Forecasting Prevalence and Mortality of Alzheimer’s Disease using the Partitioning Models

Abstract

Objectives:

Health forecasting is an important aspect of ensuring that the health system can effectively respond to the changing epidemiological environment. Common models for forecasting Alzheimer’s disease and related dementias (AD/ADRD) are based on simplifying methodological assumptions, applied to limited population subgroups, or do not allow analysis of medical interventions. This study uses 5%-Medicare data (1991–2017) to identify, partition, and forecast age-adjusted prevalence and incidence-based mortality of AD as well as their causal components.

Methods:

The core underlying methodology is the partitioning analysis that calculates the relative impact each component has on the overall trend as well as intertemporal changes in the strength and direction of these impacts. B-spline functions estimated for all parameters of partitioning models represent the basis for projections of these parameters in future.

Results:

Prevalence of AD is predicted to be stable between 2017 and 2028 primarily due to a decline in the prevalence of pre-AD-diagnosis stroke. Mortality, on the other hand, is predicted to increase. In all cases the resulting patterns come from a trade-off of two disadvantageous processes: increased incidence and disimproved survival. Analysis of health interventions demonstrates that the projected burden of AD differs significantly and leads to alternative policy implications.

Discussion:

We developed a forecasting model of AD/ADRD risks that involves rigorous mathematical models and incorporation of the dynamics of important determinative risk factors for AD/ADRD risk. The applications of such models for analyses of interventions would allow for predicting future burden of AD/ADRD conditional on a specific treatment regime.

Introduction

Health forecasting is an important aspect of ensuring that the health system can effectively respond to the changing epidemiological environment. Alzheimer’s disease and related dementias (AD/ADRD) are a well-recognized current challenge, whose impact on population health is likely to increase with time¹. Therefore, improving the forecasting models used for AD/ADRD is of practical value for the health system and policy makers. Several types of forecasting (or projection) models are currently used to predict the future prevalence of AD/ADRD and/or the number of individuals living with the disease. The simplest approach^2–4 includes: estimation of disease prevalence in population subgroups, calculation of the future values of disease prevalence using information about current trends, and computing the number of affected individuals by multiplying the estimated future prevalence values by the projected population given by the Census^5–7. In some approaches predictions are based on information (or assumptions) about the future dynamics of incidence, total mortality, and the death hazard ratio of being in an unhealthy state⁸. More advanced approaches use multistate models involving healthy and unhealthy states.^9,10

Recently, Akushevich et. al. developed a partitioning model^11–14 and applied it to the analyses of time trends in the prevalence and mortality of AD/ADRD¹. In this model observed changes in disease prevalence and incidence-based mortality (IBM) are generated by two simultaneously occurring processes: changes over time in disease incidence, and patient survival. Lower incidence decreases disease prevalence while better survival increases it. The resulting change in the prevalence proportion can then increase, decrease, or remain constant, depending on the relative magnitude of the two effects. In contrast, improvements in incidence and survival act concordantly to reduce IBM: improved incidence reduces the total number of people with the disease while improved survival further reduces the number of associated deaths.

The forecasting model that is developed in this study is based on the partitioning model¹. Projections are constructed by predicting time patterns of incidence, survival, and prevalence at 65 using a B-spline-based technique. This is in contrast to other popular forecasting models where future cohort-specific epidemiologic measures are constructed based on current estimates using a life-table technique⁸ or by solving the respective equations¹⁵; age-adjusted estimates in a given calendar year are then constructed through the weighted sum of the cohort-specific estimates. Such models could be sensitive to estimates of the initial disease prevalence and incidence. The potential impact of the respective bias is likely to be high for AD onset which is relatively rare at ages <65. The approach utilized in this study does not share this limitation.

The epidemiological components of prevalence and mortality partitioning^1,11–14 are natural targets of health interventions. Therefore, it is reasonable to construct forecasting approaches in which models of trends in incidence, survival and other partitioning components are used for the forecasting of future prevalence and mortality levels under alternative epidemiological scenarios. In this study, we utilize such trend partitioning to forecast AD prevalence and IBM for the next ten years. The forecasts are constructed for AD both independently and in concert with patterns of influential diseases prevalent in an individual prior to the diagnosis of AD.

Data and Methods

This study uses 5%-Medicare data (1991–2017) to identify, partition, and forecast time trends of age-adjusted prevalence and IBM of AD as well as their causal components such as trends in incidence and survival. Using Medicare enrollment files, we identified the first and last month/year during which an individual was enrolled in a traditional Medicare fee-for-service plan with both Parts A and B coverage. These two points in time served as the bounds over which an individual was followed. Then, we calculated the proportion of time (in months) within these bounds that the individual did not have Medicare fee-for-service coverage. If this time accounted for more than 20% of the total duration then the individual was dropped from the analysis. We required the presence of two distinct claims with a diagnosis of AD(ICD-9: 331.0; ICD-10: G30.x)/ADRD(ICD-9: 290.x or 294.2x; ICD-10: F01.x, F03.x) within 90 days of each other with the earliest date in the pair designated as the date of onset. If death occurred within 90 days of the first diagnosis of AD/ADRD this was treated as a confirmatory record^1,16. AD/ADRD was treated as an absorbing state. Sensitivity testing found that our results were not sensitive to variation in the 20% fee-for-service gap cutoff or to variation in the 90-day verification period. The same procedure was applied to identify the day at onset of risk-related diseases such as diabetes mellitus¹⁷, arterial hypertension^18,19, renal disease²⁰, cerebrovascular disease²¹, lung diseases²², and traumatic brain injury²³. Disease prevalence proportions, incidence rates, incidence-based mortality, and all-cause mortality were then calculated in two-dimensional age-time-specific bins spanning the age-time plane for which data was available (see Supplementary Figure 2 of ref. [1]). Specifically, all measures in each age-time-specific bins are estimated as the ratio of i) the number of person-years of sick individuals for prevalence and ii) number of respective cases for all other measures, to the number of person-years.

Partitioning-based forecasting model

Partitioning analysis: i) predicts trends in prevalence and mortality, ii) decomposes (or partitions) them into their constituent components, and iii) calculates the relative impact each component has on the overall trend as well as intertemporal changes in the strength and direction of these impacts.

The constituent components are AD/ADRD incidence, relative survival, morbidity at the bounds defined by the availability of data (65 years of age and the year 1992), and mortality for the general population (for incidence-based mortality only). Formally, age-adjusted prevalence, $P(y)=P_0(y)+P_{00}(y)+P_{is}(y)$, is represented as the sum of three positive contributions from individuals with pre-existing prevalence at the age-65-boundary, $\left(P_0(y)\right)$, pre-existing prevalence at the time-1992-boundary $(P_{00}(y))$, and, disease onset after 1992 and age $65(P_{is}(y))$. Simllarly, age-adjusted incidence-based mortality, $M(y)=M_0(y)+M_{00}(y)+M_{is}(y)+M_{P\mu }(y)$, is represented as the sum of the contributions of mortality among individuals with pre-existing prevalence at the age-65-boundary $(M_0(y))$, pre-existing prevalence at the time-1992-boundary $(M_{00}(y))$, as well as two other contributions associated with individuals with disease onset after 1992 and age 65: mortality because of the presence of the disease $(M_{is}(y))$ and mortality interpretable as non-disease-specific $(M_{P\mu }(y))$. Partitioning of age-adjusted prevalence and incidence-based mortality is formally obtained by differentiation over calendar time (year). Thus, partitioning for prevalence, $P\text{'}(y)/P(y)=T_0(y)+T_{00}(y)+T_{inc}(y)+T_{sur}(y)$ is determined by four components. Two major contributions reflect the effects of change in disease incidence $(T_{inc}(y))$ and relative survival $(T_{sur}(y))$ over time. The two remaining contributions reflect the effects from age $(T_0(y))$ and time $(T_{00}(y))$ boundaries. Similarly, partitioning for mortality is given by five components: $M^{\mathrm{\text{'}}}(y)/M(y)=T_{\mu }(y)+{\hat{T}}_0(y)+{\hat{T}}_{00}(y)+{\hat{T}}_{\text{inc}}(y)+{\hat{T}}_{sur}(y)$, with specific components reflecting the change over time in disease incidence $({\hat{T}}_{inc}(y))$, relative survival $({\hat{T}}_{sur}(y))$, morbidity at 65 $({\hat{T}}_0(y))$ and $1992({\hat{T}}_{00}(y))$, and mortality in the general population $(T_{\mu }(y))$. Change over time in the latter component influences the trend in the incidence-based mortality through its effect on other-cause mortality.

The analysis involves the design and estimation of separate models for i) the incidence rate, ii) relative survival after AD/ADRD diagnosis as well as of individuals prevalent at 65 and/or 1992, iii) prevalence at the age boundary (age 65), iv) prevalence at the time boundary (year 1992), and v) mortality in the general population. The distribution of age (and time after onset for relative survival) is modeled using i) the generalized 4-parameter Armitage-Doll model ^24,25 with additional individual predisposition parameterized by gamma or inverse Gaussian distributions ²⁶ (for incidence), ii) the Weibull model ^27,28 for time after disease onset with the quadratic function of age for the shape and scale parameters (for relative survival), and iii) the Gompertz model (for mortality in the general population). These models are either standard for these outcomes or demonstrated better goodness-of-fit characteristics in respect to the alternatives (specifically, we tested survival models from refs. ^29–34 as well as all models defined in Section 7.3 of ref. ³⁵). The model for any function contains parameters to describe the function for each year of diagnosis, and the parameters of B-splines that are used to model the relationships between year-specific model parameters and evaluate the $y$ -dependences of the function. The B-spline model contains equidistant knots including 4 inner knots at 1993, 2001, 2009, and 2017 years and boundary knots on the year boundaries, the parameters of the boundary knots being equal to those of the closest inner knot. The partitioning components are expressed in terms of derivatives of these functions with respect to time. B-splines allow explicit calculation of derivatives without requiring additional simplifying assumptions. The model parameters are estimated using non-linear least squares for age-specific incidence, mortality, and prevalence at boundaries, and the likelihood-based approach for relative survival ³⁶.

B-spline functions estimated for all parameters represent the basis for projections of these parameters in future. The idea of these projections is illustrated in Supplementary Figure S1. We use B-splines of 3rd degree defined by equidistant set of knots. Any model parameter is represented $p(y)=b_0{B}_0(y)+b_1{B}_1(y)+b_2{B}_2(y)+b_3{B}_3(y)+b_4{B}_4(y)+b_5{B}_5(y)$ for the estimation region $1993\le y\le 2017$. The peaks of B-splines $B_{1-4}(y)$ are located at years 1993, 2001, 2009, and 2017, respectively. We use the restriction $b_0=b_1$ and $b_4=b_5$ which is justified by the fact that $B_0(y)$ and $B_5(y)$ are nonzero only in a small part of the estimation region. For $y>2017$ the parameters are evaluated using the extended model

$$p(y)=b_0{B}_0(y)+b_1{B}_1(y)+b_2{B}_2(y)+b_3{B}_3(y)+b_4{B}_4(y)+b_5{B}_5(y)+b_6{B}_6(y)+b_7{B}_7(y)+\mathrm{K}$$

where new B-splines $B_6(y)$, $B_7(y)$, etc. with peaks in 2033, 2041, etc. were added. The coefficients at these new B-splines are calculated through weighed combinations of estimated coefficients $b_{1-4}$, e.g., $b_6=w_1{b}_1+w_2{b}_2+w_3{b}_3+w_4{b}_4$. The set of weights $(w_1,w_2,w_3,w_4)$ defines a specific project scenario. For example, the scenario $(0,0,0,1)$ assumes that future trends are similar to the trends in 2017, but the scenario $(1/\mathrm{4,},1/4,\mathrm{,1}/1,1/4)$ is based on the trends averaged over the entire estimation period.

Thus, we propose extrapolation technique for model parameters as linear combinations of B-splines in which adding one or more B-splines outside of the range of the available data allows the calculation of the B-spline function at the next time period(s). We note that this approach projects all parameters into the future, conserving the relationships between projected functions and their derivatives in each time point.

Morbidity profiles of AD onset

Each case of AD is associated with a set of risk factor diseases^37,38 diagnosed prior to AD diagnosis. Unique (mutually exclusive) combinations of pre-existing diseases known to be AD risk factors represent the morbidity profile of an individual with separate projection models developed for each such profile. The following approach was used for identification of these profiles: i) estimation of a multivariable Cox model with AD onset as the outcome and indicators of diseases included as the only explanatory variables; ii) identification of the disease with the highest significant hazard ratio, exclusion of all individuals with that disease from the sample, and the reestimation of the model; iii) iterative repeating the procedure to finally rank all diseases according their effects on AD risk, iv) empiric analysis of frequency distributions of obtained morbidity profiles to identify and combine disease groups with low prevalence, and iv) evaluation of the effects of the constructed morbidity profiles in the Cox model.

Results

The 2028 projections for AD prevalence and IBM as well as the partitioning components of the forecasting model with one knot outside of the range of the available data is shown in Figure 1. In the estimation region (before 2017), the primary determinant responsible for the observed trends in prevalence and IBM, was the deceleration and eventual decrease in incidence rates of AD/ADRD; though changes in relative survival began to have a noticeable impact after 2008 (see detailed discussion in ref.¹). The forecasted region (2017–2028) based on the B-spline forecasting approach developed in this study, demonstrates that age-adjusted prevalence is predicted to remain unchanged while IBM is expected to increase. Incidence is expected to continue to be the primary driving factor for the predicted changes.

Figure 1.

Projected prevalence and IBM of AD with partitioning components.

Recall that our model had four B-spline coefficients associated with 1993, 2001, 2009, and 2017 respectively. The projections in Figure 1 are constructed under the assumption that the future trends of incidence and survival will be as estimated by the 2017 B-spline. That is, the coefficients associated with the B-spline in the prediction region are equal to those of the 2017 B-spline of the estimation region. Alternative scenarios assume that the forecast region B-spline is determined by alternative combinations (e.g., weighted contributions) of the coefficients associated with the estimation region B-splines included in the model. The 2050 projections generated using four alternative scenarios with (1993, 2001, 2009, 2017) weights of $(0,0,0,1)$, $(0,0,1/2,1/2)$, $(0,1/3,1/3,1/3)$, and $(1/4,1/4,1/4,1/4)$ are shown in Figure 2. These scenarios assume that the future trends are respectively as in 2017, in periods of 2009–2017, 2001–2017, and in the entire estimation period.

Figure 2.

Projections of AD prevalence and mortality based on four scenarios when coefficients at B-splines in future are defined by coefficients estimated for four B-splines during estimation period, i.e., in 1993, 2001, 2009, 2017 as their weighted sum, such that a projection curve is defined by four weights $(w_{1993},w_{2001},w_{2009},w_{2017})$.

Figure 3 shows projections for three age groups: 65–74, 75–84, and 85+. Our method models each age group separately and aggregates them to add up to the total age-adjusted prevalence and mortality for ages 65+. The highest prevalence was predicted for the age group 75–84, however the difference between this group and older adults age 85+ will disappear by 2030. The highest mortality was detected for ages 85+ and the gap between this and younger age groups is expected to increase with time.

Figure 3.

Projection of age-specific groups (65–74, 75–84, and 85) as they contribute to age-adjusted prevalence and mortality.

Figure 4 shows projections for five mutually exclusive morbidity profiles identified using the procedure described in Methods Section. Each successive profile excluded the diseases associated with all previously created profiles: i) profiles with a history of traumatic brain injury (TBI) (prevalence: 15.40%, AD diagnosis odds ratio (OR): 12.4), ii) profiles with history of cerebrovascular disease including stroke (39.35%, OR:8.7), iii) profiles with history of pneumonia/influenza (16.87%, OR:6.3), iv) profiles with diabetes/renal (11.98%, OR:4.0), and v) profiles with arterial hypertension (12.06%, OR:3.7).

Figure 4.

Projected prevalence/IBM and main partitioning components of specific morbidity profiles.

The morbidity profile associated with cerebrovascular diseases has the leading contribution and is the only profile whose prevalence declines as predicted (Figure 4). All other profiles demonstrate an increase in the projected prevalence. All profiles including cerebrovascular disease associated show increases in mortality from AD in the near future. In all profiles the resulting patterns come from a trade-off of two disadvantageous processes: increased incidence and disimproved survival. Except for the cerebrovascular disease associated profile trend in incidence overpower respective trends in survival. Time patterns of partitioning components are similar. Deceleration in incidence rates was detected for all profiles, but only two of them did not have periods of increase in incidence rates: profiles associated with TBI and diabetes/renal disease.

Discussion and Conclusion

We developed a new forecasting model for prevalence and mortality based on trend partitioning¹ which models trends in age-adjusted prevalence and incidence-based mortality in terms of changes in interpretable epidemiologic quantities, such as disease incidence and survival, that generate time trends in prevalence and mortality of AD/ADRD and represent reasonable targets for public health interventions. The models for incidence and survival incorporated in the approach are estimated using a nationally representative 5% sample of U.S. Medicare beneficiaries which has higher statistical power than much of the data used in current projections. In our partitioning approach, time trends of incidence and survival as well as other components such as prevalence of AD/ADRD at 65, are modeled non-parametrically through B-spline functions that are estimated for all parameters and represent the basis for projections of these parameters into the future. The approach allows for forecasting both age-specific and age-adjusted prevalence and incidence-based mortality as well as for incorporation of individual health histories using mutually exclusive individual morbidity patterns.

Existing common approaches^2–4,8–10 to forecast prevalence of AD/ADRD and the number of people living with dementia use distinct methods for evaluation of current dementia prevalence proportions in population subgroups using different types of data. However, future prevalence of AD/ADRD in such subgroups (usually age- and sex-specific) is usually assumed to be unchanged. What is changed is the size of the population at risk (due to aging of the U.S. population), and therefore the number of individuals living with AD/ADRD. For example, a Medicare-based study⁴ constructed projections by applying the 2014 national-level prevalence of diagnosed ADRD among fee-for-service beneficiaries to the subgroup-specific population in a given year. Multistate modeling⁹ was used to evaluate transitions between preclinical and clinical AD and to estimate potential impacts of primary and secondary preventions in U.S.¹⁰ Future transition rates between health states are defined by modifying factors that characterize the effectiveness of the interventions and specify upon which transition rates the interventions act. The approach of Hebert et al.⁸ constructed the observed prevalence based on incidence and survival. In this aspect, their forecasting approach is closest to ours. The authors used combinations of weighted logistic regressions with predictors including sex, race, education, and several functions of age. Prevalence of AD dementia was evaluated using a life-table technique that required the estimates of relative risk of dying with AD dementia (the value of 2.13 was used). Projected total U.S. mortality by age, race, and sex; and future population in such strata were used to calculate future prevalence and the number of individuals living with AD dementia. The study was hampered by insufficient generalizability as it was based on data limited to the Chicago area³⁹.

The forecasting approach developed and applied in this study has several important advantages. We used large sample individual-level data nationally representative of older U.S. adults. The available statistical power allowed for high precision estimates of all components of the model such as incidence and survival and other contributing components such as prevalence at boundaries of our data availability (age 65 and year 1992). The formula connecting these components to age-adjusted prevalence is mathematically exact. Time patterns of the components are modeled using the B-spline functions for model coefficients and extrapolations of them into the future. The approach developed in this paper methodologically improves the approach of Hebert et. al.⁸, because it avoids using life tables incorporated in the approach⁸, instead, survival functions are parametric models, and time dependence is introduced through B-splines. Speaking formally, extrapolation of a function is always based on our knowledge about future function behavior. This knowledge may define a function uniquely (for example, if a function is described by a differential equation), or it may be subtler (for example, we may anticipate emergence of a new cure for a disease that may significantly change the incidence of the disease or mortality caused by the disease). Alternatively, we can assume that such a function will change in the future in a way identical to observed changes on a given part of the fitting interval (approximated by B-splines). The exact parts of the fitting interval used for predicting future behavior of the function is determined by weights reflecting assumptions on how past trends contribute to the future trends.

Predictions of prevalence and mortality in partitioning theory are determined by projections of incidence and survival—specific targets of health care interventions. The future diffusion of an AD treatment, for example, can influence the slope of survival, to the extent that it is able to slow or stop AD progression.⁴⁰ To reflect this in forecasts, the coefficients associated with B-splines at the times of the presumptive interventions can be modified to reflect the predicted effect of the treatment. For example, in Figure 5, we assume full diffusion of AD treatment by 2025. The relative survival is transformed as: $S_r\to f+(1-f)S_r$ and two interpretations of the coefficient $f$ are possible. Assuming $100\mathrm{\%}$ effectiveness of the treatment (e.g., use of treatment equalizes survival between individuals with and without AD) $f$ represents access to treatment, i.e., the fraction of patients receiving treatment. Alternatively, under perfect access $f$ can be interpreted as treatment effectiveness. Six projections are made under alternative scenarios ranging from a zero percent improvement $(f=0)$ to the complete mitigation for the additional mortality associated with AD $(f=1)$. As demonstrated, the resulting projected burden of AD differs significantly and leads to alternative policy implications.

Figure 5.

Projected effects of AD treatment 2025–2050 on of AD prevalence and mortality. Projections are made for six alternative values for the $f$ coefficient: i) treatment not effective ( $f=0$; black line); ii) $10\mathrm{\%}$ improvement in AD relative survival ( $f=0.1$; blue line); iii) $20\mathrm{\%}$ improvement in AD relative survival ( $f=0.2$; red line); iv) $50\mathrm{\%}$ improvement in $\mathrm{A}\mathrm{D}$ relative survival ( $f=0.5$; green line); v) $80\mathrm{\%}$ improvement in AD relative survival ( $f=0.8$; purple line); vi) $100\mathrm{\%}$ improvement in AD relative survival ( $f=1$; brown line). The $f$ coefficient can also be interpreted as percent of the population with access to the treatment under the assumption of full treatment effectiveness.

Theoretically the explicit expressions in the partitioning-based model used in this study coincide with the solution of McKendrick-von Foerster equations, previously used for projections of diabetes prevalence¹⁵. Both approaches are methodologically equivalent. However, the schemes for numeric implementations of projections are different. In the McKendrick-von Foerster approach, projections of future prevalence are constructed by solving the ordinary differential equations along characteristic lines that coincide with cohort evolution lines in the Lexis diagrams. This approach requires well-defined initial states at age 65 and so is appropriate for diseases with a high prevalence at 65 (e.g., diabetes). However, in the case of AD/ADRD with a small initial prevalence, relative uncertainties in initial prevalence are always substantial and could be propagated over entire age region when the cohort-specific differential equations are numerically solved. The approach based on B-splines is not sensitive to the accuracy of the prevalence at 65, and therefore is more appropriate for diseases like AD or ADRD.

Incorporation of trends of AD risk-related diseases can account for the dynamics in incidence and prevalence of comorbid conditions—an important time-variant factor determining dynamics of AD epidemiologic measures³⁸. For mutually exclusive disease-specific states, the predictions include two components: projections of the prevalence of the comorbid disease in the population and its effect on AD risk. AD is a multifactorial disease, so such decomposition results in a reduction in heterogeneity in AD risk, and therefore more accurate predictions of AD as a sum of homogeneous contributions, which allows the total disease burden to have different associations with specific disease-related states. The approach we use permits further generalization to include other risk factors and their trends into analysis and projection models. Future forecasts will be constructed based on a detailed consideration of the process of AD/ADRD development involving the incremental introduction of multiple-risk factor diseases, their treatment, interdependence of risk between treatment, disease and AD/ADRD as well as the genetic variation that influences such dependencies. Incorporation of risk factors with their time trends to forecasting models including partitioning approaches is not straightforward, however, such developments are vital because they can further clarify the mechanisms generating the risks and their time trends of AD/ADRD. Stochastic process models of aging, mortality and disease development that were suggested in the biodemographic literature^41,42 incorporate several relevant aging-related mechanisms that can underlie the connections of risk of AD/ADRD development with age dynamics and time trends of various risk factors. Despite the recognition of the importance of such models for health forecasting⁴³, their applications in this area are still sorely lacking.

An important step in development of forecasting models that incorporate disease risk factors was recently made by the Global Burden of Diseases, Injuries, and Risk Factors Study that forecasted the prevalence of dementia attributable to three dementia risk factors: high body-mass index, high fasting plasma glucose, and smoking.³⁷. Although they found large increases in the projected number of people living with dementia, such increases were claimed to be due to population ageing and population growth, while estimated prevalence changes attributable to variation in risk factor prevalence were small. However, these findings were based on assumptions resulting in a significant oversimplification of the underlying mechanisms of dementia development and were met with criticism from some researchers⁴⁴. We note that some level of simplification will always exist in health forecasting models depending on the objective of the study and the scope of the available data. However, such simplifications increase the overall uncertainty of the resulting projections and reduce the practical applicability of the results. A recent study⁴⁵ reported 12 modifiable risk factors that jointly could reduce the onset of dementia by up to 40%; our study builds upon this foundation by focusing on the mechanisms tying together pre-existing (and co-morbid) conditions, their treatments and the onset of AD/ADRD. Even though some of these risk factors (e.g., cardiovascular disease) are known, the pathways by which they impact AD/ADRD onset as well as the interactions and tradeoffs between the presence of multiple such risk factors are less clear.

The forecasting model developed in this study is an important building block for the next generation of forecasting approaches that involve rigorous mathematical models and incorporation of the dynamics of important determinative risk factors for AD/ADRD risk. Further model extensions will allow for estimating the future size of AD/ADRD-related burden while simultaneously accounting for multiple time trends in new and established risk factors. The ability of our forecasting models to account for the effects of interaction between diseases will allow us to model plausible scenarios of future demographic and epidemiologic changes in AD/ADRD patterns and trends and identify areas with potential for health care interventions in the short-term. The applications of such models for analyses of interventions would allow for evaluating new possibilities that do not require direct investments in clinical trials for testing new medications but use hidden connections among existing diseases with well-established treatment regimens as well as for identifying new factors and mechanisms of disease development responsible for dependence between AD/ADRD and diseases common in older adults and produce new information about how the robustness (incidence rate) and resilience (case fatality rate) characteristics of AD/ADRD depend on the presence or absence of other diseases, as well as on associated genetic and non-genetic factors. The results will produce new insights on aging-related health decline and the role of AD/ADRD in this process.

We acknowledge several study limitations. The approach assumes that AD/ADRD diagnosis in the Medicare claim closely approximates the clinical conditions of the beneficiaries. Medicare data is administrative billing data not designed for research and is not linked to medical charts, autopsy reports or other ways of independent validation of a diagnosis present in the claim^46–49. Combined with the difficulty of placing an AD diagnosis this introduces a certain level of uncertainty in any estimate based on such data. However, Medicare claims fully reflect the billing habits of the U.S. healthcare system and AD/ADRD as well as many of the co-morbidities used in the morbidity profiles are incentivized for reporting. Given the length of the available time series most of the errors resulting from this limitation are those of timing (e.g., the data of onset is identified later than optimal) and are unlikely to affect the identified trends. Inclusion of ADRD in our analysis provides a less stringent set of criteria for selection which can mitigate this problem. Our results are valid for the general population of the U.S. Individual population subgroups such as those associated with race/ethnicity and/or geographic locality) can have differing efficiencies of reconstruction of AD/ADRD onset⁵⁰, disparities in risks^51–53, and/or decomposition patterns. The forecasting approach is directly applicable to population subgroups, so future studies focusing on vulnerable subgroups are needed to address this limitation.

Highlights.

• AD exhibits rising incidence and post-onset mortality; falling survival.

• Mortality post AD onset expected to continue to grow through 2028.

• AD prevalence post-stroke, is expected to decrease drastically.

• AD prevalence for all non-stroke chronic morbidity profiles will increase.

• Due to conflicting trends, total AD prevalence will appear stable through 2028.