Reproduce “Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ce Ju, Antoine Collas, Florent Bouchard, Bertrand Thirion
A six-dataset benchmark tests whether SPD-based resting-state fMRI connectome models for age prediction generalize to entirely unseen cohorts. Leave-one-dataset-out evaluation reveals a much larger performance drop than pooled cross-validation, with no single method consistently outperforming simpler correlation-vector and tangent-space baselines; SPDNet performs well in several settings but does not eliminate cohort-shift effects.
Key insights
- Pooled cross-validation is substantially more optimistic than holding out an entire dataset; external performance is strongly affected by age-range overlap, clinical composition, and acquisition differences, not only by model choice.
- SPDNet often achieves the best within-dataset or selected external results, but its advantage is inconsistent and does not establish universal superiority over CorrVec or tangent-space ridge regression.
- Harmonization helps the linear SPD baselines more consistently than SPDNet, while increasing network depth or width does not reliably improve generalization under dataset shift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.