# Prediction of continuous amyloid positron emission tomography with fluid measures of phosphorylated tau and β-amyloid

## TL;DR

Researchers used machine learning to predict brain amyloid levels in Alzheimer's disease using fluid biomarkers from cerebrospinal fluid and blood.

## Contribution

A novel machine learning pipeline combining CSF and plasma biomarkers to accurately predict quantitative amyloid PET burden.

## Key findings

- Combined CSF and plasma biomarkers achieved R² = 0.79 in predicting Aβ-PET burden.
- Plasma P-tau217 and CSF Aβ42/Aβ40 were the most significant predictors of amyloid pathology.
- Plasma-only models performed nearly as well as CSF-based models, suggesting non-invasive alternatives.

## Abstract

Brain amyloid-β (Aβ) pathology is a core feature of Alzheimer disease (AD) and can be quantified using positron emission tomography (PET). Cerebrospinal fluid (CSF) and plasma biomarkers detect abnormal Aβ, but it is unclear to what degree they can predict quantitative Aβ-PET. We explored plasma and CSF biomarkers in relation to Aβ-PET in the BioFINDER-2 study (N = 1053), and the BioFINDER-1 study (N = 238). We developed a machine learning pipeline to predict Aβ-PET using CSF and plasma measures. The best models achieved R2 = 0.79. Plasma P-tau217 and CSF Aβ42/Aβ40 contributed the most. CSF Aβ42/Aβ40 contributed most to identify Aβ-positivity, while continuous Aβ-PET load within the positive range was best predicted by plasma P-tau217. Models using only plasma measures approached performance of CSF models. Altered metabolism of soluble Aβ may be highly associated with presence of Aβ plaques, while soluble P-tau217 levels may continue to change during build-up of Aβ pathology.

Machine-learning models using CSF and plasma biomarkers accurately predicted quantitative Aβ-PET burden. Distinct contributions of soluble Aβ42/Aβ40 and plasma P-tau217 reveal their role in different stages of Aβ-pathology.

Prediction of quantitative Aβ-PET reached R² = 0.79 and 7% mean error using combined CSF and plasma biomarkers.CSF Aβ42/Aβ40 was identified as the strongest marker of amyloid plaque presence.Plasma P-tau217 was shown to best track increasing amyloid plaque burden once positivity was established.Plasma-only models approached CSF-based models (R² = 0.73 and 8% mean absolute error), supporting their potential as inexpensive, less invasive alternatives.

Prediction of quantitative Aβ-PET reached R² = 0.79 and 7% mean error using combined CSF and plasma biomarkers.

CSF Aβ42/Aβ40 was identified as the strongest marker of amyloid plaque presence.

Plasma P-tau217 was shown to best track increasing amyloid plaque burden once positivity was established.

Plasma-only models approached CSF-based models (R² = 0.73 and 8% mean absolute error), supporting their potential as inexpensive, less invasive alternatives.

Machine-learning models using CSF and plasma biomarkers accurately predicted quantitative Aβ-PET burden. Distinct contributions of soluble Aβ42/Aβ40 and plasma P-tau217 reveal their role in different stages of Aβ-pathology.