A HIDDEN MARKOV MODEL’S AGREEMENT WITH CLINICAL DIAGNOSES AND ITS INDICATION OF ADDITIONAL PRECLINICAL COGNITIVE DEFICITS IN THE WISCONSIN REGISTRY FOR ALZHEIMER’S PREVENTION

Abstract

Clinicians determine cognitive status based on review of multidimensional data (e.g., neuropsychological, medical history, daily functioning). Reliable determination of statuses, including potentially at preclinical stages of decline, is essential. In this study, we used a constrained hidden Markov model (HMM) to investigate the consistency of clinicians' cognitive status assignments in the Wisconsin Registry for Alzheimer's Prevention (WRAP). Participants with up to 5 assessment waves (N=1408, total visits=4714) were drawn from WRAP. Cognitive data included: (i) demographically-adjusted z-scores from 18 neuropsychological tests (X); and (ii) a consensus label (Y) of four cognitive stages [cognitively normal (0), early MCI (1), clinical MCI (2), and dementia (3)]. A four-state hidden Markov model (HMM) was built to characterize the hidden states(Q) of disease progression using both X and Y as independent observations. Constraints on hidden state transitions and emission of labels(Y) were imposed to ensure the alignment between Y and Q. Missing values were allowed for both X and Y variables. There is clear correspondence between consensus labels(Y) and hidden states(Q) (Table 1). Most visits with impaired consensus labels(Y) were captured by HMM with recall rates of 0.849, 0.889 and 1.00 for early MCI, clinical MCI and dementia, respectively. A mismatch was observed with 715 of the normal diagnoses (Y=0, 26.5%) being decoded by HMM as early MCI (Q=1). In addition, the mean parameters of hidden states do not differ to the empirical means on both clinical MCI and dementia (Table 2). In contrast, “normal” HMM state (Q=0) has consistently higher scores than Y=0 across all cognitive tests. The “early MCI” HMM state (Q=1) also has higher scores on 4 memory tests (Table 2). While the constrained hidden Markov model built to examine the cognitive status diagnoses in WRAP appears to match consensus conference diagnoses on clinical MCI and dementia, the model suggests an even milder onset of early MCI and a stricter standard of normal than is currently recognized. This modeling approach is a cost-effective means of identifying potentially missed cases of cognitive decline and may be applied to other longitudinal AD datasets.