on data collected prior to enrolment in KardioPro. Thus the intervention did not affect the risk group categorisation.

Control groups were created by retrospective propensity score matching. All individuals insured by the health insurance company who did not participate in KardioPro were eligible as potential controls. However, as there are severe regional differences regarding health expenditures, only subjects from the regions where KardioPro has been offered were considered as controls. Depending on the year of enrolment, the number of KardioPro participants and potential controls were as follows: 2007: 2,953 new KardioPro participants, 131,109 potential controls, 2008 5,745 new KardioPro participants, 163,756 potential controls, 2009: 4,418 new KardioPro participants, 180,612 potential controls.

In our analysis, the propensity score was defined as the probability of the individual participating in KardioPro. The propensity score was computed by a logistic regression model based on the individual’s characteristics in the previous year (one propensity score model per calendar year). This model was selected from different logistic regression models (backward and stepwise variable selection and a full model) which were evaluated by cross-validation. Stepwise variable selection was used in the model and considered more than 140 potential explanatory variables (see Additional file 1 for a complete list of all variables included). We used the approximate nearest neighbour 1:1 without replacement approach by modifying a published macro [12]. The matching macro started to select pairs with an identical propensity score on the 10th decimal and decreased step by step until the first decimal. An additional criterion was being insured for the whole matching calendar year. Matching was done separately for each subgroup and also by year 2007, 2008 and 2009.

The primary health outcome was defined as event-free time and measured by counting the days until the first event (event-free days). The observation time for each matched pair started from the enrolment date of the participant in KardioPro. The observation time ended for each individual either with the first event or by censoring. Events were defined by a combined endpoint of death (all causes), myocardial infarction (MI) and stroke. MI and stroke were defined via the international classification of diseases (ICD) codes in a hospital (i.e. acute myocardial infarction (ICD-10, I21), subsequent myocardial infarction (ICD-10, I22), cerebral infarction (ICD-10, I63) and stroke (ICD-10, I64)). Censoring took place at the end of the observation period (i.e. 31 December 2010). Furthermore, the observation time was censored for both partners in a matched pair when either of them cancelled their SBK health insurance. Therefore, differences in observed time were a result of events only.

As the health outcome was calculated on a daily basis, daily health expenditures would have been desirable to calculate incremental costs. However, the routine data used for analysis reported costs for whole calendar years only, thus guiding the presentation of results.

The considered costs for each matched pair were defined as follows. For the year of enrolment, we assumed that, on account of propensity score matching, the expected health expenditures prior to enrolment did not differ systematically between the participant and the control subject. Thus, the costs for the whole calendar year were considered for each partner. In the case of censoring as a result of cancelling the SBK health insurance, which could occur on any day within the calendar year, the total costs of the particular calendar year were considered for the subject who (first) cancelled their SBK health insurance. For the matched partner however, costs were only accounted proportional to the time considered for the evaluation. When a subject died, the total reported costs of the follow-up period for both partners were considered. In cases of non-fatal events all subsequent costs for both matching partners were considered.

All costs were inflated to the year 2010 based on inflation rates provided by the German Federal Statistical Office [5]. Costs and effects were also discounted by using an annual discount rate of 5% [13]. Furthermore, costs were stratified into the following categories: hospital costs, pharmaceutical costs, physician costs and other costs. The category ‘other costs’ included, for example, physiotherapy, laboratory resources, the costs of additional services such as acupuncture and the cost of sickness benefits.

We assessed the cost-effectiveness of KardioPro from the statutory health insurance (SHI) perspective. The incremental cost-effectiveness ratio (ICER = (cost of intervention–cost of control)/(effect of intervention–effect of control)) was determined for KardioPro as a whole and for the individual risk groups. To illustrate which strategies are efficient, if KardioPro were to be offered to selected risk groups, we also considered combined strategies, i.e. offering KardioPro to more than one risk group simultaneously. This resulted in eight strategies overall (these are explicitly listed in Table 1). These combined strategies were graphically represented via an efficiency frontier [14, 15].

Based on the number of days until death, a secondary health outcome analysis was conducted. The discount rate has been varied (at 0%, 3% and 10%) via deterministic sensitivity analysis [13]. Finally, to derive standard errors, we performed a bootstrapping approach for all the strategies by drawing 10,000 samples with replacement and calculating the ICER for each sample.

Statistical analysis was performed with the software package SAS, version 9.2.

The aim of this study was to evaluate the real-world health care programme KardioPro. The intervention KardioPro was offered to all persons insured by SBK as part of the health care basket provided by a German statutory health insurer, and is covered by the German Social Code §§97, 80 SGB X and §§67, 43 SGB V. KardioPro did not introduce new treatment methods, but worked by strengthening guideline-specific treatment. This type of support falls within the common responsibilities of a statutory health insurance company. Only subjects who gave written consent to participation were included in KardioPro. Evaluation of KardioPro was authorised by SBK, designed retrospectively and commissioned to an independent research group. The evaluation was based on anonymised, patient-specific routine data from SBK and was approved by the SBK data protection officer.