Unlocking Tau PET Accessibility: a Machine Learning‐Based Prediction of Tau Pathology from Plasma, MRI and Clinical Variables

Abstract Background A key characteristic of Alzheimer’s disease (AD) is cerebral aggregation of tau. These aggregates can be quantified and localized with positron emission tomography (PET), which improves the diagnostic and prognostic work‐up of AD. However, tau‐PET is expensive and not available in clinical settings globally. With emerging disease‐modifying therapies for AD, comprehensive and accessible ways to estimate the load and distribution of tau pathology are becoming crucial. Here, we address this problem with machine learning (ML) models trained to predict tau‐PET outcomes from accessible features only (clinical variables, plasma biomarkers and MRI). Method The study included 1195 participants from the BioFINDER‐2 (BF2) cohort. BF2 was split into 80% train and 20% test sets. We created a rigorous ML pipeline that, with 10‐fold cross‐validation, evaluated different feature selection steps and estimators in a combined grid‐search and Bayesian optimization set‐up (Figure 1). The three variable blocks (clinical, plasma and MRI) were used as features (together and separately) to predict two tau‐PET characteristics: 1) tau load in the temporal cortex 2) hemispheric asymmetry of temporal tau load in tau positive participants (as an example of clinically relevant spatial information). External validation was performed in the BF2 test set, BioFINDER‐1 (BF1, n=147), ADNI (n=136), OASIS (n=46) and A4 (n=45). Note that ADNI, OASIS and A4 did not include plasma biomarkers. Result Feature combinations including plasma (particularly plasma p‐tau217) resulted in the best prediction of tau‐PET load (Figure 2). The best performing model was a combination of plasma and MRI variables in a CatBoost regressor, which achieved SUVR mean absolute error (MAE) = 0.155 and R‐squared = 0.700 on unseen data. For prediction of hemispheric asymmetry, models using MRI features performed best, and regions in the temporal lobe had the greatest contribution to the models (Figure 3). Top performance on unseen data was laterality index MAE = 6.74 and R‐squared = 0.423 in a SVR model including MRI and clinical variables. Conclusion Generalizable machine learning models reveal the potential in predicting tau PET from more accessible variables (particularly plasma biomarkers and MRI), which can unlock global accessibility to one of the most informative in vivo diagnostic techniques of AD.