Forecasting Tunisian type 2 diabetes prevalence to 2027: validation of a simple model

Background Most projections of type 2 diabetes (T2D) prevalence are simply based on demographic change (i.e. ageing). We developed a model to predict future trends in T2D prevalence in Tunisia, explicitly taking into account trends in major risk factors (obesity and smoking). This could improve assessment of policy options for prevention and health service planning. Methods The IMPACT T2D model uses a Markov approach to integrate population, obesity and smoking trends to estimate future T2D prevalence. We developed a model for the Tunisian population from 1997 to 2027, and validated the model outputs by comparing with a subsequent T2D prevalence survey conducted in 2005. Results The model estimated that the prevalence of T2D among Tunisians aged over 25 years was 12.0% in 1997 (95% confidence intervals 9.6%–14.4%), increasing to 15.1% (12.5%–17.4%) in 2005. Between 1997 and 2005, observed prevalence in men increased from 13.5% to 16.1% and in women from 12.9% to 14.1%. The model forecast for a dramatic rise in prevalence by 2027 (26.6% overall, 28.6% in men and 24.7% in women). However, if obesity prevalence declined by 20% in the 10 years from 2013, and if smoking decreased by 20% over 10 years from 2009, a 3.3% reduction in T2D prevalence could be achieved in 2027 (2.5% in men and 4.1% in women). Conclusions This innovative model provides a reasonably close estimate of T2D prevalence for Tunisia over the 1997–2027 period. Diabetes burden is now a significant public health challenge. Our model predicts that this burden will increase significantly in the next two decades. Tackling obesity, smoking and other T2D risk factors thus needs urgent action. Tunisian decision makers have therefore defined two strategies: obesity reduction and tobacco control. Responses will be evaluated in future population surveys. Electronic supplementary material The online version of this article (doi:10.1186/s12889-015-1416-z) contains supplementary material, which is available to authorized users.


The Model
The purpose of the MEDCHAMPS IMPACT Diabetes model is to provide estimates of future diabetes prevalence and offer a modeling platform for policy decision making.
We estimated the overlaps using the following approach: First, we estimated the conditional probability of being a smoker and obese, and assigned those individuals to the smoker group.
Second, we estimated the proportion of diabetes persons among the obese population using a population attributable risk approach and discount them from the obese pool. Finally the healthy group was calculated as the population -obese -smoker-diabetes pools.

Methods Overview
The model integrates information on population, obesity and smoking trends at a given point in time to estimate diabetes prevalence in the future.
The population is partitioned in three states (healthy, obese and smokers) and from them, number of diabetes patients and diabetic and non-diabetic deaths are estimated for subsequent time periods using a Markov approach.
The effect of policy decisions can be modelled by the estimated effect on risk factors trends, and the trend parameter can be modified to model increasing, decreasing or stable trends in the prevalence of obesity and or smoking. Another way of exploring policy options (like, for example, multifactorial lifestyle diabetes prevention interventions) can be modelled through the modifications of the diabetes incidence parameter.

The Model Structure
Models are simplifications of reality. In order to keep the model simple but at the same time useful, many compromises on the way the disease epidemiology is modelled are necessary. More complex models usually require different approaches and an amount of data that probably will not be available for the participating countries. A graphical description of the model is presented in figure 1 in the text.
We assume that the population can be divided in several pools: Diabetes mellitus, Obese, Smoker and "healthy" (eg: non obese, non smokers, non diabetics). A proportion of the population in each pool moves through pathways to other states as described in figure 1.
Population demographic trends are used to inform the relative size of the "starting states", and transition probabilities are used to estimate the proportion of persons moving from the starting states to the diabetes and death states. There are two "absorbing states": Diabetes Mellitus (DM) related death and Non DM related deaths. In this way, mortality competing risks are modelled. Potential overlaps between the healthy, obese and smoking group are managed by calculating the conditional probabilities of membership.

The Model Workbook
The workbook is a MS Excel spreadsheet, structured in tabs. These tabs serve different purposes, but for the end-user the key tabs are the Data Input, Dashboard and Validation tabs.
The following is a more detailed description of each tab and its purpose, with a more thorough description of the key ones.

The Data Input tab:
The following sections are: Population data: used to input the age and gender structure of the population being modelled, and the populations projections Morbidity data: cross sectional data on diabetes prevalence, obesity and smoking is provided. Trend parameters can be set up here also or use the defaults. Currently, linear trends can be modelled, but any other type of trend can be implemented.

The Dashboard tab:
Figures and tables presenting the model estimated diabetes prevalence are presented here.

The validation tab:
This tab summarizes important validation information for the country, if the validation exercise has been conducted. .

Outputs:
The raw outputs of the model and the sensitivity analysis is stored here. Available information: numbers of diabetes patients, diabetes prevalence, minimum and maximum estimates.

SA layer:
Calculations for estimating transition probabilities are performed here. Markov Chains: They are implemented in separate tabs for each gender and age group. Two sets of chains are implemented here for the baseline and scenario runs.

Sensitivity Analysis:
We used the analysis of the extremes method (Briggs), consisting in running the model with all parameters set to a minimum and maximum realistic values. This is a very conservative approach, but allows a more transparent understanding of the weight of each parameter regarding model outcomes. The Sensitivity Analysis is updated by running a macro. Details of the data used are found in the data sources section (Section 4).

Deriving model parameters
One of the key aims of the model is to use the minimum data requirements possible. We

Diabetes incidence and specific mortality
The MEDCHAMPS Diabetes Markov model use diabetes mellitus incidence and mortality as one of the critical data inputs that need to be provided by the participants countries, to help localize and calibrate the model to each population.
Since reliable and country specific sources of incidence data are probably not available, an estimate of it is needed.
We adapted a method to estimate it, and provide as an example, the estimation of baseline diabetes incidence for Tunisia.

The method
Incidence, mortality and prevalence are closely related to each other, in a way that only some values for each parameter are consistent with the other parameters at a given time. This The technique use as input whatever parameters that are known and using a multistate generic disease model using a lifetable markov approach, estimate revised parameters for the inputed ones and estimates for those unknown. This method has been implemented in software called DISMOD II (http://www.who.int/healthinfo/global_burden_disease/tools_software/en/) For the MEDCHAMPS project, it is expected that the only available parameter is probably diabetes mellitus prevalence (either self reported or using ADA/NHANES definitions).
However, diabetes excess mortality can be estimated from total mortality data (See Barendregt) using literature based estimates of mortality relative risk and disease prevalence, and we can safely assume that the remission rate for diabetes in effectively 0.
Thus, the only parameters needed (by age and gender) are diabetes mellitus prevalence, population structure and population general mortality.
An important assumption is that this method requires a population in equilibrium, since the consistency between epidemiological estimates depends on the underlying trends in each parameter. However it is difficult to disentangle these effects from data inaccuracy. The robustness of the approach to violations of these assumptions is not known.
This method produces a "population incidence", eg, the incidence both for exposed and unexposed people to diabetes risk factors.
However, the MEDCHAMPS diabetes model needs incidence in the non exposed, since incidence for obese persons and smokers is derived from that baseline incidence by using literature based relative risks.
It has been proposed that the incidence of a disease in a population is a weighted sum of the incidence among the exposed and the incidence among the unexposed to a risk factor(Epidemiology By Moyses Szklo, F. Javier Nieto, equation 3.8 in page 101) (equation 1).

(Equation 1)
, Where ip is the population incidence, ie is the incidence amongst the exposed, iu is the incidence amongst the unexposed and p is risk factor prevalence.
Since the incidence in the exposed is the incidence in the unexposed times the relative risk (RR) (Equation 2),

(Equation 2)
, it is possible to derive from this two ideas the value for the unexposed incidence from the incidence in the population. Replacing equation 2 in equation 1

Estimation of the incidence, case fatality and mortality parameters for Tunisia, 1997
This section describes the method used to estimate diabetes mellitus type II incidence for the Tunisian population in 1997.
DISMOD need at least 3 inputs. For this case, we used diabetes mellitus prevalence, diabetes mellitus remission rate and diabetes mellitus relative risk for mortality.

Diabetes prevalence was obtained from Tunisian National Nutrition Survey 1996/97
Tunisian National Nutrition Survey 1996/97: The survey was cross-sectional from June 1996 to December 1997 on a nationally representative sample (5669, 200 inhabitants in 1995): 1735 households with a total of 5815 adults over 20 years old.
 Definition: Type 2 diabetes cases were defined as subjects with measured Fast Plasma Glucose >=7 mmol/l or having a previously diagnosed diabetes.
We can safely assume that diabetes mellitus remission rate is 0, and diabetes mellitus relative risk for mortality can be estimated as proposed by Barendregt et al, based in the usual RR for mortality (mortality in diseased/mortality in non diseased) and disease prevalence. The formula is Where RR adj is the relative risk mortality, RR is the usual relative risk for mortality (mortality diseased/mortality healthy) and p is disease prevalence.