Many brain-imaging questions of clinical interest involve a rare group, such as the few people with an unusual cognitive profile or a low-prevalence diagnosis. A common fix is SMOTE (Synthetic Minority Over-sampling Technique), which invents new minority-class examples by drawing a point on the straight line between a real minority example and one of its nearest minority neighbours. Neuroimaging studies apply it routinely. Brain data, however, is not a flat table. Each person contributes two different networks. One is a structural connectome: a sparse, non-negative map of white-matter fibre counts between brain regions, measured with diffusion MRI. The other is a functional connectome: a dense matrix of correlations between regional activity time series, measured with resting-state fMRI. Drawing a straight line between two people's matrices can produce correlation matrices that are not valid correlation matrices. If the two modalities are oversampled separately, it can also produce "chimeric" subjects whose anatomy comes from one pair of people and whose activity comes from another.
This note proposes a controlled study on the Human Connectome Project (HCP). It asks one question: does synthetic oversampling of multimodal connectomes ever beat simply reweighting the loss, and if so, which design choices matter? The key methodological move borrows from purposive sampling, the practice of selecting subjects deliberately by criteria rather than at random. We build the imbalance ourselves, by deliberately removing minority-class subjects from a near-balanced HCP label. The removed subjects then serve as real held-out examples against which synthetic ones can be scored. This turns an untestable question ("are these fake brains realistic?") into a measurable one.
Core Claim
We claim that synthetic oversampling helps only under two conditions. First, the synthetic samples must be generated in a space whose intrinsic dimension is small. Second, the structural and functional networks must be interpolated jointly, meaning the same neighbour and the same mixing weight are used for both. When those conditions hold, oversampling should beat reweighting for flexible classifiers at severe imbalance. When they fail, plain reweighted linear models should match or beat it. The prediction comes from theory. Concentration and excess risk bounds for imbalanced classification with synthetic oversampling shows that the gap between the SMOTE-generated distribution and the true minority distribution shrinks at order
$$n_1^{-1/2} + (k/n_1)^{1/d},$$
where $n_1$ is the number of real minority samples, $k$ is the number of neighbours, and $d$ is the feature dimension.
On a 3,160-edge-per-modality connectome taken at face value, this bound is useless. However, Rethinking Functional Brain Connectome Analysis: Do Graph Deep Learning Models Help? reports that functional connectivity profiles have an effective rank of only 7–38%. Our hypothesis is that the intrinsic dimension, not the number of edges, governs SMOTE's fidelity on connectomes, and that concatenating two partly independent modalities raises that intrinsic dimension unless they are coupled.
The same theory paper also constrains us. It finds parametric models such as logistic regression to be robust to oversampling parameters. We therefore expect little gain for linear models, and we treat that expectation as a hypothesis to confirm rather than a weakness to hide.
What Exists and What Does Not
On the multimodal side, there are three relevant methods. Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome couples structural and functional layers in a $2N$-node graph with learnable cross-layer weights, but it was evaluated on NCANDA rather than HCP and never addresses imbalance. Structural connectivity-informed dynamic estimation (STRiDE): Multimodal connectivity constrained ICA uses structural components as priors for functional decomposition, on a near-balanced schizophrenia cohort (162 vs. 149). Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens compresses morphology and function into compact tokens. All three supply candidate low-dimensional joint spaces that we can reuse. None of them generates synthetic subjects.
On the oversampling side, Selectively Augmented Decision Tree for Explainable Dementia Detection applies SMOTE with Edited Nearest Neighbours cleaning to a few scalar OASIS features (MMSE, normalised brain volume). It shows that oversampling can help a small tabular neuroimaging problem, but it says nothing about network-valued data. We found no paper in the file that applies synthetic oversampling to structural–functional connectome pairs.
Purposive sampling in the file is a manual, offline curation step. Examples are the 15 spectrum-selected stroke cases in Observer reliability of CT angiography in the assessment of acute ischaemic stroke: data from the Third International Stroke Trial and the deliberate over-representation of particular users in Qilin: A Multimodal Information Retrieval Dataset with APP-level User Sessions. We use the same logic as an explicit, repeatable algorithmic stage of evaluation: criteria-driven depletion to create imbalance, and criteria-driven selection of a hard extreme-phenotype subgroup.
Research Program
We will take the HCP S1200 subjects who have both diffusion and resting-state data. This should be roughly the 1,050 used by A Deep Network Model on Dynamic Functional Connectivity With Applications to Gender Classification and Intelligence Prediction. For each subject we build a structural connectome (log-transformed streamline counts) and a functional connectome (regularised correlation) on a shared parcellation.
All splits respect HCP family structure, because HCP contains many twins and siblings. The split must keep families intact, and SMOTE must run only inside training folds. Otherwise a synthetic sample interpolated between two siblings leaks family identity into evaluation.
There are two tasks. The first is sex classification, which is close to balanced in HCP. We purposively deplete the female class in training to 20%, 10%, and 5% of the training set, matching the imbalance ratios used in the concentration-bound paper. The depleted real minority subjects are set aside as a fidelity reference. The second task is a naturally imbalanced, purposively defined phenotype: identifying the bottom decile of fluid intelligence. Fluid intelligence is a weak target overall (the file reports r = 0.31 for the best model), so this task tests whether conclusions from the easy task survive a hard one.
We will compare synthetic generators along three factors.
The first factor is the geometry of interpolation:
- Euclidean: straight lines on vectorised edges.
- Tangent-space: functional matrices are mapped to a flat space with the matrix logarithm, interpolated there, and mapped back, so they remain valid positive-definite matrices. Structural matrices are interpolated on log counts, with the sparsity pattern preserved.
The second factor is dimensionality:
- interpolation in the ambient edge space;
- interpolation in a learned joint low-dimensional space, either principal components of concatenated tangent-space features or a structural basis in the style of STRiDE. Synthetic points are decoded back to matrices where needed.
The third factor is coupling:
- Paired: one neighbour and one mixing weight drive both modalities.
- Decoupled: neighbours and weights are chosen per modality.
The coupling factor is the ingredient this proposal adds. Decoupled sampling is exactly what a practitioner gets by running off-the-shelf SMOTE on each modality's feature block separately.
Every generator is scored in two ways. Fidelity is measured by a classifier two-sample test (can a classifier tell synthetic minority from held-out real minority?), by a kernel maximum mean discrepancy, and by validity checks (positive-definiteness, structural sparsity). Downstream utility is measured by training classifiers on the augmented data. We also estimate intrinsic dimension in each space, using the number of components explaining 90% of variance and a nearest-neighbour estimator. We then test whether the advantage of oversampling over reweighting tracks that estimate, as the $(k/n_1)^{1/d}$ term predicts.
Baselines
The naive reference point is training on the imbalanced data with no correction and reporting balanced accuracy. Any method that cannot beat it at 5% imbalance is uninteresting. We also report minority recall and area under the precision–recall curve.
The strongest competitor, and the one we expect to be hardest to beat, is class-reweighted logistic regression or ridge regression on concatenated tangent-space structural and functional edges. Rethinking Functional Brain Connectome Analysis: Do Graph Deep Learning Models Help? shows that classical linear and kernel models match or exceed graph deep learning on connectomes. The concentration-bound paper predicts that parametric models gain little from oversampling. Reweighting adds no synthetic data and has one hyperparameter. For reference on balanced HCP sex classification, A Deep Network Model on Dynamic Functional Connectivity With Applications to Gender Classification and Intelligence Prediction reports 93.05% accuracy for its CNN+LSTM and approximately 87% for an SVM. The closest multimodal figure is 84.9% accuracy with AUC 0.907 for Joint-GCN on NCANDA sex classification, a different dataset. We found no reported result on Lacuna for multimodal HCP sex classification under induced imbalance. We will therefore first reproduce balanced-data performance for every model, so that the imbalanced results are anchored.
Two further baselines represent flexible models: a kernel SVM and a two-layer MLP on the same features. A simple but tough-to-beat baseline for fMRI time-series classification shows that a minimal MLP is competitive with recurrent and transformer models across consortia including HCP. A Joint-GCN reimplementation is included as the strongest existing multimodal architecture.
Every classifier is run four ways, with identical family-aware folds, identical hyperparameter search budgets, and nested cross-validation:
- uncorrected;
- reweighted;
- oversampled with each generator;
- undersampled.
Beating the strong baseline means that some paired, low-dimensional generator gives a flexible classifier higher balanced accuracy than reweighted logistic regression achieves, with confidence intervals across repeated depletion draws that exclude zero. It does not count to beat only the uncorrected model.
The key ablation is paired versus decoupled interpolation, holding geometry and dimension fixed. A second ablation compares ambient with low-dimensional interpolation, holding coupling fixed.
Why not just…?
Why not just reweight the loss? This is the real threat. The concentration-bound paper explicitly suggests that parametric models gain little from oversampling. If reweighted logistic regression wins at every imbalance level, we will report that, and the paper's contribution becomes a measured negative result plus a fidelity diagnostic showing why. Reweighting cannot add minority diversity, though, and the theory says flexible non-parametric learners are the ones sensitive to how the minority distribution is estimated. The comparison is therefore run for both model families, and the outcome is decided by the interaction between model family and generator, not by averages.
Why not just use one modality with a linear model? Functional-only linear models are strong on HCP; the file's evidence is that even an SVM reaches roughly 87% on balanced sex classification. We include functional-only and structural-only versions of every pipeline. If adding structure gives no gain under imbalance, the coupling question becomes moot. In that case the result constrains multimodal claims such as Joint-GCN's to balanced settings.
Why not just run standard SMOTE on concatenated features? This is the decoupled, ambient, Euclidean corner of our design, so it is tested directly. The file gives two reasons to expect it to fail: the $1/d$ exponent in the concentration bound, and the loss of positive-definiteness under straight-line interpolation of correlation matrices. The fidelity tests will show whether those failures actually matter for downstream accuracy or are only cosmetic.
Why not just undersample the majority class? At 5% imbalance this throws away most of HCP. It is nonetheless cheap, and it is included under the same budget.
Decisive Checks
Three results would settle the proposal.
The first is whether a paired, low-dimensional, tangent-space generator gives a two-sample-test AUC near chance (synthetic indistinguishable from held-out real minority) while decoupled Euclidean SMOTE does not.
The second is whether the gain of oversampling over reweighting, for kernel SVM and MLP, grows as imbalance worsens and shrinks as the generator's intrinsic dimension rises. That is the dose–response pattern the concentration bound predicts.
The third is whether the conclusions reached on depleted sex labels transfer to the naturally imbalanced bottom-decile fluid-intelligence task.
If the first check holds but the second does not, realistic synthetic connectomes are achievable but not useful for classification. That would still be a clear, citable answer for the many neuroimaging studies that apply SMOTE without checking.
References
- Concentration and excess risk bounds for imbalanced classification with synthetic oversampling
- Rethinking functional brain connectome analysis: do graph deep learning models Help
- Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome
- Structural connectivity-informed dynamic estimation (STRiDE): Multimodal connectivity constrained ICA
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens
- Selectively Augmented Decision Tree for Explainable Dementia Detection
- Observer reliability of CT angiography in the assessment of acute ischaemic stroke: data from the Third International Stroke Trial
- Qilin: A Multimodal Information Retrieval Dataset with APP-level User Sessions
- A Deep Network Model on Dynamic Functional Connectivity With Applications to Gender Classification and Intelligence Prediction
- A simple but tough-to-beat baseline for fMRI time-series classification