|
Abstract:
|
Gold-standard dementia ascertainment, involving neuropsychological testing, clinical exams, and diagnosis adjudication by a panel of clinicians, is resource-intensive and infeasible in large population-representative studies, presenting a barrier to understanding population-level burden and determinants of dementia. Algorithmic dementia classification methods are an alternative, but lack key measures used for clinical dementia ascertainment (e.g., detailed neuropsychological assessments). The first step to strengthening algorithmic dementia classification in large cohorts is developing frameworks that can make use of these important measures. We developed a Bayesian latent class mixture modeling framework for algorithmic dementia classification that incorporates neuropsychological measures and sociodemographic, health, and health behavior information.
Methods were demonstrated and validated using detailed neuropsychological measures from the Aging, Demographics, and Memory Study (ADAMS, n = 520), a substudy of the US Health and Retirement Study that included detailed neuropsychological assessments and clinical dementia adjudication (unimpaired n=211, cognitive impairment due to other conditions (other) n=86, mild cognitive impairment n = 65, dementia n=158). Utilizing a portion of the ADAMS data as a training sample (n=364), Bayesian methods were used to fit latent class mixture models and generate synthetic datasets with subgroups distinguished based on impairment characteristics, a process that overlaps with what has been termed “algorithmic dementia classification” in the epidemiologic and aging literature. The remaining hold-out portion of the ADAMS data (n=156) was used to evaluate the model’s external validity.
We generated one thousand synthetic version of ADAMS training and hold-out samples. Data quality checks showed that synthetic samples reproduced characteristics of the ADAMS cohort (e.g., similar covariate distributions between synthetic datasets and observed data). Algorithmic dementia classification within the ADAMS training sample yielded 95% credible intervals that captured the observed count in all ADAMS impairment classes (unimpaired, other, MCI, dementia) with the mean count for the dementia class possessing the largest discrepancy compared with ADAMS observed counts (25 individuals out of 364 in the sample). Algorithmic dementia classification in the ADAMS hold-out sample yielded 95% credible intervals that captured the observed count in all ADAMS impairment classes with the largest mean-count discrepancy being five individuals in the MCI class (out of 156 in the sample). Thus, our model for algorithmic dementia classification was highly concordant with ADAMS clinical diagnoses. This work demonstrates opportunities for improving algorithmic dementia classification in large cohorts where clinical diagnoses are unavailable.
|
