Few-shot prediction of amyloid β accumulation from mainly unpaired data on biomarker candidates.

The pair-wise observation of the input and target values obtained from the same sample is mandatory in any prediction problem. In the biomarker discovery of Alzheimer's disease (AD), however, obtaining such paired data is laborious and often avoided. Accumulation of amyloid-beta (Aβ) in the brain precedes neurodegeneration in AD, and the quantitative accumulation level may reflect disease progression in the very early phase. Nevertheless, the direct observation of Aβ is rarely paired with the observation of other biomarker candidates. To this end, we established a method that quantitatively predicts Aβ accumulation from biomarker candidates by integrating the mostly unpaired observations via a few-shot learning approach. When applied to 5xFAD mouse behavioral data, the proposed method predicted the accumulation level that conformed to the observed amount of Aβ in the samples with paired data. The results suggest that the proposed model can contribute to discovering Aβ predictability-based biomarkers.