Evaluating linked ICD‐10 Medicare claims data as a method of dementia case ascertainment in research settings

Abstract

INTRODUCTION

US Medicare claims can be used to identify dementia cases for research. Our objective was to evaluate the performance of International Classification of Diseases, 10th Revision (ICD‐10) code definitions versus research‐based dementia ascertainment.

METHODS

Participants of five Rush Alzheimer's Disease Center (RADC) cohorts with study visits between October 2015 and December 2019 and fee‐for‐service Medicare contributed observations. For each observation, we compared research‐based dementia status to dementia status based on six ICD‐10 code definitions.

RESULTS

A total of 1869 participants contributed 5309 observations (mean age 82.9 years, 21.0% Black, 9.3% met research‐based dementia criteria). The accuracy of ICD‐10 code definitions was high (87%–90%); five of six code definitions favored specificity over sensitivity. All ICD‐10 code definitions were less accurate among subgroups defined by older age, minoritized race, increased depressive symptoms, and history of stroke.

DISCUSSION

Performance of ICD‐10 code definitions mirrored that of ICD‐9 code definitions. Awareness of differential performance by participant characteristics can improve the robustness of research.

Highlights

• We report the performance of the International Classification of Diseases, 10th Revision (ICD‐10) code versus research‐based dementia ascertainment.

• ICD‐10 performed worse with age, depressive symptoms, minoritized race, and stroke.

• Awareness of accuracy and differential performance can improve research robustness.

1. BACKGROUND

In‐person examinations followed by expert ascertainment have long been the standard procedure used to identify participants with dementia in research settings; these examinations typically include cognitive testing, a neurological exam, and participant interviews. Experts evaluate the resulting data and determine each participant's dementia status. This data collection method is expensive and time consuming, and this inefficiency limits researchers’ ability to ascertain dementia status at scale. An alternative to in‐person examination for dementia case ascertainment in research settings is linkage to administrative data. In the United States, this can be accomplished using claims data from Medicare, which enrolls the majority of Americans over the age of 65. ¹ Previous research compared dementia ascertainment efforts using International Classification of Diseases, 9th Revision (ICD‐9) US Medicare claims to in‐person examination followed by expert ascertainment and found that overall agreement on dementia status was ≈ 85% to 90%. ² , ³ , ⁴ , ⁵ However, there is evidence that accuracy of claims may differ across subgroups; in particular, dementia identification using claims data systematically under‐identifies dementia in persons of Black race or Hispanic ethnicity, those living in rural areas, and those with low education levels, ⁴ , ⁵ , ⁶ , ⁷ while those with more severe symptoms of dementia or worse overall health appear more likely to be correctly identified as having dementia in claims data. ³ , ⁴ , ⁶ , ⁸

In October of 2015, Medicare transitioned from ICD‐9 codes to International Classification of Diseases, 10th Revision (ICD‐10) codes. The ICD‐10 coding system reflects new etiological understandings of many conditions, including dementia; as such, correspondence between ICD‐9 and ICD‐10 codes is often poor, including for dementia. ⁹ Since the switch, researchers have used several different ICD‐10 code definitions (i.e., lists of codes) for ascertaining dementia using Medicare data. ¹⁰ Here, our objective was to evaluate performance metrics of several ICD‐10 code definitions created for use in Medicare data to ascertain dementia status against research‐based in‐person examination and expert ascertainment in data from the Rush Alzheimer's Disease Center (RADC) cohorts, overall and by participant characteristics.

2. METHODS

2.1. Source data

Data were collected from five cohorts of aging adults from the RADC. The Religious Orders Study (ROS) started data collection in 1994 and includes older Catholic nuns, priests, and brothers from across the United States. ¹¹ The Rush Memory and Aging Project (MAP) started data collection in 1997 and includes members of retirement communities and subsidized senior housing in the Chicago metropolitan area. ¹¹ The Minority Aging Research Study (MARS) started data collection in 2004 and includes Black older adults living in the Chicago metropolitan area. ¹² The African American Clinical Core (AA Core) transitioned from a primarily clinic‐based study in 2008 to include community‐dwelling Black older adults living in the Chicago metropolitan area. ¹³ The Latino Core Study (LATC) started data collection in 2015 and includes older adults who identify as Latino and/or Hispanic living in the Chicago metropolitan area. ¹⁴ Participant enrollment and data collection is ongoing for all cohorts. All participants are ≥ 60 when enrolled in a cohort and agree to annual study visits. Data collection is standardized across cohorts and includes cognitive exams, motor testing, risk factor assessment, and dementia screening/ascertainment.

2.2. Cohort‐based dementia ascertainment

Dementia status is determined independently at each annual study visit without knowledge of previous dementia status determination. As part of the standard RADC data collection procedures, at each annual study visit, all participants complete a structured interview in which they self‐report demographic characteristics, medical history, and experiential risk factors. Participants then complete a battery of 19 cognitive tests, the scores of which are used to rate the severity of overall cognitive impairment including impairment in five cognitive domains: episodic memory, working memory, semantic memory, perceptual speed, and visuospatial ability. ¹⁵ A clinical neuropsychologist reviews the test results and interview data, blinded to the age, sex, and race of participants, and evaluates each participant's cognitive impairment. All collected data and the cognitive impairment evaluation are then evaluated by a neurologist, who determines the dementia status (yes/no) of each participant. ¹¹ Important to the current analysis, information about dementia status is not routinely disclosed to participants by the RADC, although information is available if requested.

2.3. Medicare linkage

With the exception of MARS, all RADC study participants are asked for consent to access their Medicare claims records as part of the main study consent process; in MARS, the Medicare linkage consent process began after cohort inception and was separate from main study consent. Linkage to Medicare fee‐for‐service (FFS) claims data for consenting participants is currently available for the period of 1991 to 2021. For this study, we used Medicare Parts A and B files from 2015 to 2019.

2.4. Inclusion criteria

RADC cohort participants were eligible for this analysis if they had FFS Medicare coverage for at least 1 month of the 12‐month period (± 6 months) surrounding a study visit between October 2015 and December 2019 (hereafter referred to as an “index visit”) and have non‐missing data on RADC cohort‐determined dementia status and relevant covariate data at that index visit. Restricting to potential study index visits within the period of December 2015 to June 2019 ensured that we would use all available Medicare data between the time of the switch to ICD‐10 in October 2015 through the end of 2019, the end of the calendar year preceding the onset of the COVID‐19 pandemic. Participants could contribute multiple observations anchored to multiple index visits during this period, providing relevant criteria were met.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional sources (e.g., PubMed). Studies to date have compared International Classification of Diseases, 9th Revision (ICD‐9) dementia code definitions to a research standard, little has been reported on the performance of commonly used ICD‐10 dementia code definitions. Relevant citations are appropriately referenced.

2. Interpretation: We found that the performance of ICD‐10 code definitions was similar to previously reported performance metrics for ICD‐9 code definitions, and that performance varied by participant characteristics. All ICD‐10 code definitions were less accurate among persons of older age and minoritized race, as well as those with increased depressive symptoms or a history of stroke.

3. Future directions: Awareness of overall accuracy and differential performance by participant characteristics can improve the use of ICD‐10 dementia code definitions when used in recruitment and analyses, improving the robustness of research.

2.5. Dementia ascertainment using Medicare ICD‐10 claims

A systematic review conducted by members of our research team ¹⁰ identified a total of 21 ICD‐10 code definitions used previously in the literature for identification of dementia in US Medicare claims data through 2022. Many of these code definitions are similar—for example, some code definitions differ only in the inclusion or exclusion of a single code—and to our knowledge, several have only been used in a single publication. Here, we consider a subset of six ICD‐10 code definitions for dementia (Table 1). Four of these code definitions have been used in recent high‐impact studies that evaluated ICD‐10 dementia codes from Medicare claims and/or are in wide use among researchers: Bynum Standard, ⁴ CCW (Chronic Conditions Warehouse), ⁷ Moura et al., ⁹ and Jain et al. ¹⁶ The additional two code definitions were developed as the outcome of the systematic review based on synthesis of the totality of ICD‐10 definitions used previously in the literature and distills previous decision making across studies into a flexible and inclusive set of case definitions algorithm which allow researchers to prioritize specificity (highly likely and likely dementia) or allow a more expansive set of codes (highly likely, likely, and possible dementia). ¹⁰ Specifically, the NORC highly likely, likely, and probably code definition contains an expanded list of ICD‐10 codes with corresponding requirements for number and location of codes compared to the NORC highly likely and likely code definition (Table 1), recognizing that specific codes reflect a diagnosis of dementia, rather than other related conditions, with varying degrees of certainty.

TABLE 1.

ICD‐10 code dementia definitions used to identify dementia diagnosis in research settings.

		ICD‐10 code definition
ICD‐10 code	ICD‐10 description	CCW a	Bynum standard b	Moura et al. c	Jain et al. d	NORC highly likely and likely e	NORC combined highly likely, likely, and possible e
F01.50	Vascular dementia without behavioral disturbance	x	x	X	x	x	x
F01.51	Vascular dementia with behavioral disturbance	x	x	x	x	x	x
F02.80	Dementia in other diseases classified elsewhere without behavioral disturbance	x	x	x	x	x	x
F02.81	Dementia in other diseases classified elsewhere with behavioral disturbance	x	x	x	x	x	x
F03.90	Unspecified dementia without behavioral disturbance	x	x	x	x	x	x
F03.91	Unspecified dementia with behavioral disturbance	x	x		x	x	x
F04	Amnestic disorder due to known physiological condition	x	x		x		x
F05	Delirium due to known physiological condition	x
F06.1	Catatonic disorder due to known physiological condition	x					x
F06.8	Other specified mental disorders due to known physiological condition	x					x
G13.8	Systemic atrophy primarily affecting central nervous system in other diseases classified elsewhere	x					x
G30.0	Alzheimer's disease with early onset	x	x	x	x	x	x
G30.1	Alzheimer's disease with late onset	x	x	x	x	x	x
G30.8	Other Alzheimer's disease	x	x	x	x	x	x
G30.9	Alzheimer's disease, unspecified	x	x	x	x	x	x
G31.0	Frontotemporal dementia					x	x
G31.01	Pick's disease	x	x	x		x	x
G31.09	Other frontotemporal dementia	x	x	x		x	x
G31.1	Senile degeneration of brain, not elsewhere classified	x	x	x		x	x
G31.2	Degeneration of nervous system due to alcohol	x	x	x			x
G31.83	Dementia with Lewy bodies		x			x	x
G31.84	Mild cognitive impairment, so stated						x
G31.89	Other specified degenerative diseases of nervous system						x
G31.9	Degenerative disease of nervous system, unspecified						x
G91.4	Hydrocephalus in diseases classified elsewhere			x
G94	Other disorders of brain in diseases classified elsewhere	x					x
R41.81	Age‐related cognitive decline	x	x	x	x	x	x
R54	Age‐related physical debility	x					x
Code definition requirements		1+ Inpatient/SNF/HHA/HOP/carrier claims	1+ Inpatient/SNF/HHA/hospice claim, 2+ carrier claims that are 7+ days apart, or 2+ HOP (federally qualified health centers, rural health centers, or critical access hospitals—option II only) claims that are 7+ days apart	1+ claim from Medicare Parts A or B	1+ inpatient/HOP/carrier/SNF/HHA claim in at least 2 different years, or 1+ claim plus a total stay of 6 months in a nursing home in MDS	Highly likely: 2+ inpatient/HOP/carrier/SNF/HHA claim; likely: 1+ inpatient/HOP/carrier/SNF/HHA claim	Highly likely: 2+ inpatient/HOP/Carrier/SNF/HHA claims; Likely: 1+ Inpatient/HOP/carrier/SNF/HHA claim; probably: not in highly likely or likely and 1+ inpatient/HOP/carrier/SNF/HHA claim

Notes: Our analysis used a reference period of 1 year surrounding the index date. However, all algorithms except for the Bynum Standard were designed to be used with a reference period of 3 years prior to the index date; Bynum Standard was designed to be used with a reference period of 3 years surrounding the index date.

Abbreviations: CCW, Chronic Conditions Warehouse; HHA, home health agency; HOP, hospital outpatient; ICD, International Classification of Diseases; MDS, minimum data set; SNF, skilled nursing facility.

^a See McCarthy et al. ⁷

^b See Grodstein et al. ⁴

^c See Moura et al. ⁹

^d See Jain et al. ¹⁶

^e See Gianattasio et al. ¹⁰

All available Medicare FFS claims data from the 12 month period (± 6 months) surrounding a study visit were used to determine participants’ dementia status according to each ICD‐10 code definition, which differs based on the specific codes required, Medicare source files in which the codes should be found, and number of instances of qualifying codes required (Table 1). Although each code definition has a different lookback period (ranging from 1 to 3 years), we used a standard 1 year lookback period for all code definitions in analyses so that we could make direct comparisons.

2.6. Statistical analyses

We calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy of Medicare claims dementia status for each of the six code definitions relative to RADC cohort dementia status at each index visit, overall and by participant characteristics at the time of the index visit. All participant characteristics were defined based on study visit data and included self‐reported age (years); sex (men/women); race (minoritized/White); Hispanic/Latino ethnicity (yes/no); education (< 16 years, ≥ 16 years); lifetime history (yes/no) of hypertension, cancer, diabetes, head injury with loss of consciousness, and stroke; and current depressive symptoms (score from 0 to 10 symptoms, Center for Epidemiologic Studies Depression Scale, 10‐item version [CES‐D‐10] score). ¹⁷ Minoritized race includes all participants who did not self‐report White race. Race was binarized due to small numbers of participants in non‐White racial groups.

We then used adjusted logistic regression using generalized estimating equations (GEEs) to account for the fact that a single individual could contribute multiple observations to our sample to identify covariates associated with higher odds of inaccuracy relative to research‐based classification (i.e., false positive or false negative), false positive classification (in a sample restricted to correct and false positive classified observations), or false negative classification (in a sample restricted to correct and false negative classified observations).

3. RESULTS

3.1. Sample characteristics

Our eligible sample included 1869 unique RADC cohort participants, who contributed 5309 observations after exclusions for missing covariate data. Overall, 543 observations had positive RADC dementia status, and 4766 had negative RADC dementia status. Characteristics of the sample observations by dementia status are shown in Table 2. Observations come from a population that is relatively old (average age of 82.9 years, standard deviation [SD] 7.6) and well educated (average of 16.4 years of education, SD 3.7). The majority of observations are from participants who are women (78.3%), White (76.3%), and non‐Hispanic (94.5%). The most common comorbidities are history of hypertension (69.4%) and history of cancer (45.1%). Notably, observations with RADC‐ascertained dementia occur at an older age (mean age of 90.1 [SD 6.3] years, compared to 82.1 [SD 7.3] years for observations without dementia) and have a lower average Mini‐Mental State Examination (MMSE) score (15.0 [SD 8.2], compared to 28.0 [SD 2.1] for observations without dementia.)

TABLE 2.

Sample observation characteristics (N = 5309) of Rush Alzheimer's Disease Center (RADC) cohorts with Medicare linkage from October 2015 to December 2019.

	RADC dementia status
Variable	Total (N = 5309)	Yes (N = 543)	No (N = 4766)
Age in years, mean (SD)	82.9 (7.6)	90.1 (6.3)	82.1 (7.3)
Education, N (%)
Higher education (≥16 years)	3457 (65.1)	363 (66.9)	3094 (64.9)
Lower education (<16 years)	1852 (34.9)	180 (33.1)	1672 (35.1)
Sex, N (%)
Female	4158 (78.3)	431 (79.4)	3727 (78.2)
Male	1151 (21.7)	112 (20.6)	1039 (21.8)
Race, N (%)
Minoritized	1257 (23.7)	69 (12.7)	1188 (24.9)
White	4052 (76.3)	474 (87.3)	3578 (75.1)
Latino ethnicity, N (%)
Yes	291 (5.5)	32 (5.9)	259 (5.4)
No	5018 (94.5)	511 (94.1)	4507 (94.6)
RADC study cohort, N (%)
MAP	2388 (45.0)	288 (53.0)	2100 (44.1)
ROS	1932 (36.4)	232 (42.7)	1700 (35.7)
Other	989 (18.7)	23 (4.3)	966 (20.3)
History of hypertension, N (%)
Yes	3684 (69.4)	396 (72.9)	3288 (69.0)
No	1625 (30.6)	147 (27.1)	1478 (31.0)
History of cancer, N (%)
Yes	2392 (45.1)	245 (45.1)	2148 (45.1)
No	2916 (54.9)	298 (54.9)	2618 (54.9)
History of diabetes, N (%)
Yes	863 (16.3)	64 (11.8)	799 (16.8)
No	4446 (83.7)	479 (88.2)	3967 (83.2)
History of head injury with loss of consciousness, N (%)
Yes	602 (11.3)	91 (16.8)	511 (10.7)
No	4707 (88.7)	452 (83.2)	4255 (89.3)
History of stroke, N (%)
Yes	577 (10.9)	127 (23.4)	450 (9.4)
No	4732 (89.1)	416 (76.6)	4316 (90.6)
Depressive symptoms (CES‐D‐10 score), mean (SD)	1.1 (1.6)	1.4 (2.0)	1.0 (1.6)
Cognitive performance (MMSE score), mean (SD)	26.7 (5.1)	15.0 (8.2)	28.0 (2.1)

Notes: Mean and standard deviation are shown for continuous variables, while N and percentage (%) are shown for categorical variables. All histories of medical conditions are self‐reported. CES‐D‐10 scores range from 0 to 10, with higher scores indicating increased depressive symptoms. MMSE scores range from 0 to 30, with higher scores indicating higher cognitive performance.

Abbreviations: CES‐D‐10, Center for Epidemiologic Studies Depression Scale, 10‐item version; MAP, Memory and Aging Project; MMSE, Mini‐Mental State Examination; ROS, Religious Orders Study; SD, standard deviation.

Table S1 in supporting information provides characteristics of observations from one or more RADC study visits between October 2015 and December 2019 from RADC participants by whether they provided consent to Medicare data linkage. Of note, observations from those who did not give consent for linkage were more likely be among those of minoritized race and to be at visits at which they were classified by the RADC as having dementia. Overall, 93% of potentially eligible observations were among participants with consent to Medicare linkage, including 99% in MAP, 97% in ROS, 79% in MARS, 90% in AA Core, and 85% in LATC.

3.2. ICD‐10 dementia code definition performance

Table 3 shows performance metrics for each ICD‐10 dementia code definition, overall and by subgroups. All six code definitions show similar overall accuracy (range: 87%–90%) and NPV (range: 96%–98%) compared to the RADC research standard. Five of the six code definitions have high specificity at the expense of lower sensitivity; one, Bynum Standard, ⁴ has higher sensitivity (80%) at the expense of lower specificity (88%) relative to the 58% to 62% sensitivity and 92% to 94% specificity observed for the other code definitions. When performance metrics are evaluated in subgroups, results are similar. All code definitions show similar accuracy within subgroups, all code definitions favored specificity over sensitivity except the Bynum Standard, ⁴ and the performance metrics for Moura et al., ⁹ Jain et al., ¹⁶ and NORC highly likely and likely ¹⁰ code definitions were extremely close or identical.

TABLE 3.

Performance metrics of ICD‐10 code definitions compared to Rush Alzheimer's Disease Center standard from October 2015 to December 2019, overall and by subgroup.

	Accuracy	Sensitivity	Specificity	PPV	NPV	RADC dementia status prevalence	Accuracy	Sensitivity	Specificity	PPV	NPV	RADC dementia status prevalence
Overall						0.10
CCW a	0.89	0.60	0.92	0.44	0.96
Bynum standard b	0.87	0.80	0.88	0.40	0.98
Moura et al. c	0.90	0.58	0.94	0.48	0.96
Jain et al. d	0.90	0.58	0.94	0.49	0.96
NORC highly likely and likely e	0.90	0.59	0.94	0.48	0.96
NORC highly likely, likely, and probably e	0.87	0.62	0.89	0.37	0.96
Age	Younger age (<80)					0.01	Older age (≥80)					0.14
CCW	0.96	0.61	0.96	0.19	0.99		0.85	0.60	0.89	0.47	0.93
Bynum standard	0.94	0.71	0.94	0.15	0.99		0.83	0.80	0.83	0.43	0.96
Moura et al.	0.97	0.58	0.97	0.23	0.99		0.87	0.58	0.91	0.52	0.93
Jain et al.	0.97	0.58	0.97	0.23	0.99		0.87	0.58	0.91	0.52	0.93
NORC highly likely and likely	0.97	0.58	0.97	0.22	0.99		0.87	0.59	0.91	0.52	0.93
NORC highly likely, likely, and probably	0.95	0.64	0.95	0.16	0.99		0.82	0.62	0.85	0.40	0.93
Education	Lower education (<16)					0.09	Higher education (≥16)					0.10
CCW	0.88	0.61	0.90	0.36	0.96		0.90	0.59	0.93	0.48	0.96
Bynum standard	0.86	0.82	0.86	0.35	0.98		0.88	0.79	0.89	0.43	0.98
Moura et al.	0.89	0.58	0.92	0.41	0.96		0.91	0.58	0.94	0.53	0.96
Jain et al.	0.89	0.58	0.92	0.41	0.96		0.91	0.58	0.95	0.54	0.96
NORC highly likely and likely	0.89	0.58	0.92	0.41	0.96		0.91	0.59	0.94	0.53	0.96
NORC highly likely, likely, and probably	0.84	0.62	0.87	0.30	0.96		0.88	0.63	0.90	0.41	0.96
Sex	Male					0.09	Female					0.09
CCW	0.88	0.57	0.91	0.38	0.96		0.89	0.61	0.92	0.45	0.96
Bynum standard	0.85	0.73	0.86	0.33	0.97		0.88	0.82	0.88	0.42	0.98
Moura et al.	0.89	0.55	0.92	0.41	0.96		0.91	0.59	0.94	0.51	0.96
Jain et al.	0.89	0.54	0.92	0.40	0.95		0.91	0.59	0.94	0.52	0.96
NORC highly likely and likely	0.89	0.55	0.92	0.40	0.96		0.91	0.60	0.94	0.51	0.96
NORC highly likely, likely, and probably	0.85	0.59	0.88	0.32	0.96		0.87	0.63	0.89	0.38	0.96
Race f	Minoritized race					0.05	White race					0.11
CCW	0.91	0.46	0.93	0.24	0.97		0.88	0.62	0.92	0.48	0.95
Bynum standard	0.89	0.61	0.91	0.24	0.98		0.86	0.83	0.87	0.43	0.98
Moura et al.	0.92	0.44	0.94	0.26	0.97		0.90	0.61	0.93	0.53	0.95
Jain et al.	0.92	0.44	0.94	0.26	0.97		0.90	0.60	0.94	0.54	0.95
NORC highly likely and likely	0.92	0.44	0.94	0.26	0.97		0.90	0.61	0.93	0.53	0.95
NORC highly likely, likely, and probably	0.89	0.46	0.91	0.20	0.97		0.86	0.65	0.88	0.40	0.95
Ethnicity	Hispanic ethnicity					0.09	Non‐Hispanic ethnicity					0.09
CCW	0.90	0.50	0.94	0.46	0.95		0.90	0.60	0.92	0.44	0.96
Bynum standard	0.88	0.59	0.91	0.41	0.96		0.87	0.81	0.87	0.40	0.98
Moura et al.	0.90	0.44	0.95	0.47	0.94		0.90	0.59	0.94	0.48	0.96
Jain et al.	0.90	0.41	0.95	0.43	0.94		0.91	0.59	0.94	0.49	0.96
NORC highly likely and likely	0.90	0.44	0.95	0.45	0.94		0.90	0.60	0.94	0.48	0.96
NORC highly likely, likely, and probably	0.89	0.53	0.93	0.43	0.95		0.86	0.63	0.89	0.37	0.96
Hypertension	History of hypertension					0.10	No history of hypertension					0.08
CCW	0.89	0.57	0.92	0.45	0.95		0.89	0.67	0.91	0.40	0.97
Bynum standard	0.87	0.78	0.88	0.42	0.97		0.87	0.83	0.87	0.35	0.98
Moura et al.	0.90	0.56	0.94	0.51	0.95		0.90	0.65	0.92	0.42	0.97
Jain et al.	0.91	0.56	0.94	0.52	0.95		0.90	0.65	0.92	0.43	0.97
NORC highly likely and likely	0.90	0.56	0.94	0.51	0.95		0.90	0.65	0.92	0.42	0.97
NORC highly likely, likely, and probably	0.87	0.59	0.90	0.38	0.95		0.87	0.71	0.88	0.34	0.97
Cancer	History of cancer					0.10	No history of cancer					0.09
CCW	0.89	0.58	0.93	0.45	0.95		0.89	0.61	0.92	0.42	0.96
Bynum standard	0.87	0.79	0.88	0.41	0.98		0.86	0.80	0.87	0.39	0.98
Moura et al.	0.91	0.57	0.94	0.52	0.95		0.90	0.60	0.93	0.46	0.96
Jain et al.	0.91	0.57	0.95	0.53	0.95		0.90	0.59	0.93	0.46	0.96
NORC highly likely and likely	0.91	0.57	0.94	0.51	0.95		0.90	0.60	0.93	0.46	0.96
NORC highly likely, likely, and probably	0.86	0.61	0.89	0.36	0.96		0.87	0.64	0.89	0.37	0.96
Diabetes	History of diabetes					0.07	No history of diabetes					0.10
CCW	0.90	0.60	0.92	0.35	0.97		0.89	0.59	0.92	0.45	0.95
Bynum standard	0.88	0.77	0.89	0.32	0.98		0.87	0.80	0.88	0.41	0.98
Moura et al.	0.92	0.60	0.94	0.42	0.97		0.90	0.58	0.94	0.49	0.95
Jain et al.	0.92	0.58	0.94	0.41	0.97		0.90	0.58	0.94	0.50	0.95
NORC highly likely and likely	0.92	0.60	0.94	0.41	0.97		0.90	0.58	0.94	0.49	0.95
NORC highly likely, likely, and probably	0.88	0.62	0.90	0.30	0.97		0.86	0.62	0.89	0.38	0.96
Head injury with loss of consciousness	History of head injury with loss of consciousness					0.15	No history of head injury with loss of consciousness					0.09
CCW	0.87	0.55	0.92	0.56	0.92		0.89	0.61	0.92	0.44	0.96
Bynum standard	0.84	0.73	0.85	0.48	0.95		0.87	0.81	0.87	0.40	0.98
Moura et al.	0.87	0.54	0.93	0.60	0.92		0.90	0.60	0.93	0.49	0.96
Jain et al.	0.88	0.54	0.94	0.60	0.92		0.90	0.60	0.94	0.49	0.96
NORC highly likely and likely	0.87	0.54	0.93	0.60	0.92		0.90	0.60	0.93	0.49	0.96
NORC highly likely, likely, and probably	0.85	0.60	0.89	0.50	0.93		0.87	0.63	0.89	0.38	0.96
Stroke	History of stroke					0.21	No history of stroke					0.08
CCW	0.78	0.57	0.84	0.49	0.88		0.90	0.61	0.93	0.42	0.97
Bynum standard	0.78	0.78	0.78	0.48	0.93		0.88	0.81	0.89	0.38	0.98
Moura et al.	0.80	0.54	0.86	0.51	0.88		0.92	0.60	0.94	0.48	0.96
Jain et al.	0.80	0.54	0.86	0.51	0.88		0.92	0.60	0.94	0.48	0.96
NORC highly likely and likely	0.80	0.55	0.86	0.52	0.88		0.92	0.60	0.94	0.47	0.96
NORC highly likely, likely, and probably	0.75	0.58	0.79	0.43	0.88		0.88	0.64	0.90	0.35	0.97
Depressive symptoms (CES‐D‐10 score)	CES‐D‐10 score of 0					0.06	CES‐D‐10 score above 0					0.07
CCW	0.91	0.49	0.94	0.34	0.96		0.88	0.54	0.91	0.31	0.96
Bynum standard	0.89	0.75	0.90	0.33	0.98		0.85	0.74	0.86	0.29	0.98
Moura et al.	0.92	0.48	0.95	0.38	0.96		0.90	0.52	0.93	0.35	0.96
Jain et al.	0.92	0.48	0.95	0.38	0.97		0.90	0.51	0.93	0.36	0.96
NORC highly likely and likely	0.92	0.48	0.95	0.38	0.96		0.90	0.52	0.93	0.35	0.96
NORC highly likely, likely, and probably	0.89	0.54	0.91	0.27	0.97		0.85	0.56	0.87	0.26	0.96

^a See McCarthy et al. ⁷

^b See Grodstein et al. ⁴

^c See Moura et al. ⁹

^d See Jain et al. ¹⁶

^e See Gianattasio et al. ¹⁰

^f Minoritized race includes all participants who did not self‐report White race. Race was binarized due to small numbers of participants in non‐White racial groups.

Abbreviations: CCW, Chronic Conditions Warehouse; CES‐D‐10, Center for Epidemiologic Studies Depression Scale, 10‐item version; MMSE, Mini‐Mental State Examination; NPV, negative predictive value; PPV, positive predictive value.

However, although performance was similar across algorithms by subgroup, performance varied substantially across subgroups regardless of the algorithm. For example, accuracy ranged from 75% to 97%, and NPV ranged from 88% to 99%. The younger age subgroup (< 80 years old) had the highest accuracy (range: 94%–97%) and highest NPV (range: 99%–99%) compared to any other subgroup, but this group also had the lowest PPV (range: 15%–23%). The subgroup with a history of stroke had the lowest accuracy (range: 75%–80%) and lowest NPV (range: 88%–93%).

3.3. Characteristics associated with accuracy of ICD‐10 dementia code definitions

Factors associated with accuracy, false positive, and false negative classification were reasonably consistent across all code definitions (Tables 4, 5, 6). ICD‐10 code definitions were significantly more accurate with higher MMSE score and in persons with a history of hypertension or cancer, but less accurate with increased age, increased depressive symptom levels, in minoritized racial groups, and in persons with a history of stroke (Table 4). Older age, more depressive symptoms, and history of stroke were significantly associated with increased odds of false positive classification, while the association between minority race and false positive classification was substantially elevated, but only significant or marginally significant for half of the observed code definitions (Table 5). Conversely, persons with a history of hypertension or cancer were less likely to receive false positive classification (Table 5). Older age, higher education, minoritized race, Hispanic ethnicity, and history of stroke were significantly associated with increased odds of false negative classification, while higher cognitive performance was associated with reduced risk (Table 6).

TABLE 4.

Adjusted odds ratios of covariates associated with accuracy of ICD‐10 dementia code definitions.

	CCW a		Bynum standard b		Moura et al. c		Jain et al. d		NORC highly likely and likely e		NORC highly likely, likely, and probably e
Variable	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p
Age at visit	0.92 (0.90, 0.94)	<0.01	0.93 (0.91, 0.94)	<0.01	0.92 (0.90, 0.94)	<0.01	0.92 (0.90, 0.94)	<0.01	0.92 (0.90, 0.94)	<0.01	0.92 (0.91, 0.94)	<0.01
Male sex	0.96 (0.69, 1.33)	0.79	0.86 (0.64, 1.17)	0.34	0.94 (0.67, 1.33)	0.73	0.92 (0.65, 1.31)	0.65	0.94 (0.67, 1.32)	0.71	0.95 (0.71, 1.29)	0.75
≥16 years education	0.99 (0.74, 1.31)	0.92	1.04 (0.79, 1.37)	0.79	0.90 (0.66, 1.21)	0.48	0.88 (0.65, 1.19)	0.40	0.90 (0.67, 1.22)	0.50	1.08 (0.83, 1.40)	0.57
Minority race	0.67 (0.47, 0.95)	0.02	0.83 (0.60, 1.14)	0.25	0.58 (0.40, 0.83)	<0.01	0.57 (0.40, 0.82)	<0.01	0.58 (0.40, 0.83)	<0.01	0.71 (0.52, 0.98)	0.04
Hispanic ethnicity	0.87 (0.46, 1.66)	0.67	0.98 (0.52, 1.84)	0.94	0.74 (0.38, 1.44)	0.38	0.70 (0.37, 1.37)	0.30	0.74 (0.38, 1.44)	0.37	1.13 (0.64, 2.00)	0.67
History of hypertension	1.37 (1.02, 1.83)	0.04	1.41 (1.07, 1.87)	0.01	1.62 (1.19, 2.19)	<0.01	1.63 (1.20, 2.21)	<0.01	1.61 (1.18, 2.18)	<0.01	1.30 (0.98, 1.71)	0.07
History of cancer	1.33 (1.01, 1.74)	0.04	1.28 (0.99, 1.66)	0.06	1.42 (1.06, 1.90)	0.02	1.47 (1.09, 1.98)	0.01	1.41 (1.05, 1.89)	0.02	1.14 (0.89, 1.47)	0.30
History of diabetes	0.88 (0.60, 1.29)	0.51	0.93 (0.66, 1.32)	0.69	0.97 (0.64, 1.48)	0.88	0.94 (0.62, 1.43)	0.77	0.97 (0.64, 1.47)	0.87	0.90 (0.64, 1.27)	0.55
History of head injury	0.89 (0.61, 1.30)	0.54	0.85 (0.59, 1.23)	0.38	0.85 (0.58, 1.24)	0.40	0.85 (0.58, 1.26)	0.42	0.84 (0.58, 1.24)	0.38	0.98 (0.67, 1.42)	0.90
History of stroke	0.50 (0.35, 0.71)	<0.01	0.59 (0.42, 0.84)	<0.01	0.45 (0.31, 0.66)	<0.01	0.45 (0.31, 0.65)	<0.01	0.46 (0.32, 0.67)	<0.01	0.51 (0.36, 0.71)	<0.01
Depressive symptoms (CES‐D‐10 score) f	0.92 (0.85, 0.98)	0.01	0.90 (0.85, 0.96)	<0.01	0.92 (0.86, 1.00)	0.04	0.92 (0.85, 0.99)	0.04	0.92 (0.86, 1.00)	0.04	0.90 (0.85, 0.96)	<0.01
Cognitive performance (MMSE score) f	1.12 (1.10, 1.16)	<0.01	1.06 (1.03, 1.09)	<0.01	1.15 (1.12, 1.18)	<0.01	1.15 (1.12, 1.18)	<0.01	1.15 (1.11, 1.18)	<0.01	1.10 (1.07, 1.12)	<0.01

Abbreviations: CCW, Chronic Conditions Warehouse; CES‐D‐10, Center for Epidemiologic Studies Depression Scale, 10‐item version; CI, confidence interval; ICD‐10, International Classification of Diseases, 10th Revision; MMSE, Mini‐Mental State Examination.

Bolded values indicate statistical significance, p < 0.05.

^a See McCarthy et al. ⁷

^b See Grodstein et al. ⁴

^c See Moura et al. ⁹

^d See Jain et al. ¹⁶

^e See Gianattasio et al. ¹⁰

^f Odds ratios correspond to the change per 1‐unit increase in CES‐D‐10 or MMSE score. Higher scores indicate increased depressive symptoms and cognitive performance, respectively.

TABLE 5.

Adjusted odds ratios of covariates associated with false positivity of ICD‐10 dementia code definitions.

	CCW a		Bynum standard b		Moura et al. c		Jain et al. d		NORC highly likely and likely e		NORC highly likely, likely, and probably e
Variable	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p
Age at visit	1.07 (1.05, 1.10)	<0.01	1.07 (1.05, 1.09)	<0.01	1.07 (1.04, 1.09)	<0.01	1.07 (1.04, 1.09)	<0.01	1.07 (1.04, 1.09)	<0.01	1.08 (1.05, 1.10)	<0.01
Male sex	1.18 (0.80, 1.76)	0.41	1.16 (0.83, 1.62)	0.38	1.21 (0.79, 1.86)	0.37	1.22 (0.79, 1.88)	0.37	1.21 (0.79, 1.86)	0.37	1.15 (0.81, 1.62)	0.44
≥16 years education	0.72 (0.52, 1.00)	0.05	0.81 (0.60, 1.08)	0.15	0.78 (0.54, 1.12)	0.18	0.81 (0.56, 1.16)	0.24	0.78 (0.54, 1.12)	0.18	0.73 (0.54, 0.97)	0.03
Minority race	1.28 (0.86, 1.92)	0.23	1.01 (0.71, 1.44)	0.95	1.52 (1.00, 2.32)	0.05	1.57 (1.02, 2.40)	0.04	1.52 (1.00, 2.32)	0.05	1.24 (0.87, 1.77)	0.24
Hispanic ethnicity	0.77 (0.35, 1.70)	0.52	0.62 (0.28, 1.38)	0.24	0.80 (0.33, 1.90)	0.60	0.79 (0.33, 1.90)	0.60	0.80 (0.33, 1.89)	0.60	0.63 (0.29, 1.36)	0.24
History of hypertension	0.67 (0.47, 0.96)	0.03	0.67 (0.50, 0.91)	<0.01	0.54 (0.37, 0.78)	<0.01	0.52 (0.35, 0.76)	<0.01	0.54 (0.37, 0.78)	<0.01	0.71 (0.52, 0.97)	0.03
History of cancer	0.74 (0.53, 1.02)	0.07	0.78 (0.59, 1.03)	0.08	0.67 (0.47, 0.98)	0.04	0.64 (0.44, 0.93)	0.02	0.67 (0.47, 0.98)	0.04	0.90 (0.67, 1.19)	0.45
History of diabetes	1.18 (0.75, 1.86)	0.47	1.11 (0.76, 1.60)	0.60	1.10 (0.66, 1.82)	0.72	1.12 (0.68, 1.87)	0.65	1.10 (0.66, 1.82)	0.72	1.14 (0.77, 1.67)	0.52
History of head injury	0.87 (0.52, 1.46)	0.60	1.04 (0.68, 1.59)	0.86	0.90 (0.53, 1.53)	0.70	0.89 (0.52, 1.53)	0.67	0.90 (0.53, 1.53)	0.70	0.87 (0.55, 1.40)	0.57
History of stroke	1.88 (1.20, 2.96)	<0.01	1.65 (1.12, 2.43)	0.01	2.03 (1.25, 3.30)	<0.01	2.11 (1.29, 3.43)	<0.01	2.03 (1.25, 3.30)	<0.01	1.83 (1.22, 2.73)	<0.01
Depressive symptoms (CES‐D‐10 score) f	1.16 (1.08, 1.25)	<0.01	1.14 (1.07, 1.22)	<0.01	1.15 (1.06, 1.25)	<0.01	1.15 (1.06, 1.25)	<0.01	1.15 (1.06, 1.25)	<0.01	1.15 (1.08, 1.24)	<0.01
Cognitive performance (MMSE score) f	1.00 (0.98, 1.03)	0.79	0.99 (0.96, 1.01)	0.22	0.99 (0.96, 1.02)	0.50	0.99 (0.96, 1.02)	0.51	0.99 (0.96, 1.02)	0.50	1.02 (1.00, 1.05)	0.10

Abbreviations: CCW, Chronic Conditions Warehouse; CES‐D‐10, Center for Epidemiologic Studies Depression Scale, 10‐item version; CI, confidence interval; ICD‐10, International Classification of Diseases, 10th Revision; MMSE, Mini‐Mental State Examination.

Bolded values indicate statistical significance, p < 0.05.

^a See McCarthy et al. ⁷

^b See Grodstein et al. ⁴

^c See Moura et al. ⁹

^d See Jain et al. ¹⁶

^e See Gianattasio et al. ¹⁰

^f Odds ratios correspond to the change per 1‐unit increase in CES‐D‐10 or MMSE score. Higher scores indicate increased depressive symptoms and cognitive performance, respectively.

TABLE 6.

Adjusted odds ratios of covariates associated with false negativity of ICD‐10 dementia code definitions.

	CCW a		Bynum standard b		Moura et al. c		Jain et al. d		NORC highly likely and likely e		NORC highly likely, likely, and probably e
Variable	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p
Age at visit	1.12 (1.09, 1.16)	<0.01	1.15 (1.10, 1.20)	<0.01	1.12 (1.08, 1.16)	<0.01	1.12 (1.08, 1.16)	<0.01	1.12 (1.08, 1.16)	<0.01	1.11 (1.08, 1.15)	<0.01
Male sex	0.92 (0.56, 1.52)	0.75	1.34 (0.71, 2.52)	0.37	0.93 (0.56, 1.56)	0.79	0.98 (0.58, 1.63)	0.92	0.95 (0.57, 1.58)	0.83	0.95 (0.56, 1.59)	0.83
≥16 years education	1.93 (1.20, 3.10)	<0.01	2.24 (1.18, 4.25)	0.01	1.86 (1.16, 3.00)	0.01	1.81 (1.13, 2.91)	0.01	1.84 (1.15, 2.96)	0.01	1.77 (1.08, 2.89)	0.02
Minority race	2.27 (1.18, 4.35)	0.01	3.44 (1.57, 7.54)	<0.01	2.18 (1.15, 4.15)	0.01	2.12 (1.11, 4.08)	0.02	2.19 (1.15, 4.18)	0.02	2.36 (1.22, 4.57)	0.01
Hispanic ethnicity	2.51 (0.97, 6.51)	0.06	4.45 (1.79, 11.08)	<0.01	2.92 (1.23, 6.94)	0.01	3.18 (1.34, 7.56)	<0.01	2.92 (1.23, 6.94)	0.02	2.26 (1.12, 4.59)	0.02
History of hypertension	0.84 (0.54, 1.31)	0.45	0.90 (0.49, 1.65)	0.72	0.81 (0.53, 1.25)	0.35	0.84 (0.55, 1.30)	0.44	0.83 (0.54, 1.28)	0.41	0.95 (0.60, 1.52)	0.84
History of cancer	0.79 (0.52, 1.23)	0.30	0.75 (0.42, 1.33)	0.32	0.78 (0.51, 1.20)	0.26	0.78 (0.51, 1.20)	0.26	0.79 (0.51, 1.22)	0.28	0.78 (0.50, 1.22)	0.27
History of diabetes	1.00 (0.52, 1.92)	0.99	1.00 (0.43, 2.32)	0.99	0.92 (0.48, 1.78)	0.81	0.97 (0.50, 1.86)	0.92	0.93 (0.48, 1.80)	0.83	1.01 (0.52, 1.94)	0.98
History of head injury	1.63 (0.97, 2.75)	0.06	1.83 (0.98, 3.42)	0.06	1.60 (0.95, 2.67)	0.08	1.60 (0.96, 2.68)	0.07	1.61 (0.96, 2.69)	0.07	1.47 (0.87, 2.48)	0.15
History of stroke	2.14 (1.28, 3.59)	<0.01	1.80 (0.87, 3.72)	0.11	2.30 (1.39, 3.81)	<0.01	2.19 (1.33, 3.62)	<0.01	2.22 (1.34, 3.65)	<0.01	2.25 (1.36, 3.72)	<0.01
Depressive symptoms (CES‐D‐10 score) f	0.91 (0.81, 1.04)	0.16	0.92 (0.79, 1.08)	0.31	0.95 (0.83, 1.08)	0.39	0.95 (0.83, 1.08)	0.41	0.95 (0.84, 1.08)	0.42	0.94 (0.83, 1.06)	0.28
Cognitive performance (MMSE score) f	0.81 (0.79, 0.84)	<0.01	0.87 (0.84, 0.90)	<0.01	0.81 (0.78, 0.84)	<0.01	0.81 (0.78, 0.84)	<0.01	0.81 (0.78, 0.84)	<0.01	0.81 (0.78, 0.84)	<0.01

Abbreviations: CCW, Chronic Conditions Warehouse; CES‐D‐10, Center for Epidemiologic Studies Depression Scale, 10‐item version; CI, confidence interval; ICD‐10, International Classification of Diseases, 10th Revision; MMSE, Mini‐Mental State Examination.

Bolded values indicate statistical significance, p < 0.05.

^a See McCarthy et al. ⁷

^b See Grodstein et al. ⁴

^c See Moura et al. ⁹

^d See Jain et al. ¹⁶

^e See Gianattasio et al. ¹⁰

^f Odds ratios correspond to the change per 1‐unit increase in CES‐D‐10 or MMSE score. Higher scores indicate increased depressive symptoms and cognitive performance, respectively.

4. DISCUSSION

ICD‐10 code definitions for dementia performed similarly to expectations set by prior experience with ICD‐9 code definitions for dementia. All ICD‐10 code definitions considered had similar accuracy (87%–90%) and NPV (96%–98%) against RADC dementia classification. Most favored specificity over sensitivity except the Bynum Standard, which had higher sensitivity at the expense of slightly lower specificity. To meet the criteria for dementia using the Bynum Standard, a participant must have at least one claim with a qualifying ICD‐10 code from MedPAR (i.e., inpatient or skilled nursing facility), home health, or hospice files, or at least two claims with qualifying ICD‐10 codes in carrier or outpatient files (at least 7 days apart). ⁶ These criteria may allow for the higher sensitivity of the Bynum Standard ICD‐10 code definition.

Reflecting the high degree of overlap across code definitions, demographic correlates of false positives and false negatives were consistent across code definitions. Higher depressive symptoms were associated with false positive classification; minoritized race was associated with false negative classification; and older age and history of stroke were associated with both false positive and false negative classification. As such, the switch to ICD‐10 has not appeared to meaningfully improve some of the bias previously identified in the classification of persons with dementia through Medicare claims data. Aligning with our findings using ICD‐10–based code definitions, previous studies focusing on ICD‐9–based code definitions found that higher MMSE scores and White race were associated with higher accuracy, and that older age was associated with lower accuracy, particularly false positives. ⁴ , ¹⁸

Our findings suggest that using Medicare claims to ascertain dementia cases continues to have limitations for older persons, minoritized racial groups, persons with a history of stroke, or those with elevated depressive symptoms. If we use these ICD‐10 dementia code definitions for case ascertainment, we may generate case samples enriched for persons with characteristics that were found to be associated with false positivity, such as depressive symptoms, while systematically under‐representing those with characteristics found to be associated with false negativity, such as Hispanic ethnicity or minoritized race. Using claims data to ascertain dementia cases in research settings has advantages and disadvantages. Access to claims records is powerful because it allows for analysis of large sample sizes and does not require the costs, resources, and time commitment required for in‐person assessment; this can extend the reach of dementia research to new settings or understudied populations. At the same time, the use of these data may perpetuate a lack of generalizability and diverse representation in research samples. Because performance varies by demographic characteristics, using claims data to compare outcomes based on these characteristics (e.g., race) may bias conclusions.

This study extends our current understanding of dementia identification in claims by suggesting that medical conditions may influence the accuracy of dementia identification in claims data. A history of hypertension was associated with higher accuracy, while history of stroke was associated with lower accuracy. These relationships between dementia identification and comorbid health conditions are unlikely to remain static over time, especially as dementia prevention, treatment, and management are constantly evolving within the medical landscape. It is possible that over time, greater recognition of dementia by the medical community and changing incentives for diagnosis could alter the associations between dementia identification and the comorbid health conditions identified in this study.

Our study expands upon work by a previous group that evaluated ICD‐10 dementia codes commensurate with ICD‐9 dementia code definitions and evaluated the updated code definitions against expert adjudication of dementia status based on review of data abstracted from electronic medical records. ⁹ They found that ICD‐10 dementia code definitions performed similarly to ICD‐9 dementia code definitions compared to expert adjudication, and that performance varied with participant age, with overcounting of dementia cases occurring most frequently in ages 75 to 79 and other demographic subgroups. Our study adds to these findings because we evaluated the performance of multiple code definitions taken from numerous sources against a research‐based reference standard which relied on in‐person study‐based ascertainment according to a standardized protocol and examined performance across additional subgroups defined by history of comorbid health conditions, providing a novel contribution to the field. Our use of a standardized lookback period deviates from the original intended use of the code definitions used in this study; however, this uniformity allows for direct comparisons of code definitions for research purposes. We recognize that the length of a lookback period offers a trade‐off between the high PPV conferred by a short lookback period and higher identification of less severe disease conferred by a longer lookback period, as discussed by McCarthy et al. ⁷

This study has several strengths. Our use of a large and robust dataset of five harmonized cohorts with in‐person, rigorous dementia ascertainment and linkage to Medicare claims data is unique because it allows for analysis of a large sample size, including members of groups that previous research has shown are less likely to be accurately captured by claims data, such as older Black adults. ⁴ , ⁵ , ⁷ , ¹⁹ The RADC aims to minimize attrition by reducing barriers to participation, minimizing selection bias due to attrition as a threat to internal validity. For example, data collection occurs at a participant's home. By evaluating performance across several established code definitions ⁴ , ⁹ , ¹⁶ , ²⁰ and participant subgroups, our work may aid researchers in selecting the ICD‐10 dementia code definition(s) most appropriate for their study objectives and sample. Finally, our selection of ICD‐10 code definitions was based on a systematic review of those previously used in the literature.

Our study also has limitations. First, the majority of participants come from the Chicago metropolitan area, while the only nation‐wide cohort, ROS, exclusively enrolls Catholic nuns, priests, and brothers. Further research may focus on populations in additional geographic regions to confirm our findings. In addition, we only consider ICD‐10–based dementia ascertainment in those with FFS Medicare coverage. Whether these patterns hold in those enrolled in Medicare Advantage is unclear, particularly given the different incentives that may influence coding practice in FFS versus Medicare Advantage. ²¹ , ²² We also recognize that differences may exist between participants who did and did not consent to Medicare data linkage; in particular, persons from racial minority groups and persons with dementia were less likely to be included in our sample. Finally, our findings may not extend into the future as diagnostic procedures evolve. In the United States, advocacy organizations, pharmaceutical companies, and multiple levels of government are all involved in efforts to increase awareness and diagnosis of dementia. Recent approvals of anti‐amyloid monoclonal antibodies that require earlier recognition of cognitive changes and dementia may influence care provision and claims, and Medicare is currently piloting a new dementia care delivery and payment model, Guiding an Improved Dementia Experience (GUIDE); ²³ , ²⁴ , ²⁵ as such, the relationship between dementia status and dementia claims in Medicare data may change.

While the use of ICD‐10 codes in Medicare FFS claims for dementia ascertainment provides a valuable approach to ascertaining dementia at scale, it has limitations that may constrain utility for specific research questions. When making use of Medicare claims data to ascertain dementia cases in research settings, analysts should be sensitive to the limitations of this classification approach relative to their research goals. Our results have practical application. For example, knowledge of measurement properties can be used by researchers during participant recruitment to identify subgroups that might need to be oversampled to get good representation given knowledge of relatively poor sensitivity when relying on ICD‐10 codes to identify cases. During analysis, researchers can use measurement error correction techniques such as regression calibration if algorithm performance differs significantly between subgroups to reduce bias due to differential misclassification. In addition, our results suggest that researchers should incorporate data sources with different ascertainment approaches to triangulate findings and confirm associations found in Medicare samples; confirmation of conclusions across study designs with different known potential biases strengthens inference. Awareness of overall accuracy and differential performance by participant characteristics can improve the use of ICD‐10 dementia code definitions when used in recruitment and analyses, improving the robustness of research.

CONFLICT OF INTEREST STATEMENT

Kan Z. Gianattasio and David B. Rein report grants from the US National Institute on Aging. Christina Prather reports research grants from the US National Institutes of Health and Agency for Healthcare Research and Quality and is/was a paid consultant for the Alzheimer's Association. Raj C. Shah reports being the site principal investigator or sub‐investigator for Alzheimer's disease clinical trials for which his institution (Rush University Medical Center) is/was compensated during the past 24 months (Athira Pharma, Inc., Edgewater NEXT, Eisai, Inc., Eli Lilly & Co., Inc., and Genentech, Inc.). Melinda C. Power reports research grants from the US National Institutes of Health and US Department of Defense, and prior service as a paid member of the Biogen Healthy Climate, Healthy Lives Scientific Advisory Council. Other authors declare no conflicts of interest. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All studies were approved by an institutional review board at Rush University Medical Center. Informed consent for study participation and Medicare linkage were obtained as a repository consent.

Supporting information

Supporting information

Supporting information

Bhattacharyya J, Barnes LL, Chen Y, et al. Evaluating linked ICD‐10 Medicare claims data as a method of dementia case ascertainment in research settings. Alzheimer's Dement. 2025;21:e70200. 10.1002/alz.70200