Identification of cognitive impairment using the Lancet Commission's risk factors and Medicare administrative data

Abstract

INTRODUCTION

Prediction models based on administrative data may present a scalable opportunity to identify risk of cognitive impairment, but their accuracy relative to models using richer information is uncertain.

METHODS

We developed and validated models to identify the likelihood of mild cognitive impairment (MCI) and dementia using the Health and Retirement Study linked to Medicare data from 2000 to 2016 (N = 63,740). Predictors covered most risk factors identified by the 2024 Lancet Commission. Model performance was assessed using multiple metrics, including the area under the receiver operating characteristic curve (AUC).

RESULTS

Probit models with demographics and chronic conditions yielded high AUCs of 71.3% (MCI) and 82.1% (dementia). Adding individual level education provided the largest improvement in AUCs, whereas dual eligibility status offered smaller gains (p  <  0.001). Air pollution exposure, obesity, and interaction terms did not enhance prediction.

DISCUSSION

Predictors in administrative data can be used to generate reasonably accurate, well calibrated models predicting likelihood of cognitive impairment.

Highlights

• Prediction models using only administrative data yield reasonably high accuracy.

• Adding individual level education provides the largest improvement in accuracy.

• Census tract‐level education and dual eligibility status yields small improvements.

1. BACKGROUND

Cognitive impairment is a major health concern among older adults because of aging populations and the global rise in established risk factors such as hypertension, ¹ diabetes, ² , ³ and poor lifestyle behaviors. ⁴ Timely diagnosis allows individuals and families to plan for future dementia care ⁵ and allows for better management of modifiable risk factors, ⁶ and the consideration of emerging disease modifying therapies that slow down disease progression. ⁷ However, proactive cognitive assessment in the United States is not commonly implemented ⁸ and the disease is frequently undiagnosed or identified late, even in high‐income countries with robust healthcare systems. ⁹ , ¹⁰

Risk‐based detection approaches that identify individuals at higher risk of existing cognitive impairment hold promise as an alternative. Numerous risk factors for mild cognitive impairment (MCI) and dementia have been identified, and the 2024 Lancet Commission report has recognized the following 14 factors: hypercholesterinemia, vision and hearing loss, hypertension, diabetes, alcohol abuse, obesity, smoking, depression, traumatic brain injury, low education, physical inactivity, social isolation, and air pollution. ¹¹ Integrating these variables into a model can help prioritize at risk individuals, who may benefit from early assessment and intervention, potentially even before they experience subjective memory complaints.

Several such models have been developed that use data from electronic health records (EHRs), registries, and cohort studies, often applying Natural Language Processing (NLP) and Machine Learning (ML) methods. For example, Barnes et al developed a LASSO regression using EHR data to estimate the probability of unrecognized dementia, achieving a 81% receiver operating characteristic (ROC) area under the curve (AUC). ¹² Akter et al. achieved an AUC of 80.9%–83.3% in predicting AD by using EHR data to test six ML classifier models. ¹³ Li et al. used coded and NLP derived EHR data to estimate the risk of MCI (AUC = 77%), ¹⁴ while another ML study using the National Alzheimer's Coordinating Center data achieved an AUC of over 75% for prevalent MCI. ¹⁵

RESEARCH IN CONTEXT

1. Systematic review: We reviewed existing prediction models for cognitive impairment focusing on tools developed with electronic health records, registries, cohort studies, and administrative data. Prior work demonstrates that complex machine‐learning models can achieve high discrimination but often rely on data sources unavailable or impractical in many health systems. We further assessed the predictive value of a simple model using administrative data, including risk factors identified by the 2024 Lancet Commission.

2. Interpretation: Our findings show that widely available administrative data can generate reasonably accurate and well‐calibrated models for predicting mild cognitive impairment and dementia. Scalable detection systems may not need to rely on complex or unstructured clinical data.

3. Future directions: Future work should incorporate additional risk factors (e.g., social isolation, physical inactivity) and compare performance against models using clinical diagnoses or biomarkers.

While scientifically interesting, the complexity behind those models may limit their real‐world scalability, especially in lower‐resource settings. Extracting information from unstructured physicians notes with NLP requires tools, skill and experience, and EHR systems vary substantially across healthcare settings, limiting portability. ¹⁶ , ¹⁷ Although patient registries offer rich clinical detail, they are expensive to maintain and often have idiosyncratic variable definitions. ¹⁸ Moreover, compared with ML‐based models, the relationship between independent and dependent variables in standard statistical models are explicitly specified. The automated selection of predictors may also harm face validity such as conditions not known to be risk factors showing to be strong predictors. ¹⁴

For those reasons, the current study explores how well a simple model, based on widely available data and guided by established dementia research, could estimate an individual's probability of cognitive impairment. Specifically, the model uses established risk factors from administrative data that include standardized diagnostic and procedure codes and are ubiquitously collected for billing and service documentation ¹⁹ –in this case, data from Medicare, the U.S. federal health insurance program for people aged 65 years and older–and applies probit regression, a type of generalized linear modeling approach, which is transparent and interpretable. Variable selection is based on the list of 14 risk factors for dementia that were identified by the 2024 Lancet Commission report. ¹¹ Our research questions include: which of these risk factors can be implemented with administrative data, and how accurate are predictions based on different levels of details (e.g., demographic vs. clinical information) compared to one another.

2. METHODS

2.1. Data sources

Data came from the Health and Retirement Study (HRS), ²⁰ linked to Medicare enrollment, claims and encounter data. The HRS is a biannual, nationally representative and longitudinal survey of U.S. adults aged 50 years and older that is conducted by the University of Michigan. It collects data on demographics, employment and wealth, health conditions and healthcare use, and has a cognitive assessment module. For the current work, we restricted the analysis to HRS participants aged 65 years or older without missing data, to align with the age eligibility for Medicare. Additionally, we linked the American Community Survey ²¹ and the Global Burden of Disease data ²² to construct contextual variables based on respondents’ residential locations.

2.2. Outcome variables

Cognitive impairment statuses were derived from the cognitive assessment data using the widely used Langa–Kabeto–Weir algorithm, ²³ , ²⁴ which has a published accuracy of 83%. ²⁵ The algorithm computes a summary score based on cognitive tests administered in each HRS wave. Respondents were then categorized as being cognitively normal if the score ranged from 12 to 27, having cognitive impairment but no dementia (CIND) if it ranged from 7 to 11, and having dementia if it ranged from 0 to 6. For respondents who did not participate in a specific wave, a summary score based on proxy reports was used. ²⁴ Prior studies show that CIND corresponds closely to clinically diagnosed MCI; ²³ for simplicity, we refer to MCI only hereinafter.

2.3. Predictor variables

We used diagnosis codes in the HRS‐linked Medicare data to operationalize all 10 chronic conditions identified as risk factors by the 2024 Lancet Commission based on validated algorithms published by the Chronic Condition Data Warehouse: ²⁶ chronic kidney disease, chronic obstructive pulmonary disease, depression, diabetes, heart failure and ischemic heart disease, hyperlipidemia, hypertension, obesity, peripheral vascular disease, tobacco use disorders, alcohol use disorders, deafness and hearing impairment, blindness and visual impairment, traumatic brain injury, and stroke. Because obesity is known to be under‐documented in administrative data, ²⁷ we tested an alternative measure using individually corrected self‐reported weight data from the HRS, with details noted in the Methods section of the Appendix. Appendix Table A1 shows the crosswalk between these conditions and the Lancet Commission's risk factors.

Different measures of low education and air pollution were included as predictors. Although individual‐level educational attainment would be most accurate, such information is typically not documented in administrative data. We therefore approximated it with two alternative ways and separately tested predictors of individual‐level education, education composition in the census tract of residency, and dual eligibility status. Air pollution was based on population‐weighted PM 2.5 concentrations at the ZIP code‐level. Additional details are available in the Methods section of the Appendix.

In sum, we were able to operationalize all but two (social isolation and physical inactivity, which are not reliably captured in claims data or additional datasets) of the 14 risk factors identified by the 2024 Lancet Commission using administrative data and location‐linked contextual data. Additionally, we included demographic information available in both survey and Medicare data: sex (male/female), age groups (65–69, 70–74, 75–79, 80–84, and ≥ 85 years), and race and ethnicity (non‐Hispanic [NH] White, NH Black, Hispanic, NH Asian and Pacific Islander, and NH Other).

2.4. Analytical data

Our analytic data comprised of two subsets of the HRS’ biannual waves: waves from 2000 to 2014 to develop predictive models, and the 2016 wave for validation. We disregarded the most recent HRS waves (2018 and 2020), as prior work indicated that their partial transition from telephonic to web‐based cognitive assessments may have overestimated scores. ²⁸ Data from the calibration waves were stacked, to incorporate a secular trend in the model development as the incidence of cognitive impairment changed across years, ²⁹ and an earlier version of our predictive model was robust against the stacking when compared with a bootstrapped nonoverlapping cohort. ⁹ Therefore, the calibration data comprised of 55,501 observations, and the validation data included 8239 observations.

2.5. Statistical analysis

Our prediction of MCI and dementia was built upon a previously published predictive model that relied solely on demographics and dual eligibility status. ⁹ , ¹⁰ The method used two probit models to separately predict the likelihood of having MCI (vs cognitively normal) and dementia (vs cognitively normal).

For calibration, predictors were added, one block at a time in the following order, to evaluate their unique contribution to model performance: age only (Model [M]1); sex (M2); chronic conditions (including health‐related behaviors) (M3); race and ethnicity (M4); education operationalized in three ways, namely, individual‐level education (M5), census tract‐level education (M6), or dual eligibility status (M7); air pollution (M8), and an alternative measure of obesity using self‐reported data (M9). We included a linear year effect to account for the declining age‐adjusted incidence of cognitive impairment in high‐income countries, ²⁹ and clustered the standard error at the individual level to account for the longitudinal nature of the calibration sample. Additionally, we tested models using a quadratic trend (not presented) that resulted in similar results.

To assess the model performance, we applied the estimated regression coefficients to the validation data and considered several criteria. First, we computed the AUC and tested whether adding predictors increased the AUC significantly (e.g., M2 vs. M1) based on a chi‐square test. ³⁰ Secondly, we computed the Brier score, ³¹ equivalent to the mean squared error of the probabilistic forecasts ³² and ranging between 0 and 1 with a lower value implying better performance (i.e., smaller error). A Brier score can be normalized as the Spiegelhalter's z‐statistic ³³ for further inferences on model performance: when the z‐statistic is not significantly different from 0, the calibration is considered sufficient. Additionally, we compared the predicted rates of MCI and dementia with the observed rates, applying sampling weights derived by the HRS to both. We then examined calibration plots to assess the agreement between prediction and observation over the full range of risk groups and evaluated our work following the PROBAST checklist. ³⁴

In addition, we investigated whether two sets of interaction terms with chronic conditions would improve model accuracy. The first interacted all chronic condition variables with age group variables (M10) to explore whether the association of chronic illness with cognitive impairment is moderated by age. The second interacted chronic conditions with race and ethnicity (M11) to incorporate the potential difference in disease severity by different race/ethnic groups.

The study was approved by the Institutional Review Board of the University of Southern California (UPS 14‐00148) under expedited review and with a waiver for informed consent and Health Insurance Portability and Accountability Act (HIPAA) authorization. Claims and other data were linked and accessed through the NIA Data LINKAGE Program. ³⁵ Data were processed and analyzed using SAS 7.15 (SAS Institute Inc) and Stata 16 (StataCorp LLC).

3. RESULTS

3.1. Descriptive statistics

The calibration sample included 55,501 observations from 11,982 unique individuals and the validation sample 8,239 observations or individuals. Table 1 presents characteristics of both unweighted and population‐weighted samples. For both the calibration and validation weighted sample, more than 80% identified as NH White, more than half reported having a high school degree or less (60.1% and 50.8% for the calibration and validation sample, respectively), and around 9%–10% were dually eligible for Medicare and Medicaid. The most common chronic conditions were hypertension (69.1% and 74.4%, respectively) and hyperlipidemia (56.9% and 69.7%, respectively), whereas the least common were traumatic brain injury (0.3% for both samples) and vision loss (0.7% and 0.4%, respectively). Table 2 reports the distribution of cognitive status. For the calibration observations, 68.9% were classified as having normal cognition, 20.6% as having MCI, and 10.5% as having dementia, with similar rates for the validation observations.

TABLE 1.

Characteristics of observations in calibration and validation data.

	Unweighted sample			Weighted sample
Variables	Calibration (N = 55,501)	Validation (N = 8239)	p‐value	Calibration	Validation
Age(%)			< 0.001
65–69	23.2	19.9		24.4	26.3
70–74	24.7	19.2		23.6	25.5
75–79	21.0	25.5		20.6	19.0
80–84	15.8	18.7		16.2	14.0
85 and older	15.4	16.7		15.2	15.2
Sex(%)			0.031
Female	40.7	39.4		40.7	42.6
Male	59.3	60.6		59.3	57.4
Race and ethnicity(%)			< 0.001
NH White	79.9	71.8		85.4	80.8
NH Black	12.2	14.7		7.7	8.4
Hispanic	6.3	11.0		5.0	8.2
NH API	0.6	0.9		0.7	0.9
NH Other	1.1	1.6		1.2	1.8
Individual‐level education(%)			< 0.001
Below high school	27.7	22.0		25.1	17.9
High school degree	34.7	33.8		35.0	32.9
Some college	18.5	21.6		19.2	22.5
College and above	19.1	22.6		20.8	26.7
Dually eligible for Medicare and Medicaid(%)			0.651
No	12.5	10.6		9.6	9.1
Yes	89.3	89.4		90.4	90.9
Chronic conditions(%)
Chronic kidney disease	11.3	23.8	< 0.001	11.1	21.5
Chronic obstructive pulmonary disease	16.0	17.4	0.001	15.8	16.1
Depression	10.1	17.6	< 0.001	10.5	17.5
Diabetes	26.4	32.7	< 0.001	25.6	30.2
Poorly controlled diabetes	26.2	28.6	< 0.001	25.4	26.0
Heart failure	13.5	13.8	0.454	13.1	12.5
Hyperlipidemia	56.5	71.3	< 0.001	56.9	69.7
Hypertension	69.9	78.1	< 0.001	69.1	74.4
Poorly controlled hypertension	10.6	18.9	< 0.001	10.1	16.9
Ischemic heart disease	29.0	26.5	< 0.001	28.4	25.3
Peripheral vascular disease	12.0	18.3	< 0.001	11.5	16.2
Tobacco use disorders	3.7	7.6	< 0.001	3.9	7.8
Alcohol use disorders	0.9	1.7	< 0.001	1.0	1.7
Deafness and hearing impairment	3.4	6.5	< 0.001	3.6	6.2
Vision loss	0.8	0.6	0.227	0.7	0.4
Traumatic brain injury	0.3	0.3	0.857	0.3	0.3
Stroke	5.2	7.4	< 0.001	4.8	6.6
Obesity	5.1	18.6	< 0.001	5.2	18.1

Abbreviations: API, Asian and Pacific Islander; NH, non‐Hispanic.

TABLE 2.

Distribution of cognitive status for observations in calibration and validation data.

	Calibration observations (n = 55,501)		Validation observations (n = 8239)
Cognitive status	Count	Percent (%)	Count	Percent (%)
Cognitive normal	52,790	68.9	5834	67.7
MCI	15,783	20.6	1861	21.6
Dementia	8045	10.5	922	10.7

Abbreviation: MCI, mild cognitive impairment.

3.2. Model performance using claims‐only predictors

Table 3 showcases the performance of the prediction of the first four models that only includes predictors available in claims data. Using age alone (M1) achieved an AUC of 62.4% when predicting MCI and 72.7% for dementia. Adding sex (M2) yielded an improvement in AUC, reaching 62.7% for MCI and 72.9% for dementia. Further adding chronic conditions (M3) significantly improved AUC (p < 0.001 compared to M2) for both MCI (AUC = 66.8%) and dementia (AUC = 77.4%). The Brier score also showed a small decline, from 0.179 to 0.174 for MCI and from 0.109 to 0.102 for dementia, but the Spiegelhalter's z‐test suggested that M3 was not sufficiently well calibrated (p < 0.001 for MCI and p < 0.05 for dementia). Subsequently, including race and ethnicity (M4) significantly improved performance (p < 0.001 compared to M3) reaching an AUC of 71.3% for MCI and 82.1% for dementia; the Brier scores continued to drop to 0.166 for MCI and 0.095 for dementia, and both models achieved satisfactory calibration (p > 0.05 based on Spiegelhalter's z‐test).

TABLE 3.

Performance of the predictive models using demographics and chronic conditions, evaluated based on the validation sample.

	Predictors								Prevalence of impairment status
Parameter	Age	Sex	Chronic conditions	Race / Ethnicity	AUC (95% CI)	χ 2‐statistic a	Brier score b	Spiegelhalter's z‐statistic c	Predicted rate	Difference from observed rate d
Predicting MCI vs. cognitive normal (n = 7326)
M1	X				0.624		0.179	4.954 ***	0.190	−0.002
					(0.609, 0.639)
M2	X	X			0.627	3.41	0.179	4.992 ***	0.190	−0.002
					(0.612, 0.641)
M3	X	X	X		0.668	55.24 ***	0.174	3.602 ***	0.191	−0.001
					(0.654, 0.682)
M4	X	X	X	X	0.713	60.66 ***	0.166	1.358	0.187	−0.005
					(0.700, 0.726)
Predicting dementia vs. cognitive normal (n = 6421)
M1	X				0.727		0.109	3.252 ***	0.096	0.007
					(0.709, 0.745)
M2	X	X			0.729	1.61	0.109	3.180 ***	0.096	0.007
					(0.710, 0.748)
M3	X	X	X		0.774	42.77 ***	0.102	2.146 *	0.103	0.014
					(0.757, 0.791)
M4	X	X	X	X	0.821	78.62 ***	0.095	−0.990	0.101	0.011
					(0.806, 0.836)

Note: X implies the corresponding predictor was included in the model.

Abbreviations: AUC, area under the curve; CI, confidence interval; M, model; MCI, mild cognitive impairment.

^a The chi‐squared test infers whether the corresponding and prior model (e.g., M2 vs M1) have similar AUCs with a significant p‐value suggesting better performance.

^b The Brier score is ranged between 0 and 1 with a smaller value indicating smaller prediction errors.

^c The Spiegelhalter's z‐statistic supplements the Brier score with a non‐significant result suggesting sufficient calibration.

^d The observed rates of MCI and dementia were 0.192 and 0.089, respectively.

^* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

Figure 1 depicts the ROC curves and illustrates the gains in AUC as additional variables were introduced from M1 to M4, and Appendix Figure A1 further presents calibration plots using M4. Both figures show that M4–by using demographic and chronic condition information alone—could achieve good calibration. The predictive equations for an individual's risk of having cognitive impairment can be constructed using the regression coefficients in Appendix Table A2. Additional metrics based on risk thresholds are presented in the Appendix (Table A3), and a decision curve analysis is presented in Appendix Figure A2.

FIGURE 1.

Receiver operating characteristics curve predicting cognitive impairment using demographics and chronic conditions, evaluated based on the validation data. MCI, mild cognitive impairment; M, model; AUC, area under the receiver operating characteristic curve.

3.3. Model performance adding contextual predictors

We further tested models that augmented the model with claims‐only predictors (M4) by linking data from various sources. Table 4 presents predictive performance when incorporating different indicators of education. Including individual‐level education (M5) significantly improved AUC (p < 0.001), to 75.7% from 71.3% with M4 for MCI and to 85.6% from 82.1% for dementia. Replacing individual‐level education with census tract–level education (M6) or dual eligibility status (M7) also statistically significantly increased the AUC compared to M4, albeit smaller absolute increases: including census tract–level education improved AUC to 72.2% for MCI (p < 0.001) and 82.5% for dementia (p < 0.01), while including dual eligibility status achieved an AUC of 71.9% for MCI (p < 0.001) and 82.6% for dementia (p < 0.05). The Brier score decreased with the inclusion of individual‐level education (to 0.157 with M5 from 0.166 with M4 for MCI, and to 0.088 from 0.095 for dementia), but less so when the census tract education (0.164 for MCI and 0.095 for dementia, M6) or dual eligibility status (0.165 for MCI and 0.093 for dementia, M7) was used. The Spiegelhalter's z‐statistic was not significant because M4 already achieved satisfactory calibration.

TABLE 4.

Performance of the predictive models using demographics, chronic conditions, and education, evaluated based on the validation sample

	Predictors								Prevalence of impairment status
Parameter	Demographics + chronic conditions	Individual‐level education	Tract‐level education	Dual eligibility status	AUC (95% CI)	χ 2‐statistic a	Brier score b	Spiegelhalter's z‐statistic c	Predicted rate	Difference from observed rate d
Predicting MCI vs. cognitive normal (n = 7316)
M4	X				0.713		0.166	1.278	0.187	−0.005
					(0.699, 0.726)
M5	X	X			0.757	121.22 ***	0.157	0.697	0.186	−0.006
					(0.745, 0.770)
M6	X		X		0.722	26.63 ***	0.164	1.639	0.184	−0.007
					(0.709, 0.736)
M7	X			X	0.719	14.94 ***	0.165	1.456	0.186	−0.005
					(0.706, 0.733)
Predicting dementia vs. cognitive normal (n = 6414)
M4	X				0.821		0.095	−0.932	0.100	0.011
					(0.806, 0.836)
M5	X	X			0.856	79.35 ***	0.088	−0.989	0.101	0.012
					(0.842, 0.869)
M6	X		X		0.825	7.99 **	0.095	−0.445	0.100	0.010
					(0.811, 0.840)
M7	X			X	0.826	3.99 *	0.093	0.516	0.100	0.011
					(0.812, 0.841)

Note: X implies the corresponding predictor was included in the model.

Abbreviations: AUC, area under the curve; CI, confidence interval; M, model; MCI, mild cognitive impairment.

^a The chi‐squared test infers whether the corresponding model and model 4 have similar AUCs with a significant p‐value suggesting better performance.

^b The Brier score is ranged between 0 and 1 with a smaller value indicating smaller prediction errors.

^c The Spiegelhalter's z‐statistic supplements the Brier score with a non‐significant result suggesting sufficient calibration.

^d The observed rates of MCI and dementia were 0.192 and 0.089, respectively.

^* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

Additional predictors did not lead to a statistically significant improvement in model prediction. Including air pollution (M8; Appendix Table A4) resulted in similar AUC and Brier scores (AUC = 71.4% for MCI and 82.1% for dementia; Brier score = 0.166 for MCI and 0.095 for dementia), compared with the model without it (M4). Using self‐reported obesity (M9; Appendix Table A5) did not improve prediction performance, despite 42.9% of respondents being classified as obese based on self‐reported obesity (versus 6.2% based on claims). Interacting age groups with chronic conditions (M10; Appendix Table A6) did not enhance predictive performance (AUC = 71.0% for MCI and 81.9% for dementia), despite a statistically significant joint F‐test (p < 0.001) suggesting that some interaction terms might be significant. Neither did the interaction between race and ethnic groups and chronic conditions (M11; Appendix Table A6; AUC = 71.2% for MCI and 82.0% for dementia).

3.4. Predicted versus observed prevalence rates of impairment status

For all models, we compared the estimated prevalence rates of MCI and dementia derived from the predictive models to the observed rates derived from the cognitive scores. The observed prevalence rates of MCI and dementia in the validation sample were around 19.2% and 8.9%, respectively, and were not substantially different from the predicted rates, which ranged from 18.4% to 19.1% for MCI and 9.6% to 10.4% for dementia across M1–M7.

4. DISCUSSION

In this study, we tested the ability of a model containing most risk factors identified by the Lancet Commission to estimate the presence of MCI and dementia, using variables that can be exclusively derived from administrative data sources. Our model with demographic variables and chronic condition diagnoses (M4) achieved an AUC of 71.3% for MCI and 82.1% for dementia with good calibration and close alignment of the predicted and observed rates. In addition to AUCs, we employed a variety of metrics to assess different aspects of the model performance, as prediction at the individual level could still be unreliable even with good overall discrimination. ³⁶ This includes the Brier score and associated Spiegelhalter's z‐statistics to evaluate the mean squared predictive error, comparison between predicted and observed impairment rates to evaluate the mean prediction, and the calibration plot to depict individual‐level prediction. Most metrics suggest a better prediction for dementia compared to MCI, possibly explained by the stronger correlation between predictors and cognitive decline as impairment proceeds.

Our findings demonstrate that a detection model using only administrative data such as Medicare claims and encounter data can achieve accuracy that is comparable to models using richer datasets. For example, the eRADAR model, designed to detect unrecognized dementia, achieved an AUC of 80.9%. ¹² Beyond demographic variables and chronic conditions, the model incorporated data on blood pressure, healthcare utilization, and electronic pharmacy dispensing records on medications for dementia symptoms. While eRADAR's outcome of interest differs from ours, the reasonably high AUCs of both models imply variables derived from administrative data sources can also be used to construct screening models with high accuracy. Additionally, healthcare utilization was a key predictor in eRADAR and is also available in claims data, suggesting future studies using administrative data could test the contribution of such metrics.

Another predictive model employing Cox regression to predict time to dementia onset in adults aged 65 and older included predictors such as age, sex, education levels, chronic conditions, and functional limitations. ³⁷ Tested across four cohort studies, including the HRS, the model's Harrell's C statistic–a metric commonly used in survival analysis but comparable to the AUC–ranged from 68% to 78%, with 76% specifically based on the HRS cohorts, similar to our model that included age, sex, and chronic conditions (M3) but not race and ethnicity.

Poorer performance when the prediction only relies on age and chronic conditions from claims data further underlines the potential value of incorporating demographic and socioeconomic factors. A model to detect MCI with a machine learning algorithm using information on 25 chronic conditions in the Merative MarketScan Commercial and Medicare databases reported AUCs of 75% (ages 50–64), 70% (ages 65–79), and 66% (ages 80 +). ³⁸ Our findings align closely with these results, in spite of using a simpler regression model. Using age, sex, and chronic conditions as predictors yielded an AUC of 67% for adults aged 65 and older. The fact that we achieved higher AUCs when including race and ethnicity (71.3%) and different measures of educational attainment (72.2%–75.7%) demonstrate that the addition of such demographic and socioeconomic data enhances predictive accuracy. Of note, the cited study suggested that effects of certain chronic conditions on MCI prediction were age‐dependent, ³⁸ but our analyses found no significant benefit from adding interaction terms between age and chronic conditions.

Another distinct contribution of our study is the integration of external data sources with traditional administrative data. Incorporating census tract‐level education data (AUC of 72.2% and 82.5% for MCI and dementia, respectively) and dual eligibility status (71.9% and 82.6%, respectively) showed a statistically significant, yet small, improvement in the model's AUC compared to the baseline claims‐only model (71.3% and 82.1%, respectively), whereas inclusion of individual‐level education data performed considerably better (75.7% and 85.6%, respectively). These results confirmed that information on educational attainment can contribute to prediction of cognitive impairment, but only if it is based on precise, person‐level data. Nevertheless, our findings suggest that tract‐level education or dual eligibility status could be a feasible alternative in settings where individual‐level education data are not available. While historic air pollution data linked to the patient's ZIP code of residence and additional measures of obesity did not improve predictive performance (as presented in Appendix Table A4 and Table A5, respectively), it is conceivable that longitudinal data for exposure to air pollutants and more precise location information might.

Our study comes with limitations. First, we were unable to operationalize social isolation and physical inactivity due to data constraints. Their effects may have been partially captured by included measures, such as depression and obesity, which are correlated with social isolation and physical inactivity, respectively. Secondly, our analysis did not include the most recent waves of the HRS due to methodological inconsistencies arising from the transition to web‐based cognitive testing. Updating our models with more contemporary data will be essential once appropriate methods are developed to address these inconsistencies. Third, our dependent variables rely on cognitive test scores rather than clinical diagnosis to determine MCI and dementia. We also used CIND as a proxy for MCI, which may imply potential misclassification; however, our estimated observed CIND rate of 19.2% (Table 3, table noted) was reasonably close to other studies’ estimate of 22%, ³⁹ suggesting such approximation is unlikely to induce systematic bias in our prediction. Future studies may use alternative measures or clinical diagnosis from claims data to construct and validate similar models. Fourth, our sample analysis was limited to individuals with linked Medicare administrative data. Of the 109,356 HRS respondents aged 65 and older from 2000 to 2020, only a small number were excluded due to missing demographic, education, or cognitive data. However, around 33% individuals did not agree to having their Medicare administrative data linked to their survey responses, which made it impossible to construct indicators for chronic conditions, and were excluded, yielding a final sample of 63,720 individual‐years. The loss of respondents may introduce selection bias if individuals without linked claims differ systematically from those with complete linkage; for instance, if those without linkage were healthier, excluding them may potentially inflate predictive performance by restricting the sample to those with clearer comorbidity patterns. While the direction of the bias is unknown, we acknowledge such possibility and suggest future studies to validate the model in other cohorts to assess differences in model performance. Lastly, tract‐level education, dual eligibility status, and ZIP code‐level air pollution exposure are approximate measures that cannot fully substitute for individual‐level education and pollution measures. Yet, our use and comparison of different proxy measures provided valuable insights into the feasibility and accuracy of incorporating risk factors beyond claims data.

In sum, this study demonstrates that MCI and dementia can be predicted with reasonably high accuracy and satisfactory calibration using robust evidentiary basis, administrative data, and well‐established statistical methods. The relative simplicity of the approach makes it highly scalable even in lower‐resource settings. Future research should focus on refining these models by exploring additional data sources and methods that will allow additional risk factors.

AUTHOR CONTRIBUTIONS

Andrew Becker (Data curation); Hankyung Jun (Data curation; Formal analysis; Writing—Original draft preparation); Ying Liu (Conceptualization; Methodology; Writing—Review & Editing); Soeren Mattke (Conceptualization; Writing—Review & Editing; Supervision; Project administration; Funding acquisition); Wei Ye (Data curation; Visualization; Formal analysis; Writing—Review & Editing); Hao Yin (Data curation).

CONFLICTS OF INTERESTS STATEMENT

Soeren Mattke serves on the board of directors of Senscio Systems, Inc., and the scientific advisory board of AiCure Technologies, Alzpath, and Boston Millennia Partners. He has received consulting or speaker fees from Biogen, C2N, Eisai, Eli Lilly, Novartis, Novo Nordisk and Roche/Genentech. All other authors have nothing to disclose. Author disclosures are available in the supporting information.

ETHICS STATEMENT

The study was approved by the Institutional Review Board (IRB) of the University of Southern California (UPS 14‐00148) under expedited review and with a waiver for informed consent and Health Insurance Portability and Accountability Act (HIPAA) authorization.

CONSENT STATEMENT

The study was approved by the Institutional Review Board of the University of Southern California (UPS 14‐00148) under expedited review and with waivers for informed consent and HIPAA authorization. All procedures were in accordance with the principles expressed in the Declaration of Helsinki.

Supporting information

Supporting information: dad270363‐sup‐0001‐Appendix.docx

Supporting information: dad270363‐sup‐0002‐ICMJE.pdf

DATA AVAILABILITY STATEMENT

The data supporting the findings of this study are not publicly available due to privacy or ethical restrictions. Data were accessed through the NIA Data LINKAGE Program.