Explaining the Decline in Coronary Heart Disease Mortality in the Netherlands between 1997 and 2007

Objective We set out to determine what proportion of the mortality decline from 1997 to 2007 in coronary heart disease (CHD) in the Netherlands could be attributed to advances in medical treatment and to improvements in population-wide cardiovascular risk factors. Methods We used the IMPACT-SEC model. Nationwide information was obtained on changes between 1997 and 2007 in the use of 42 treatments and in cardiovascular risk factor levels in adults, aged 25 or over. The primary outcome was the number of CHD deaths prevented or postponed. Results The age-standardized CHD mortality fell by 48% from 269 to 141 per 100.000, with remarkably similar relative declines across socioeconomic groups. This resulted in 11,200 fewer CHD deaths in 2007 than expected. The model was able to explain 72% of the mortality decline. Approximately 37% (95% CI: 10%-80%) of the decline was attributable to changes in acute phase and secondary prevention treatments: the largest contributions came from treating patients in the community with heart failure (11%) or chronic angina (9%). Approximately 36% (24%-67%) was attributable to decreases in risk factors: blood pressure (30%), total cholesterol levels (10%), smoking (5%) and physical inactivity (1%). Ten% more deaths could have been prevented if body mass index and diabetes would not have increased. Overall, these findings did not vary across socioeconomic groups, although within socioeconomic groups the contribution of risk factors differed. Conclusion CHD mortality has recently halved in The Netherlands. Equally large contributions have come from the increased use of acute and secondary prevention treatments and from improvements in population risk factors (including primary prevention treatments). Increases in obesity and diabetes represent a major challenge for future prevention policies.


Contents
Overview of the IMPACT-SEC model Page

3.5
Accounting for poly-pharmacy Potential overlaps between patient groups: avoiding double counting Overlap between pharmacological and nonpharmacological contributions to risk factor DPPs Negative DPPs Cumulative risk-reduction: adjusting DPPs to calculate cumulative benefit of multiple risk factor changes 7 9 9 10 10

INTRODUCTION
This technical appendix is based on the technical report for the IMPACT-SEC model created using English and Scottish data. 1,2 We have adapted their model to create the Dutch IMPACT-SEC model. However, much of the theory and methods remain. IMPACT is a deterministic, cell-based policy model. The IMPACT model examines the effects of changes in treatment uptake and risk factor trends on changes in mortality from coronary heart disease (CHD) among adults in the Netherlands aged 25 years and over. It uses epidemiological information to estimate the contributions of population-level risk factor changes (impacting mainly on incidence) and changes in the uptake of evidencebased treatments (impacting mainly on case fatality) on mortality decline between two points in time (the start-year and the end-year). The primary outcome measure of the model is the deaths prevented or postponed (DPPs). The extended IMPACT-SEC model accommodates sub-national variation in CHD mortality trends by socioeconomic circumstances (SEC groups). The tables included in this supplementary appendix provide details about the sources and methods that were used in extending the IMPACT model to accommodate socioeconomic circumstances (IMPACT-SEC model).
The starting point for the model is to calculate the 'target' number of CHD deaths the model needs to explain. This target number is obtained by linking the Dutch population register with the Dutch cause of death register to calculate the difference between the actual observed CHD deaths recorded in the end-year and the deaths expected in the end-year had the CHD mortality rates remained the same as in the start-year (i.e. simple direct standardisation).
The calculation of the modelled estimate of DPPs rests on utilising two well-studied relationships: firstly, that between risk factor change and the relative reduction in CHD mortality; secondly, that between treatment uptake and reductions in one-year mortality in patients with a specific form of CHD. The model applies the relative risk reduction quantified in previous randomised controlled trials and meta-analyses to estimate the mortality reduction attributable to: a) temporal change in risk factor prevalence (in those without diagnosed CHD) to calculate the DPPs 'explained' by specific risk factor trends; b) net change over the period in the uptake of specific treatments in patients with each specific form of CHD to estimate DPPs 'explained' owing to improved one-year mortality rates. Great care is taken to avoid double counting of the same individuals.
The mortality benefits from the risk factor reduction in the population, and the treatment benefits in patient groups are then summed. Thus summing uses a cumulative approach (rather than an additive approach), in order to avoid double-counting of benefits in the same individual. (This approach is detailed in Section 3.5). This mortality sum represents the deaths prevented or postponed (DPPs) 'explained' by the model. At the end of the modelling process, the total DPPs 'explained' by the model is then compared with the observed fall in deaths (the 'target' to be explained). Model fit is therefore calculated as the difference between the observed deaths and model DPPs, and expressed as the percentage explained. This measures the extent to which the model was successful in explaining the observed change in CHD mortality in the population.
A policy model like IMPACT thus stands in contrast to a typical multivariate regression model. A typical multivariate regression model represents a statistical approach to describing a single data-set, for instance generated by a single cohort or randomised controlled trial. In contrast, a policy model such as IMPACT seeks to integrate and synthesise best estimates from a variety of sources to reliably estimate the extent to which a range of factors, acting in combination, explain or predict an outcome. We did not obtain the parameters for this model by running regressions. Rather, the model incorporates the best coefficients from the largest meta-analysis or randomised controlled trials of the reduction in case fatality attributed to treatment or the independent effect sizes of a unit change in each risk factor on CHD mortality.
Examples of the calculation method used for estimating the DPPs due to treatment uptake (Example 1, page 97) and for continuous and binary risk factor change (Examples 2 and 3, respectively, page 98) are provided below. Earlier versions of the IMPACT mortality model have been previously applied to national data from Europe, United States, Ontario, New Zealand and China. [3][4][5][6][7] The methodology has previously been described in detail online and elsewhere. 4-6

Socioeconomic groups
We used a socioeconomic indicator by postal code (SCP 2002(SCP -2006) as a proxy indicator of socioeconomic circumstances. 8 Socioeconomic scores for a total of 3,965 postal codes were calculated by SCP (Netherlands Institute for Social Research). The mean number of inhabitants was 4,126 per four-digit postal code in 2007. The socioeconomic scores were based on a principal component analysis of the following items: (1) mean annual income per household, (2) percentage of households with low income, (3) percentage of households with low education and (4) percentage of unemployed inhabitants. Rank numbers of socioeconomic scores per postal code were used to make three socioeconomic groups in the Netherlands; lowest socioeconomic group (20% most deprived Dutch inhabitants), middle group (60% of Dutch inhabitants) and highest socioeconomic group (20% most affluent Dutch inhabitants). Socioeconomic circumstances were defined separately in every age-sex stratum. By doing so, the age and sex distribution of the three socioeconomic groups was comparable.

Changes in mortality rates from CHD, Netherlands 1997 to 2007: Expected and observed number of deaths from CHD
Mortality rates from CHD were calculated using the underlying cause of death (ICD9 code 410-414). Both unadjusted and age-adjusted mortality rates were calculated and presented in Table A. The expected number of CHD deaths in 2007 was calculated by multiplying the age-sex-socioeconomic group specific mortality rates from CHD in 1997 by the population counts for 2007 in that age-sex-socioeconomic stratum. Summing over all strata then yielded the expected number of deaths in 2007 had mortality rates remained unchanged. The difference between the number of expected and observed deaths from CHD represented the mortality fall, or the total number of deaths prevented or postponed (DPP), in 2007 relative to 1997. Population counts, CHD mortality rates, observed and expected numbers of deaths are shown in Table A. The data sources used to estimate the size of each disease group (stratified by age-sexsocioeconomic) are shown in Table B. The general approach to calculating the number of DPPs from an intervention among a particular disease group was first to stratify by age, sex and socioeconomic, then to multiply the estimated number of patients in 2007 by the proportion of these patients receiving a particular treatment, by the one-year mortality rate, and by the relative reduction in the mortality rate due to the administered treatment. Sources for treatment uptake are shown in Table D. Sources for estimates of treatment efficacy (relative risk reductions) are shown in Table F. We obtained the relative risks based on the most recent published systematic reviews and meta-analyses of epidemiological studies. Each treatment relative risk value in the model was based on a meta-analysis comparison with an older therapy, or in some cases with a placebo if relevant. Age-sex specific one-year mortality rates for each patient group are presented in Table G. Linked hospital admission and death data were used to calculate historical (1997) one-year mortality rates in the Netherlands where possible (DG1 AMI, DG2 Unstable angina and DG6 Hospital HF). Previously published data 7 was used for the remaining disease groups where Dutch data was not available to calculate rates.

Treatment component of CHD patients in the IMPACT-SEC model
It was assumed that compliance (adherence), i.e. the proportion of treated patients actually taking therapeutically effective levels of medication, was 100% among hospital patients, 70% among symptomatic community patients, and 50% among asymptomatic community patients taking cholesterol-lowering drugs or blood pressure lowering medication for primary prevention. An adjustment was also made in certain cases for sub-optimal dose.

Risk factor component of IMPACT-SEC model
The second part of the IMPACT-SEC model estimated the number of DPPs related to changes in cardiovascular risk factor levels in the population. The risk factors considered were smoking, total cholesterol, systolic blood pressure, body mass index, diabetes and physical inactivity. Two approaches to calculating DPPs from changes in risk factors were used: the regression approach and change in the Population Attributable Risk Fraction (PARF) approach. These are illustrated below.

Estimating DPPs from risk factor change -regression approach for continuous risk factors
In the regression approach -used for systolic blood pressure (SBP), total cholesterol and body mass index -the number of CHD deaths in 1997 (the start year) after adjusting for population change between 1997 and 2007 were multiplied by the absolute change in risk factor level, and by a regression coefficient ('beta') quantifying the estimated relative change in CHD mortality that would result from a one-unit change in risk factor level (Table I). Natural logarithms were used, as is conventional, in order to best describe the log-linear relationship between absolute changes in risk factor levels and relative change in mortality. Levels of risk factors in 1997 and 2007 by sex and socioeconomic group are shown in Table K. This calculation was then repeated for each age-sex-socioeconomic group. Data sources for the number of CHD deaths are shown in Table B and data sources for risk factors trends in Table H. Sources for the regression (beta) coefficients used in these analyses are listed in Table I. The regression coefficients were assumed equal across socioeconomic groups.

Estimating DPPs from risk factor change -PARF approach for binary risk factors
The PARF approach was used for smoking, diabetes, and physical inactivity. PARF, which can be interpreted as the proportion by which the mortality rate from CHD would be reduced if the exposure were eliminated, 10 was calculated as:

PARF = [P × (RR -1)] / [1 + P × (RR -1)]
Where P is the prevalence of the risk factor and RR is the relative risk for CHD mortality associated with risk factor presence. A RR of 3.3 associated with smoking, for example, expresses the ratio of risk of CHD mortality in smokers to that in non-smokers. DPPs were then estimated as the expected CHD deaths in 2007 (had 1997 mortality rates remained constant) multiplied by the difference in PARF for 1997 and 2007. A negative sign for the DPPs denotes deaths increased or brought-forward due to the increase in diabetes prevalence. The calculation was then repeated for each age-sexsocioeconomic group.
Relative risks estimated by expert working groups for the World Health Organization's Global Burden of Disease 2001 Study were used for smoking and physical activity. 11 Effect estimates were based on systematic reviews of cohort studies (adjusted for regression dilution bias) and meta-analyses of randomised controlled trials. Age-variation in the relative risks for diabetes were taken from the DECODE study. 12 These were then applied to the sex-variation in relative risks estimated by Huxley et al. 13 The published relative risk values for smoking, physical activity and diabetes are shown in Table J. These were adjusted in our study to: a) match the 10-year age bands used in IMPACT-SEC and b) employ a dichotomous rather than trichotomous measure of physical activity. Detailed information on how RRs were modified to fit to the age-sex distributions used in the IMPACT-SEC model can be found in the Scottish IMPACT-SEC supplementary appendix (page 41-47). 2 RRs were assumed constant across socioeconomic groups. Agesex specific RRs are given in Table J.

OTHER METHODOLOGICAL CONSIDERATIONS
Other than calculations to take into account change in treatments and risk factors over time, several other adjustments had to be made.

Accounting for poly-pharmacy
Persons with or at high risk of developing CHD may take a number of different medications. However, data from randomised clinical trials on efficacy of treatment combinations are sparse. Mant and Hicks suggested a method to estimate mortality reduction by poly-pharmacy. 9 The adjustment is carried out in a step-by-step manner as set out in the example below. First the total effect is calculated using an inappropriate additive model, which is then adjusted using effect size calculation with an appropriate multiplicative model.

EXAMPLE 4: Estimation of reduced benefit if patient taking multiple medications (Mant and Hicks approach)
Adjustment for poly-pharmacy in secondary prevention post revascularization in men aged [55][56][57][58][59][60][61][62][63][64] in the most affluent group Taking the example of secondary prevention post myocardial infarction, good evidence (Table F) suggests that, for each intervention, the relative reduction in mortality is approximately: aspirin 15%, beta-blockers 23%, ACE inhibitors (ACE I) 20%, cholesterol lowering drugs 22%, acenocoumarol 22%, and rehabilitation 26%. Our best estimates for uptake were respectively 56%, 46%, 40%, 53%, 10%, and 29%. Assuming a oneyear mortality rate of 3% for men aged 55-64 and a total of 3,060 men aged 55-64 residing in the most affluent group in 2007 the total DPPs, with no adjustment for polypharmacy, would be calculated as shown in the The Mant and Hicks approach suggests that in individual patients receiving all these interventions, mortality reduction is very unlikely to be simply additive. Instead, having considered the 15% mortality reduction achieved by aspirin, the next medication, in this case a beta-blocker, can only reduce the residual mortality (1-0.15). Likewise, the subsequent addition of an ACE inhibitor can then only decrease the remaining mortality, which will be (1-0.15) × (1-0.23). The Mant and Hicks approach therefore suggests that a cumulative relative benefit can be estimated as follows:

Potential overlaps between patient groups: avoiding double counting
To avoid double counting, potential overlaps between different groups of patients were identified and appropriate adjustments made by record linkage or subtracting one group from another. For instance, we can subtract the number of severe heart failure patients treated in hospital from the total number of heart failure patients in the community (because community heart failure patients could be admitted to hospital on one or more occasions). As far as possible record linkage has been used to assign individual patients to only one of the eight disease states; thus avoiding overlaps. A hierarchy of allocation based on mortality was created to assign an individual patient (existing in multiple patient groups) to just one patient group (the one with the highest one-year mortality). The hierarchy structure used was hospitalized HF>hospitalized AMI>hospitalized UA>community HF>community post AMI>community post revascularization. Where this process is not possible then assumptions on overlap adjustment were made showing how potential overlaps were accounted for; these are shown in Table C.

Pharmacological and non-pharmacological contributions to risk factor DPPs
Risk factor improvements, such as lower blood pressure or total cholesterol, may be achieved through medications, lifestyle changes, or a combination. First we calculated the overall number of DPPs due to changes in mean SBP and total cholesterol levels. Then, we calculated the proportion of DPPs that was due to pharmacological contributions. The estimated effect of blood pressure and cholesterol lowering drugs for primary prevention was calculated in a similar way as treatment effects in the treatment component were calculated (eligible population for primary prevention therapy × treatment uptake × relative risk reduction × one-year mortality rate). All DPPs due to risk factor changes were counted in the risk factor component. The proportion of DPPs due to pharmacological contributions was presented separately in the risk factor component.

Negative DPPs (treatments)
In a small number of cases, "negative" DPPs were apparently generated reflecting a decrease in treatment uptake or numbers. For instance with thrombolysis treatments (a larger proportion receiving angioplasty instead of thrombolysis). These negatives were mostly trivial, and were zeroed to reflect the reality: harmful treatments were not being administered. This approach was applied only to disease group (DG)1 AMI in relation to treatment using thrombolysis, DG2 UA in relation to heparin treatment and in DG2 UA and DG5 chronic stable angina in relation to CABG surgery.
3.5 Cumulative risk-reduction: adjusting DPPs to calculate cumulative benefit of multiple risk factor changes CHD deaths are usually caused by multiple risk factors acting simultaneously. Hence, part of the effect of one risk factor may be mediated through another. For example, physical inactivity may have a direct effect on CHD but may also partly be mediated through its effects on BMI and blood pressure. It is recommended therefore that mortality benefits attributable to risk factors which may be causally related, or which overlap in population groups, should not be combined by simple addition. Ideally, their effects should instead be jointly estimated. 14- 18 We do not currently have sources that allow joint estimation of relative risks for combinations of risk factors in this Dutch population. However, several large cohort studies and meta-analyses have published independent risk reduction coefficients for each risk factor included in this study. These are detailed in Tables I and J for continuous and dichotomous risk factors, respectively. One approach commonly used is to calculate the cumulative risk-reduction. 19 This approach accounts for risk factor prevalence overlap but assumes independence of effects. 15,16 The general equation for cumulative risk-reduction is stated as: Thus for CHD risk factors, the specific equation is stated as: where R denotes the mortality change attributable to a specific risk factor. This is in contrast to additive risk-reduction: The adjustment factor is calculated as: Combined effect/Additive risk-reduction. The adjustment factor would always be expected to be less than 1. In other words, cumulative risk factor reduction would be smaller than the mortality benefits arrived at by a simple summation of the benefits of each risk factor in turn. In order to avoid positive and negative R values cancelling each other out in the mathematical application, with the perverse effect of the cumulative benefits being apparently greater than the additive in some instances, we first converted all R values into absolute (i.e. sign-free) numbers. We did this on the understanding that the proportional change in CHD mortality associated with risk factor change was independent of the direction of change. The agesex-socioeconomic adjustment factors fell within the range of 0.78 to 0.97.

Record linkage in the Netherlands
We linked data between national registers using a record identification number assigned to each resident in the Netherlands with a unique combination of birth date, sex and postal code (about 84% of population). Registries and linking procedures used in this study have been described in detail previously. 20 The quality of the national Dutch registers has been previously investigated -the overall quality is high. 21,22 Linkage of individual data between registers was performed in accordance with the privacy legislation in the Netherlands.

Dutch hospital discharge register
Hospital discharge records included information concerning primary and secondary diagnoses, performed medical procedures, and dates of hospital admission and discharge. Hospital discharge diagnoses were coded according to ICD-9. Performed medical procedures were coded using 'Classificatie van Verrichtingen' codes. The Dutch hospital discharge register was available electronically from 1995 onwards.

PHARMO
PHARMO is a database network, i.e. a dynamic cohort study of over 1.4 million persons aged 25 years and over, based on a record linkage system containing drug-dispensing records from community and hospital pharmacies linked with hospital discharge records, as previously described. 23 The drug-dispensing records from hospital and community pharmacies contained information concerning the dispensed drug, dispensing date and the prescription length. All prescription drugs were coded according to the Anatomical Therapeutic Chemical (ATC) classification system. Data was available from 1994 onwards.

HNU
'Huisartsen Netwerk Utrecht' (HNU) is a general practitioner (GP) registry of 5 GP practices comprising around 60,000 patients. The registry data we used were collected from 1996 to up to 2013. Information was collected from HIS data (HIS: GP information system).

RIVM
The National Institute for Public Health and the Environment (RIVM) was the main data source for risk factor values in those aged <65 years. We used data from the Doetinchem Cohort Study. 56 The Doetinchem Cohort study started in 1987-1991 (N=7,768, aged 20-59 years at baseline). The study comprised a physical examination for measurements of body weight, height, body mass index (BMI), systolic and diastolic blood pressure, a nonfasting blood sample (total cholesterol and glucose) and several questionnaires about lifestyle and diet. The overall response rate was 62%. Follow-up examinations were carried out every 5years. The response rates for all follow-up measurements varied between 75% and 80%. Blood pressure was measured twice in each examination in sitting position after 2 minutes of rest. The mean value of two measurements was used in the analyses. We used data from examination 2 (1993-1997) and examination 4 (2003-2007). Blood pressure in examination 4 was measured with a different device and participants sat in a slightly different position during the measurement compared with previous examinations. Therefore, blood pressure measurements in examination 4 were statistically adjusted to make blood pressure values in the different examinations comparable. HNU 'Huisartsen Netwerk Utrectht' (HNU) is a general practitioner (GP) registry of 5 GP practices comprising around 60,000 patients. The registry data we used were collected from 1996 to up to 2013. Information used from HNU for the IMPACT-NL model was diabetes prevalence, defined as an ICPC-code T90 mentioned in the electronic patient record. Information was collected from HIS data (HIS: GP information system).

Systolic blood pressure
We used RIVM data for the age groups 25-54 years and LASA data for the age group of ≥55 years. For the age group of 25-34 years we had no trend data available, therefore we assumed similar trends as in the age group 35-44 years. Due to the small numbers we used a weighted average for all persons above the age of 75 years. Trends in the age groups 75-84 years and 85+ were assumed equal.

Total Cholesterol
We used RIVM data for the age groups 25-64 years and LASA data for the age group of ≥65 years. For the age group of 25-34 years we had no trend data available, therefore we assumed similar trends as in the age group 35-44 years. In LASA the time period between available cholesterol data from wave C and G was 13 year. Because our trend 1997-2007 was 10 years we applied a 10/13 adjustment factor to the change in total cholesterol.

BMI
We used RIVM data for the age groups 25-64 years and LASA data for the age group of ≥65 years. For the age group of 25-34 years we had no trend data available, therefore we assumed similar trends as in the age group 35-44 years.     SBP, systolic blood pressure. a Eligible persons for primary prevention treatment (n= 9,747,083) were defined as all persons who did not have a cardiovascular-related hospital admission during the 5 years and 9 months prior to October 1 in the index year, and did not use nitrates, digitalis glycosides or antithrombotic drugs in the index year. 23 b Despite an increase in BMI in women (age-standardized to the Dutch population), DPPs are almost 0. Changes in BMI are inconsistent in women: all age groups showed an increase in BMI, except for the age groups 55-64 years and 75-84 years.