Machine learning composite of remote cognitive tests for screening of preclinical AD cohorts

Abstract

AbstractBackgroundComputerized cognitive tests are critical in the screening of preclinical Alzheimer’s Disease (AD) given their ability to be conducted remotely at low cost. Here, we explore whether machine‐learning derived composites of remote cognitive tests can improve sensitivity when detecting brain amyloid (Aβ) accumulation in cognitively normal (CN) cohorts than traditional pen‐and‐paper alternatives.MethodBased on the CogState Brief Battery tests in CN individuals from the A4 dataset (n = 4230, Aβ‐ = 3094), we construct a data‐driven composite, denoted the Remote Empirical Cognitive composite for Alzheimer’s Disease (RECAD), designed to be sensitive to Aβ‐PET positivity. Specifically, the composite uses a random feature elimination logistic regression to identify a subset of predictive tests. We construct and evaluate the model using 10 repeat, 5‐fold cross‐validation in the A4 baseline data. We conduct temporal validation with the A4 second visit and external validation using CN participants from the ADNI (n = 389, Aβ‐ = 143), comparing against the pen‐and‐paper‐based PACC composite and previous computerized composites CBB and C3.ResultRECAD includes three accuracy scores (Identification, One‐Back and One‐card) and one reaction time (Detection). In cross‐validation in the A4 baseline, RECAD show a significantly higher ability to discriminate Aβ status [mean Cohen’s d = 0.26] than PACC [mean d = 0.25, t‐test p<0.001] and computerized composites CBB [median d = 0.13, p<0.001] and C3 [mean d = 0.18, p<0.001]. RECAD significantly improves beyond C3 in the A4 second visit (z‐test p<0.001). Moreover, these trends are replicated in the ADNI at baseline (RECAD d = 0.215, PACC d = 0.061) and 24 months (RECAD d = 0.295, PACC d = ‐0.026) but the small sample and effect size limits our ability to demonstrate statistical significance.ConclusionRECAD, as a machine‐learning derived composite of remotely conducted cognitive tests, shows a comparable or better sensitivity to the established pen‐and‐paper composite test when screening for Aβ+ among CN cohorts. While effect sizes are small, conducting such remote tests repeatedly may be a useful pre‐screening for downstream biofluid evaluation.