A Machine learning classification framework using fused fractal property feature vectors for Alzheimer's disease diagnosis.

Abstract

Alzheimer's disease (AD) profoundly affects brain tissue and network structures. Analyzing the topological properties of these networks helps to understand the progression of the disease. Most studies focus on single-scale brain networks, but few address multiscale brain networks. In this study, the renormalization group approach was applied to rescale the gray matter brain networks of AD patients and cognitively normal (CN) into three scales: the original, once-renormalized, and twice-renormalized networks. Based on the fractal property of these networks at different scales, a novel framework for classifying Alzheimer's disease using fractal and renormalization group was proposed. We integrated the fractal metrics across different scales to create fused feature vectors, which served as inputs for the classification framework aimed at diagnosing Alzheimer's disease. The experimental result indicates that the original and once-renormalized networks of both CN and AD exhibit the fractal property. The classification framework performed best when using the fused feature vector, including the average connection ratio of the original and once-renormalized networks. Using the fused feature vector of the average connection ratio, the One-Dimensional Convolution Neural Network model achieved an accuracy of 92.59% and an F1 score of 91.19%. This marks an improvement of approximately 10% in accuracy and 5% in F1 score compared to results using feature fusion of the average degree, average path length, and clustering coefficient.