Multi-dimensional Persistent Feature Analysis Identifies Connectivity Patterns of Resting-State Brain Networks in Alzheimer’s Disease

Jin Li,Chenyuan Bian,Haoran Luo,Dandan Chen,Luolong Cao,Hong Liang

DOI: https://doi.org/10.1088/1741-2552/abc7ef

IF: 5.043

2021-01-01

Journal of Neural Engineering

Abstract:Objective. The characterization of functional brain network is crucial to understanding the neural mechanisms associated with Alzheimer’s disease (AD) and mild cognitive impairment (MCI). Some studies have shown that graph theoretical analysis could reveal changes of the disease-related brain networks by thresholding edge weights. But the choice of threshold depends on ambiguous cognitive conditions, which leads to the lack of interpretability. Recently, persistent homology (PH) was proposed to record the persistence of topological features of networks across every possible thresholds, reporting a higher sensitivity than graph theoretical features in detecting network-level biomarkers of AD. However, most research on PH focused on zero-dimensional features (persistence of connected components) reflecting the intrinsic topology of the brain network, rather than one-dimensional features (persistence of cycles) with an interesting neurobiological communication pattern. Our aim is to explore the multi-dimensional persistent features of brain networks in the AD and MCI patients, and further to capture valuable brain connectivity patterns. Approach. We characterized the change rate of the connected component numbers across graph filtration using the functional derivative curves, and examined the persistence landscapes that vectorize the persistence of cycle structures. After that, the multi-dimensional persistent features were validated in disease identification using a K-nearest neighbor algorithm. Furthermore, a connectivity pattern mining framework was designed to capture the disease-specific brain structures. Main results. We found that the multi-dimensional persistent features can identify statistical group differences, quantify subject-level distances, and yield disease-specific connectivity patterns. Relatively high classification accuracies were received when compared with graph theoretical features. Significance. This work represents a conceptual bridge linking complex brain network analysis and computational topology. Our results can be beneficial for providing a complementary objective opinion to the clinical diagnosis of neurodegenerative diseases.