Introduction: Cerebrospinal fluid (CSF) tau phosphorylation at multiple sites is associated with cortical amyloid and other pathologic changes in Alzheimer's disease. These relationships can be non-linear. We used an artificial neural network to assess the ability of 10 different CSF tau phosphorylation sites to predict continuous amyloid positron emission tomography (PET) values.
Methods: CSF tau phosphorylation occupancies at 10 sites (including pT181/T181, pT217/T217, pT231/T231 and pT205/T205) were measured by mass spectrometry in 346 individuals (57 cognitively impaired, 289 cognitively unimpaired). We generated synthetic amyloid PET scans using biomarkers and evaluated their performance.
Results: Concentration of CSF pT217/T217 had low predictive error (average error: 13%), but also a low predictive range (ceiling 63 Centiloids). CSF pT231/T231 has slightly higher error (average error: 19%) but predicted through a greater range (87 Centiloids).
Discussion: Tradeoffs exist in biomarker selection. Some phosphorylation sites offer greater concordance with amyloid PET at lower levels, while others perform better over a greater range.
Highlights: Novel pTau isoforms can predict cortical amyloid burden. pT217/T217 accurately predicts cortical amyloid burden in low-amyloid individuals. Traditional CSF biomarkers correspond with higher levels of amyloid.
Keywords: CSF tau occupancy; PET; biomarker concordance; machine learning; novel biomarkers.
© 2024 The Author(s). Alzheimer's & Dementia published by Wiley Periodicals LLC on behalf of Alzheimer's Association.