Abstract
BACKGROUND: Cognitive dysfunction is central to clinicopathological models of Alzheimer’s disease (AD). While AD prospective studies assess similar cognitive domains, the neuropsychological tests used vary between studies, limiting potential for aggregation. We examined a machine learning (ML) data harmonisation method(1) for neuropsychological test data to develop a harmonised PACC score for the Alzheimer’s Dementia Onset and Progression in International Cohorts (ADOPIC) consortium. METHOD: ADOPIC included longitudinal clinicopathological data from AIBL (N = 1765), ADNI (N = 1779) and OASIS (N = 440) cohorts (Table 1). Harmonization involved three stages. First, cognitive domains of interest and the neuropsychological tests assessing each were defined. Second, a standardized scoring and naming convention was established for demographic and neuropsychological outcomes. Third, the ML harmonisation approach(1) was applied. Test scores present in one cohort, but not another, were treated as missing and imputed using ML before calculating a PACC(1). Imputation utilized data from neuropsychological tests, age, sex, years of education, and APOE‐ɛ4 status. We validated the harmonized PACC (H‐PACC) by randomly simulating missing cognitive scores and analysing percentage error of imputed scores versus actual data(1). Validity of harmonized scores was determined by their sensitivity to decline associated with amyloid positivity and clinical disease stage. Linear mixed models (LMMs) modelled trajectory of change on H‐PACC and AIBL PACC scores (including identical tests), with time, CDR‐global score, and amyloid status (Aβ‐/Aβ+) interactions as fixed effects. Sex and age at baseline were covariates. Variation in baseline and decline in PACC scores was modelled using random intercepts and slopes. Effect sizes for LMMs were computed with pseudo‐R‐squared. RESULT: Percentage errors for imputed neuropsychological test scores showed high accuracy with low standard deviations (Figure 1). Sensitivity to clinicopathological disease stage was qualitatively similar for the H‐PACC and AIBL PACC (Figure 2), both discriminating (p<0.001) annual decline rates of Aβ+ CDR 0.5 and CDR≥1 groups from Aβ‐ CDR 0 group. Effect sizes were substantial, with H‐PACC and AIBL PACC data explaining 95% and 93% of variance, respectively. CONCLUSION: The H‐PACC, developed via ML harmonization, was precise and has utility for aggregating neuropsychological test data to be used in prospective AD studies. References: (1) https://doi.org/10.1002/alz.044302 (2) https://doi.org/10.1093/bioinformatics/btr597