The potential impact of diabetes prevention on the future burden of dementia and disability

Aims/hypothesis Diabetes is associated with an increased risk of dementia. We estimated the potential impact of trends in diabetes prevalence upon mortality and the future burden of dementia and disability in England and Wales. Methods We used a probabilistic multi-state, open cohort Markov model to integrate observed trends in diabetes, cardiovascular disease and dementia to forecast the occurrence of disability and dementia up to the year 2060. Model input data were taken from the English Longitudinal Study of Ageing, Office for National Statistics vital data and published effect estimates for health-state transition probabilities. The baseline scenario corresponded to recent trends in obesity: a 26% increase in the number of people with diabetes by 2060. This scenario was evaluated against three alternative projected trends in diabetes: increases of 49%, 20% and 7%. Results Our results suggest that changes in the trend in diabetes prevalence will lead to changes in mortality and incidence of dementia and disability, which will become visible after 10–15 years. If the relative prevalence of diabetes increases 49% by 2060, expected additional deaths would be approximately 255,000 (95% uncertainty interval [UI] 236,000–272,200), with 85,900 (71,500–101,600) cumulative additional cases of dementia and 104,900 (85,900–125,400) additional cases of disability. With a smaller relative increase in diabetes prevalence (7% increase by 2060), we estimated 222,200 (205,700–237,300) fewer deaths, and 77,000 (64,300–90,800) and 93,300 (76,700–111,400) fewer additional cases of dementia and disability, respectively, than the baseline case of a 26% increase in diabetes. Conclusions/interpretation Reducing the burden of diabetes could result in substantial reductions in the incidence of dementia and disability over the medium to long term. Electronic supplementary material The online version of this article (10.1007/s00125-019-05015-4) contains peer-reviewed but unedited supplementary supplementary material, which is available to authorised users.


IMPACT-BAM model diagram
ESM Fig. 1 presents IMPACT-BAM model structure. Detailed description of the baseline model, states definition, and calculation of transition probabilities have been previously described [1,2].

Mortality projections
For the present study, future projections for cardiovascular and non-cardiovascular mortality rates by age and sex were estimated based on observed mortality rates reported by the ONS up to 2016. P-spline smoothed lines [3][4][5] were fitted to logarithmic transformed CVD and non-CVD mortality rates in each 5-year age band from 1990 to 2016 by sex using the pspline function is Stata (StataCorp. 2017. Stata Statistical Software: Release 15. College Station, TX: StataCorp LLC). Change in P-spline smoothed values of log-transformed mortality rates over subsequent years followed a normal distribution. To obtain missing values for change in mortality rates in future years, linear regression models were fitted to the change in log-transformed mortality rates over subsequent years by sex with interaction terms for 5-year age groups. Linear predictions calculated from the linear regression coefficients represented predicted change in log-mortality rates for subsequent years which were used to project mortality rates to the future until 2060. For all age-groups P-spline smoothed log-transformed mortality rates declined over time but the magnitude of decline also declined, resulting in projected mortality rates reaching a plateau in the decades to come. Standard error for the linear predictor was calculated as time in years multiplied by the standard error of the linear prediction.This method closely matches the method used for the official mortality projections used by the ONS [6]. All-cause mortality rates projected to the future using the described method closely matched all-cause mortality rate projections from the ONS (not shown) The projected CVD and non-CVD mortality rates men and women are presented in ESM figures 2-5.

Validation of the model against observed data
We carried out partially-dependent validation of our estimates of CVD and Non-CVD deaths with observed ONS mortality data reported for England & Wales for the period 2006-2016. Using the definition suggested by the ISPOR-SMDM Modeling Good Research Practices Task Force [7], partially-dependent validation occurs when the external source to which the output is being compared to was used to build a part of the model, but it does not wholly determine the outcome to be validated. Because of above, this validation confirms internal consistency of the model rather than real-world validity of the projections. The model provided a good match to the ONS estimates of the number of CVD and Non-CVD deaths (ESM Fig. 6 and ESM Fig. 7).
We carried out independent validation (i.e. no information from these sources was used to build the model) of our model estimates of the prevalence of CVD and dementia. Our estimates of CVD in 2011 for men fall within the 95% confidence intervals reported by the HSE [8]. However, our model estimates a slightly higher prevalence of CVD in women (ESM Fig. 8).
Our age-specific estimates of dementia prevalence in 2011 were akin to those reported in CFAS II for the same year (ESM Fig. 9). Most of our estimates of age-specific prevalence fall within the 95% confidence interval reported by CFAS II. The only exceptions were for women 85+, where our estimates were lower than those from CFAS.

Policy layer
This latest version of IMPACT-BAM evaluates the impact of changes in risk factors at the population level (due to hypothetical policy interventions) on future cases of dementia, disability, CVD and mortality.

Basic concept
We modified relevant transition probabilities in the baseline model according to assumed changes in specific risk factors using a population attributable risk fraction (PARF) approach. The PARF calculates the proportion by which disease burden would be reduced if the prevalence of a risk factor was reduced to zero. Symbolically, PARF is the following 6 : Where P is the diabetes prevalence, and RR is the disease risk ratio. P and RR are age and sexspecific. In this paper, we are interested in how this PARF varies because of changes in diabetes prevalence. Symbolically, this would be: Where P' is the prevalence of the disease after the intervention. This equation is equivalent to the potential impact fraction (PIF) equation for discrete risks factors generally reported in publications [9].
To obtain the RRs describing the association between diabetes and incidence of dementia, the incidence of recovery from functional impairment, CVD incidence/mortality and non-CVD mortality, we conducted a systematic review and meta-analyses (See section 4.2). The RRs obtained from these meta-analyses were then adjusted by the duration of diabetes (see section 4.3)

Methods
PubMed was searched using the following search strategy to identify studies reporting the association between diabetes and the incidence of dementia or incidence of recovery from functional impairment:

AND Humans[Mesh] AND English[lang]
References for retrieved relevant publications were hand searched for any papers that may have been excluded from the PubMed search. Studies were included in the meta-analysis if they were prospective, cohort, or longitudinal studies; published in English; conducted in Europe, North America, or Australia; dementia was ascertained using DSM III, DSM-IV, or NINDES-AIRNEN criteria; disability/functional impairment was ascertained by impairment in independently conducting one or more basic activities of daily living (getting in or out of bed, cutting food and eating, using the toilet, bathing/showering, putting on clothes including shoes and socks, walking across the room) and the study reported the relative risk or hazard ratio of incident dementia or incidence functional impairment in individuals with diabetes compared to those without. The relative risk or hazard ratio with the maximum level of adjustment was included in the metaanalysis. Studies were excluded if the follow up ended before 1990 or if an updated version of the study was later published.
Meta-analyses were conducted using the metan function of the STATA software, version 15, to obtain a pooled estimate of the association between diabetes and dementia or functional impairment. Study weights in the meta-analysis were assigned in proportion to the person-years of follow up in each study. Heterogeneity between studies was assessed using the I2 statistic. Although the I2 statistic was high for the meta-analysis related to the dementia outcome, there was not considerable heterogeneity in terms of the design and quality of the studies, method for assessment of exposure or outcome, ethnicity, age and sex structure, or the results obtained. A fixed effects meta-analysis was thus conducted for both outcomes. The difference between a fixed versus random effects meta-analysis on the pooled estimate was small.

Dementia
The search strategy returned 732 titles published by December 2017. Sixty-three titles were found to be relevant, including two studies identified by hand searching the references, and the full texts were examined by two independent reviewers (HW and SAA). Twenty-two studies met the inclusion and exclusion criteria and were included in the meta-analysis .
The results of the meta-analysis are presented in ESM Fig. 10

Functional impairment
The search strategy returned 521 titles published by December 2017. Twenty-six titles were found to be relevant, and the full texts were examined by two independent reviewers (HW and SAA). Twelve studies met the inclusion and exclusion criteria and were included in the metaanalysis [32][33][34][35][36][37][38][39][40][41][42][43]. The measures of association in all studies were adjusted for age, sex and education or socioeconomic status. Several studies additionally adjusted for the confounding effects of body mass index (BMI) and other known risk factors [35-38, 40, 42]. The results of the meta-analysis are presented in ESM Fig. 11. The pooled relative risk of functional impairment associated with diabetes was 1.46 (95% CI 1.33, 1.90) in a meta-analysis of all 12 studies. Among studies that additionally adjusted for BMI, the pooled relative risk was 1.52 (95% CI 1.30, 1.74).
ESM Fig. 11. Forest plot is summarising studies investigating the association of diabetes and functional impairment in independently conducting one or more activities of daily living.

Recovery from functional impairment
One study (UK: English Longitudinal Study of Ageing) was identified that reported the RR of recovery from functional impairment comparing individuals with and without diabetes (RR of recovery 0.93 (95% CI 0.86-1.00)) [33].

CVD incidence / CVD Mortality
The Emerging Risk Factors Collaboration reported individual-level meta-analysis for 698 782 people (52 765 non-fatal or fatal vascular outcomes; over 8·49 million person-years of follow up) from 102 prospective studies [44]. The Pooled RR for Non-fatal myocardial infarction in individuals with diabetes compared to those without was 1.82 (95% CI 1.64-2.03). The corresponding figure for death from coronary heart disease was 2.31 (95% CI 2.05-2.60), similar to the pooled RR of vascular deaths (2.32 (95% CI, 2.11 to 2.56)) reported in a separate metaanalysis [45].

Non-CVD Mortality
The Emerging Risk Factors Collaboration reported an individual level data meta-analysis on the association of diabetes with cause-specific deaths among 820,900 individuals over a total of 12.3 million person-years of follow up in 97 prospective studies [45].

Multilevel exposure
There is evidence suggesting the risk of unfavourable outcomes of diabetes strongly depends on the duration of the disease. For example, ADVANCE study reported increase in risk among diabetic subjects with the longer duration of diabetes: [47] • for macrovascular events: 17% (12%-22%) for each 5-years of diabetes duration • for microvascular events: 31% (26%-36%) for each 5-years of diabetes duration • for all-cause death: 21% (15%-26%) for each 5-years of diabetes duration Therefore, in IMPACT-BAM, we treated diabetes as a multilevel exposure risk factor, using different categories of diabetes duration as different exposure levels. We explained how we project diabetes prevalence and duration the multilevel PARF formula in section 4.4. The previous PARF equation in section 4.1 considers only two levels of risk factor exposure (exposure =0 and exposure ≠0). Thus, we used the following extension to consider multilevel exposure 5 : Subscript i refers to the ith exposure level. Pi=prevalence of the risk factor in ith exposure level, RRj= relative risk comparing ith exposure level with unexposed group. Pi and RRi are age and sexspecific.
Then, ∆ is The RRs obtained from our literature review and meta-analyses described in the previous section were not stratified by diabetes duration. Therefore, to account for diabetes duration, we corrected RRs for longer than 5 years duration of diabetes by expected increase in risk based on data from ADVANCE (see above). This assumption might result in overestimation of the effect size.
The following table (ESM Table 1) reports the unadjusted RRs from our literature review and meta-analyses and the adjusted relative risks values used in the PARF approach.

Projection of future diabetes trends
We evaluated three potential future scenarios of trends in diabetes duration and compared them to a baseline scenario which assumes the continuation of the current obesity trend. Increasing prevalence of diabetes is mostly driven by the obesity epidemic. To obtain reasonable scenarios of possible future trends in diabetes, we calculated the expected change in diabetes prevalence due to possible changes in the obesity trend in England.
We used the already existing Diabetes Prevalence Model, published by Public Health England to translate changes in obesity into trends in diabetes prevalence [48].
The baseline scenario assumes the continuation of current obesity trends at the rate of 1% per 5 years, and the corresponding expected relative increase in diabetes prevalence is 26% between 2015 and 2060. The other scenarios assume: 1. Acceleration of obesity with the trend increasing to 5% per 5 years, increasing diabetes prevalence by 49% between 2015 and 2060 2. Halt to any further increase in obesity, resulting in a slower increase in diabetes prevalence of 20% between 2015 and 2060 3. Reversing current obesity trend (decrease obesity at 3% per 5 years); this will not immediately decrease the prevalence of diabetes, resulting in a relative increase in diabetes prevalence of 7% between 2015 and 2060 (see ESM Fig. 12).
The projection of diabetes prevalence trend was then smoothed to obtain diabetes prevalence values stratified by single year. As Diabetes Prevalence Model allowed to forecast up to 2035, We extended the projections until 2060 using local polynomial regression (loess() function of R package, with span parameter = 10 and degree of polynomials = 2). The result of this process is shown in ESM Fig. 12 and manuscript Table 1: PHE model does not allow to project diabetes prevalence stratified by age and sex. We used then age/sex gradient for the prevalence of diabetes from another source, assuming this gradient will be similar across calendar years of the projections. A Canadian study data was used to obtain this gradient since no English national studies known to authors reported diabetes prevalence for older age groups [49]. Then the distribution was smoothed across age groups to obtain single age-stratified values.

Projecting the distribution of duration of diabetes
Notice that the extended PARF formula also requires age and sex-specific estimates of the risk factor prevalence at each exposure level. For any given year, sex and age, the number of diabetics is composed of individuals with different time spans living with the disease. For example, the overall prevalence of diabetic men aged 60 in 2040 is composed of individuals who have lived with diabetes less than five years, 5-9 years, 10-14 years, 15-19 years, 20-24 years and more than 25 years.
We used the 2014 Health Survey for England (HSE) data to obtain age and sex-specific distributions of diabetes prevalence across these six categories of diabetes duration.
For the baseline scenario, we assumed that the prevalence in each of the categories would remain constant in the future. For example, that the prevalence of diabetic men aged 60 in 2040 with a diabetes duration of 10 years is equal to the prevalence of diabetic men aged 60 in 2014 (HSE data) with a diabetes duration of 10 years.
For the scenarios, we assumed that as the result of changes in prevalence of obesity, there would be an "excess" of cases of diabetes compared to the baseline. This is visualized on the ESM Fig.  13 as the shaded area between red and black curves. These new cases will propagate across time following the ageing of their cohort. For example, the prevalence of diabetics among men aged 60 in 2040 living with the disease for more than 25 years is equal to the new cases diabetes in men aged 35 in 2015. The prevalence of diabetics among men aged 60 in 2040 living with the disease for 20 years is equal to the new cases of diabetes in men aged 40 in 2020, and so on. Formally, this is: Where , , , is the prevalence of diabetics sex s, age a, in year t and with d years living with the disease. To calculate the new cases, we calculate first the difference between the scenariospecific diabetes prevalence and the baseline diabetes prevalence by sex, age and calendar year: , , . Then we assumed that new cases are equal to: