We constructed a Markov model12
to assess the health effects and associated costs of daily consumption of plain dark chocolate compared with no chocolate in a population with metabolic syndrome without diabetes and initially without cardiovascular disease. The figure depicts the health states included in the model: “alive without cardiovascular disease,” “alive with cardiovascular disease,” “dead from cardiovascular disease,” and “dead from other causes.” We used decision analysis13
to compare two health strategies: no dark chocolate (control) with dark chocolate (treatment). The perspective adopted was that of the Australian healthcare system.
Markov model of effect of dark chocolate consumption versus no consumption and its effects on cardiovascular outcomes in a population with metabolic syndrome
All individuals entered the model in the initial health state of being alive without cardiovascular disease. With each annual cycle, we used risk prediction algorithms and population life tables to determine the probability of an individual transitioning to the other health states—that is, developing non-fatal cardiovascular disease or dying from cardiovascular disease or non-cardiovascular causes. All events were assumed to occur half way through a cycle. Individuals continued to cycle through the model, with more and more moving into diseased or dead states as the period of follow-up increased, until the period of interest (10 years) was reached or death occurred.
In any cycle, we generated random numbers of between 0 and 1.0 to determine if any event occurred. For example, the 100th person in our model was a 44 year old, non-smoking male, whose baseline characteristics were: systolic blood pressure 143 mm Hg, total cholesterol concentration 7.0 mmol/L, high density lipoprotein cholesterol concentration 1.0 mmol/L, glycated haemoglobin (HbA1c) 32.2 mmol/mol, body mass index 33.9, and absence of diabetes. In the first cycle (year 1), his probability of a non-fatal cardiovascular event was 0.27%, fatal cardiovascular event was 0.006%, and non-cardiovascular fatal event was 0.15%. The random number generated for this man in cycle 1 was 0.046. He was therefore assumed to have experienced a non-fatal cardiovascular event and was subsequently moved to the health state of being alive with cardiovascular disease for the start of cycle 2. Had the random number been between 0.27 and 0.27+0.006%, he would have been assumed to have experienced a fatal cardiovascular event and moved to the health state of dead from cardiovascular disease. Had the random number been between 0.27+0.006% and 0.27+0.006%+0.15%, he would have been assumed to have experienced a non-cardiovascular fatal event and moved to the health state of dead from other causes. Had the random number been greater than 0.27+0.006%+0.15%, he would have been assumed to have survived the cycle without a cardiovascular event and been returned to the health state of being alive without cardiovascular disease.
Modelled population and subject data
The population used in the model comprised participants selected from the Australian Diabetes Obesity and Lifestyle (AusDiab) study, among whom cardiovascular risk was estimated individually. Detailed descriptions of the Australian Diabetes, Obesity and Lifestyle study have been published elsewhere.14
Only participants free of cardiovascular disease or diabetes, or both, at baseline, and classified as having metabolic syndrome according to the joint interim guidelines15
(published in 2009) were included in the model. The joint interim guidelines define metabolic syndrome on the basis of three of five risk factors: increased waist circumference with population specific and country specific definitions; triglyceride concentrations ≥150 mg/dL (1.7 mmol/L) or drug treatment for increased triglyceride levels; high density lipoprotein cholesterol concentration <40 mg/dL (1.0 mmol/L) in males and <50 mg/dL (1.3 mmol/L) in females, or drug treatment for reduced high density lipoprotein cholesterol levels; systolic blood pressure ≥130 mm Hg or diastolic blood pressure ≥85 mm Hg or both, or treatment for hypertension; and fasting glucose concentration ≥100 mg/dL or drug treatment for increased glucose levels. We also excluded those who were receiving antihypertensive therapy.
Risks of cardiovascular disease and death
We used Framingham algorithms16
to calculate the baseline risk of non-fatal cardiovascular disease, comprising myocardial infarction and stroke, as well as cardiovascular death. These risks were calculated according to individual specific data on age, sex, systolic blood pressure, total cholesterol level, high density lipoprotein cholesterol level, smoking status, presence or absence of diabetes, and presence or absence of left ventricular hypertrophy. With each annual cycle we recalculated cardiovascular risk according to increases in age and expected changes in systolic blood pressure, total cholesterol level, and high density lipoprotein cholesterol level. Age and sex specific changes to blood pressure and lipid levels were determined according to baseline data stratified by sex and five year age bands. From these we derived annual changes within age bands. We assumed all changes to be linear.
We calculated the risk of death among people with cardiovascular disease using one year mortality data from the Reduction of Atherothrombosis for Continued Health (REACH) registry.17
This prospective cohort study followed people with at least three atherothrombotic risk factors or a history of atherothrombotic disease, or both, for a period of two years, collecting data on morbidity and mortality. As mortality data from the Reduction of Atherothrombosis for Continued Health registry were not specified for subgroups, we made the assumption that all people with cardiovascular disease shared the same risks of death.
We calculated the risks of dying from non-cardiovascular causes using national long term, age and sex specific mortality data from Australia.18
The most recent available data were from 2007. Owing to lack of data stratified by history of cardiovascular disease, we made the assumption that the data were the same for participants with and without cardiovascular disease.
Changes in cardiovascular risk associated with the treatment arm were calculated by application of expected effects of dark chocolate on systolic blood pressure and lipid levels. Data on the blood pressure lowering effects of dark chocolate consumption were gathered from a meta-analysis of 13 randomised controlled trials studying the effect of chocolate or cocoa on blood pressure.9
All trials were longer than 14 days but heterogeneous for levels of flavonoids (range 30-1008 mg/day) administered in various forms (table ). This meta-analysis found that cocoa products rich in flavonols, such as dark chocolate, had a blood pressure lowering effect compared with the control (systolic blood pressure −3.2 mm Hg, 95% confidence interval −5.1 to −1.2 mm Hg, P=0.001; diastolic blood pressure −2.0 mm Hg, −3.4 to −0.7 mm Hg, P=0.003). When participants were stratified according to hypertension, blood pressure was significantly reduced in those with hypertension (systolic blood pressure −5.0 mm Hg, −8.0 to −2.1 mm Hg, P<0.001; diastolic blood pressure −2.7 mm Hg, −4.9 to −0.6 mm Hg, P=0.01), but not significantly reduced in those in a normotensive state (systolic blood pressure −1.6 mm Hg, −3.8 to 0.7 mm Hg, P=0.17; diastolic blood pressure −1.3 mm Hg, −2.9 to 0.3 mm Hg, P=0.12). Some heterogeneity existed among the treatment effects observed between trials (systolic blood pressure I2
=74%; diastolic blood pressure I2
=62%). This remained high in the hypertensive subgroup (I2
=79%) but was reduced in the normotensive state (I2
Table 1 Characteristics of trials included in meta-analyses from which the blood pressure lowering and cholesterol lowering effects of dark chocolate therapy were drawn
Effects of dark chocolate consumption on lipid profiles were informed by a meta-analysis of eight short term trials investigating the effect of cocoa on healthy participants (table 1).11
It showed that short term (2-18 weeks) consumption of dark chocolate decreased low density lipoprotein cholesterol concentrations (−0.15 mmol/L, 95% confidence interval −0.29 to −0.02 mmol/L, P=0.03) but had no significant effect on total cholesterol concentrations (−0.15 mmol/L, −0.32 to 0.02 mmol/L, P=0.08) or high density lipoprotein cholesterol concentration (0.03 mmol/L, −0.07 to 0.13 mmol/L, P=0.56). Subgroup analyses according to health status, however, showed that dark chocolate consumption could significantly reduce both total cholesterol and low density lipoprotein cholesterol levels in those with high cardiovascular risk (total cholesterol 0.21 mmol/L, 95% confidence interval −0.35 to −0.06 mmol/L, P=0.007; low density lipoprotein cholesterol −0.20 mmol/L, −0.38 to −0.01 mmol/L, P=0.04). In subgroup analyses, high density lipoprotein cholesterol levels did not change significantly.
Table 2 summarises the treatment effects assumed in the model.
Table 2 Data inputs used in Markov model of the effect and cost effectiveness of dark chocolate consumption in a population with metabolic syndrome free of cardiovascular disease and diagnosed diabetes
Costs of cardiovascular events were taken from a review on the cost of cardiovascular complications in a “healthy” population.19
They are summarised in table 2. Costs included direct costs of myocardial infarction and stroke, measured for the first year of the event and after the first year. We took the cost of a fatal cardiovascular event from an economic analysis of chronic diseases in Australia.20
All costs were inflated to reflect 2012 costs according to the Australian national health price index.21
The outcomes of interest were number of events prevented, number of life years saved, and potential monies available for prevention strategies provided the incremental cost effectiveness ratios met arbitrary thresholds, in terms of Australian dollars per year of life saved. We calculated the number of deaths prevented, by determining the difference in number of deaths between those consuming and not consuming dark chocolate. Similarly, we calculated the years of life saved by determining the number of life years gained from dark chocolate consumption compared without.
We calculated incremental cost effectiveness ratios by comparing the difference in net costs between treatment (a diet rich in cocoa products) and no treatment (control), divided by the difference in years of life lived by each cohort. The time horizon of the modelled analysis was 10 years.
All future benefits (years of life lived) and costs were discounted at 5% per annum.22
All analyses were also done using compliance levels of 80%, 90%, and 100%.
Probabilistic sensitivity analyses
Probabilistic sensitivity analyses23
were undertaken using 95% confidence intervals surrounding the point estimates for blood pressure lowering effect and lipid change and 10% uniform distributions around cost inputs. We assessed the effects of the uncertainty surrounding point estimates simultaneously by 1000 iterations of Monte Carlo simulation.24
Table 2 outlines the uncertainty ranges applied to key model inputs.