Predicting regional tau accumulation with machine learning-based tau-PET and advanced radiomics

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Alzheimer's disease (AD) is characterized by tau tangles, neurodegeneration, and cognitive decline.
• Predicting individual tau accumulation is crucial for AD understanding and clinical trials.
• Current tools lack robustness for predicting personalized tau trajectories.

Purpose of the Study:

• To assess if flortaucipir-PET imaging biomarkers, combined with clinical and genomic data, can predict future tau accumulation.
• To develop a machine learning model for forecasting tau progression in AD patients.

Main Methods:

• Quantified participant data (N=276) including clinical, genetic (APOE-ε4), and flortaucipir-PET imaging measures.
• Extracted regional and radiomic texture features from flortaucipir-PET scans.
• Trained an AdaBoost machine learning algorithm to create a prognostic index for stratifying patients and predicting tau accumulation rates.

Main Results:

• The model achieved high accuracy in classifying slow vs. fast tau progressors (AUC 0.82-0.86).
• Successfully predicted annualized tau accumulation rates (Pearson's r = 0.69-0.73).
• An adaptive model using intermediate timepoints improved prediction accuracy (r = 0.74-0.87).

Conclusions:

• A robust machine learning approach can predict future tau accumulation in Alzheimer's disease.
• This prognostic index can enhance patient enrollment, stratification, and efficacy assessment in clinical trials.
• Multimodal features improve tau prediction accuracy in global and lobar brain regions.