Robust Group PCA for Separable Noise: An Argument for Subject-Level PCA

Samuel Oriola, Calvin McCurdy, Bradley Thomas Baker, Vince Calhoun, Rogers F Silva

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Group PCA, ICA, Data Analysis, Functional MRI, Multivariate

TL;DR: We tested which group PCA approach better approximated the ground-truth source mean over subjects in a carefully designed synthetic dataset.

Abstract:

Functional magnetic resonance imaging (fMRI) captures whole-brain function with high spatial resolution and has driven discoveries in brain connectivity across age, gender, mental illnesses, and developmental stages.
As fMRI datasets grow in number, aggregation methods such as averaging or low-rank approximations are increasingly more likely to lose subject-specific details, potentially biasing group estimates and misrepresenting individuals, which in turn limits replication and reduces the translational utility of findings—--especially among minorities.
Group principal component analysis (PCA) is the \emph{de facto} tool for aggregating datasets, with tools like FSL and GIFT supporting various implementations that have shaped neuroimaging studies for decades.
Yet, the impact of subject variability on the group-level estimate, as well as the ensuing subject specificity, remain unquantified.
This study aims to identify computational strategies that improve the accuracy and robustness of group-level representations.
Three common group PCA implementations are considered: 1) simple concatenation, 2) concatenation with subject sum of squares normalization, and 3) concatenation with subject-level PCA whitening.
Simulated scenarios test these methods to identify optimal approaches for group dimensionality reduction while preserving the ground-truth group mean information.
Our results demonstrate that concatenation with subject-level PCA whitening achieved the best overall approximation of the ground-truth group mean, with performance differences largely driven by separable noise.

Track: 7. General Track

Registration Id: JKNLNG2RYRS

Submission Number: 342