Conditioned connectomes outperform the standard correlation kernel for predicting cognition in an independent cohort
Year: 2026
Type: preprint
Abstract: Connectome-based cognition predictions are scored differently across papers. We re-scored a public Human Connectome Project benchmark (753 subjects, six fMRI modalities, 60 family-aware splits) as its authors do, with a public age band in place of exact age, and evaluated a kernel ridge pipeline that conditions each subject's connectome and filters the between-subject kernel's spectrum. With fixed settings it reached r = 0.7215, against 0.6779 for the benchmark paper's best model and 0.6798 for our re-run of its kernel (corrected t = 3.23). This margin is transductive: the filter uses all subjects' unlabelled features, helps only after conditioning, and gives no significant gain fitted on training subjects alone; without it, the difference was not significant. On 263 HCP subjects outside the benchmark, most used in development, a pre-registered comparison favoured the filtered pipeline, computed with them present, over the benchmark's kernel by 0.064 (95% interval 0.030 to 0.099). In an independent cohort never used in development (1,290 adults aged 36 to 89, resting state), the conditioned connectome, fitted on training subjects only, beat the benchmark's kernel as pre-registered (r = 0.5861 against 0.5257; corrected t = 5.47); the filter added nothing there. Against a frozen fMRI foundation model on identical splits, the unfiltered pipeline was ahead in 7 of 8 settings, significantly on resting state (corrected t = 3.88), held-out subjects and ABIDE autism diagnosis, and nominally in AOMIC. Of 18 pre-registered single-decision contrasts, 9 were met, 5 by counting repeats. Two scoring choices each reversed a method ordering.
Source: Research Square
Author: Chulmin Pyo
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7236787316
