Premium
Predicting future regional tau accumulation in preclinical Alzheimer’s disease
Author(s) -
Giorgio Joseph,
Jagust William J.,
Baker Suzanne L.,
Landau Susan M.,
Tino Peter,
Kourtzi Zoe
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.039827
Subject(s) - dementia , cohort , medicine , oncology , psychology , alzheimer's disease neuroimaging initiative , multivariate statistics , neuroscience , disease , machine learning , computer science
Background Current models of AD pathogenesis hypothesise that b‐amyloid induces downstream tau deposition leading to cognitive decline. Here, we use machine learning with baseline data to predict dynamic changes in cortical tau in preclinical AD populations. Methods Using individuals from the ADNI (n=307) we train a machine learning algorithm that derives a continuous index using baseline medial temporal atrophy, amyloid‐PET, and APOE genotype. The index is generated to discriminate between cognitively normal (CN) individuals who remained normal over ≥ 3 years (n=145) (34 amyloid positive at baseline) and non‐demented progressing (NDP) individuals (CN/MCI at baseline) who developed dementia (n=162) (135 amyloid positive at baseline) The index combines information from multivariate data to maximise between class separation and minimise within class variability. We test the hypothesis that this index is predictive of future tau accumulation in two out‐of‐sample cohorts. In a separate ADNI cohort (ADNI‐3; 72CN/43MCI) with longitudinal flortaucipir PET(FTP) scans we derived the index and related it to future tau accumulation (FTP‐PET slope over time in Desikan‐Killiany ROIs normalised to eroded subcortical white matter) for individuals classified as NDP. Next, for cognitively normal individuals from the Berkeley Aging Cohort Study (BACS; n=56) we derived the index and tested predictions of regional future rate of tau accumulation for individuals classified as NDP. Results The algorithm using baseline data separated CN from NDP individuals with high accuracy (86.4%). Testing on ADNI‐3 individuals classified 54/115 individuals as NDP. These NDP individuals accumulated global cortical tau 2.8 times faster than non‐NDP individuals (Figure 1); for NDP individuals this index also predicted individual variability in regional future tau accumulation (Figure 2a). For BACS participants classified as NDP (n=23), the individualised regional predictions explain up to 38.5% of the observed variance in longitudinal tau accumulation in the temporal cortex (mean=27%) and 32% of the variance in superior and medial regions of the posterior parietal cortex (mean=23%) (Figure 2b). Conclusion Machine learning derived a continuous prognostic index from baseline data that was capable of predicting changes in regional tau accumulation in both MCI and cognitively normal people. This could have clinical applications in developing clinical trials or furthering our understanding of AD pathophysiological mechanisms.