AI Framework Predicts Alzheimer's Amyloid and Tau Burden Without PET Scans

Alzheimer's disease (AD) is biologically defined by progressive accumulation of amyloid beta (Aβ) plaques and neurofibrillary tau (τ) tangles, which develop years before symptom onset. PET imaging is the gold standard for detecting these pathologies, but it is expensive, inaccessible in routine care, and logistically demanding. The estimated $42.5 billion spent on AD drug development from 1995–2021—with a 95% failure rate—underscores how costly the current PET-dependent screening process is for clinical trials. A scalable, cost-effective alternative is urgently needed.

The research team built a two-stage transformer-based machine learning framework integrating seven independent cohorts: NACC, A4, OASIS3, AIBL, FHS, ADNI, and HABS. Input modalities included demographics, medical history, neuropsychological assessments, genetic markers (including APOE status), physical/neurological exam findings, and multi-sequence MRI-derived brain volumes. Features were harmonized to the Uniform Data Set 3 (UDS3) format. Stage one trained the model to jointly predict global Aβ positivity and meta-temporal tau (meta-τ) positivity. Stage two fine-tuned the model to predict regional tau PET positivity across anatomically distinct brain regions. The framework was explicitly designed to handle missing data using modality-specific embeddings, reflecting real-world clinical heterogeneity.

The model achieved an AUROC of 0.79 for amyloid classification and 0.84 for tau classification on external validation datasets (ADNI and HABS) and a held-out NACC subset. Predicted PET statuses were consistent across established biomarker profiles including CSF and plasma markers, and correlated with postmortem neuropathology grades. Shapley (SHAP) value analysis of MRI-derived regional brain volumes revealed clusters of important brain regions that aligned with known spatial patterns of tau deposition—particularly in medial and neocortical temporal regions. A graph network community detection algorithm confirmed that model-identified regional clusters closely matched communities derived from actual regional tau PET SUVr values.

The joint prediction of Aβ and τ is a key scientific advance, as these pathologies interact synergistically in AD progression. By outputting calibrated probabilities, the framework can map individuals onto established biological staging criteria (e.g., AT staging), enabling continuous rather than binary risk stratification. This has direct relevance to trials like TRAILBLAZER-ALZ 2, where tau burden predicted differential response to Donanemab—an amyloid-lowering therapy.

The authors acknowledge that plasma biomarkers such as p-tau 217 offer promising standalone performance for amyloid detection, but lack the spatial resolution and cross-population generalizability needed to replace PET for tau staging. Their multimodal framework, by contrast, captures regional tau topology non-invasively and tolerates incomplete data. The approach represents a practical pre-screening pathway that could reduce PET imaging costs, broaden clinical trial eligibility assessment, and facilitate earlier intervention in populations with limited access to advanced neuroimaging.