Development and validation of predictive models for undiagnosed Alzheimer's disease and related dementias in the emergency department.

Abstract

Introduction: The electronic health record (EHR) offers a promising resource to improve recognition of Alzheimer's disease and related dementias (ADRD), yet existing approaches have been developed within cohorts that underrepresent individuals prone to underdiagnosis and do not differentiate disease severity.

Methods: We applied a validated computational phenotype to identify probable ADRD (p-ADRD) and its severity within a random sample of 5000 individuals aged ≥65 years who presented to one of nine emergency departments in the Yale New Haven Health System (2014 to 2022). The dichotomous outcome was undiagnosed ADRD, assigned to those meeting criteria for p-ADRD but lacking validated diagnostic codes. We applied logistic regression with elastic net regularization to predict undiagnosed status for the overall cohort and within each severity stratum, assessing performance using 10-fold cross-validation.

Results: Of 1358 individuals with p-ADRD (mean age 86.7 years, standard deviation 7.4), 934 (68.8%) were undiagnosed. The overall model discriminated undiagnosed ADRD with an area under the receiver operating characteristic curve (AUROC) of 0.85 (95% confidence interval [CI] 0.83 to 0.87). The severity-stratified models achieved AUROC of 0.80 (95% CI 0.77 to 0.83) for mild and 0.91 (95% CI 0.88 to 0.94) for moderate to severe disease. Although the predictors of undiagnosed status differed by severity, fall-related injury was protective against underdiagnosis within each stratum.

Discussion: We developed and validated predictive models using structured EHR elements, which demonstrated robust discrimination of undiagnosed ADRD within a clinical cohort prone to underdiagnosis. The heterogeneous predictors across severity strata underscore the importance of aligning prediction with disease severity, given the distinct clinical needs that attend ADRD progression.