Expected and diagnosed rates of mild cognitive impairment and dementia in the U.S. Medicare population: observational analysis

The analyses include the 100% sample for beneficiaries aged 65 and older enrolled in Medicare FFS or a Medicare Advantage Plan, who were nearly continuously enrolled either for at least 3 years or until death. Following the coverage definition used by the Chronic Conditions Data Warehouse (CCW), our definition of nearly continuous enrollment requires an average of 11 months of both parts A and B or part C coverage each year (at least 33 out of a possible 36 months) or, if the beneficiary died during the third year of the surveillance period, with fully continuous parts A and B or part C coverage and no interruption until the month of death. Across all 3-year windows from 2015 to 2019, this restriction excluded 13.2% (n = 6,890,000) of beneficiaries aged 65 and older during each of the 3 years.

Advantage Plan enrollment was defined as having at least 2 months of enrollment [15], as switching outside of the open enrollment period is uncommon [16, 17]. We analyzed claims and encounter data for inpatient and outpatient facilities, carriers, and skilled nursing facilities and the corresponding enrollment data for 2015 to 2019. Data were accessed through the Centers for Medicare & Medicaid Services Virtual Research Data Center, and the study protocol and data protection procedures were approved by our Institutional Review Board (UP-21-00441) under expedited review and with a waiver for informed consent and HIPAA (Health Insurance Portability and Accountability Act) authorization. Data were processed and analyzed with SAS 7.15 (SAS Institute Inc.), and all statistical analyses were conducted with Stata 16 (StataCorp LLC).

To identify persons diagnosed with dementia, we modified the algorithm published by the CCW [18] based on a recent study by Festa et al., [19] which found that noncognitive and unspecific codes in the CCW algorithm, as well as diagnoses made during a single inpatient or skilled nursing facility episode, were likely to be false positives. The list of included and removed codes is shown in Additional file 1: Table S1 and Table S2. We required 2 claims with the remaining codes on separate days in either setting over rolling windows of 3 years. We also used the original CCW algorithm for comparison. No definitions for MCI are currently published by the CCW, so we identified the diagnosis based on ICD-10-CM code G31.84 (i.e., mild cognitive impairment of uncertain or unknown etiology) and the ICD-9-CM code 331.83 (i.e., mild cognitive impairment), also requiring 2 claims on separate days. The rates were calculated for each 3-year rolling window for which we had complete data—that is, 2015 to 2017, 2016 to 2018, and 2017 to 2019.

As some persons may meet the criteria for both MCI and dementia, we applied the following assignment rules to avoid double counting: If a person was uniquely assigned to either MCI or dementia during the midpoint year, we used that assignment. If not, we based the assignment on the latest claim with a diagnosis of MCI or dementia in the midpoint year. If neither diagnosis was documented during the midpoint year, we based the assignment on the claim closest to the midpoint year or, for ties, in the earlier year. Among this group, accounting for about 10% of those who had either diagnosis, around 40% were assigned to MCI and 60% to dementia.

To estimate the underlying prevalence of MCI and dementia based on cognitive assessments, we used HRS data from 2000 to 2016 [11]. We applied the Langa-Kabeto-Weir Classification of Cognitive Function [20] to categorize participants as cognitively normal, as having cognitive impairment but no dementia ([CIND] with CIND representing MCI), or as having dementia. These classifications were based on cognitive assessments of self-respondents and proxy interviews, usually with a spouse or other family member, if a respondent was unable or unwilling to do an interview.

We used probit models to separately predict CIND (vs cognitively normal) and dementia (vs cognitively normal) using variables that are present both in the HRS and Medicare data to allow using the model to generate expected rates in the Medicare population. Predictors included sex, age groups (50–54, 55–59, 60–64, 65–69, 70–74, 75–79, 80–84, and ≥ 85 years), race and ethnicity (non-Hispanic [NH] White, Black, Hispanic, and others), dual eligibility status (individuals covered by both Medicare and Medicaid), and a linear trend for year to account for the secular decline in dementia incidence [21]. Our analytic sample consisted of HRS participants aged 65 and older.

Model calibration was conducted using the 2000 to 2014 data. To validate the calibration, we applied the derived regression weights to the 2016 HRS participants, calculated their probabilities of having CIND or dementia, and used receiver operating characteristic curves to quantify the prediction accuracy of the model. The derivation of national estimates of the rate of Medicare beneficiaries with CIND and dementia based on these predicted probabilities is detailed in Additional file 1: Methods. Sampling weights were used in the analysis.

As the person-level predictors in the HRS-based models are also available in Medicare enrollment data, the estimated regression weights can be applied to Medicare data to generate expected diagnosis rates for MCI and dementia, accounting for potentially different demographic compositions and a secular trend if additional years of data are used. The ratio between the observed rates based on diagnoses documented in the claims data and the expected rates predicted by the model provides a measure for potential gaps in diagnoses. Such observed to expected (O/E) ratios are frequently used in quality measurement [22]. The O/E ratio can be interpreted as the proportion of expected cases that were diagnosed or the detection rate.

This research was funded by a contract from Genentech, a member of the Roche Group, to the University of Southern California. The sponsor provided comments on a draft of the paper, but the authors had full control of the design, analysis, final manuscript, and decision to submit.

Table 1 shows the observed rates of MCI and dementia diagnosis in the Medicare population. Rates for MCI increased by about 10% for each observation window or 0.3 percentage points from 0.012 to 0.015, whereas dementia rates declined slightly from 0.098 to 0.095. From 2017 to 2019, rates for both MCI and dementia increased with age (Table 2). Dementia (0.108 vs 0.079), but not MCI (0.016 vs 0.015), was more frequently diagnosed in women than in men (Table 3). Dementia, but not MCI, was also more frequently diagnosed in Black and Hispanic individuals than in NH White individuals (Table 4). Dually eligible beneficiaries had substantially higher diagnosis rates for dementia compared with their Medicare-only counterparts (0.213 vs 0.079), but similar rates for MCI (0.017 vs 0.015) and diagnosis rates for both stages of cognitive impairment were similar for individuals enrolled in Medicare FFS and Medicare Advantage Plans (Table 5).

The estimated regression weights derived from the 2000 to 2014 HRS data are listed in Additional file 1: Table S3. Applying these estimates to 2016 HRS data showed that the predicted rate of MCI in 2016 was 0.176 compared with an observed rate of 0.175; for dementia, the predicted and observed rates were 0.082 and 0.078, respectively. The area under the receiver operating characteristic curve was 0.670 for separating MCI from being cognitively normal and 0.772 for separating dementia from being cognitively normal; accuracy rates show that 68.6% and 79.2% of the sample were correctly classified for MCI and dementia, respectively (Additional file 1: Table S4 and Fig. 1).

The expected prevalence shows a slight decline for both MCI (0.199 to 0.196) and dementia (0.090 to 0.086) from 2015 to 2019 (Table 1). The detection rates for both stages increased during the same period, albeit faster for MCI than dementia with average increases of 20% (1.7 percentage points) and 0.85% (1.8 percentage points), respectively. Put differently, of the estimated 8 million Medicare beneficiaries with MCI, 7.4 million were undiagnosed, while diagnosis rates for dementia were about 10% too high. Table 2 documents that detection rates increase with age, except for the oldest cohort, and Table 3 shows that these ratios are higher for women than for men. Detection rates for MCI in NH White Medicare beneficiaries were 2- to 3-fold those of Black and Hispanic beneficiaries and those with other or unknown race or ethnicity, and detection rates for dementia in NH White Medicare beneficiaries were 2.5-fold those of Black and Hispanic beneficiaries (Table 4). In other words, only around two-thirds to three-quarters of the expected non-White individuals with dementia were diagnosed, whereas diagnosis rates for NH White individuals were higher than predicted. Dually eligible individuals had lower estimated detection rates than their Medicare-only counterparts for MCI (detection rate 0.056 vs 0.085) and dementia (detection rate 0.855 vs 1.243). Beneficiaries in Medicare FFS were more likely to have their cognitive impairment detected than those in Medicare Advantage Plans, with a detection rate of 0.084 vs 0.073 for MCI and 1.167 vs 1.034 for dementia (Table 5).

These results should be interpreted within the context of the study limitations. We acknowledge that a predictive model based on demographic information alone has only adequate accuracy and it could be improved by incorporating clinical characteristics of the individuals. Also, we estimated the expected prevalence of MCI based on cognitive test scores, which is not the same as a true clinical diagnosis. However, our predicted number of 8.06 million cases is close to the 7.95 million predicted based on a widely recognized meta-analysis by Petersen et al., [37] suggesting our estimated detection rate should be reasonably accurate. In contrast, it is possible that our model underpredicts dementia prevalence, because our expected rate of 0.086 is slightly lower than the rates reported for the US population aged 65 and older by Rajan et al. [38] (0.113) and Manly et al. [39] (0.100), which are closer to the observed rate of 0.096. Our classification algorithm for the cognitive state using the HRS data may overclassify dementia in minority populations [33], which could explain part of their lower detection rates. Similarly, the classification algorithm based on claims data may have misclassified some individuals [40]. We did not account for switching between FFS and Medicare Advantage within an observation window, as switching rates tend to be low, between 1 and 6% [16, 17]. Diagnoses may have been communicated but not documented in claims data because of concerns for stigma and loss of driver’s license, but lack of documentation still represents a problem, as it may limit the clinician’s ability to prescribe symptomatic medications and connect patients with support services. We would also expect the reluctance to document to be greater for dementia than for MCI, which is the opposite of what the data show.