GRFBrain: Graph-Structured Rectified Flows for EEG Dynamic Modeling

EEG patterns are not confined to individual electrodes: important events such as seizures involve changing synchrony and communication across channels. This paper proposes GRFBrain, a model for forecasting those evolving relationships while representing uncertainty rather than producing only one predicted future. The authors combine a deterministic history-based forecast with a stochastic flow that models the structured residual variation left over. In the paper’s graph-structured residual flow framework, the source of that stochastic process is shaped by recent functional connectivity rather than being ordinary isotropic noise.

What the paper is trying to model

A multichannel EEG recording can be represented as a graph. Channels are nodes, while edges describe functional associations between channel pairs. Both parts change over time: channel-level spectral activity may evolve, and the relationships among channels may strengthen, weaken, appear, or disappear.
Many predictive models map an observed history to one future state. That can be useful for learning the average or most likely trajectory, but it may hide meaningful alternatives. Near a transition, similar EEG histories might be followed by different futures. In such cases, deviations from the average could contain information relevant to seizure detection or abnormal-EEG classification.
GRFBrain separates these two aspects. A history-only graph forecaster first estimates a reference future: the mean historical node state is used for node features, while a frozen graph forecaster produces an edge reference. The model then defines the actual future as this reference plus a residual. Flow matching is used to learn how to transport samples from a conditional source distribution to the future residual distribution.
The transport coordinate, written as τ, runs from 0 to 1 and is not physical EEG time. It describes movement through a probability distribution. This distinction matters: the learned velocity field is a mechanism for distributional forecasting, not a claim that the brain evolves according to the flow’s artificial time coordinate.

How the approach works

Standard conditional flow matching often begins with an independent, identically distributed Gaussian source. Every coordinate is treated as unrelated at the start, leaving the learned velocity network to reconstruct channel and edge dependencies during transport.
GRFBrain instead builds a graph-conditioned Gaussian source from the recent EEG history. The authors average recent functional graphs, construct a normalized graph Laplacian, and use its spectrum to define a covariance matrix. High graph-frequency modes are suppressed, encouraging source perturbations to reflect smoother relationships over the observed graph. The covariance is trace-normalized so that the graph-structured source has the same overall energy as an IID Gaussian control. This is intended to isolate the effect of relational structure from the effect of simply adding more or less noise.
The source is applied separately to node and edge residuals. Node innovations are transformed using the graph-derived matrix. Edge innovations are formed by applying the same graph geometry to both endpoints of each channel pair, followed by an edge-energy normalization. The node and edge sources are conditionally independent, but they share the history-derived graph structure.
A joint node–edge Transformer then predicts the flow velocity. It maintains one token for each channel and one symmetric token for each unordered pair of channels, including pairs whose observed edge weight is zero. Edge tokens influence node updates through attention biases and incident-edge messages. Updated node representations then revise the edge tokens. In this way, the model supports both edge-to-node and node-to-edge interactions rather than fixing the graph or learning two unrelated flows.
During training, source and target residuals are independently paired given the EEG history, and a shared τ is sampled. The model learns to match the velocity from each interpolated state to the target displacement. For downstream prediction, the velocity network can be evaluated once at the history-derived reference state, corresponding to zero residual coordinates, without sampling or numerical integration.

What results stand out

The main reported evaluation focuses on seizure detection in the TUH EEG Seizure Corpus, or TUSZ, using 12-second and 60-second settings. In the 12-second setting, GRFBrain reports an AUROC of 0.877 ± 0.003 and an F1 score of 0.523 ± 0.014. The corresponding results for standard flow matching are 0.813 ± 0.002 AUROC and 0.372 ± 0.013 F1. DCRNN, a graph-based recurrent baseline, reports 0.825 ± 0.002 AUROC and 0.416 ± 0.009 F1.
For the 60-second setting, GRFBrain reports 0.831 ± 0.004 AUROC and 0.488 ± 0.032 F1, compared with 0.725 ± 0.006 AUROC and 0.311 ± 0.028 F1 for standard flow matching. These results support the authors’ claim that the combination of reference-centered residuals, graph-conditioned source structure, and joint node–edge transport can improve seizure-detection representations over the listed baselines.
The ablations also point to the importance of bidirectional coupling. At 12 seconds, the full model reaches an F1 of 0.523. Removing edge-to-node interactions lowers F1 to 0.488, while removing node-to-edge interactions lowers it to 0.462. Removing both cross-interaction routes produces a much larger drop, to 0.374. This pattern is consistent with the paper’s argument that channel activity and functional connectivity should be modeled together.
The broader motivation resembles the contrast described in UQ-LOB’s treatment of point forecasting and uncertainty: a single forecast can conceal uncertainty about plausible outcomes. GRFBrain addresses that issue specifically through conditional transport in EEG graph space, rather than by attaching a separate uncertainty module to a point predictor.

What to keep in mind

The paper does not claim that every gain comes solely from the graph-shaped source. Its experiments are designed to distinguish improvements due to deterministic forecasting, learned representations, source geometry, sampling, and node–edge coupling. This is important because GRFBrain changes several components at once: it centers the target on a forecast, replaces isotropic noise with a history-conditioned covariance, and uses a shared node–edge velocity network.
The graph itself is also treated cautiously. The authors describe it as an empirical conditioning structure, not as an anatomical or causal brain network. The model therefore uses observed functional associations to define a useful inductive bias, but the reported method does not establish that those associations represent direct neural causation.
Finally, flow matching’s artificial transport time should not be interpreted as a physiological time scale. The framework is designed to learn conditional distributions of future graph states. Its value lies in whether those learned representations help downstream tasks and whether controlled comparisons isolate the contribution of structured residual transport.