Classification of the Tau Staging Status with MR-Based and Amyloid PET-Based Imaging Features in Machine Learning

The emergence of anti-amyloid therapies, such as Lecanemab and Donanemab, has brought new lights to treat Alzheimer’s disease (AD). Advanced tau pathology may degrade the treatment efficacy. Tau-PET imaging, while effective for staging tau burden, is limited by costs and availability. This study investigates machine learning models using cognitive tests, MRI and amyloid-PET features to predict tau staging in amyloid-positive patients, offering a cost-effective surrogate for tau staging. Utilizing data from ADNI-3 (n=380) and OASIS-3 (n=120), we processed the brain structural MRI with FreeSurfer and applied feature selection to identify key brain regions, showing significant differences in those regions’ volume between tau-positive and tau-negative groups (p<0.001). Logistic regression, SVM, and random forest models were trained, with the SVM model achieving the highest performance: AUCs of 0.89 (ADNI for training and cross-validation) and 0.90 (OASIS as external validation), sensitivities of 75.7% and 80.0%, and specificities of 87.9% and 95.3%, respectively. Our model demonstrated competitive performance in classifying the tau status and may serve as a screening tool or surrogate tau staging biomarker. By focusing exclusively on amyloid-positive patients, the proposed machine learning model may assist the pre-treatment evaluation and planning for personalized AD therapies.