Psychiatry sorts people into diagnostic boxes. Brain data keeps showing that the underlying biology varies along continuous, shared axes such as general psychopathology, mood, fear, or impulsivity. The field has learned to find those axes with linear methods that pair brain measures with symptom measures. What it cannot yet do reliably is say which brain circuits carry a given axis in a way that holds up in a new sample. In the largest multimodal study in this area, the symptom dimensions replicated but the brain-side weights did not generalize to unseen data. We propose to assign that credit with the Shapley value, the unique rule from cooperative game theory that splits a joint outcome fairly among cooperating parts. Here the parts are large-scale brain networks, and the outcome is a person's predicted position on a continuous symptom dimension.
The key design choice is how a network is "removed" when we measure what it adds. We will not zero it out or replace it with a population average, because both produce brain-connectivity patterns that no real person has. Instead, we swap in values that a typical, low-symptom brain would show given the rest of that person's connectivity. The claim is that this kind of credit, computed over a handful of canonical networks, replicates across independent cohorts where linear loadings fail. It also measures something linear methods cannot: whether two circuits matter only in combination.
Core Claim
Our hypothesis concerns a model that predicts a continuous, transdiagnostic symptom score from functional connectivity. If its prediction is decomposed into Shapley credits over 7 or 17 brain networks, and absent networks are filled with values drawn from a healthy-reference distribution conditioned on the networks still present, then two things should follow. First, the network-level credit profile for each dimension will replicate between discovery and replication samples better than linear imaging loadings. Second, pairwise Shapley interaction terms will reveal synergies that additive decompositions miss. One candidate synergy is that loss of separation between the default mode network and executive networks raises mood or psychosis scores mainly when somatomotor connectivity is also abnormal.
Either outcome is informative. If the interaction terms are negligible and simple fill-in rules replicate just as well, then the field's linear tools were adequate and the replication problem lies elsewhere. The experiments are designed so that this result can appear.
Where This Sits
Work that motivates the proposal. Three lines of work established that the target is dimensional and identified the gap.
- Linked dimensions of psychopathology and connectivity in functional brain networks used sparse canonical correlation on 3,410 connectivity edges and 111 symptoms in the Philadelphia Neurodevelopmental Cohort. It found four dimensions with a shared default-mode/executive segregation loss.
- Somatosensory-Motor Dysconnectivity Spans Multiple Transdiagnostic Dimensions of Psychopathology used partial least squares on the UCLA Consortium for Neuropsychiatric Phenomics cohort. It placed somatomotor dysconnectivity under all three of its components.
- Multimodal neural correlates of childhood psychopathology applied partial least squares to the Adolescent Brain Cognitive Development (ABCD) study. It reported the decisive weakness: symptom dimensions replicate, imaging loadings do not.
All three are linear and additive by construction. That makes the two reported substrates, default-mode/executive segregation and somatomotor dysconnectivity, natural candidates for an untested interaction.
Closest existing work. Characterizing dynamic functional connectivity subnetwork contributions in narrative classification with Shapley values treats Yeo 7- and 17-network subnetworks as Shapley players to explain a trained classifier. That classifier separates audio from video and restaurant from airport narratives in 31 healthy participants. We differ in four ways. Our target is a continuous clinical dimension, not a healthy binary task. We test whether credits replicate across cohorts instead of reporting them once. We compute interaction terms. And we make the fill-in rule for absent networks an explicit, ablated variable. The file does not describe how that paper represents an absent subnetwork, so we cannot claim it uses out-of-distribution masking, only that the choice is not examined there.
Work that supplies the method and constrains the design. Robust Root Cause Diagnosis using In-Distribution Interventions shows a failure in anomalous systems. Explaining a failure by counterfactual reasoning means first estimating the hidden noise that produced the anomaly, and that estimate becomes unreliable precisely when the observed state is far from normal operation. The paper fixes this by setting a candidate cause to an in-distribution value and checking whether the downstream failure resolves, then splitting blame across multiple causes with Shapley values. We borrow that idea and treat a high-symptom individual's connectome as the "anomaly."
We do not borrow its assumption of a known directed acyclic causal graph, which brain networks violate. Our interventions are therefore on the predictive model, not on a causal model of the brain, and the resulting credits should be read as model attributions that are robust to off-manifold inputs, not as causal lesion effects.
Work treated as a secondary lens. Evaluate then Cooperate: Shapley-based View Cooperation Enhancement for Multi-view Clustering observed that some inputs receive negative Shapley values, meaning they actively hurt the output. We will check whether any brain network has reliably negative credit for a dimension. We are not adopting that paper's practice of steering training with Shapley values; the file's own evidence for that comes from image and text clustering, and adding it would confound the attribution question.
Research Program
Targets. The symptom side uses dimensions known to replicate. In ABCD, these are general psychopathology and the other components derived from the 119 Child Behavior Checklist items. In the Philadelphia cohort, they are the Mood, Fear, and Externalizing dimensions; we drop Psychosis because it did not replicate. Dimension scores are fixed from symptoms alone in the discovery split and then projected onto the replication split, so the brain side never shapes the target.
Inputs and models. The brain side is resting-state connectivity on a 400-region cortical parcellation plus subcortex. It is built with rank correlation and high-amplitude signal retention, following Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction and its DataCentricBrainGraphs code. We train two regressors so that conclusions are not tied to one architecture:
- a kernel ridge model, a simple nonlinear regressor;
- the graph transformer from Biologically Plausible Brain Graph Transformer, with its classification head replaced by a regression head, using the BioBGT code. Its quadratic cost in the number of regions is manageable at 419 nodes.
Players and the fill-in rule. Players are the Yeo 7 networks, with the 17-network version as a finer replication. A coalition $S$ is a set of networks whose connectivity is kept as observed. For every edge touching a network outside $S$, we draw values from a healthy-reference model of connectivity, conditioned on the edges inside $S$. The reference model is a low-rank Gaussian fitted to low-symptom participants in the discovery split. The value of the coalition is
$$v_x(S) = \mathbb{E}_{\tilde{x} \sim p_{\text{ref}}(\cdot \mid x_S)}\big[f(x_S, \tilde{x}_{\bar S})\big],$$
where $f$ is the trained regressor, $x_S$ are the person's observed edges belonging to coalition $S$, $\tilde{x}_{\bar S}$ are the filled-in remaining edges, and $p_{\text{ref}}$ is the conditional healthy-reference distribution.
With 7 players there are only 128 coalitions, so the Shapley values are exact. For 17 players we use permutation sampling and report sampling error. We also compute the pairwise Shapley interaction index for each pair of networks, which measures how much two networks contribute jointly beyond their separate contributions.
Large cohorts carry a healthy-volunteer bias, as Breaking the Norm: Population-Scale Normative Modeling of Brain Structure in Depression and Anxiety acknowledges. We therefore test whether results move when the low-symptom threshold that defines the reference group changes.
Analysis. Credit profiles are averaged within symptom-severity bins. The question is whether a dimension's circuit profile changes shape along the continuum or only scales up. A change in shape would be a dimensional analogue of what categorical diagnosis tries, and fails, to capture.
Synthetic check. We generate connectomes with a planted additive effect and a planted two-network synergy. This verifies that the in-distribution credits recover both, and that marginal masking produces measurable spurious credit when inputs are correlated.
Baselines
Naive reference: predicting the discovery-sample mean. This fixes the floor for predictive accuracy. Any attribution analysis of a model that does not beat the floor on the replication split is abandoned.
Strongest existing method for brain-side interpretation: linear decomposition with bootstrapped loadings. This means sparse canonical correlation as in Linked dimensions of psychopathology and connectivity in functional brain networks and partial least squares as in Multimodal neural correlates of childhood psychopathology. Both are run on identical connectivity features, identical splits, and an equal hyperparameter search. Network-level loadings are obtained by summing absolute edge loadings within each network.
The reported reference points from the file are as follows. In the Philadelphia cohort, sparse canonical correlation reached brain–behavior correlations of r = 0.68 to 0.71, with clinical-loading replication of r = 0.59 to 0.85 for the three replicating dimensions. For replication of ABCD imaging loadings there is no reported result found on Lacuna; the paper reports only that generalization was limited. We will therefore recompute that number ourselves. Beating this baseline means a higher discovery-to-replication correlation of network credit profiles than of network loadings, with a bootstrap interval on the difference that excludes zero, in both cohorts.
Strongest nonlinear predictor: BioBGT. Its reported results are classification only: 74.00% accuracy on the Autism Brain Imaging Data Exchange and 71.06% accuracy with 74.63% F1 on ADHD-200. No reported result found on Lacuna for BioBGT regression on continuous dimensions. We report its out-of-sample correlation and R² alongside kernel ridge under a matched tuning budget.
Ablations that isolate the key ingredient. The same regressors and the same 128 coalitions are evaluated under three other fill-in rules: zeroing, unconditional population means, and unconditional draws from the reference distribution. The fill-in rule is the only thing that changes. If the in-distribution credits do not replicate better than these alternatives, the proposal's central ingredient fails.
Why not just…?
Why not just keep partial least squares and bootstrap the loadings? Bootstrapping stabilizes loadings within a sample, but the documented failure is across samples, and a linear additive model cannot represent a synergy at all. This alternative is a genuine threat if the synthetic and real interaction terms turn out to be near zero. In that case a well-regularized linear model may replicate just as well, and the head-to-head replication comparison in Baselines is designed to show it.
Why not just retrain on each subset of networks, or drop one network at a time? Retraining a model on every subset of 7 networks takes 128 fits, which is feasible. It never feeds the model unrealistic inputs, so it is the cleanest competitor for population-level questions, and we include it. However, it answers a different question: which networks can predict a dimension, not which networks drive this person's score. Leave-one-out also misses synergy because it measures only one context per network. We will report where retraining credit and per-person in-distribution credit agree and where they diverge. If they agree everywhere, the simpler method suffices at the group level.
Why not just run standard Shapley explanations with marginal masking, or read attention weights? Off-the-shelf Shapley tools commonly fill absent features with background samples drawn independently of the rest of the input. This creates connectomes no subject has. Robust Root Cause Diagnosis using In-Distribution Interventions shows that this kind of off-manifold evaluation distorts blame in anomalous states. Attention weights measure where the model looks, not what each part adds to the output. Our fill-in ablation tests the first argument directly. For the second, we report rank agreement between attention mass per network and Shapley credit.
Proposed Experiments and Decisive Checks
The primary data are ABCD (3,504 discovery, 1,747 replication) and the Philadelphia cohort (663 discovery, 336 replication), with the published splits respected. The UCLA cohort (224 participants spanning controls, bipolar disorder, ADHD, and schizophrenia) is too small for training. It serves as a transfer test: models fitted on ABCD are applied to it unchanged, and we check whether general-psychopathology credit concentrates in the same networks across diagnostic groups, a direct test of the transdiagnostic claim.
There are four decisive checks.
- Replication. Do in-distribution credit profiles replicate better than linear loadings, and better than masking-based credits?
- Synergy. Does the interaction between the default-mode/executive pair and the somatomotor network significantly exceed a permutation null on the replication split?
- Agreement across models and parcellations. Are credits stable across the two regressors and across the 7- and 17-network schemes? If not, the credits mostly describe the model, not the brain.
- Negative credit. Does any network receive reliably negative credit, and does it persist after retraining?
If check 1 fails, we conclude that the off-manifold argument does not matter at network granularity. If check 2 fails, we conclude that the linear methods were missing nothing important at this scale. Both are publishable outcomes that would redirect the field's attention to sample size or measurement noise as the source of non-replication.
References
- Linked dimensions of psychopathology and connectivity in functional brain networks
- Somatosensory-Motor Dysconnectivity Spans Multiple Transdiagnostic Dimensions of Psychopathology
- Multimodal neural correlates of childhood psychopathology
- Characterizing dynamic functional connectivity subnetwork contributions in narrative classification with Shapley values
- Robust Root Cause Diagnosis using In-Distribution Interventions
- Evaluate then Cooperate: Shapley-based View Cooperation Enhancement for Multi-view Clustering
- Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction
- Biologically Plausible Brain Graph Transformer
- Breaking the Norm: Population-Scale Normative Modeling of Brain Structure in Depression and Anxiety