A comparison of machine learning based‐composite cognitive test scores to track cognitive decline in early stages of Alzheimer’s disease dementia: ADOPIC study

Abstract

AbstractBackgroundClinical trials of early Alzheimer’s disease (AD) dementia use cognition as a primary outcome and therefore cognitive measures that accurately reflect disease progression and represent multiple cognitive domains are required. Current cognitive endpoints are computed by averaging standardized change from baseline scores (i.e., preclinical Alzheimer cognitive composite (PACC)). Here we compare the PACC to composite scores for which the combination of multiple performance scores has been optimized using machine learning (ML) algorithms in a large, harmonised data set, namely ADOPIC.MethodsA dataset harmonised across various cognitive scores (Table 1) was used to construct composite scores using ML‐based algorithms: a manifold learning dimension reduction technique (UMAP), principal component analysis (PCA) and Latent variable analysis (LVA). Data from ADNI (n = 1470), AIBL (n = 1105) and OASIS participants (n = 412, Table 2) with ≥3 assessments ≤5 years before clinical progression/last visit were included. Participants were classified clinically as; stable cognitively unimpaired (CU), CU progressing to mild cognitive impairment (MCI) or dementia, stable MCI, MCI progressing to dementia, or AD dementia. For UMAP, all stable participants (including dementia) were used in 4‐fold cross‐validation training sets. However, for testing UMAP and unsupervised models (PCA and LVA) all clinical groups (except for dementia) were used. The validity of ML‐based composites was challenged by modelling mean (SD) change over time in the non‐demented clinical groups, using linear mixed model (LMM) analysis. Signal‐to‐noise ratios (SNRs) were calculated (mean change/SD change) for each composite. The mean change was measured in progressives relative to stable participants of each group.ResultsEach ML‐based cognitive composite showed sensitivity to cognitive decline in the progressor groups (Figure 1 A). For the MCI progressors, PCA and UMAP composites had significantly higher SNRs than PACC (P<0.01; Figure 1B), however, LVA performance was not significantly better than PACC. For the CU progressor group, SNRs for PCA and LVA did not show significant differences with PACC and UMAP performed significantly worse than PACC (P<0.01).ConclusionML‐based cognitive composite score computed using PCA provides a practical solution to track cognitive decline with improved performance in tracking cognitive decline in MCI progressors compared to PACC, while being comparable in CUs.