Rank-Indexed Persistent Laplacians for Connectomes That Travel Across Scanners

Brain-imaging models that score well inside one dataset often collapse when tested on a dataset from another hospital. In one benchmark of brain-age prediction from resting-state functional MRI, the explained variance goes negative under leave-one-dataset-out testing. The models learn how a particular scanner and site distort connectivity values rather than how brains differ. This note proposes a representation built to ignore one common class of those distortions. We describe each brain's connectivity graph by how its shape evolves as edges are added in order of strength. The ordering uses the rank of each connection's strength, never its raw value. At each step we record the spectrum of the persistent Laplacian, a matrix that summarizes which clusters and loops survive from one stage of this growth to a later one and how tightly they are held together. Many site effects rescale or bend connectivity values monotonically: global signal differences, scanner-dependent inflation of correlations, differing preprocessing. Any such effect leaves the rank order, and therefore this representation, exactly unchanged. The representation is also a small, fixed feature vector with no learned parameters, so a ridge model on top can be trained on cohorts of a few dozen subjects.

The contribution is a test as much as a method. We ask whether a multi-scale topological and spectral description of individual connectomes narrows the gap between within-dataset and external performance at small sample sizes, and why. We separate what comes from rank invariance, what comes from tracking structure across scales, and what comes from the non-zero spectrum rather than topology alone. Each ablation has a simpler competitor that could plausibly win, and the design is built so that such an outcome would be visible.

Core Claim

Build the filtration of a connectome on edge ranks and describe it with persistent Laplacian spectra. The prediction is that this reduces how well the acquisition site can be decoded from the features, and shrinks the drop from within-dataset to leave-one-dataset-out accuracy, compared with correlation-matrix representations at matched training size.

The mechanism is stated precisely. Let $W$ be a subject's region-by-region correlation matrix, and let each edge $e$ get filtration value $\rho(e) \in (0,1]$: its quantile among all edges, strongest first. Then for any strictly increasing function $g$ applied entrywise,

$$ \rho_{g(W)}(e) = \rho_W(e) \quad \text{for every edge } e, $$

so every complex $K_a$ in the filtration (all edges with $\rho \le a$, plus the triangles they close) is identical for $W$ and $g(W)$. Features are the smallest $k$ non-zero eigenvalues and the zero-eigenvalue count of $\Delta_q^{K_a,K_b}$, the $q$-th persistent Laplacian between densities $a<b$, for $q\in\{0,1\}$ over a fixed grid of density pairs. The zero count equals the persistent Betti number (clusters for $q=0$, loops for $q=1$), as proved in Persistent Laplacians: Properties, Algorithms and Implications. The non-zero eigenvalues measure how close a cluster is to splitting or a loop to closing, which The Persistent Laplacian for Data Science found to be the most discriminative part of the spectrum.

The claim is falsifiable in a specific way. The invariance covers monotone distortions only. Head motion and site-specific artifacts that reorder edges will leak through, and the experiments measure how much of real cross-site shift is of each kind.

Relation to Prior Work

Three papers supply the method. Persistent Laplacians: Properties, Algorithms and Implications defines the operator, proves stability under interleaving distance, and gives the Schur-complement algorithm we reuse. The Persistent Laplacian for Data Science provides the eigenvalue truncation-and-flattening pipeline and an open-source package. On MNIST it reports 82.1% accuracy against 62.5% for persistent homology and 76.8% for a static graph Laplacian. Graph Persistence goes Spectral (SpectRe) attaches Laplacian eigenvalues to persistence events and proves only local stability, with instability where several edges share a filtration value.

All three evaluate on standard single-dataset splits with thousands of samples (MNIST, QM7, ZINC, MUTAG, OGB-MOLHIV). None asks about external shift or cohorts under a hundred subjects, and none designs the filtration for invariance. Our rank filtration is also injective by construction, which removes the non-injective failure locus SpectRe identifies. The closest low-sample work, PL-MBO, uses persistent Laplacians for label propagation on point clouds under random within-dataset splits. TopLapGBT adds persistent Laplacian features to protein descriptors and reports a 1–2% gain over homology alone, again within one dataset. We borrow neither's learner; we borrow their evidence that the non-zero spectrum adds signal beyond topology.

Two lines of work motivate the proposal and also constrain it. Benchmarking External Generalization of SPD Matrix Learning documents the external-generalization cliff across six cohorts. It finds that the deep Riemannian network SPDNet degrades more than Tangent-Space Ridge, and that the representations it tests are all static correlation-matrix representations. It also shows that some failures come from population mismatch: holding out pediatric ABIDE for brain age gives MAE above 30 years. No representation can fix that kind of extrapolation, so our protocol must separate it from acquisition shift.

Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction shows that rank correlations and high-amplitude signal retention raise GCN accuracy on ABIDE from 60.00% to 65.86%. This is supporting evidence that rank-based construction matters, and also the source of our strongest "no topology needed" competitor. Brain-graph transformers with biological priors (BioBGT, Biologically Plausible Brain Graph Transformer; Com-BrainTF, Community-Aware Transformer) report higher ABIDE accuracies, but under pooled cross-validation rather than held-out sites. Their numbers are not comparable to ours and are not the bar we aim at.

Research Program

We build the rank-indexed persistent Laplacian featurizer for connectomes. We use the Harvard-Oxford atlas at about 110 regions as the primary parcellation, as in the federated ABIDE study, and Craddock 200 as a robustness check. Densities run from 1% to 30%, because the loop Laplacian's size equals the number of edges and dense complexes are where the cost cited by SpectRe bites. We compute $k=10$ smallest non-zero eigenvalues per density pair with the Schur-complement reduction and sparse eigensolvers.

The learner is deliberately plain: ridge regression for age, logistic regression for diagnosis. Any gain then belongs to the representation, and the method stays usable at $N\approx 25$. A secondary variant feeds the features to a small DeepSets encoder, as SpectRe does, to test whether learned pooling helps or re-introduces overfitting.

The analysis has three parts. First, site decodability: the chance-normalized balanced accuracy of a classifier predicting acquisition site from each representation, on age- and sex-matched subjects. Second, the within-to-external gap: within-dataset cross-validation score minus leave-one-dataset-out score. Third, learning curves over per-dataset training sizes of 25, 50, 100, and all subjects.

We also run a controlled injection study on one site. We apply synthetic monotone transforms (power laws, sigmoid compression), then non-monotone perturbations (motion-like distance-dependent inflation, additive noise). This checks where the invariance holds and where it breaks, before we interpret the real multi-site results.

Baselines

The naive reference is predicting the training-set mean age, or the majority diagnosis, for the held-out dataset. The SPD benchmark shows that complex models often fail to beat this under leave-one-dataset-out testing, so any method with negative $R^2$ is reported as failing outright.

The strongest existing method for external brain age is Tangent-Space Ridge on correlation matrices, which Benchmarking External Generalization of SPD Matrix Learning found as good as or better than SPDNet out of dataset. We run it with and without ComBat harmonization. Its reported leave-one-dataset-out behavior is negative $R^2$ and MAE above 30 years when ABIDE is held out. The file has no per-cohort numbers, so we re-run it on identical subjects, preprocessing, and folds. Beating it means lower held-out MAE and a smaller within-to-external gap on the age-matched protocol in at least four of six held-out cohorts, with paired bootstrap intervals over subjects.

For ABIDE diagnosis under leave-one-site-out, the reference is the inter-site accuracy of 67% reported by Deriving reproducible biomarkers from multi-site resting-state data. We re-implement its best pipeline (atlas plus tangent-space connectivity plus linear classifier) on our splits. Beating it means higher mean held-out-site AUROC with lower variance across sites.

Every method gets the same subjects, preprocessing, nested inner cross-validation for regularization strength, and the same hyperparameter budget of 50 configurations.

Four ablations isolate the ingredients. The first replaces ranks with raw correlation thresholds, testing whether rank invariance matters. The second replaces the persistent Laplacian with static Laplacian spectra computed independently at each density, testing whether tracking structure across scales matters. The third keeps only zero eigenvalues, which is persistent homology, testing whether the non-zero spectrum matters. The fourth feeds rank-transformed correlation matrices directly into Tangent-Space Ridge, testing whether topology matters at all once ranks are used.

Why not just…?

Why not just use persistent homology? It is cheaper and already multi-scale. Within-dataset evidence favors the spectrum: 82.1% versus 62.5% on MNIST in The Persistent Laplacian for Data Science, and a 1–2% gain in TopLapGBT. There is no external-shift evidence either way. It is possible that Betti curves, being coarser, are more site-invariant and the extra spectral detail is exactly where site noise lives. The zero-eigenvalue-only ablation settles this directly. If persistent homology matches our method externally, we report that and recommend it.

Why not just harmonize with ComBat and use Tangent-Space Ridge? It is the field's default and our strongest competitor. The SPD benchmark reports that ComBat helps only when target label distributions are known during harmonization. It also cannot be applied to a truly unseen site without that site's data. Our representation needs no target-site statistics. We still include ComBat on held-out sites (fit on unlabeled target data) as a stronger-than-fair comparison. If it closes the gap as well as ranks do, the topology is not earning its cost.

Why not just rank-transform the connectome, or threshold it at several densities, and skip the persistent Laplacian? This is the most serious threat. The data-centric benchmark already shows rank correlations help, and multi-density thresholding is standard in network neuroscience. If most site shift is monotone, rank-transforming alone may capture the whole benefit. The rank-transformed Tangent-Space Ridge baseline and the static multi-density Laplacian ablation exist to test this. The proposal's specific claim survives only if persistent spectra beat both under leave-one-dataset-out at small $N$. If they do not, the finding reduces to "use ranks," which we would report as the result.

Why not just augment the small cohorts? Signal-level augmentation multiplies the samples of one site, which does not teach invariance to another site. Multi-site fMRI Analysis Using Federated Learning found multi-site exposure beat single-site training with augmentation. the MEMD augmentation study found gains plateau and even hurt temporally sensitive models. We include time-series mixing within diagnostic group as an extra baseline on the smallest training sizes only.

Proposed Experiments

The primary experiment is leave-one-dataset-out brain-age regression across COBRE, ADNIDOD, Cam-CAN, ABIDE, OASIS-3, and ADNI, following the SPD benchmark protocol. We run it twice. The full protocol reproduces the published cliff. The age-overlap protocol restricts training and test subjects to shared age ranges, so acquisition shift is measured apart from extrapolation. Metrics are MAE, $R^2$, the within-to-external gap, and site decodability.

The secondary experiment is leave-one-site-out ASD classification on ABIDE's 17 sites, with AUROC and accuracy. The four largest sites (NYU, UM, USM, UCLA) are also run with training capped at 25 and 50 subjects per site, to probe the small-sample regime.

Three results would be decisive. First, in the injection study, our features must be exactly unchanged under monotone transforms while raw-threshold persistent Laplacian and tangent features shift. How fast our features degrade under non-monotone perturbation then bounds what the real-data gains can mean. Second, on the age-overlap protocol, site decodability must drop and the external gap must shrink relative to Tangent-Space Ridge at $N=50$ per dataset. Third, the static multi-density ablation and the rank-transformed ridge must both fall short of the full method. The first two results with a failed third would still yield a useful, smaller conclusion about rank filtrations. Failure of the second means topological spectra do not travel better than correlations, which the field currently assumes without testing. Runtime per subject is reported alongside, so the cost of the loop-dimension spectra is weighed against any gain.