Case Definition for Diagnosed Alzheimer Disease and Related Dementias in Medicare

This cross-sectional study uses Medicare claims and encounter data as well as diagnostic and prescription drug codes to develop case definitions for diagnosed Alzheimer disease and related dementias.


Introduction
Surveillance is a fundamental public health activity.Lack of a US dementia surveillance system hinders public health efforts to support persons living with Alzheimer disease and related dementias (ADRD), address health disparities, and plan ADRD care resources.
Medicare administrative data are an attractive source upon which to build a dementia surveillance system and are commonly used to identify persons living with ADRD, but a consensus diagnostic code case definition does not exist.Perhaps the most widely used definition (the Centers for Medicare and Medicaid Services [CMS] Chronic Conditions Warehouse [CCW] algorithm) uses 22
The impact of using different ICD-10-CM or prescription codes on the number of people identified or their characteristics is unknown.Because ICD-10-CM codes are used for billing (rather than diagnostic) purposes, specific codes may not be sensitive nor specific to dementia, and coding practices may differ systematically by health care practice, patient characteristics, and geography.
We examined how choices of ICD-10-CM and prescription drug codes used to identify persons with clinically recognized ADRD in Medicare fee-for-service (FFS) claims and Medicare Advantage (MA) encounter data affect dementia prevalence estimates and characteristics of the people identified.We synthesized this information to develop a new case definition using diagnostic and prescription drug codes that can be applied to administrative data to support surveillance of persons with diagnosed dementia in the Medicare system.

Methods
This cross-sectional study was deemed exempt from review and the requirement of informed consent by the NORC Institutional Review Board.The reporting of this research follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

Statistical Analysis
To categorize beneficiaries with or without evidence of dementia as of 2019, we conducted a crosssectional analysis of January 2017 to December 2019 FFS and MA data to identify all claims and encounters with a relevant ICD-10-CM code listed in any position, and all PDE claims for a relevant NDC.We classified beneficiaries hierarchically, first with a tier 1 ICD-10-CM code, then with a tier 2 code among remaining beneficiaries, and so on, identifying only the incremental beneficiaries in each tier if they had not been classified earlier.We compared distributions of age, sex, race and ethnicity (as indicated by the Research Triangle Institute race code 22 ), MA enrollment, LTC use, and 2019 mortality across tiers.Race and ethnicity categories included American Indian or Alaska Native, Asian or Pacific Islander, Hispanic, non-Hispanic Black, non-Hispanic White, unknown, and other (defined as any race or ethnicity not otherwise specified); race and ethnicity were included because existing evidence shows that there are disparities in dementia prevalence across race and ethnicity groups.
We compared cross-tier beneficiary frailty using a claims-based frailty index (CFI), 23 an adapted CFI that excludes ADRD codes in tiers 1 to 4, and per-member-per-month (PMPM) spending, averaged across all months of 2019 FFS coverage.
Using these data (eTable 2 in Supplement 1), we found that beneficiaries in tiers 1 and 2 were older, more frail, more likely to be female, in LTC, and die than those in tiers 3 to 5.There were minimal differences in race and ethnicity across tiers, with exception of a higher-than-expected representation of Hispanic and Asian and Pacific Islander beneficiaries in tier 5; however, the overall size of the sample categorized as tier 5 was very small, at just 0.1% (52 338 of 60 000 869 beneficiaries).Based on the findings from the cross-tier comparison and author consensus, we a Although these codes appear in fewer than 10 algorithms, they are root codes and are categorized alongside their associated detailed codes.
b We dropped root codes G31 and G31.8 (which appeared in 1 algorithm) due to their associated detailed codes being categorized into different tiers.
further aggregated codes into 3 categories with decreasing confidence of having a true ADRD diagnosis: a highly likely ADRD category requiring at least 2 claims or encounters on different dates with ICD-10-CM codes from tiers 1 or 2; a likely ADRD category requiring 1 claim or encounter with an ICD-10-10-CM code from tiers 1 or 2; and a possible ADRD category requiring at least 1 claim or encounter with an ICD-10-CM or NDC code from tiers 3, 4, or 5 over a 3-year lookback period.We categorized beneficiaries and reevaluated group demographics, health insurance type, frailty and mortality, and rural residency.We then computed prevalence of highly likely, likely, and possible ADRD within population subgroups defined by these characteristics.We age-standardized to the full analytical population to evaluate differences unconfounded by age.
All analyses were conducted in SAS Enterprise Guide 7.1 and SAS Studio version 3.81 (SAS Institute).Data analysis was conducted from September 2022 to March 2024.

Results
Of  Age-standardizing subgroups to the age distribution of the Medicare population resulted in changes in ADRD prevalence estimates in some groups (Table 4).Relative differences in ADRD prevalence narrowed across sex but widened across race and ethnicity groups.Most notably, non-Hispanic White beneficiaries became less likely to have any evidence of ADRD (12.9% across categories), while racial and ethnic minority groups became more likely to have evidence of ADRD (non-Hispanic Black beneficiaries, 16.5%; Hispanic beneficiaries, 15.3%).Among non-Hispanic Black beneficiaries, age standardization resulted in a substantial increase in the proportion of those with highly likely or likely ADRD (9.9% to 12.0%).Age standardization also reduced ADRD prevalence   We then added a highly likely category to describe beneficiaries who received 2 or more likely codes on different dates of service.We posit that these categories are superior to previous definitions for provisional use in surveillance systems, but caution that validation is necessary.To our knowledge, this is the first application of claims identification algorithms to all-age FFS and MA beneficiaries.We have used this case definition to compute provisional national-, state-, and countylevel estimates of ADRD prevalence and incidence in 2020 Medicare and published them on our dementia surveillance website. 24Estimates will be refined pending validation and updated with additional years of data as they become available.
Our 3-level case definition is novel in that it was driven by researcher-consensus as well as data analysis and identifies dementia with varying degrees of certainty.Of note, ICD-10-CM codes used to identify possible ADRD have lower researcher consensus and less specific code descriptions (ie, do not contain dementia or Alzheimer).6][27] Our definition also excludes several previously used codes that were determined to not indicate ADRD by expert clinicians.Compared with the commonly used CCW algorithm, which similarly uses a 3-year look-back period, our case definition is more specific when limited to the highly likely and likely categories, but broader when also including the possible ADRD category.The CCW algorithm estimated prevalence of 10.7% in 2019 Medicare FFS beneficiaries 28 falls between our estimates for FFS beneficiaries of 8.6% for highly likely or likely ADRD and 12.6% for all 3 categories.
Importantly, we saw expected and meaningful differences between beneficiaries identified in each ADRD category.0][31][32][33][34][35] Notably, prevalence of highly likely ADRD was 29.1% in beneficiaries aged 85 years or older, 72.8% in LTC users, and 36.2% in decedents, compared with 7.2% in the general Medicare population.Higher rates of dual-eligibility in ADRD groups may be driven by ADRD beneficiaries spending down assets to qualify for Medicaid and obtain LTC coverage.These differences persisted after age standardization and lend confidence to our case definitions.
Application of our case definitions also showed disparities in diagnosis rates by race in the expected direction-higher dementia risk among non-Hispanic Black beneficiaries relative to non-Hispanic White beneficiaries 36,37 -after age standardization to account for lower life expectancy among non-Hispanic Black individuals. 38However, because non-Hispanic Black individuals also have a greater risk of under-diagnosis of ADRD than non-Hispanic White individuals, 39 disparities in true underlying rates may be higher than observed.Additionally, we found higher-than-expected representation of Hispanic and Asian and Pacific Islander beneficiaries among those that had an ADRD-targeting drug without diagnostic (ICD-10-CM) evidence.We hypothesize that differences in cultural perceptions around dementia and cognitive decline (eg, memory loss as a normal aging process) 40,41 may result in lower utilization of diagnosis codes when providers suspect dementia.
Using PDE claims may result in higher and more accurate rates of ADRD among Hispanic and Asian and Pacific Islander individuals despite the overall small number of beneficiaries identified by PDE claims alone.
Finally, also consistent with past research, 29,35,42 PMPM FFS spending was substantially higher for beneficiaries with evidence of ADRD compared with those with no evidence of ADRD.Medicare FFS PMPM spending was relatively similar across the highly likely, likely, and possible ADRD groups despite differences in frailty and mortality.Medicare FFS spending may not be generalizable to those with MA (for whom costs cannot be computed) and is only part of the economic story.Medicaid is the primary US payer of LTC; higher rates of dual-eligibility and LTC use among the highly likely ADRD group indicate that differences in total federal and state spending between the highly likely ADRD and other groups are likely larger.We also did not capture patient and family health-related out-ofpocket expenses and informal care costs ($203 117 in families caring for a patient living with dementia vs $102 955 in families caring for a patient without dementia over the last 7 years of the patient's life 42 ), forgone wages, or other impacts on informal caregivers, and payments made by other assistance programs.Finally, we caution that our spending measure represents total Medicare FFS spending, rather than the incremental ADRD costs.

Limitations
This study is limited by at least the following.First, our ADRD case definition was driven by researcher-consensus, and validation against other dementia ascertainment methods (including ascertainment based on in-person clinical and neuropsychological assessments) is necessary.Both over-and under-diagnosis of ADRD have been documented in Medicare claims, 35,39 and the 8.0 million beneficiaries identified as having some evidence of ADRD by our case definition will include some without ADRD, especially those in the possible category.Similarly, this method only captures documented cases of dementia in Medicare administrative records and cannot capture beneficiaries with unrecognized and/or undocumented ADRD.If we assume a 60% rate of undetected dementia in the US 43 our estimates would suggest an additional 12 million beneficiaries may be living with ADRD.Additionally, our data show a marginally higher rate of ADRD in MA than in FFS enrollees (14.5% vs 12.6% across the 3 categories), which may reflect beneficiary selection in MA plans, MA vs FFS differences in clinical ADRD assessment and diagnosis rates, differences in claims or encounter documentation, or a combination thereof.Given the rapid rise in MA participation (from 33% in 2017 to 51% in 2023) and variation in MA penetration across counties, 44,45 it is also important to understand potential differences in performance of this case definition between MA and FFS beneficiaries.As such, validation of this case definition against in-person clinical and other ascertainment methods to assess performance (including sensitivity, specificity, positive predictive value, and negative predictive value), separately for Medicare FFS and MA, is critical for refining and calibrating estimates to accurately capture the diagnosed prevalence and incidence of dementia.
Pending validation, our case definitions should be considered provisional.Notably, we expect the possible ADRD category to identify a higher proportion of individuals who do not have ADRD.Thus, it is important to report the possible ADRD category separately from the likely and highly likely ADRD categories in research and surveillance efforts using these case definitions.
Second, evidence for ADRD documented in electronic health or insurance records outside the Medicare system is not captured by our method; this is particularly problematic for beneficiaries without Medicare Parts B or D (7.5% and 25.6% of Medicare enrollees, respectively 43,46 ).Third, we deliberately used data from 2017 to 2019 to avoid the COVID-19 pandemic years, which resulted in secular shocks, including excess senior deaths, forgone or deferred care, and increased telehealth, which may have impacted dementia diagnosis.Research is necessary to understand these effects but will necessarily be delayed pending new data.Fourth, Namzaric, a memantine and donepezil combination drug approved in 2014, was not included by any prescription-drug based identification strategy; while the impact of including this drug necessitates further investigation, we anticipate a negligible effect given that just 0.1% of the sample had an ADRD-targeting prescription drug without ICD-10-CM evidence.Similarly, ICD-10-CM code updates from October 2022 added 29 highly specific codes each under code roots F01 (vascular dementia) F02 (dementia in other diseases classified elsewhere), and F03 (unspecified dementia) (eTable 5 in Supplement 1). 47We recommend that applications of our approach to Medicare records beginning in October 2022 include these for identifying highly likely and likely ADRD.Fifth, in developing our case definitions, we only considered use of ICD-10-CM codes and prescription drugs but did not consider other criteria of existing ADRDidentification algorithms, including look-back period, types of claims or encounter data considered, number of claims or encounters with relevant ICD-10-CM codes required, and time elapsed between claims and encounters; sensitivity analyses around these different criteria are beyond the scope of this paper.

Conclusions
In

Figure .
Figure.Demographic Distributions Across Alzheimer Disease and Related Dementias (ADRD) Categories (Raw and Age-Standardized)

Table 1 .
ICD-10-CM Codes and Prescription Names Used to Categorize Individuals Across Tiers and ADRD Categories (continued)
Abbreviations: CFI, Claims-Based Frailty Index; FFS, fee-for-service.aOtherincluded any race or ethnicity not otherwise specified.bHaddual-eligibility for at least 1 month in 2019.cEnrolled in Medicare Advantage (Medicare Part C) for at least 1 month in 2019.dBeneficiaries with a long-term care stay qualifying for the Centers for Medicare & Medicaid Services quality measure (Ն100 days in the facility without a gap of Ն30 days in between), and which spans January 1, 2019 (identified using the minimum data set).ehRural and urban do not sum to 100% due to invalid zip codes in claims or zip codes with missing rural-urban commuting area codes.of

Table 3 .
Characteristics of Beneficiaries Identified as Having Highly Likely ADRD, Likely ADRD, Possible ADRD, and No Evidence of ADRD a Row percentages (ie, percentage of sample categorized into each ADRD group).bSamplesubset with beneficiary age restricted to age 65 to 117 years (N = 50 833 592).cSamplesubset with beneficiary age restricted to age 65 to 117 years and insurance enrollment restricted to those with both Medicare Part A and Medicare Part B enrollment in January 2019 (N = 46 767 159).dOtherincluded any race or ethnicity not otherwise specified.eHaddual-eligibility for at least 1 month in 2019.fEnrolled in Medicare Advantage (Medicare Part C) for at least 1 month in 2019.gBeneficiaries with a long-term care stay qualifying for the Centers for Medicare & Medicaid Services quality measure (Ն100 days in the facility without a gap of Ն30 days in between), and which spans January 1, 2019 (identified using minimum data set).kRural and urban do not sum to 100% due to invalid zip codes in claims or zip codes with missing ruralurban commuting area codes.

Table 4 .
Prevalence of Highly Likely ADRD, Likely ADRD, and Possible ADRD in Beneficiary Subpopulations Defined by Demographics, Health Care Insurance and Care Utilization, Mortality, and Rural Residency Before and After Standardization to the Age Distribution of the General Population Beneficiaries with an LTC stay qualifying for the Centers for Medicare & Medicaid Services quality measure (Ն100 days in the facility without a gap of Ն30 days in between), and which spans January 1, 2019 (identified using minimum data set).amongLTC users (from 72.8% to 62.8% with highly likely ADRD) and decedents (from 36.2% to 23.8% with highly likely ADRD) but had minimal impact in ADRD prevalence among non-LTC users and nondecedents; this is because LTC-users and decedent groups were heavily skewed toward older ages, while the age distribution of the non-LTC users and nondecedent groups mimicked that of the general Medicare population (eTable 4 in Supplement 1).NDC code ADRD case definitions informed by a systematic review of previous algorithms, author and expert input, and analyses of Medicare data.The review identified 43 ICD-10-CM codes and 5 prescription drugs used by the CCW and 20 researcher-developed algorithms to identify ADRD in Medicare data.We divided codes into categories that were likely to indicate ADRD vs those that were possibly ADRD based on past frequency of use by other researchers, characteristics of beneficiaries identified by codes, and author and expert consensus around code definitions.
Abbreviations: ADRD, Alzheimer disease and related dementias; LTC, long-term care; NA, not applicable.a Prevalence of highly likely ADRD, likely ADRD, possible ADRD, and no evidence of ADRD within population subgroups.Percentages are by row.b Other includes Asian or Pacific Islander, American Indian or Alaska Native, any race or ethnicity not otherwise specified, or unknown.c Had dual-eligibility for at least 1 month in 2019.d Enrolled in Medicare Advantage (Medicare Part C) for at least 1 month in 2019.e

eReferences SUPPLEMENT 2. Data Sharing Statement
this cross-sectional study, our novel case definition for ADRD identified approximately 5.4 million Medicare beneficiaries with evidence of at least likely ADRD in 2019.Pending validation against in-person clinical and other ascertainment methods, this definition can be adopted for provisional use in national surveillance efforts.ICD-10-CM Codes and Prescription Drugs Used in the CCW and 21 Unique Researcher-Developed Claims-Based Dementia Identification Algorithms eTable 2. Characteristics of Beneficiaries Categorized Into Each Tier of ICD-10-CM Codes (as Defined by Frequency of Use Across the CCW and Researcher-Developed Algorithms) and NDCs eTable 3. Raw and Age-Adjusted Characteristics of Beneficiaries Identified as Having Highly Likely ADRD, Likely ADRD, Possible ADRD, and No Evidence of ADRD eTable 4. Beneficiary Age Distribution in the Full Sample, Within LTC Users and Non-Users, and Within Decedents and Non-Decedents eTable 5. New Subcodes Associated With F01, F02, and F03 That Went Into Effect in October 2022 SUPPLEMENT 1. eAppendix.Literature Review Protocol eTable 1.