Alpha-Z divergence unveils further distinct phenotypic traits of human brain connectivity fingerprint

Abstract

The accurate identification of individuals from functional connectomes (FCs) is critical 12 for advancing individualized assessments in neuro/psychiatric research. Traditional 13 methods, such as using Pearson's correlation, have limitations in capturing the com- 14 plex, non-Euclidean geometry of FC data, leading to suboptimal performance in identi- 15 fying individual differences. Recent developments have introduced geodesic distance 16 as a more robust metric; however, its performance is highly sensitive to regulariza- 17 tion choices of FCs, which vary by spatial scale and are task-condition specific.To 18 address these challenges, we propose a novel distance measure and divergence- 19 based distance metric, the Alpha-Z Bures-Wasserstein divergence, which provides a 20 more flexible and geometry-aware framework for FC comparison. Unlike prior meth- 21 ods, our approach does not require meticulous parameter tuning and maintains strong 22 identification performance across multiple task conditions and spatial resolutions. We 23 evaluate our approach against both traditional (e.g., Euclidean, Pearson) and state- 24 of-the-art manifold-based distances (e.g., affine-invariant, log-Euclidean, and Bures- 25 Wasserstein) and systematically investigate how varying regularization strengths affect 26 the performance of geodesic distance measures on the Human Connectome Project 27 dataset, spanning multiple cognitive tasks and parcellation resolutions from low to 28 high. Our results demonstrate that the proposed method significantly improves the 29 identification rates over traditional and existing geodesic distance measures, particu- 30 larly when optimized regularization is applied, and notably in high-dimensional settings 31 where matrix rank deficiencies degrade the performance of existing metrics.