Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2026 Jul 6;74(8):2326–2337. doi: 10.1111/jgs.70562

Assigning Probable Dementia Status Using Routinely Collected Electronic Health Record Data

Natalia Festa 1,✉, Kelsey Alexovitz 1, Natalia Sifnugel 2, Inessa Cohen 3,4, Isaac V Faustino 4, Siddarth Khasnavis 5,6, Juan Young 5,6, Mark Iscoe 3,4, Adam P Mecca 5, Ling Han 1, Ula Hwang 2,7
PMCID: PMC13496366  PMID: 42410282

ABSTRACT

Introduction

More than half of older adults with Alzheimer's Disease and Related Dementias (ADRD) are undiagnosed, limiting timely access to person‐centered care. Therefore, clinicians, researchers, and population health managers need scalable, reproducible approaches to monitor both prevalence and diagnostic gaps. We evaluated whether a decision‐analytic modeling framework can translate a limited number of clinician‐adjudicated cases of ADRD into a probabilistic computational phenotype for accurate, population‐level assignments of probable ADRD in the emergency department (ED) setting using routinely collected electronic health record (EHR) data.

Methods

Retrospective cohort study of 5000 adults aged ≥ 65 years from nine EDs within a large integrated health system (2014–2022). We randomly selected 500 individuals for clinician adjudication of dementia status (reference cohort), reserving the remaining 4500 as a phenotyping cohort. We developed the phenotype as a logistic regression model trained on adjudicated cases, embedding pattern‐mixture multiple imputation to address information bias. We applied decision‐curve analysis to evaluate clinical utility across probabilistic thresholds. We applied the phenotype to assign dementia status to 4500 unadjudicated patients and compared clinical characteristics to adjudicated cases.

Results

The mean (SD) age was 77.4 (9.0) years; 55.4% were women; 102 individuals (20.4%) had clinician‐adjudicated ADRD. The model demonstrated good discrimination (AUROC 0.87; 95% CI 0.82–0.91). Decision‐curve analysis revealed net clinical benefit across examined thresholds (predicted probabilities 12%–32%), identifying an additional 16–18 probable ADRD cases per 100 older adults. Among those without ADRD‐related diagnosis codes, net benefit ranged from 8 to 13 additional correct identifications per 100. Phenotype‐assigned cases closely resembled clinician‐adjudicated cases (standardized mean differences ≤ 0.20).

Conclusions

A probabilistic computational phenotype derived from routinely collected EHR data accurately reproduced clinician‐adjudicated ADRD status and demonstrated net clinical benefit, including among ED patients whose ADRD was not captured by diagnosis codes. Adoption of this replicable framework may enable healthcare organizations to strengthen ADRD surveillance and reduce underdiagnosis.

Keywords: dementia, electronic health record, phenotype, surveillance

Summary

  • Key Points
    • ○
      We developed a probabilistic computational phenotype using routinely collected electronic health record data that requires only modest clinical adjudication to enable scalable, population‐level surveillance of Alzheimer's Disease and Related Dementias (ADRD) risk.
    • ○
      The framework offers a transparent workflow to address information bias and classification uncertainty inherent to routinely collected healthcare data, supporting replication across distinct populations and healthcare settings.
    • ○
      This approach enables detection of probable ADRD in individuals without ADRD‐related diagnosis codes.
  • Why does this paper matter?
    • ○
      This research provides a transparent and replicable approach with which to address the critical challenge of ADRD underdiagnosis in real‐world clinical settings. Using routinely collected electronic health record data, the framework identifies individuals with probable ADRD, while accounting for variation in clinical documentation and enabling healthcare organizations to select classification thresholds aligned with their specific clinical and population health management objectives.

The Analytic Workflow for the Development, Validation, and Application of a Computational Phenotype for Probable Dementia. Each analytic step is mapped to the corresponding manuscript section, table, figure, or appendix. TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) is a reporting standard for prediction model studies; the applicable domains are indicated within each phase. *37 records excluded due to insufficient EHR documentation for cognitive adjudication, yielding 463 adjudicated cases. Abbreviations. ADRD: Alzheimer's Disease and Related Dementias; AUROC: Area Under the Receiver Operating Characteristic Curve; CITL: Calibration‐in‐the‐Large; EHR: Electronic Health Record; ED: Emergency Department; MAR: Missing at Random; NMAR: Not Missing at Random; NNS: Number Needed to Screen; SMD: Standardized Mean Difference.

graphic file with name JGS-74-2326-g004.webp

1. Introduction

Persons living with dementia (PLWD) have poorer health‐related quality of life, more frequent contact with the healthcare system, and are at increased likelihood of institutional long‐term care placement compared to age‐matched individuals with intact cognition [1, 2, 3]. Despite the substantial personal and societal burden of Alzheimer's Disease and Related Dementias (ADRD), more than half of affected individuals lack a formal diagnosis [4, 5, 6]. The costs of underdiagnosis are increasing as advances in disease‐modifying therapies and care management have the potential to influence the cognitive and functional trajectories of affected individuals [7, 8, 9]. Moreover, undiagnosed ADRD may itself confer distinct harms, including adverse hospitalization outcomes and continuation of potentially injurious activities, such as unsupervised medication management [10, 11].

Much of the current understanding of ADRD underdiagnosis has been informed by prospective longitudinal cohort studies; however, several factors impede the translation of this evidence to healthcare settings [12, 13]. First, cohort studies are subject to nonrandom attrition due to cognitive or functional impairments among participants, introducing selection bias that underrepresents individuals with more advanced disease and compromising external validity [14]. Second, cohort studies often exclude older adults without proxy respondents, representing an estimated 3% to 10% of patients in acute and long‐term care settings, further limiting generalizability [14, 15]. Third, cohort studies collect detailed demographic and functional measures that are relevant predictors of ADRD, such as educational attainment and gait speed, which are often absent from routinely collected healthcare data [16].

To address factors contributing to underdiagnosis, clinicians, researchers, and population health managers need scalable methods to infer probable ADRD status from routinely collected data. The electronic health record (EHR) is a promising resource with which to improve ADRD recognition, due to its representativeness of real‐world clinical populations and aggregation of measures from routine clinical encounters [17]. Nonetheless, the utility of EHR‐based approaches to ADRD recognition has been constrained by several important limitations. First, prevailing EHR‐based approaches to ADRD recognition heavily weight ADRD‐related diagnosis codes and medication prescriptions, favoring the identification of individuals with previously diagnosed disease [18, 19, 20]. Second, information bias arising from systematic data missingness in the EHR is frequently overlooked, despite its potential to affect the accurate recognition of ADRD for populations that are most at risk of underdiagnosis [20, 21, 22, 23]. Third, inconsistent or inadequate reporting of data‐handling procedures, including operational variable definitions, compromises reproducibility across healthcare organizations [17, 24]. Furthermore, the few algorithms validated to address ADRD underrecognition were developed in ambulatory care settings, which underrepresent individuals who rely upon episodic acute care and are predisposed to underdiagnosis [25, 26, 27].

The emergency department (ED) constitutes an opportune and critical setting for ADRD identification because it serves a population that is prone to underdiagnosis due to episodic or fragmented healthcare access [27, 28, 29, 30]. PLWD present to the ED at twice the rate of their cognitively intact counterparts, with utilization intensifying in the months preceding a formal diagnosis, rendering the encounter a sentinel event for individuals who might otherwise remain undiagnosed [27, 28, 29, 30]. Timely identification in this setting should enable interventions ranging from the prevention of iatrogenic complications, such as delirium superimposed on ADRD, to the facilitation of referrals for formal cognitive evaluations needed to access disease‐modifying therapies and care management interventions [31, 32, 33]. In recognition of these clinical and population health management imperatives, the American College of Emergency Physicians (ACEP) Geriatric Emergency Department Guidelines and the Geriatric Emergency Department Accreditation (GEDA) program recommend the implementation of valid, reliable, and feasible instruments to improve the identification of ADRD within the ED [34].

To address the methodological limitations of existing EHR‐based approaches to ADRD recognition with a feasible and scalable approach that is suited to the ED environment, we developed a probabilistic computational phenotype, defined as an automated approach to classifying probable ADRD status from routinely collected EHR data. A computational phenotype has the potential to address ADRD underdiagnosis as it can improve disease recognition among individuals lacking a formal diagnosis [17, 35]. In so doing, we present a validation framework that interrogates and addresses multiple potential sources of bias to ensure robust classification and to facilitate replication across distinct healthcare settings.

2. Methods

2.1. Design and Sampling Approach

We generated a random sample of 5000 unique index ED encounters for individuals aged ≥ 65 years, across 9 hospitals within the Yale New Haven Health System between 2014 and 2022 [18]. From this cohort, we abstracted demographic and clinical attributes, including cognitive status, as well as information regarding their health services utilization, up to and including their index ED encounter. The computational phenotype, therefore, characterizes ADRD status at the time of the ED encounter, based upon data generated both prior to and during the ED visit. We selected a random subsample of 500 unique person‐encounters for adjudication of their cognitive status by qualified clinicians (reference cohort). The remaining 4500 individuals comprised the phenotyping cohort, defined as patients without clinician adjudication whose probable ADRD status would subsequently be assigned using a probabilistic computational phenotype (Figure 1). We applied a decision‐analytic framework to extend established computational phenotyping methods by explicitly modeling uncertainty arising from multiple sources, including missingness in the EHR, diagnostic ambiguity, and parameter estimation [36, 37, 38].

FIGURE 1.

FIGURE 1

Open in a new tab

The analytic workflow for the development, validation, and application of a computational phenotype for probable dementia. Each analytic step is mapped to the corresponding manuscript section, table, figure, or supplement. TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) is a reporting standard for prediction model studies; the relevant domains are indicated within each phase banner. *37 records excluded due to insufficient EHR documentation for cognitive adjudication, yielding 463 adjudicated cases. Abbreviations: ADRD: Alzheimer's disease and related dementias; AUROC: Area under the receiver operating characteristic curve; CITL: Calibration‐in‐the‐large; EHR: Electronic health record; ED: Emergency department; MAR: missing at random; NMAR: not missing at random; NNS: Number needed to screen; SMD: standardized mean difference; TRIPOD: transparent reporting of a multivariable prediction model for individual prognosis or diagnosis.

2.2. Reference Standard Definition of ADRD

Five clinicians, including one senior clinician, were recruited to adjudicate the ADRD status of individuals within the reference cohort. The senior clinician (APM), a geriatric psychiatrist with expertise in the diagnosis and management of ADRD, oversaw the modification and adaptation of a previously published EHR‐based adjudication protocol, as well as the training of each of four clinician‐adjudicators, including two geriatric psychiatrists (S.K. and J.Y.) and geriatricians (N.F. and E.G.) [36]. During the training period, the clinician‐adjudicators independently reviewed cases until achieving a pre‐established threshold for interrater agreement (kappa ≥ 0.80). The clinician‐adjudicators then independently classified cognitive status for the 500‐person reference cohort, using available documentation up to the date and time of the index ED encounter. The clinician‐adjudicators classified cognitive status as consistent with normal cognition, mild cognitive impairment, or dementia. Cases of clinician‐adjudicated dementia were further classified as mild, moderate, or severe. The clinician‐adjudicators rated their confidence in each cognitive classification and deferred abstraction for records with insufficient documentation for chart review (n = 37), resulting in a total of 463 successfully adjudicated cases. The complete adjudication protocol, including diagnostic criteria for stage and severity classification, is presented in Supporting Information S1 (Tables S1–S5).

2.3. Step 1. Variable Selection and Definitions

2.3.1.

To support a replicable modeling framework, candidate variables were selected a priori based on previously established associations with ADRD or its severity, clinical relevance, and limited to structured EHR fields available across healthcare organizations [18, 39, 40, 41, 42]. We organized these variables into four domains—demographic characteristics, dementia‐specific clinical factors, general clinical factors, and healthcare utilization (Table S6)—and applied standardized, reproducible operational definitions for each (available in Yale Dataverse) [43]. Individuals with and without qualifying ADRD diagnoses were distinguished using the Bynum Standard, a validated algorithm requiring ≥ 2 qualifying dementia‐related diagnostic codes recorded ≥ 7 days apart within a 3‐year lookback period [44, 45]. To ensure concordance with the duration of chart abstraction, we extended the standard 3‐year lookback window to include longitudinal data during the adjudication period for individuals with established care histories. Comorbidities were captured using validated diagnostic‐code sets from the Centers for Medicare & Medicaid Services (CMS) Chronic Conditions Warehouse and the published literature [46, 47]. Pharmaceutical classes pertinent to both dementia‐specific and general clinical factors were mapped through code lists from the National Library of Medicine's Value Set Authority Center and the American Geriatrics Society Beers Criteria [48, 49]. We defined arrival source as the setting from which an individual presented to the index ED encounter. To support the application of this modeling framework across healthcare settings, we retained all candidate variables irrespective of their statistical association with ADRD in the current sample. The only exceptions were thyroid‐stimulating hormone, thiamine, and vitamin B12, which were excluded because missingness exceeded 30%, a level beyond which imputation was considered unreliable.

2.4. Step 2. Information Bias Assessment

As missingness within the EHR is often influenced by multiple factors, including healthcare access and the quality and completeness of clinical documentation, we inspected and summarized patterns of missingness for each retained variable. To test whether patterns of missingness conveyed information about unmeasured patient characteristics, we defined four distinct missing‐data scenarios using pattern‐mixture models, ranging from the conventional assumption that data were missing at random (MAR) to scenarios reflecting systematic under‐ or overestimation of imputed values (Supporting Information S2) [21, 22, 23]. The discrimination of clinician‐adjudicated ADRD by the computational phenotype was robust to each of the four missing‐data assumptions, with AUROC estimates ranging from 0.87 to 0.88. As we did not discern appreciable bias introduced by missingness within our sample, the MAR assumption was selected for parsimony and used in subsequent analyses.

2.5. Step 3. Model Coefficient Estimation

Within the clinician‐adjudicated cohort, we fit logistic regression models to each of the 10 imputed datasets under the best imputation strategy, using the same variable set described in Step 1. To account for temporal trends in ADRD screening and documentation practices, we incorporated a fixed effect for the calendar year of each individual's index ED encounter [42].

2.6. Step 4. Classification Performance Assessment

4A. Compute Predicted Probabilities of ADRD. We applied the fitted logistic regression coefficients to generate a predicted probability of ADRD for each person within each of the 10 imputed versions of the clinician‐adjudicated cohort.

4B. Validate Performance across Classification Thresholds. To assess classification performance, we estimated the area under the receiver operating characteristic curve (AUROC) and corresponding 95% confidence intervals (CI), quantifying the ability of the computational phenotype to discriminate between individuals with and without clinician‐adjudicated ADRD. We assessed calibration by examining the relationship between predicted probabilities and observed proportions of clinician‐adjudicated ADRD and reported calibration‐in‐the‐large (CITL) and calibration slope. Because clinical utility depends on calibration within the range of predicted probabilities that governs classification decisions and because global calibration slope is attenuated in internal validation studies of modest sample size, we report calibration both across the full range of predicted probabilities and restricted to predicted probabilities at or below the maximum classification threshold evaluated in the decision‐analytic framework (32.0%) [50, 51]. To ensure robust estimates, we applied 10‐fold cross‐validation within each imputed dataset and then pooled the resulting metrics across imputations, incorporating both within‐ and between‐imputation variance.

We then transformed subject‐specific predicted probabilities into dichotomous ADRD classifications using the cohort‐level mean predicted probability (22.0%) as the reference threshold. Individuals at or above this threshold were classified as ADRD‐positive, and those below it as ADRD‐negative. To evaluate clinical tradeoffs, we adjusted the classification threshold in five‐percentage‐point increments above and below the reference threshold. At each threshold, we conducted decision‐curve analysis to measure the net clinical benefit of probabilistic ADRD classification and estimated the number needed to screen (NNS) to identify one additional true positive case. The decision‐curve analysis compares phenotype‐based classification against two reference strategies, evaluating all individuals regardless of predicted probability or evaluating none, while penalizing for false positive classifications at each threshold [52].

4C. Evaluate Performance among Individuals without Documented ADRD Diagnoses. We repeated the evaluation of discrimination, calibration, and net clinical benefit after restricting the clinician‐adjudicated cohort to individuals without ADRD‐related diagnostic codes, enabling assessment of the computational phenotype among individuals whose ADRD would otherwise remain undetected by standard classification algorithms. The reference threshold was centered on the mean predicted probability (14.3%) for this subgroup.

2.7. Step 5. Phenotype Assignment and Assessment of Face Validity

5A. Assign Probable ADRD Status. Within each of the 10 imputed datasets, we computed subject‐specific predicted probabilities of ADRD for the 4500 individuals without clinician adjudication. We then classified these individuals as ADRD‐positive or ADRD‐negative using the cohort‐level mean predicted probability (22.0%) as the classification threshold, consistent with the reference threshold applied in Step 4B.

5B. Compare Individuals with Clinician‐Adjudicated versus Phenotype‐Assigned ADRD. We compared the characteristics of individuals with clinician‐adjudicated ADRD against those with phenotype‐assigned ADRD. Because the reference standard is unavailable for the phenotyping cohort, we compared phenotype‐assigned and clinician‐adjudicated cases across measured characteristics as an assessment of face validity. We calculated standardized mean differences (SMDs) for the same demographic, clinical, and health services utilization characteristics used as model inputs and shown in Table S6, with smaller values indicating greater similarity between groups and values below 0.20 indicating balance. We then stratified the sample into patients with and without ADRD‐related diagnosis codes and repeated the analysis to obtain separate estimates for each group [44, 45].

2.8. Step 6. Sensitivity Analysis

To assess the robustness of our findings to variation in EHR documentation, we repeated the phenotyping procedure using a modified reference standard that excluded individuals whose clinician‐adjudicated cognitive status was assigned with low certainty, treating adjudicator confidence as an indicator of record completeness.

We used R version 4.4.2 to conduct the analysis. We followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines [53]. The code and supporting documentation are available on the Yale Dataverse [43].

3. Results

Consistent with the random subsample design, the clinician‐adjudicated and phenotyping cohorts were closely matched in sociodemographic, clinical, and healthcare utilization characteristics (Table 1). In the clinician‐adjudicated cohort, the mean age was 77.4 years (standard deviation (SD) 9.0); 55.4% were women, and 19.2% identified as members of underrepresented racial or ethnic groups. ADRD‐related diagnosis codes were present for 10% of the clinician‐adjudicated cohort and 8.8% of the phenotyping cohort. Within the clinician‐adjudicated cohort, 102 individuals (20.4%) were classified as having ADRD, of whom 31 (30.4%) were classified as having severe impairment. Among the 5000 sampled individuals, 10% or fewer were missing sociodemographic, clinical, and healthcare utilization characteristics; 30% or fewer were missing lab values (Table S7).

TABLE 1.

Cohort characteristics a .

Clinician‐adjudicated cohort b Phenotyping cohort
No. Individuals 500 4500
Demographic
Age 77.4 (9.0) 77.4 (9.0)
Sex
Female 277 (55.4%) 2496 (55%)
Male 223 (44.6%) 2004 (45%)
Race
African American/Black 48 (9.6%) 370 (8.4%)
Asian e 61 (1.4%)
Pacific Islander e e
White 404 (80.8%) 3633 (83%)
Other 26 (5.2%) 319 (7.3%)
Ethnicity
Hispanic or Latino 31 (6.2%) 348 (7.9%)
Arrival source
Community (Residential) 461 (92.2%) 4102 (91.2%)
Institutional (Nursing Home) e 317 (7.0%)
Other (Alternative Acute Care Setting) e 81 (1.8%)
Dementia‐specific clinical factors
Presence of a qualifying diagnostic code c 50 (10.0%) 395 (8.8%)
Clinician‐adjudicated dementia 102 (20.4%) —
Clinician‐adjudicated dementia severity
Mild 39 (7.8%) —
Moderate 32 (6.4%) —
Severe 31 (6.2%) —
Prescribed a dementia‐specific medication d 26 (5.2%) 230 (5.1%)
Received secondary memory care evaluation 24 (4.8%) 193 (4.3%)
General clinical factors
Elixhauser comorbidity index score 9.1 (10.8) 8.5 (10.6)
Comorbid cardiovascular disease 188 (37.6%) 1540 (34.2%)
Comorbid depression 173 (34.6%) 1385 (30.8%)
Comorbid sensory impairment
Hearing 113 (22.6%) 793 (17.6%)
Visual 131 (26.2%) 1155 (25.7%)
Stroke 106 (21.2%) 977 (21.7%)
Fall‐related injury 303 (60.6%) 2784 (61.9%)
Parkinson's disease 18 (3.6%) 170 (3.8%)
Head CT in ED 0.2 (0.4) 0.3 (0.5)
Potentially inappropriate medications 211 (42.2%) 1856 (41.2%)
First‐generation antihistamines 21 (4.2%) 272 (6.0%)
Antimuscarinics 19 (3.8%) 175 (3.9%)
Muscle relaxants and anti‐spasmodics e 86 (1.9%)
Tricyclic antidepressants e 69 (1.5%)
Antipsychotics 32 (6.4%) 261 (5.8%)
Sedative/Hypnotics 76 (15.2%) 720 (16.0%)
Gabapentinoids 45 (9.0%) 447 (9.9%)
Opiates 98 (19.6%) 733 (16.3%)
Health services utilization
Primary insurance
Medicare 248 (49.6%) 2232 (49.6%)
Dual‐eligible for medicare and medicaid 82 (16.4%) 808 (18.0%)
Annual emergency department encounters 1.0 (1.6) 1.0 (1.8)
Annual hospitalizations 0.4 (1.0) 0.4 (1.0)
Annual outpatient encounters 18.2 (35.4) 15.7 (29.4)
Annual home health encounters 0.2 (3.8) 0.3 (4.1)
Lab values
Serum sodium 139.3 (3.1) 139.3 (3.3)
Albumin 3.8 (0.6) 3.7 (0.6)
Mean platelet volume 10.0 (1.3) 10.1 (1.3)
Open in a new tab

Abbreviations: ADRD: Alzheimer's disease and related dementias; CT: Computed tomography; ED: Emergency department.

a

The mean and standard deviation are reported for continuous variables, number and percent are reported for categorical variables.

b

Characteristics are shown for all 500 individuals randomly selected for inclusion in the clinician‐adjudicated cohort; 37 records lacked sufficient documentation for cognitive adjudication and were therefore excluded from subsequent analyses.

c

Qualifying diagnostic codes are per the Bynum Standard.

d

Dementia‐specific medications include memantine, donepezil, galantamine, rivastigmine.

e

Redacted for cell size < 11.

3.1. Model Discrimination, Calibration, and Threshold Testing in the Overall Cohort

Using data imputed under the MAR assumption, a logistic regression model produced predicted probabilities of ADRD for every individual (Figure S1). The computational phenotype demonstrated good discrimination in the overall clinician‐adjudicated cohort, achieving an AUROC of 0.87 (95% CI 0.82–0.91) (Figure 2A). Assessment of calibration across the full distribution of predicted probabilities demonstrated a CITL of 0.08 and a calibration slope of 0.69 (Figure 2). Within the range of predicted probabilities evaluated in the decision‐analytic framework (≤ 32.0%), which retained 78.6% (364 of 463) of the cohort, the calibration slope improved to 0.86 with a CITL of 0.47.

FIGURE 2.

FIGURE 2

Open in a new tab

Discrimination and calibration of the computational phenotype for clinician‐adjudicated ADRD. (A) displays the AUROC and corresponding 95% CI, estimated using 10‐fold cross‐validation in the clinician‐adjudicated cohort (n = 463); the diagonal dashed line represents an AUROC of 0.50, corresponding to classification performance no better than chance. (B) displays the relationship between predicted probability (x‐axis) and observed proportion (y‐axis) of clinician‐adjudicated ADRD; the diagonal dashed line represents perfect calibration and the shaded region indicates the 95% CI. Abbreviations: ADRD: Alzheimer's disease and related dementias; AUROC: Area under the receiver operating characteristic curve; CI: Confidence interval.

The decision‐curve analysis demonstrated that model‐based classification of ADRD status resulted in net clinical benefit for each threshold examined (predicted probabilities 12%–32%) (Figure 3). At the lowest threshold (12%), the model correctly identified an additional 18 cases of probable ADRD per 100 older adults compared with a baseline scenario in which no individuals are classified as having ADRD. At the highest threshold (32%), the model correctly identified an additional 16 cases per 100 older adults. The NNS increased with higher classification thresholds, ranging from 6 individuals at the lowest threshold (12.0%) to 7 individuals at the highest threshold (32.0%) (Table S8).

FIGURE 3.

FIGURE 3

Open in a new tab

Net clinical benefit of ADRD classification at distinct predicted probability thresholds. We used decision‐curve analysis in the clinician‐adjudicated cohort (n = 463) to estimate net clinical benefit, defined as true positive classifications minus a threshold‐weighted penalty for false positive classifications, across classification thresholds ranging from 12.0% to 32.0%, centered on the mean predicted probability within the sample (22.0%). Abbreviations: ADRD: Alzheimer's disease and related dementias.

3.2. Model Discrimination, Calibration, and Threshold Testing Among Those Without ADRD‐Related Diagnostic Codes

The computational phenotype demonstrated fair discrimination within the subgroup of individuals lacking prior ADRD‐related diagnostic codes, achieving an AUROC of 0.79 (95% CI 0.73–0.86) (Figure S2, Panel A). Assessment of calibration across the full distribution of predicted probabilities demonstrated a CITL of −0.05 and a calibration slope of 0.59 (Figure S2, Panel B). Within the range of predicted probabilities evaluated in the decision‐analytic framework (≤ 32.0%), which retained 86.4% (357 of 413) of the subgroup, the calibration slope improved to 0.81 with a CITL of 0.40. The decision‐curve analysis, centered on the group‐specific mean predicted probability (14.3%), demonstrated net clinical benefit ranging from 8 to 13 identified cases per 100 older adults (Figure S3). The NNS increased with higher classification thresholds, ranging from 8 individuals at the lowest evaluated threshold (4.3%) to 12 individuals at the highest threshold (24.3%) (Table S8).

3.3. Assignment of ADRD Status and Assessment of Balance Between Clinician‐Adjudicated and Phenotyping Cohorts

The computational phenotype accurately reproduced the clinical profile of clinician‐adjudicated ADRD, as evidenced by small SMDs between the clinician‐adjudicated and phenotyping cohorts (Figure 4). The SMDs were below the conventional threshold of 0.20 for most variables, indicating balance. The same pattern was observed within subgroups distinguished by the presence or absence of qualifying ADRD diagnosis codes.

FIGURE 4.

FIGURE 4

Open in a new tab

Balance of patient characteristics between clinician‐adjudicated and phenotyping ADRD Cohorts, overall and by diagnostic status. Individuals in the phenotyping cohort were classified as probable ADRD based on a predicted probability threshold equal to the cohort‐level mean predicted probability (22.0%). To assess the representativeness of the phenotype‐derived cohort, we report standardized mean differences between those with clinician‐adjudicated and phenotype‐assigned ADRD, with dashed vertical lines indicating the threshold for an SMD of 0.2. Abbreviations: ADRD: Alzheimer's disease and related dementias; SMD: Standardized mean difference.

3.4. Sensitivity Analysis

When ADRD cases with the lowest clinician‐adjudicated certainty were excluded from the clinician‐adjudicated cohort, measurable characteristics remained balanced with an SMD ≤ 0.20 for most variables (Figure S4).

4. Discussion

This study demonstrates that a probabilistic computational phenotype derived from routinely collected EHR data and validated within a decision‐analytic framework can accurately reproduce clinician‐adjudicated ADRD status, even in the absence of ADRD‐related diagnostic codes. The phenotype achieved good discrimination of clinician‐adjudicated ADRD, maintaining fair performance among individuals lacking diagnostic codes whose cases would otherwise remain unrecognized. The calibration slopes below 1.0 observed in both groups are consistent with expectations for internal validation studies of this sample size and improved within the range of predicted probabilities governing most classification decisions [50, 51, 54]. A modest decrement in discrimination among individuals without prior diagnoses is expected, given the greater diagnostic uncertainty inherent to this subgroup and the smaller number of confirmed cases available for model training. The preservation of meaningful discrimination and net clinical benefit in this subgroup is an encouraging result, consistent with the performance of validated instruments for undiagnosed dementia in ambulatory care cohorts [25, 26]. Because the framework set forth in this study models the uncertainty arising from multiple potential sources of bias and translates those tradeoffs into net clinical benefit, it offers healthcare organizations a practical and adaptable template for EHR‐based ADRD surveillance. The decision‐analytic framework provides a transparent illustration of the classification tradeoffs inherent in probabilistic classification; in future prospective applications, this framework should enable healthcare organizations to identify the threshold most aligned with their specific clinical and population health management objectives.

The presented framework addresses four entrenched methodological and translational barriers to reliable case identification within health systems. First, comprehensive clinician adjudication of cognitive status remains prohibitively resource‐intensive across large patient populations, while most EDs lack the resources and administrative support to offer cognitive screening at scale [20, 55]. By embedding a modest number of clinician adjudications within an automated phenotyping framework, this approach enables scalable, population‐level inference. Second, the integrated decision‐analytic framework renders the clinical tradeoffs inherent in classification transparent. A positive classification is not a diagnosis but a probable indication for referral for further diagnostic evaluation, subject to variation with respect to local clinical objectives and infrastructure. The costs of false‐positive classification, principally referral for a secondary assessment, must be weighed against the increasing costs that attend underdiagnosis [8, 9, 10, 11]. Third, traditional rule‐based algorithms that rely on ADRD‐related diagnostic codes or prescriptions are expected to overlook more than half of ADRD cases [4, 5, 6]. Through incorporating an evidence‐based set of structured EHR elements into a probabilistic model, we enhance detection of individuals with probable ADRD who have not received a coded diagnosis. Fourth, earlier approaches have often overlooked the potential influence of information bias arising from systematic patterns of missingness in the EHR, despite its importance to accurate and replicable ADRD classification [20, 21]. Our application of pattern‐mixture models provides a replicable method for organizations to formally test missing‐data assumptions against their own documentation patterns.

Prior research has estimated the thresholds at which the benefits of ED‐based ADRD screening outweigh potential harms, concluding that screening is warranted across a broad range of pretest probabilities [56]. However, these point‐of‐care cognitive screening instruments often require clinician administration, which may preclude implementation across healthcare settings with distinct resource constraints [55, 56]. Because EHR‐based approaches to ADRD classification are automated and informed by existing data, they offer a promising alternative that avoids marginal clinical demands on ED clinicians. Earlier EHR‐based instruments for ADRD detection within the ED were not developed to distinguish diagnosed from undiagnosed ADRD or to address information bias arising from heterogeneous documentation practices across healthcare settings, representing constraints on their potential clinical benefit and scalability [18]. In contrast, the instrument and approach presented here were designed to address these limitations, incorporating a decision‐analytic framework that quantifies expected clinical benefit across a continuous range of classification thresholds, while interrogating multiple potential sources of bias. The analyses reported in this paper constitute development and internal validation; external validation within a distinct cohort is an important next step to further establish generalizability. There are multiple promising potential applications of this instrument. For example, once initial encounter data are available, the phenotype could alert clinicians to unrecognized ADRD, prompting timely delirium precautions as well as referral for cognitive evaluation at discharge.

5. Limitations

The following limitations should be noted. First, this study was designed to demonstrate the feasibility of assigning probable ADRD status using standardized, structured EHR elements, rather than to identify a minimal or optimal subset of predictor variables. Although our model performed well, it is possible that more parsimonious models would yield comparable results. Evaluating such alternative models was beyond the scope of this study; however, we plan future predictive work to complement and refine this protocol as additional data become available. Second, analyses were conducted within a single integrated healthcare system; implementation elsewhere requires retraining using data specific to individual healthcare organizations, which can be accomplished using the documentation and code available through the Yale Dataverse. While the diverse hospital settings included within our sampling frame provide a degree of representativeness, external validation across distinct geographies and healthcare organizations is needed. Third, an inherent limitation of EHR data is information bias due to variations in the availability and quality of documentation across individual patients and providers [20, 21]. We addressed this limitation through the use of pattern‐mixture models to impute missing values under multiple assumptions regarding patterns of missingness. Fourth, because several race‐ or ethnicity‐specific categories contained too few observations for stable imputation, we dichotomized whether individuals identified as members of underrepresented racial or ethnic groups to preserve their representation in the analytic cohort [57]. Fifth, because the onset of dementia follows a prolonged preclinical phase and its diagnosis is predicated on clinical judgment, there is an unavoidable risk of misclassification either in clinical practice or through EHR adjudication [20]. To address the potential impact of clinical ambiguity on our results, we conducted a sensitivity analysis in which we modified the diagnostic classifications included in our reference standard.

6. Conclusion

This study presents a replicable framework for healthcare organizations to translate routinely collected clinical data into meaningful, population‐level estimates of probable ADRD. By establishing a transparent and evidence‐based approach that is suited to the operational constraints of the ED, this framework has the potential to enable improved measurement and reduce ADRD underrecognition. As ADRD prevalence increases with population aging, scalable methods for case identification will be critical to the timely initiation of person‐centered care management and strategic planning within healthcare systems.

Author Contributions

Conceptualization: N.F., K.A., U.H.; Funding acquisition: U.H.; Project administration: N.F., K.A., N.S., I.C., I.V.F., U.H.; Supervision: A.P.M., L.H., U.H.; Drafting and revision: All (N.F., K.A., N.S., I.C., I.V.F., S.K., J.Y., M.I., A.P.M., L.H., and U.H.).

Funding

This work was supported by National Institute on Aging (NIA) R03AG088893 (N.F.), R33AG069822 (N.F., N.S., L.H., U.H.), P30AG073104 (N.F., N.S., I.C., L.H., U.H.). National Institutes of Health (NIH) KL2TR001862 (M.I.). Claude D. Pepper Older Adults Independence Center (OAIC) P30AG021342 (N.F., L.H.).

Disclosure

The authors have nothing to report.

Conflicts of Interest

U.H., I.C., N.F., and L.H. are supported by the National Institute on Aging. M,I. is supported by the National Institutes of Health. A.P.M. receives research support through grants to Yale University from the National Institutes of Health, Eli Lilly, Janssen, and Genentech, unrelated to this work. All other authors declare no competing interests.

Supporting information

Supporting Information S1: Modified dementia adjudication protocol (MDAP).

Supporting Information S2: Interrogating information bias through pattern‐mixture modeling.

Table S1: Criteria for classification of mild cognitive impairment.

Table S2: Criteria for classification of dementia.

Table S3: Severity classification using DSM‐5 criteria.

Table S4: Severity classification using the functional assessment staging test (FAST).

Table S5: Items for clinician‐adjudication.

Table S6: Adjusted odds ratios for candidate predictors of ADRD.

Table S7: Missing data patterns in the clinician‐adjudicated and phenotyping cohorts by measure.

Figure S1: Distribution of predicted probabilities of ADRD by clinician‐adjudicated ADRD status.

Figure S2: Discrimination and calibration of the computational phenotype for individuals without ADRD‐related diagnostic codes.

Figure S3: Net clinical benefit of ADRD classification at distinct predicted probability thresholds for individuals without ADRD‐related diagnostic codes.

Table S8: Number needed to screen at distinct classification thresholds, overall and for individuals without ADRD‐related diagnostic codes.

Figure S4: Balance of patient characteristics between clinician‐adjudicated and phenotyping ADRD cohorts (sensitivity analysis excluding low‐confidence adjudications).

JGS-74-2326-s001.docx (713.3KB, docx)

Acknowledgments

Drs. Natalia Festa and Kelsey Alexovitz contributed equally as co‐first authors. The authors would like to recognize Emily Gao, MD, for her earlier contributions to electronic health record review.

Festa N., Alexovitz K., Sifnugel N., et al., “Assigning Probable Dementia Status Using Routinely Collected Electronic Health Record Data,” Journal of the American Geriatrics Society 74, no. 8 (2026): 2326–2337, 10.1111/jgs.70562.

The contents of this manuscript are the sole responsibility of the authors and do not represent the views of the NIA, NIH, or Pepper OAIC.

Data Availability Statement

The code and variable definitions corresponding to this study will be made available to the public via the Yale Dataverse upon publication. The data underlying this study contain protected health information and are not available to the public.

References

  • 1. Hurd M. D., Martorell P., and Langa K., “Future Monetary Costs of Dementia in the United States Under Alternative Dementia Prevalence Scenarios,” Journal of Population Ageing 8, no. 1–2 (2015): 101–112, 10.1007/s12062-015-9112-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Feng Z., Coots L. A., Kaganova Y., and Wiener J. M., “Hospital and ED Use Among Medicare Beneficiaries With Dementia Varies by Setting and Proximity to Death,” Health Affairs 33, no. 4 (2014): 683–690, 10.1377/hlthaff.2013.1179. [DOI] [PubMed] [Google Scholar]
  • 3. Schulz R., Belle S. H., Czaja S. J., McGinnis K. A., Stevens A., and Zhang S., “Long‐Term Care Placement of Dementia Patients and Caregiver Health and Well‐Being,” Journal of the American Medical Association 292, no. 8 (2004): 961–967, 10.1001/JAMA.292.8.961. [DOI] [PubMed] [Google Scholar]
  • 4. Alzheimer's Association , “Alzheimer's Facts and Figures Report,” 2025, accessed May 19, 2025, https://www.alz.org/alzheimers‐dementia/facts‐figures.
  • 5. Amjad H., Roth D. L., Sheehan O. C., Lyketsos C. G., Wolff J. L., and Samus Q. M., “Underdiagnosis of Dementia: An Observational Study of Patterns in Diagnosis and Awareness in US Older Adults,” Journal of General Internal Medicine 33, no. 7 (2018): 1131–1138, 10.1007/s11606-018-4377-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Savva G. M. and Arthur A., “Who Has Undiagnosed Dementia? A Cross‐Sectional Analysis of Participants of the Aging, Demographics and Memory Study,” Age and Ageing 44, no. 4 (2015): 642–647, 10.1093/AGEING/AFV020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Robinson L., Tang E., and Taylor J. P., “Dementia: Timely Diagnosis and Early Intervention,” BMJ 350 (2015): 350, 10.1136/BMJ.H3029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Van Dyck C. H., Swanson C. J., Aisen P., et al., “Lecanemab in Early Alzheimer's Disease,” New England Journal of Medicine 388, no. 1 (2023): 9–21, 10.1056/NEJMOA2212948. [DOI] [PubMed] [Google Scholar]
  • 9. Boustani M. A., Ben Miled Z., Owora A. H., et al., “Digital Detection of Dementia in Primary Care: A Randomized Clinical Trial,” JAMA Network Open 8, no. 11 (2025): e2542222, 10.1001/jamanetworkopen.2025.42222. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Amjad H., Sekhon V. K., Wolff J. L., Samus Q. M., and Roth D. L., “Hospitalization Outcomes Among Older Adults Living Undiagnosed or Unaware of Dementia,” Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring 17, no. 1 (2025): e70051, 10.1002/DAD2.70051. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Amjad H., Roth D. L., Samus Q. M., Yasar S., and Wolff J. L., “Potentially Unsafe Activities and Living Conditions of Older Adults With Dementia,” Journal of the American Geriatrics Society 64, no. 6 (2016): 1223–1232, 10.1111/jgs.14164. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Banack H. R., Kaufman J. S., Wactawski‐Wende J., Troen B. R., and Stovitz S. D., “Investigating and Remediating Selection Bias in Geriatrics Research: The Selection Bias Toolkit,” Journal of the American Geriatrics Society 67, no. 9 (2019): 1970–1976, 10.1111/JGS.16022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Latham‐Mintus K., “Selective Bias in Longitudinal Studies,” in Encyclopedia of Gerontology and Population Aging (Springer, 2021), 4388–4390, 10.1007/978-3-030-22009-9_577. [DOI] [Google Scholar]
  • 14. Hernandez R., Jin H., Lee P. J., et al., “Attrition From Longitudinal Ageing Studies and Performance Across Domains of Cognitive Functioning: An Individual Participant Data Meta‐Analysis,” BMJ Open 14, no. 3 (2024): e079241, 10.1136/BMJOPEN-2023-079241. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Chamberlain S. A., Duggleby W., Teaster P., and Estabrooks C., “Characteristics and Unmet Care Needs of Unbefriended Residents in Long‐Term Care: A Qualitative Interview Study,” Aging & Mental Health 24, no. 4 (2020): 659–667, 10.1080/13607863.2019.1566812. [DOI] [PubMed] [Google Scholar]
  • 16. Collyer T. A., Murray A. M., Woods R. L., et al., “Association of Dual Decline in Cognition and Gait Speed With Risk of Dementia in Older Adults,” JAMA Network Open 5, no. 5 (2022): E2214647, 10.1001/JAMANETWORKOPEN.2022.14647. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Walling A. M., Pevnick J., Bennett A. V., et al., “Dementia and Electronic Health Record Phenotypes: A Scoping Review of Available Phenotypes and Opportunities for Future Research,” Journal of the American Medical Informatics Association 30, no. 7 (2023): 1333–1348, 10.1093/jamia/ocad086. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Cohen I., Taylor R. A., Xue H., et al., “Detection of Emergency Department Patients at Risk of Dementia Through Artificial Intelligence,” Alzheimer's & Dementia 21, no. 6 (2025): e70334, 10.1002/ALZ.70334. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Reuben D. B., Hackbarth A. S., Wenger N. S., Tan Z. S., and Jennings L. A., “An Automated Approach to Identifying Patients With Dementia Using Electronic Medical Records,” Journal of the American Geriatrics Society 65, no. 3 (2017): 658–659, 10.1111/JGS.14744. [DOI] [PubMed] [Google Scholar]
  • 20. Festa N., Moura L. M. V. R., Blacker D., Newhouse J. P., and Hsu J., “Promise and Peril of Claims‐Based Dementia Ascertainment in Causal Inference,” BMJ Evidence‐Based Medicine 28, no. 4 (2023): 222–225, 10.1136/bmjebm-2022-112134. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Haneuse S., Arterburn D., and Daniels M. J., “Assessing Missing Data Assumptions in EHR‐Based Studies: A Complex and Underappreciated Task,” JAMA Network Open 4, no. 2 (2021): e210184, 10.1001/JAMANETWORKOPEN.2021.0184. [DOI] [PubMed] [Google Scholar]
  • 22. Little R. J. A. and Wang Y., “Pattern‐Mixture Models for Multivariate Incomplete Data With Covariates,” Biometrics 52, no. 1 (1996): 98–111, 10.2307/2533148. [DOI] [PubMed] [Google Scholar]
  • 23. Little R. J., “Pattern‐Mixture Models for Missing Data,” Journal of the American Medical Association 334, no. 23 (2025): 2126–2127, 10.1001/jama.2025.18109. [DOI] [PubMed] [Google Scholar]
  • 24. Haneuse S., Bogart A., Jazic I., et al., “Learning About Missing Data Mechanisms in Electronic Health Records‐Based Research: A Survey‐Based Approach,” Epidemiology 27, no. 1 (2016): 82–90, 10.1097/EDE.0000000000000393. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Barnes D. E., Zhou J., Walker R. L., et al., “Development and Validation of eRADAR: A Tool Using EHR Data to Detect Unrecognized Dementia,” Journal of the American Geriatrics Society 68, no. 1 (2020): 103–111, 10.1111/JGS.16182. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Coley R. Y., Smith J. J., Karliner L., et al., “External Validation of the eRADAR Risk Score for Detecting Undiagnosed Dementia in Two Real‐World Healthcare Systems,” Journal of General Internal Medicine 38, no. 2 (2023): 351–360, 10.1007/S11606-022-07736-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Hwang U., Peter J. J., Health L., and York N., “GEAR 2.0 ‐ Detection of Dementia and Cognitive Impairment,” Alzheimer's & Dementia 20, no. S5 (2024): e091835, 10.1002/ALZ.091835. [DOI] [Google Scholar]
  • 28. MA L. M., Stump T. E., Messina F. C., Miller D. K., and Callahan C. M., “Emergency Department Use Among Older Adults With Dementia,” Alzheimer Disease and Associated Disorders 30, no. 1 (2016): 35–40, 10.1097/WAD.0000000000000118. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Saxena S., Carpenter C., Floden D. P., Meldon S., Taylor R. A., and Hwang U., “Novel Algorithms & Blood‐Based Biomarkers: Dementia Detection and Care Transitions for Persons Living With Dementia in the Emergency Department,” Alzheimer's & Dementia 21, no. 5 (2025): e70287, 10.1002/ALZ.70287. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Seidenfeld J., Runels T., Goulet J. L., et al., “Patterns of Emergency Department Visits Prior to Dementia or Cognitive Impairment Diagnosis: An Opportunity for Dementia Detection?,” Academic Emergency Medicine 32, no. 5 (2025): 570–573, 10.1111/acem.14832. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Gettel C., Galske J., Araujo K., et al., “Detection and Differentiation of Undiagnosed Dementia in the Emergency Department: A Pilot Referral Pathway,” Alzheimer's & Dementia 21, no. 4 (2025): e70189, 10.1002/alz.70189. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Shenvi C., Kennedy M., Austin C. A., Wilson M. P., Gerardi M., and Schneider S., “Managing Delirium and Agitation in the Older Emergency Department Patient: The ADEPT Tool,” Annals of Emergency Medicine 75 (2020): 136–145, 10.1016/j.annemergmed.2019.07.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Lee S., Chen H., Hibino S., et al., “Can We Improve Delirium Prevention and Treatment in the Emergency Department? A Systematic Review,” Journal of the American Geriatrics Society 70, no. 6 (2022): 1838–1849, 10.1111/jgs.17740. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Carpenter C. R., Bromley M., Caterino J. M., et al., “Optimal Older Adult Emergency Care: Introducing Multidisciplinary Geriatric Emergency Department Guidelines From the American College of Emergency Physicians, American Geriatrics Society, Emergency Nurses Association, and Society for Academic Emergency Medicine,” Journal of the American Geriatrics Society 62, no. 7 (2014): 1360–1363, 10.1111/jgs.12883. [DOI] [PubMed] [Google Scholar]
  • 35. Hripcsak G. and Albers D. J., “Next‐Generation Phenotyping of Electronic Health Records,” Journal of the American Medical Informatics Association 20, no. 1 (2013): 117–121, 10.1136/amiajnl-2012-001145. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Moura L. M. V. R., Festa N., Price M., et al., “Identifying Medicare Beneficiaries With Dementia,” Journal of the American Geriatrics Society 69, no. 8 (2021): 2240–2251, 10.1111/JGS.17183. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Thokala P., Goodacre S., Cooper G., et al., “Decision Analytical Modelling of Strategies for Investigating Suspected Acute Aortic Syndrome,” Emergency Medicine Journal 41, no. 12 (2024): 728–735, 10.1136/EMERMED-2024-214222. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Carrell D. S., Floyd J. S., Gruber S., et al., “A General Framework for Developing Computable Clinical Phenotype Algorithms,” Journal of the American Medical Informatics Association 31, no. 8 (2024): 1785–1796, 10.1093/JAMIA/OCAE121. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Reuben D. B., Gill T. M., Stevens A., et al., “Health System, Community‐Based, or Usual Dementia Care for Persons With Dementia and Caregivers: The D‐CARE Randomized Clinical Trial,” Journal of the American Medical Association 333, no. 11 (2025): 950–961, 10.1001/JAMA.2024.25056. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Lim M. Y. L. and Loo J. H. Y., “Screening an Elderly Hearing Impaired Population for Mild Cognitive Impairment Using Mini‐Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA),” International Journal of Geriatric Psychiatry 33, no. 7 (2018): 972–979, 10.1002/GPS.4880. [DOI] [PubMed] [Google Scholar]
  • 41. Jorgensen L. E., Palmer C. V., Pratt S., Erickson K. I., and Moncrieff D., “The Effect of Decreased Audibility on MMSE Performance: A Measure Commonly Used for Diagnosing Dementia,” Journal of the American Academy of Audiology 27, no. 4 (2016): 311–323, 10.3766/JAAA.15006. [DOI] [PubMed] [Google Scholar]
  • 42. Festa N., Price M., Weiss M., et al., “Evaluating the Accuracy of Medicare Risk Adjustment for Alzheimer's Disease and Related Dementias,” Health Affairs 41, no. 9 (2022): 1324–1332, 10.1377/hlthaff.2022.00185. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Festa N., Alexovitz K., Sifnugel N., et al., Assigning Probable Dementia Status Using Routinely Collected Electronic Health Record Data (Yale Dataverse, 2025), https://dataverse.yale.edu/dataverse/adrd_phenotype. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. EP M. C., Chang C. H., Tilton N., et al., “Validation of Claims Algorithms to Identify Alzheimer's Disease and Related Dementias,” Journals of Gerontology, Series A, Biological Sciences and Medical Sciences 77, no. 6 (2022): 1261–1271, 10.1093/gerona/glab373. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45. Grodstein F., Chang C.‐H., Capuano A. W., et al., “Identification of Dementia in Recent Medicare Claims Data, Compared to Rigorous Clinical Assessments,” Journals of Gerontology: Series A (2021): glab377, 10.1093/GERONA/GLAB377. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Centers for Medicare & Medicaid Services , “Chronic Conditions Warehouse Technical Guidance: Getting Started With CMS Medicare Administrative Research Files, Version 2.8,” 2022, https://www2.ccwdata.org/documents/10280/19002246/ccw‐medicare‐data‐user‐guide.pdf.
  • 47. Min L., Tinetti M., Langa K. M., Ha J., Alexander N., and Hoffman G., “Measurement of Fall Injury With Health Care System Data and Assessment of Inclusiveness and Validity of Measurement Models,” JAMA Network Open 2, no. 8 (2019): e199679, 10.1001/jamanetworkopen.2019.9679. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48. National Library of Medicine , “Value Set Authority Center,” 2025, https://vsac.nlm.nih.gov/.
  • 49. American Geriatrics Society , “2023 Updated AGS Beers Criteria for Potentially Inappropriate Medication Use in Older Adults,” Journal of the American Geriatrics Society 71, no. 7 (2023): 2052–2081, 10.1111/jgs.18372. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50. Mishra A., McClelland R. L., Inoue L. Y. T., et al., “Recalibration Methods for Improved Clinical Utility of Risk Scores,” Medical Decision Making 42, no. 4 (2021): 500, 10.1177/0272989X211044697. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51. Van Calster B., DJ M. L., Van Smeden M., et al., “Calibration: The Achilles Heel of Predictive Analytics,” BMC Medicine 17, no. 1 (2019): 230, 10.1186/s12916-019-1466-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52. Vickers A. J., van Calster B., and Steyerberg E. W., “A Simple, Step‐By‐Step Guide to Interpreting Decision Curve Analysis,” Diagnostic and Prognostic Research 3, no. 1 (2019): 1–8, 10.1186/s41512-019-0064-7/figures/3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53. Collins G. S., Reitsma J. B., Altman D. G., and Moons K. G. M., “Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD): The TRIPOD Statement,” Annals of Internal Medicine 162, no. 1 (2015): 55–63, 10.7326/M14-0697. [DOI] [PubMed] [Google Scholar]
  • 54. Vergouwe Y., Steyerberg E. W., Eijkemans M. J. C., and Habbema J. D. F., “Substantial Effective Sample Sizes Were Required for External Validation Studies of Predictive Logistic Regression Models,” Journal of Clinical Epidemiology 58, no. 5 (2005): 475–483, 10.1016/j.jclinepi.2004.06.017. [DOI] [PubMed] [Google Scholar]
  • 55. Shih R. D., Carpenter C. R., Tolia V., Binder E. F., and Ouslander J. G., “Balancing Vision With Pragmatism: The Geriatric Emergency Department Guidelines‐Realistic Expectations From Emergency Medicine and Geriatric Medicine,” Journal of the American Geriatrics Society 70, no. 5 (2022): 1368–1373, 10.1111/jgs.17745. [DOI] [PubMed] [Google Scholar]
  • 56. Carpenter C. R., Banerjee J., Keyes D., et al., “Accuracy of Dementia Screening Instruments in Emergency Medicine: A Diagnostic Meta‐Analysis,” Academic Emergency Medicine 26, no. 2 (2019): 226–245, 10.1111/acem.13573. [DOI] [PubMed] [Google Scholar]
  • 57. Nguyen C. D., Carlin J. B., and Lee K. J., “Practical Strategies for Handling Breakdown of Multiple Imputation Procedures,” Emerging Themes in Epidemiology 18, no. 1 (2021): 5, 10.1186/S12982-021-00095-3. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supporting Information S1: Modified dementia adjudication protocol (MDAP).

Supporting Information S2: Interrogating information bias through pattern‐mixture modeling.

Table S1: Criteria for classification of mild cognitive impairment.

Table S2: Criteria for classification of dementia.

Table S3: Severity classification using DSM‐5 criteria.

Table S4: Severity classification using the functional assessment staging test (FAST).

Table S5: Items for clinician‐adjudication.

Table S6: Adjusted odds ratios for candidate predictors of ADRD.

Table S7: Missing data patterns in the clinician‐adjudicated and phenotyping cohorts by measure.

Figure S1: Distribution of predicted probabilities of ADRD by clinician‐adjudicated ADRD status.

Figure S2: Discrimination and calibration of the computational phenotype for individuals without ADRD‐related diagnostic codes.

Figure S3: Net clinical benefit of ADRD classification at distinct predicted probability thresholds for individuals without ADRD‐related diagnostic codes.

Table S8: Number needed to screen at distinct classification thresholds, overall and for individuals without ADRD‐related diagnostic codes.

Figure S4: Balance of patient characteristics between clinician‐adjudicated and phenotyping ADRD cohorts (sensitivity analysis excluding low‐confidence adjudications).

JGS-74-2326-s001.docx (713.3KB, docx)

Data Availability Statement

The code and variable definitions corresponding to this study will be made available to the public via the Yale Dataverse upon publication. The data underlying this study contain protected health information and are not available to the public.

Articles from Journal of the American Geriatrics Society are provided here courtesy of Wiley

RESOURCES