A low-variance subspace underlies individual differences in resting state fMRI

People differ remarkably from one another, yet isolating individual differences in their brain activity remains challenging. Non-invasive whole-brain recordings of human brain activity, such as those from resting state fMRI (rs-fMRI), are complex and noisy, making it difficult to isolate stable dimensions of individual differences. Ideally, we want to find a few core dimensions that vary across people but have high test-retest reliability, giving the same value each time they are measured in the same person. However, it is still unknown whether any such reliable dimensions exist, and if they do, what could drive this reliability. Here, we show that there is a low-dimensional linear subspace of highly-reliable rs-fMRI activity. These dimensions form personal fingerprints, allowing participants to be identified with high accuracy despite fingerprints explaining only a fraction of the total variance. Many of these dimensions inherit their reliability from a single morphological, demographic, or behavioral property, and most dimensions can be predicted from the anatomical layout of cortical regions. These dimensions were identified using reliability component analysis (RCA), a new dimensionality reduction technique similar to principal component analysis (PCA) but which maximizes reliability instead of explained variance. Together, our findings suggest that stable individual signatures can be isolated from rs-fMRI. These signatures reflect persistent anatomical and physiological differences, and provide a principled low-dimensional basis for biomarker discovery.