Analyses of diagnostic patterns at 30 Alzheimer's disease centers in the US

Abstract

Background: The US Alzheimer's Disease Centers (ADCs) (n = 30) recently created a uniform data set. We sought to determine which variables were most important in making a diagnosis, and how these differed across ADCs.

Methods: A cross-sectional analysis of first visits to ADCs via polytomous logistic regression. We analyzed subjects with complete data (n = 7,555, 89%), and also used multiple imputation to infer missing data.

Results: There were 8,495 subjects; 50, 26, and 24% were diagnosed as normal, having mild cognitive impairment (MCI), or mild Alzheimer's disease [Clinical Dementia Rating (CDR) score <1], respectively. The model using 7,555 subjects was 86% accurate in predicting diagnosis. Important predictors were physician-reported decline and the CDR sum of boxes, followed by 4 cognitive tests (Mini Mental State Examination, Category Fluency Tests, Logical Memory Test, Boston Naming Test). Multiple imputation revealed Trail Making Test B to be additionally important. Consensus versus single-clinician diagnoses were 2-3 times more likely to result in MCI than normal diagnoses. Excluding clinical judgment variables, functional assessment and psychiatric symptoms were important additional predictors; model accuracy remained high (78%). There were significant differences between centers in the use of different cognitive tests in making diagnoses.

Conclusions: We recommend creating a hypothetic data set to use across ADCs to improve diagnostic consistency, and a survey on the use of raw or adjusted cognitive test scores by different ADCs.

Fig. 1

Consensus diagnoses by center.

Fig. 2

ORs for diagnosing MCI versus normal, per unit increase in MMSE (significant differences from common OR indicated by an asterisk).