Alzheimer's disease-related β-amyloid deposition prediction based on plasma biomarkers, ApoE4, and CDR with a machine...

Alzheimer's disease (AD) is an irreversible, incurable neurodegenerative disorder that imposes a substantial economic and social burden. As current therapies can only slow disease progression, early detection is of critical importance. Among existing diagnostic tools, amyloid positron emission tomography (PET) remains the most accurate non-invasive method for detecting β-amyloid deposition; however, its high cost significantly limits its applicability for early screening. To address this challenge, this study investigates the use of plasma biomarkers as a low-cost alternative for predicting cerebral β-amyloid burden.

The predictive performance of three machine learning models-Classification and Regression Trees (CART), Support Vector Machines (SVM), and Logistic Regression (LR)-was systematically compared under varying feature subsets and thresholding strategies. Data were obtained from 276 participants in the ADNI database. The qualitative status of cerebral amyloid deposition, defined using threshold-based classification of Centiloid values, was set as the prediction target. Plasma biomarkers and other AD-related risk factors were employed as model features.

Cross-domain data integration enabled effective prediction of amyloid deposition status. The SVM model, combined with feature subset selection (FSS) and a threshold value of 25, achieved the best performance, yielding an area under the curve (AUC) of 0.80.

By applying the Centiloid quantification framework for amyloid-PET and employing plasma biomarkers along with established AD-related risk factors, this study demonstrates the feasibility of machine learning models as a low-risk, cost-effective approach for predicting amyloid-PET imaging outcomes. These findings highlight the potential to support clinical decision-making and reduce reliance on costly imaging techniques for early AD detection.