Global mortality from dementia: Application of a new method and results from the Global Burden of Disease Study 2019

1While there is a wealth of evidence indicating that individuals with dementia have a higher risk for mortality, the mechanisms by which dementia leads to death are less clear. The International Classification of Disease (ICD) guidelines on the certification of causes of death define the underlying cause of death as the condition which gave rise to all other conditions leading to the death of the individual. Prior evidence indicates that conditions such as bronchopneumonia and pulmonary embolisms are often the immediate causes of death among those with dementia and that these conditions are more frequent in those with dementia than in those without dementia, suggesting that dementia may actually be the underlying cause of death in many of these cases.,, However, oftentimes, these immediate causes of death are listed on the death certificate without mention of dementia.

2Due to these challenges in the certification of deaths due to dementia, the use of vital registration data for the estimation of mortality due to dementia is highly susceptible to changes over time in coding practices that exist within and between vital registration systems and the societies in which they operate. Previous studies have shown increases over time in the certification of dementia as a cause of death on death certificates, while the estimated prevalence of the disease has remained stable or even decreased.,, Comparisons of the certification of dementia deaths on death certificates with information on dementia patients from population‐based epidemiological studies have indicated that dementia has been previously underreported on death certificates.,,,,

3Published estimates of dementia mortality are largely from the United States and Western Europe and have either focused on the estimation of mortality in those with dementia rather than mortality that can be ascribed to dementia as the underlying cause, or have relied on the generalization of prevalence and mortality risk estimates from one population‐based cohort to the entire country.,, The only available global estimates of dementia mortality are from previous iterations of the Global Burden of Disease (GBD) study. However, these estimates are subject to a number of known limitations. For GBD 2010, estimates were based on vital registration data. The recognition of the biases that exist in reporting dementia as a cause of death in vital registration data led to a new methodology for the estimation of dementia mortality, introduced in GBD 2015. For GBD 2015, 2016, and 2017, we used estimates of excess mortality derived from countries most likely to code to dementia as a cause of death per prevalent dementia case, and prevalence estimates to calculate mortality due to dementia., While this method allowed for the global estimation of dementia deaths and accounted for the under‐reporting present in vital registration data, it was sensitive to the choice of how many countries to include in the estimation of excess mortality, and it assumed that the coding practices in these countries do not under‐ or over‐assign dementia as a cause of death.

4This study aims to improve on prior methods of estimation for dementia mortality and describe the results over time and across geographies. By removing all reliance on vital registration data and instead calculating mortality due to dementia as the product of total excess deaths and the proportion of these deaths in individuals with severe, end‐stage disease, this study addresses some of the primary weaknesses in previous estimation methods.

5Systematic review: The authors reviewed the literature using traditional sources (PubMed). Prior estimates of dementia mortality are from previous iterations of the Global Burden of Disease study or are limited in geographic scope.Interpretation: Our study improves on prior methods for the estimation of mortality by removing any reliance on inconsistent vital registration data from the modeling process. We estimated there were 1.62 million (0.41–4.21) deaths due to dementia globally in 2019.Future directions: Future work should seek to strengthen the analyses by incorporating additional data sources on dementia prevalence, the excess risk of death, and the proportion of individuals with end‐stage disease, both to limit the effect of data sparsity on the precision of estimates and to increase the geographic coverage of available data. More precise and accurate estimates will increase the utility of the results for the purposes of public health planning and resource allocation.

6Systematic review: The authors reviewed the literature using traditional sources (PubMed). Prior estimates of dementia mortality are from previous iterations of the Global Burden of Disease study or are limited in geographic scope.

7Interpretation: Our study improves on prior methods for the estimation of mortality by removing any reliance on inconsistent vital registration data from the modeling process. We estimated there were 1.62 million (0.41–4.21) deaths due to dementia globally in 2019.

8Future directions: Future work should seek to strengthen the analyses by incorporating additional data sources on dementia prevalence, the excess risk of death, and the proportion of individuals with end‐stage disease, both to limit the effect of data sparsity on the precision of estimates and to increase the geographic coverage of available data. More precise and accurate estimates will increase the utility of the results for the purposes of public health planning and resource allocation.

9The category of dementia as used in this article is equivalent to the GBD disease designation of “Alzheimer's disease and other dementias.” Information on general GBD methods can be found in the GBD 2019 summary papers., This GBD study used de‐identified data, and the waiver of informed consent was reviewed and approved by the University of Washington Institutional Review Board (Study 9060).

10The goal of this study was to estimate the number of individuals globally who died of dementia as an underlying cause of death, in line with ICD‐10 principles. To conduct this analysis, we first estimated the attributable risk of all‐cause mortality via systematic review and meta‐analysis, and multiplied this estimate by the number of individuals with dementia to calculate excess deaths due to dementia. However, not all excess deaths among individuals with dementia are likely attributable to dementia as an underlying cause of death due to the presence of common comorbid conditions such as cardiovascular diseases. Therefore, to calculate deaths attributable to dementia as an underlying cause of death, we multiplied excess dementia deaths by the proportion of individuals who died with end‐stage dementia out of the total number of excess dementia deaths, using Formula 1:

11The overall analytic framework can be visualized in Figure 1. This strategy assumes that individuals who die with conditions signaling end‐stage dementia die from dementia as an underlying cause of death. This assumption is supported by prior research indicating that end‐stage conditions such as pneumonia, febrile illness, and eating problems are associated with extremely high short‐term mortality rates among those with advanced disease.

12Through a PubMed literature review focused on dementia excess mortality and relative risk spanning 1980 to 2018, we identified 4470 total hits, of which 34 studies were marked for extraction (additional details in supporting information).,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, The data sources included clinical or community‐based cohorts and excluded studies conducted solely in nursing homes. Studies reporting odds ratios, hazard ratios, or relative risks of death given dementia or similar exposure were extracted. Although our case definition was either the Diagnostic and Statistical Manual of Mental Disorders (DSM) or ICD definitions of dementia based on clinical ascertainment by a physician in a population‐representative study, we accepted a wider range of exposures including Alzheimer's disease (AD) and cognitive impairment and adjusted for differences in our modeling strategy to increase data volume (Table 1).

13Using the data on relative risks extracted from each study and the GBD estimate of all‐cause mortality rate for the given study location and time period, we calculated attributable risk of all‐cause mortality. By multiplying the excess risk in the exposed (relative risk — 1) by the event rate in the unexposed (approximated using all‐cause mortality rates) we estimated the all‐cause mortality rate difference, or attributable risk, using Formula 2:

14We then conducted a Bayesian meta‐regression on the attributable risk data. The model includes covariates to predict between‐study heterogeneity and adds uncertainty from unexplained heterogeneity to coefficient estimates and predictions. Additionally, the model incorporates automatic outlier identification as part of the likelihood function, which identifies data with implausible combinations of means and variances, such that even if a data point is far from the mean, it may not be identified as an outlier if it has a high variance (additional details in supporting information).

15The meta‐regression included covariates for the exposure used in each study (all dementia, AD, cognitive impairment), whether the study was conducted in a clinical sample, and five indicators for categories of variables commonly adjusted for in the component studies. These indicator variables described whether each component study of the meta‐regression adjusted for educational attainment, basic cardiovascular disease comorbidities (e.g., stroke and heart disease), more extensive cardiovascular disease factors (e.g., blood pressure and cholesterol), and smoking and alcohol consumption. Additionally, some studies adjusted for factors including activities of daily living, or residence in nursing homes, which may be on the causal pathway between dementia exposure and death, and these were grouped together into a dummy variable for the category of “over‐controlling.” Attributable risk was estimated from this model by generating predictions by age group for the mean population‐based study on dementia, which did not control for education, cardiovascular disease, or other factors.

16We used all available data on the prevalence of dementia from cross‐sectional studies, cohort studies, and administrative claims databases for the estimation of dementia prevalence (additional details in supporting information; data sources are available at the Global Data Health Exchange [http://ghdx.healthdata.org/]). Using DisMod‐MR 2.1, the Bayesian meta‐regression tool primarily used in nonfatal modeling for GBD with settings of no remission and no incidence before age 40, we generated estimates for the prevalence of dementia by age, sex, year, and location (additional details in supporting information). We then calculated the total number of excess deaths due to dementia as the product of prevalence and attributable risk.

17Our estimate of the total number of excess deaths in those with dementia likely includes deaths that should be attributed to other conditions, such as cardiovascular diseases, that are more common in those with dementia due to shared risk factors such as high blood pressure or lower educational attainment. To calculate the number of deaths caused by dementia from the number of excess dementia deaths, we used mortality records linked to inpatient records, covering deaths from 2003 to 2017 in the Emilia Romagna region of Italy, and looked for markers of severe, end‐stage disease up to 1 year before death. A similar strategy of identifying cases of end‐stage disease using ICD codes has been implemented using Medicare data in the United States. To select these markers, for each ICD code in the data we calculated the difference in the proportion of individuals who died with dementia and had a record of such a code in the year before death and the proportion of individuals who died without dementia and had a record of the same code in the year before death. We defined dementia as any ICD‐9 code for dementia (290–290.4, 294.1–294.2, 294.8–294.9, 331.0–331.2), in either inpatient records or death records. We reviewed the 150 codes with the highest difference and selected codes that could indicate end‐stage disease, excluding codes for conditions such as cardiovascular disease or cancers that cannot plausibly be caused by dementia. Codes for decubitus ulcer, malnutrition, sepsis, pneumonia, urinary tract infections, falling from bed, senility, dehydration, sodium imbalance, muscular wasting, bronchitis, dysphagia, hip fracture, and bedridden status were used as indicators of severe disease (ICD code lists in supporting information).

18To determine the proportion of excess deaths that could be ascribed to dementia as the underlying cause of death, we calculated the proportion of dementia deaths with markers of end‐stage disease in the year prior to death, above and beyond the occurrence of end‐stage disease markers in those who died without dementia. The subtraction of the proportions with end‐stage disease markers in those without dementia from the proportions in those with dementia we deemed to represent individuals who died with severe, end‐stage dementia.

19To apply the estimates of the proportion who died with severe, end‐stage dementia out of those who died with dementia (died with severe disease/total deaths with dementia) to the total excess deaths we adjusted these proportions using the adjustment factor (relative risk/[relative risk – 1]) to calculate the proportion of individuals who died with end‐stage dementia out of excess dementia deaths with Formula 3:

20Relative risks of mortality with dementia were estimated using a second meta‐regression model and the same studies identified through the literature review described in section 2.2.

21We calculated the number of deaths due to dementia as the product of total excess dementia deaths and the proportion of those who died with severe disease out of excess dementia deaths. These results were combined with mortality estimates from all other causes of death and scaled to sum to the total estimated all‐cause mortality in a particular age, sex, year, and location combination. Age‐standardized rates were calculated using the world population standard for GBD. Regional and global estimates were calculated by aggregating the estimates from the most‐detailed locations. The Socio‐demographic Index (SDI), a composite measure of income, education, and fertility, was calculated as the geometric mean of normalized estimates of per capita income, the fertility rate in women under 25, and the average years of education in individuals 15 and older. Countries were divided into quintiles of SDI for analyses. Throughout all steps included in the analytic process, uncertainty in the estimates was propagated by sampling 1000 draws at each step. For example, the calculation of the product of two estimates involved conducting a draw‐wise multiplication of 1000 draws from the distribution of each estimate. We defined uncertainty intervals as the 25th and 975th values of the ordered draws.

22We identified 34 studies containing 60 unique estimates (on a given, age, sex, and/or exposure category) on the excess mortality associated with dementia. These data covered 7 of the 21 world regions, most from high‐income North America and Western Europe (Table 1).

23Attributable risk increased with age, and this relationship became more pronounced after 70 years of age. In 40‐ to 44‐year‐olds, the attributable risk was 0.019 (0.003–0.087), and in the 95+ age group these estimates reached 0.169 (0.0472–0.574; Figure 2). This indicates that individuals over 95 with dementia experience a mortality rate that is 16,900 per 100,000 person‐years higher than those over 95 who do not have dementia.

24Of the 14 markers of severe, end‐stage disease considered, the fraction of people who had pneumonia in the last year before death out of those who died with dementia was the highest (0.31). This proportion remained the highest (0.13) after subtracting the fraction of people without dementia who had the same end‐stage condition in the last year before death to calculate the excess proportion with pneumonia in those who died with dementia. The next end‐stage conditions with the largest absolute and excess proportions were dehydration (absolute: 0.15; excess: 0.11), urinary tract infections (0.12; 0.06), and decubitus ulcer (0.06; 0.05). The proportion of individuals who died with dementia and had any end‐stage disease marker in the year before death was 0.59, and the proportion of individuals with an end‐stage disease marker in the year before death who died without dementia was 0.35. Therefore, we calculated the proportion of people with dementia who died with severe, end‐stage dementia (the excess proportion) as 0.24 (Figure 3).

25Overall, we estimated there were 1.62 million (0.41–4.21) deaths due to dementia globally in 2019. Far more deaths occurred in women (1.06 million [0.27–2.71]) than men (0.56 million [0.14–1.51]), giving a female‐to‐male ratio of 1.94 (1.76–2.11; Table 2). The female‐to‐male ratio for age‐standardized rates was 1.19 (1.10–1.26), suggesting that a portion of the sex difference was independent of the higher life expectancy in women. Between 1990 and 2019, there was a significant increase globally in age‐standardized mortality rates among males (5.1% [0.4–12.0]), but a non‐significant increase in age‐standardized mortality rates among females (3.0% [‐2.6–11.0]). Among males, the largest increase was observed for low SDI countries (9.6% [0.3–26.3]). Between 1990 and 2019, there were no significant changes globally in age‐standardized mortality rates for both sexes combined (3.0% [‐1.2–9.8]), but a large increase in all‐age mortality rates (100.1% [89.1–117.5]) due to population aging.

26The total number of deaths from dementia increased with SDI quintile, from 62.1 thousand (15.1–171.3) deaths in the low SDI quintile to 563.2 thousand (148.2–1394.4) deaths in the high SDI quintile. However, the age‐standardized rates were not significantly different between SDI quintiles, indicating that differences in the numbers stem from differences in population size and population age structure. The age‐standardized rates by SDI quintile ranged from a high of 23.2 per 100,000 (5.69–60.43) in the middle SDI quintile to a low of 21.61 per 100,000 (5.22–56.36) in the low‐middle SDI quintile.

27There was a 1.7‐fold variation in age‐standardized mortality rates between countries. The countries (with population over 1 million) with the highest age‐standardized mortality rates due to dementia were Afghanistan (30.79 deaths per 100,000 [7.46‐82.34]), Mozambique (29.74 [7.29–80.08]), and Vietnam (29.21 [7.24–74.60]). The countries with the lowest age‐standardized mortality rates due to dementia were Bangladesh (18.23 deaths per 100,000 person‐years [4.50–49.29]), India (19.12 [4.57–52.01]), and Nepal (19.51 [4.53–53.31]; Figure 4).

28In data from vital registration systems, huge variations in the patterns of deaths due to dementia were observed both over time and between countries. In Brazil, for example, the number of deaths due to dementia reported in vital registration data increased more than 300‐fold between 1980 and 2017. Across countries with vital registration data in 2015, the proportion of deaths attributed to dementia ranged from 0% (Qatar) to 17% (Finland). In contrast, our estimates were more stable both over time, and between countries, similar to trends observed in prevalence estimates. In the majority of locations and years, our estimated mortality due to dementia exceeded the mortality reported on death certificates in vital registration systems. Returning to the example of Brazil, our estimates were 2.12 (0.54–5.52) times the reported 21,653 deaths in vital registration data. However, in some high‐income countries in recent years our estimates were lower than those reported via vital registration systems. In the United States in 2015, our estimates were 0.54 (0.14–1.32) times the 250,863 deaths recorded in the vital registration system.

29Comparing dementia to all other diseases estimated in the GBD study (at Level 3 of the GBD cause hierarchy), in 1990 dementia was the 20th leading cause of death globally comparing all‐age mortality rates and the sixth leading cause of death among individuals 70 and older. However, by 2019, deaths due to dementia were estimated to rank seventh globally across all ages and fourth among individuals 70 and older.

30Globally, 1.62 million (0.41–4.21) people were estimated to have died of end‐stage dementia in 2019. Age‐standardized rates were stable over time but there were large increases in all‐age rates as a result of population aging. With this trend expected to increase in the future, an increase in dementia deaths can also be anticipated.

31We found that conditions such as decubitus ulcer, bronchitis, dysphagia, hip fracture, pneumonia, and bedridden status appear in those who die with dementia substantially more than those who die without dementia. This finding is consistent with previous evidence from both autopsies and vital registration data that identifies conditions such as bronchopneumonia, aspiration pneumonia, and sepsis as common among those who die with dementia., This convergent evidence also supports our assumption that these end‐stage conditions are indicative of individuals who died of dementia as an underlying cause of death, in line with ICD‐10 principles.

32Our new method of estimating dementia mortality updated the methods used in previous iterations of the GBD study. In prior iterations, we used an excess mortality regression strategy that was sensitive to the choice of which countries to include and assumed that the level of coding deaths to dementia was correct in countries that had the highest rates of coding per prevalent dementia case (i.e., there was no over‐coding)., Through the use of estimates of attributable risk from cohort studies and data on the proportion of deaths with end‐stage disease, we avoided any reliance on vital registration data and cause of death coding practices. As we have found extreme changes in mortality rates from vital registration data both over time and between countries, use of these data risks incorporating a large source of measurement error. While removing this reliance could be seen as an improvement, as we do not have a gold‐standard measure to which we can compare our estimates of dementia mortality, we do not have the ability to assess their accuracy.

33As a result of this change in strategy we no longer assume that the highest rates of coding dementia as a cause of death per prevalent case is correct and we have estimated fewer deaths due to dementia. Where in previous iterations of the GBD study our estimates were always the same or higher than dementia mortality reported in vital registration data, now our estimates are in some cases lower than what is seen in vital registration data, particularly in recent years for high‐income countries, including the United States, Finland, Sweden, and the United Kingdom. This suggests that in these countries in the most recent years there may be over‐coding to dementia as an underlying cause of death. However, our estimates suggest much higher levels of dementia mortality than reported in vital registration systems in many countries, including Brazil and Germany. Improvements in the awareness of the mortality associated with dementia and clearer guidelines for death certification may help improve the quality of data from vital registration systems.

34Our estimates are lower compared to prior studies that have estimated dementia‐related mortality in the United States without using vital registration data (114,838 95% uncertainty interval [UI; 30,031–281,180] deaths in adults 75 and older in 2010 vs. 503,400 deaths in James et al.; 119,916 [31,261–296,271] deaths in adults 65 and older in 2010 vs. 600,000 deaths in Weuve et al.). However, both of these studies used Cox proportional hazard models to estimate the risk of mortality among individuals with dementia, and did not control for common comorbidities such as vascular diseases. While James et al. did report that adjustment for vascular comorbidities did not change results, this could be due to measurement error in the vascular risk index or the specific characteristics of the sample used. Without adjusting for common risk factor profiles, comorbid conditions, or sub‐setting to deaths with end‐stage conditions, these estimates are representative of the total number of excess deaths due to dementia and not just those attributable to dementia as an underlying cause of death.

35A number of limitations still remain. First, the data on the risk of all‐cause mortality due to dementia are heterogeneous, resulting in uncertain estimates of absolute and relative mortality risk. This large source of uncertainty contributes greatly to the large uncertainty intervals for our final estimates. Second, we remain reliant on our estimates of prevalence to calculate excess deaths from data on attributable risk, and these data are also heterogeneous and sparse, particularly in low‐income settings. Third, we are assuming that the risk of dying from end‐stage dementia across locations is the same. While there are no available therapies that can slow or prevent death, differences in the composition of comorbidities and the quality of care could influence mortality.,, Fourth, we assume that administrative data can be used to accurately identify markers of end‐stage disease. Administrative data was used to classify dementia in the analysis of end‐stage disease, and prior studies have shown administrative records can have poor sensitivity, particularly when considering mild disease. However, it is more likely that individuals with severe, end‐stage dementia would be correctly classified, particularly when considering the long time span of our linked data (2003 to 2017). Fifth, due to a lack of available data sources with death record linkages, we are using data from Italy for the measurement of end‐stage conditions, and are generalizing these findings to the rest of the world. With no disease‐modifying treatments currently available, the biological drivers of disease progression are likely to be similar across locations, leading to similarities in the profiles of end‐stage conditions., However, this assumption may be violated to the extent that health‐care practices and other cultural factors influence the proportion of individuals with dementia who have end‐stage disease. The integration of data sources from other locations should these become available would strengthen the analysis. Finally, the process of scaling all cause‐specific mortality to add up to the total estimated all‐cause mortality can be sensitive to estimates of all‐cause mortality, which are challenging at the oldest ages when dementia mortality is highest, and can affect geographic variation.

36In the context of population aging and growth, the importance of dementia as a public health concern will rise. Our estimates can be used by health‐ and social‐care authorities involved in end‐of‐life and palliative care to improve planning for services addressing the needs of people who die from dementia. By helping funders and policymakers better understand the global distribution of dementia burden and mortality across locations and time, and make fair comparisons between dementia and other diseases, these estimates will help guide evidence‐driven resource allocation and health system planning.

37This work was funded by the Bill and Melinda Gates Foundation, Seattle, WA, and by Gates Ventures, Seattle, WA.

38G. J. Hankey reports personal fees from AC Immune, Lausanne, Switzerland for serving on Data Safety Monitoring Committees of ACI‐24‐701, ACI‐24‐1801, ACI‐35‐1201, and ACI‐35‐1802 trials of immune therapies (vaccines targeted to beta amyloid and tau) for Alzheimer's disease. M. Kivimäki reports grants from the UK Medical Research Council (MRC S011676), the US National Institutes on Aging (NIA R01AG056477) and NordForsk, outside the submitted work. S. Lorkowski reports personal fees from Akcea Therapeutics, Amedes, AMGEN, Berlin‐Chemie, Boehringer Ingelheim Pharma, Daiichi Sankyo, Lilly, MSD Sharp & Dohme, Novo Nordisk, Sanofi‐Aventis, Synlab, Unilever, and Upfield, and non‐financial support from Preventicus, all outside the submitted work. M. Saylan reports being an employee of Bayer AG. J. A. Singh reports personal fees from Crealta/Horizon, Medisys, Fidia, UBM LLC, Trio Health, Medscape, WebMD, Clinical Care Options, Clearview Healthcare Partners, Putnam Associates, Focus Forward, Navigant Consulting, Spherix, Practice Point Communications, the National Institutes of Health and the American College of Rheumatology, and Simply Speaking; owning stock options in Amarin, Viking, Moderna, and Vaxart pharmaceuticals; non‐financial support from FDA Arthritis Advisory Committee, Veterans Affairs Rheumatology Field Advisory Committee and from the Steering committee of OMERACT, an International organization that develops measures for clinical trials and receives arm's length funding from 12 pharmaceutical companies; and being the Editor and the Director of the UAB Cochrane Musculoskeletal Group Satellite Center on Network Meta‐analysis, all outside the submitted work. C. E. I. Szoeke reports grants from NHMRC 1062133, Alzheimer's Association, and NIA320312, during the conduct of the study. A. Wimo reports personal fees from WHO and non‐financial support from ADI, during the conduct of the study; grants from MSD and personal fees from Biogen, outside the submitted work. C. Wu reports grants from Ministry of Science and Technology in China, Suzhou Municipal Science and Technology Bureau, and the Kunshan Government, and personal fees from HealthKeepers, all outside the submitted work.