Brain-imaging studies keep reporting that a disease affects one hemisphere more than the other: Alzheimer's atrophy that favours the left hippocampus, loss of left–right "homotopic" connections, right-lateralized deficits in depression. When the evidence is a connectome, a graph whose nodes are brain regions and whose edges are white-matter or functional connections, a simple problem goes unaddressed. Almost every network measure (clustering, efficiency, the fraction of edges that cross the midline) is partly determined by how many connections each region has. A hemisphere that has lost edges will look "less clustered" or "less integrated" even if the remaining wiring is arranged exactly as in a healthy brain. With cohorts of 20 to 50 patients, this degree-driven artefact can easily be mistaken for lateralized pathology.
We propose to separate the two effects. For each subject, the observed left–right difference in a network measure is compared against a cloud of randomized versions of that same subject's brain graph. The randomization is degree-preserving edge swapping, which rewires pairs of edges so that every region keeps exactly its number of connections. Each region's within-hemisphere and across-midline connection counts are preserved separately. What remains after this comparison is asymmetry the degree sequence cannot explain. To show that the test is calibrated and has power at small sample sizes, we build synthetic two-hemisphere brains from recursive network constructions. In these graphs the two hemispheres have identical degree sequences but different higher-order wiring, and the size of that difference is known analytically. They therefore supply a ground truth that real data cannot.
Core Claim
The claim is that hemisphere-stratified, degree-preserving null ensembles turn a raw asymmetry index into a per-subject standardized score that keeps its false-positive rate at small $n$. The score also replicates across small subsamples more reliably than raw indices or degree-regressed indices. The claim is checked in two settings: recursive synthetic graphs, where the true answer is known, and ADNI-2 and OASIS3 structural connectomes, where we test whether the score adds information beyond morphometry.
For a graph measure $f$, such as mean clustering or global efficiency computed on one hemisphere's subgraph, define the raw asymmetry $A(G) = f(G_L) - f(G_R)$. Here $G_L$ and $G_R$ are the left and right intra-hemispheric subgraphs. The degree-residualized score for subject $s$ is
$$z_s = \frac{A(G_s) - \mathbb{E}_{\tilde G \sim \mathcal{N}(G_s)}[A(\tilde G)]}{\mathrm{SD}_{\tilde G \sim \mathcal{N}(G_s)}[A(\tilde G)]},$$
where $\mathcal{N}(G_s)$ is the ensemble of graphs reachable from $G_s$ by stratified swaps. These are swaps that exchange two left-intra edges, two right-intra edges, or two inter-hemispheric edges, never edges from different classes. A raw left–right difference in clustering can arise purely because the left hemisphere has a more heterogeneous degree sequence. Under the null that average difference is absorbed into the expectation term, so $z_s$ only moves when wiring departs from what those degrees imply.
Prior Work and Where This Sits
Three lines of evidence motivate the proposal by showing that lateralized network signals exist but are measured without degree control.
Learning Brain Connectivity of Alzheimer's Disease from Neuroimaging Data uses sparse inverse covariance on FDG-PET in 42 AD patients and finds disrupted homotopic connectivity ($p = 0.009$ AD vs. controls), but tests it with bootstrap resampling rather than a degree-matched null. Discriminative Network Models of Schizophrenia reports inter-hemispheric degree deficits in 22 subjects using thresholded correlations, with no randomized reference graph. Diagnosis and prognosis of Alzheimer's disease using brain morphometry and white matter connectomes builds tractography connectomes on NHIS-IH and ADNI-2 but computes no hemispheric quantities at all.
On the morphometric side, Quantifying Hippocampal Shape Asymmetry in Alzheimer's Disease Using Optimal Shape Correspondences shows that a coarse asymmetry summary can be null ($p = 0.80$ for volume asymmetry on OASIS3) while localized measures are strongly significant. We take this as a structural lesson: the summary statistic matters. Here the analogous coarse summary is a raw graph-metric difference that mixes degree with wiring. A null model of the mouse whole-neocortex micro-connectome builds a careful statistical null for a whole cortex but explicitly assumes hemispheric symmetry, which is precisely the assumption our nulls are designed not to make.
The method comes from the edge-swapping literature. The 2-switch is formalized in Oversquashing in GNNs through the lens of information contraction and graph expansion: delete edges $(A,B),(C,D)$ and add $(A,D),(C,B)$. That paper uses it to raise spectral gap in graph neural networks via the G-RLEF heuristic, while we use it to sample a null rather than to optimize. Relations between average distance, heterogeneity and network synchronizability uses degree-preserving rewiring to decouple path length from clustering, which is the same logic of isolating one property from degree. We add the hemisphere stratification, which none of these works need.
One paper constrains the interpretation rather than supplying a method. Different thresholds of bond percolation in scale-free networks with identical degree sequence proves that graphs with an identical degree sequence can have percolation thresholds anywhere from about 0.618 down to 0, depending on a recursion parameter $q$. This cuts two ways. It means a residual $z_s$ is "asymmetry beyond degree," not "asymmetry beyond everything," so we never interpret it as pathology-specific without further controls. It also supplies exactly the synthetic family we need: two hemispheres with matched degrees and a closed-form, tunable difference in global structure.
Recursive multifractal models of the brain exist. Controlling the Multifractal Generating Measures of Complex Networks fits a single generating measure to ADNI fMRI graphs and reports 94% CN vs. late-MCI accuracy, but it treats the brain as one undivided system. We use recursion as a calibration device, not as a fitted model of patients.
Research Program
The first component is the stratified null sampler. We implement a Markov chain over 2-switches restricted to three edge classes: left-intra, right-intra, and the bipartite inter-hemispheric subgraph. This preserves every node's intra-degree and inter-degree, and hence the total count of midline-crossing edges.
Because restricted swap chains can mix slowly or fail to reach parts of the constrained space, we monitor mixing directly. Diagnostics are the autocorrelation of $A(\tilde G)$ along the chain and agreement between independent chains started from the observed graph. We also report results at 10 and 100 swaps per edge.
Structural connectomes are weighted. We therefore run the primary analysis on binarized graphs at three fixed densities, and run a weighted variant in which weights travel with their edges. The weighted variant preserves strength only approximately, and we report it as secondary.
The second component is a recursive ground-truth benchmark. Each synthetic "subject" consists of two hemispheres generated by the edge-replacement rule of Different thresholds of bond percolation in scale-free networks with identical degree sequence. Each hemisphere is grown to the generation that gives roughly 40–45 nodes, matching the 85–87-region parcellations used in Multi-Head Graph Convolutional Network for Structural Connectome Classification and Graph Contrastive Learning for Connectome Classification. The two hemispheres are joined by homotopic and a few random inter-hemispheric links.
Rigid determinism is a known weakness of recursive graphs (Controlling the Multifractal Generating Measures of Complex Networks). Each subject is therefore a noisy instance: a fraction $\epsilon$ of edges is perturbed by unstratified swaps, so subjects vary while degrees stay fixed.
The benchmark has two arms. In the null-true arm, both hemispheres share the same $q$, but the left hemisphere is given a different degree sequence by pruning hub edges. Any detected structural asymmetry is then a false positive driven by degree. In the alternative-true arm, the degree sequences are identical and only $q_L \neq q_R$, so the true asymmetry is beyond-degree and its size is ordered by the closed-form threshold $\lambda_c(q)$. At finite size, we treat the analytical $\lambda_c$ as giving the ordering of effect sizes and estimate the finite-size values numerically. Groups of $n \in \{10, 20, 40, 80\}$ subjects per arm let us trace false-positive rate and power as functions of $n$ and $|q_L - q_R|$.
The third component is the clinical application. Using the tractography pipeline described by Diagnosis and prognosis of Alzheimer's disease using brain morphometry and white matter connectomes, we build ADNI-2 and OASIS3 connectomes and compute $z_s$ for clustering, efficiency, and homotopic-edge retention. We then ask two questions. First, do group differences in $z_s$ replicate in small, disjoint subsamples? Second, do $z_s$ features add discriminative information beyond morphometry? HCP's 1,048 healthy connectomes serve as a reference for normal asymmetry and as a large pool for subsampling.
Baselines
The naive reference point is the raw asymmetry index $(f_L - f_R)/(f_L + f_R)$ computed on each subject's graph and compared between groups with a Mann-Whitney test. This is the network analogue of the volumetric index that failed in Quantifying Hippocampal Shape Asymmetry in Alzheimer's Disease Using Optimal Shape Correspondences. It is run on exactly the same graphs and densities as our score. Beating it means a lower false-positive rate in the null-true arm and higher subsample replication on HCP and ADNI.
The degree-regression baseline computes the raw asymmetry and then regresses out each subject's left–right differences in mean degree and degree variance. It is the strongest cheap competitor because it targets the same confound with no sampling. It receives the same covariates our null implicitly conditions on, and beating it means controlling false positives in the null-true arm where degree enters nonlinearly.
The unstratified null is the conventional degree-preserving rewiring that swaps any two edges regardless of hemisphere class. It is the ablation that isolates our key ingredient, since it preserves total degree but lets edges migrate across the midline. It runs with the same chain length. The stratified version must show better calibration on homotopic and inter-hemispheric statistics specifically.
For classification, the strongest existing method with reported numbers on our benchmark is the morphometry-plus-connectome model of Diagnosis and prognosis of Alzheimer's disease using brain morphometry and white matter connectomes. On ADNI-2 it reports 96% AD vs. HC, 70% MCI vs. HC, and 75% AD vs. MCI accuracy. We re-run its nested cross-validation with regularized logistic regression and identical folds. We then add only the $z_s$ features, giving the same tuning budget to both versions. MCI vs. HC is the contrast with room to move. A gain over our own re-run of the 70% figure, confirmed across repeated outer folds, would count. A GNN reference uses the multi-head graph convolutional model of Multi-Head Graph Convolutional Network for Structural Connectome Classification, which reports 90.6% on PREVENT-AD sex classification. It is retrained on the ADNI-2 diagnostic task with the same splits, since no reported ADNI-2 diagnostic number for it was found on Lacuna.
Why not just…?
Why not just regress node degree out of the raw asymmetry index? This is the real threat, and the proposal is built to settle it. Graph measures such as clustering depend on the whole degree sequence, including its heterogeneity and the degree correlations that rewiring studies like Cavity analysis on the robustness of random networks against targeted attacks: Influences of degree-degree correlations show matter. A linear covariate on mean or variance may leave residual confounding. The null-true arm measures this directly. If regression holds a nominal 5% false-positive rate there, and replicates as well as $z_s$ on HCP subsamples, we will report that the expensive null is unnecessary for that measure.
Why not just use standard degree-preserving rewiring without stratification? Unstratified swaps let left-intra and inter-hemispheric edges trade places. The null graphs then have a different number of midline-crossing edges than the subject, which biases any homotopic or inter-hemispheric statistic. This is the ablation above. If the two nulls agree on intra-hemispheric measures but differ on homotopic retention, that pattern confirms where stratification earns its cost.
Why not just train a regularized classifier or GNN with augmentation and skip inference? Classifiers answer "can groups be separated," not "is the lateralized difference real." The single-site-to-multi-site drop in Diagnosis and prognosis of Alzheimer's disease using brain morphometry and white matter connectomes, from 69% to 59% for prognosis once connectome features are added, shows that high accuracy on small cohorts does not certify a biological signal. Augmentations such as edge dropping in Graph Contrastive Learning for Connectome Classification also change degrees. We still include the classification comparison, so if a GNN matches morphometry-plus-$z_s$ that is reported.
Proposed Experiments and Decisive Checks
The first decisive check is calibration on the recursive benchmark. In the null-true arm at $n = 20$ per group, the stratified score should keep its false-positive rate near nominal, while the raw index exceeds it. In the alternative-true arm, power should increase monotonically with $|\lambda_c(q_L) - \lambda_c(q_R)|$. If stratified $z_s$ is no better calibrated than degree regression, the core claim fails.
The second check is small-sample replication on HCP. We draw repeated pairs of disjoint subsamples of 20 and 40 subjects, using a split such as sex where group effects can be estimated from the full 1,048. We measure how often each method recovers the full-sample sign and significance of each asymmetry measure.
The third check repeats the replication analysis on ADNI-2 and OASIS3 AD vs. control subsamples of matching size. We test whether a degree-residualized homotopic deficit survives where the raw deficit reported in Learning Brain Connectivity of Alzheimer's Disease from Neuroimaging Data might be degree-driven. An outcome where the raw deficit is significant but $z_s$ is null would be informative in its own right: it would say the reported disconnection is largely edge loss rather than rewiring.
The final check is the ADNI-2 MCI vs. HC classification comparison against the re-run baseline under identical nested cross-validation. Chain-mixing diagnostics and the weighted variant are reported for every real-data result.
References
- Learning Brain Connectivity of Alzheimer's Disease from Neuroimaging Data
- Discriminative Network Models of Schizophrenia
- Diagnosis and prognosis of Alzheimer's disease using brain morphometry and white matter connectomes
- Quantifying Hippocampal Shape Asymmetry in Alzheimer's Disease Using Optimal Shape Correspondences
- A null model of the mouse whole-neocortex micro-connectome
- Oversquashing in GNNs through the lens of information contraction and graph expansion
- Relations between average distance, heterogeneity and network synchronizability
- Different thresholds of bond percolation in scale-free networks with identical degree sequence
- Controlling the Multifractal Generating Measures of Complex Networks
- Multi-head Graph Convolutional Network for Structural Connectome Classification
- Graph Contrastive Learning for Connectome Classification
- Cavity analysis on the robustness of random networks against targeted attacks: Influences of degree-degree correlations