Researchers built machine-learning models that classify tau status using only MRI and amyloid-PET features, according to a study published in Alzheimer's & Dementia titled "Classification of tau status with machine learning models in amyloid-positive cohorts." The models were trained and tested only in people who had already tested positive for amyloid, the first Alzheimer's-related brain change, and the goal was to see whether that group's tau burden, a marker tied to symptom severity, could be estimated without a separate tau PET scan.
Tau PET imaging is expensive, not universally available, and often excluded from insurance coverage, which has driven interest in substitutes drawn from scans patients may already have. The study is part of a broader push in Alzheimer's research toward this kind of substitution: other recent work has tried predicting tau load from plasma markers, MRI and clinical variables, and separately from deep learning applied to MRI alone, reflecting a field-wide effort to reduce reliance on a single costly biomarker.
The authors report their approach as a classification exercise, sorting participants into tau-positive or tau-negative categories, not a measurement of exact tau burden or a diagnosis. Because the models were built and evaluated within amyloid-positive cohorts, it is not established how they would perform in people who have not had an amyloid scan, in earlier or more diverse disease stages, or in routine clinical settings outside the research cohorts used to build them. Independent replication in separate cohorts, and comparison against tau PET performed on the same individuals, would be needed before any such tool could inform clinical decisions.