Band2Net: Frequency-to-Network Transport Distillation for fMRI-Privileged, EEG-Only Session-Condition Decoding

Nishi Kanta Paul, Md Shihabul Islam Shovo, Dwarikanath MahapatraarXiv 2026 · arXiv

TL;DR

Band2Net is a new deep learning framework that uses high-resolution fMRI data as a "privileged" teacher to train better EEG-only brain state decoders. By using optimal transport to map EEG frequency bands to canonical fMRI brain networks, the model achieves a significant performance boost in identifying psilocybin-induced states using only portable, low-cost EEG data.

What Problem Does This Paper Address?

Neuroscientists often face a trade-off between "where" and "when." Functional MRI (fMRI) provides excellent spatial resolution, allowing us to see large-scale brain networks in action, but it is expensive, bulky, and has poor temporal resolution. Electroencephalography (EEG) is the opposite: it’s portable, cheap, and captures millisecond-level rhythms, but it struggles to localize activity to specific brain regions or networks.

The authors want to track brain states (specifically the effects of psilocybin) in real-world settings where fMRI isn't feasible. The challenge is that EEG lacks an explicit notion of cortical networks. Current "fusion" models require both modalities at test time, which defeats the purpose of portability. Band2Net addresses this by using fMRI during training to "teach" an EEG model how to recognize network-level changes, even when the fMRI data is absent during actual deployment.

Methods

The core innovation of Band2Net is treating the relationship between EEG and fMRI as an Optimal Transport (OT) problem.

  1. The Student (EEG): The model processes EEG data into five frequency bands ($\delta, \theta, \alpha, \beta, \gamma$) and represents them as connectivity graphs. A shared Graph Neural Network (GNN) encoder creates embeddings for each band.
  2. The Teacher (fMRI): A frozen fMRI encoder groups brain activity into seven canonical "Yeo networks" (e.g., Visual, Somatomotor, Default Mode). These act as "tokens" representing high-level brain organization.
  3. Frequency-to-Network Transport: Instead of a fixed mapping, Band2Net solves a balanced entropic transport problem for every observation. It learns a "plan" to route information from specific EEG bands to specific fMRI networks. For instance, it might learn that alpha-band changes in EEG are best supervised by the fMRI visual network.
  4. Distillation: The EEG student is trained to minimize a standard classification loss while simultaneously trying to match the "transported" fMRI targets. Crucially, the fMRI branch is discarded after training; only the EEG branch is used for inference.
The Band2Net architecture showing the fMRI teacher, EEG student, and the transport mechanism.

Datasets

The study uses the PsiConnect dataset, a public multimodal study of brain dynamics under psilocybin.

  • Participants: 56 usable participants.
  • Conditions: Recordings were taken before and after a 19 mg dose of psilocybin.
  • Contexts: Four different states were recorded: eyes-closed rest, guided meditation, music listening, and movie viewing.
  • Key Detail: EEG and fMRI were acquired from the same people under the same conditions, but not simultaneously, making this a "state-matched" rather than "time-locked" problem.

Key Insights and Ideas

  • Band-Specific Routing: The model confirms that different EEG frequencies relate to different functional networks. As shown in the mean transport plan (Figure 2a below), the alpha ($\alpha$) band concentrates heavily on the visual network, while the delta ($\delta$) band maps more to the limbic network.
  • The "Ghost" of fMRI: Even though the fMRI data is gone at test time, the "privileged" training allows the EEG model to learn features that are more biologically grounded in network science.
  • Structural Importance: Interestingly, the authors found that even using label-free network descriptors (summarizing connectivity without a learned teacher) recovered much of the performance gain. This suggests that the structure of brain networks is the most critical piece of information being transferred.
Figure 2: (a) Mean transport plan over training observations, folds and seeds, normalised per band (%, rounded); seed-to-seed total variation 0.11. Yeo-7 networks as in Fig. 1: visual, somatomotor, dorsal and ventral attention, limbic, frontoparietal, default mode. (b) Change in ROC-AUC points (AUC × 100) when bands are withheld at test time, relative to each model’s full-band score.

Key Results

Band2Net was evaluated on its ability to decode whether a session was "pre-dose" or "post-dose" using only EEG.

  • Performance: Band2Net reached an ROC-AUC of 0.801, significantly outperforming a standard EEG-only GNN (0.751).
  • The "Privilege" Gain: The use of fMRI as privileged information provided a 5.0 AUC point boost over the student model trained on EEG alone.
  • Robustness: The model outperformed other common distillation techniques like Logit KD and Contrastive Representation Distillation (CRD).
  • Band Sensitivity: Withholding the alpha band at test time caused the largest drop in performance, aligning with known neurobiological effects of psilocybin on alpha rhythms.
Table 1: EEG-only decoding of session condition on the 112 session- level predictions (Sec. 3.2). AUC with 95% subject-level bootstrap CI. Gain: paired AUC difference, Band2Net minus method, in AUC points (AUC×100) with 95% CI; †primary comparison; ∗significant after Holm correction over the 16 secondary comparisons in this ta- ble. BA: balanced accuracy, F1: macro-F1 (%). Per-seed AUC SD is at most 0.012 for every method. The fMRI teacher is not an EEG- only model.

Limitations and Future Work

  • Sample Size: The study is limited to 56 participants from a single dataset. Larger, more diverse cohorts are needed to confirm generalizability.
  • Confounding Factors: The authors note that session condition is partly confounded with the order of acquisition (pre-dose always comes before post-dose), though they performed checks to mitigate this.
  • Continuous Outcomes: While this study focused on binary "pre vs. post" decoding, future work could explore if these methods can predict continuous variables, such as a participant's subjective drug intensity.
  • Routing Complexity: At this sample size, the authors found that a single learned transport plan shared across observations performed nearly as well as per-observation routing, suggesting that simpler models might suffice for smaller datasets.

Summary updated

Abstract

Functional MRI resolves large-scale brain networks but is expensive and stationary, whereas EEG is portable but poorly localised. We treat fMRI as privileged information, available during training and absent at deployment, with EEG and fMRI acquired from the same participants under matched conditions but not simultaneously. Band2Net encodes EEG into five frequency-specific connectivity graphs and fMRI into seven canonical network tokens from a frozen teacher, and solves a balanced entropic transport problem between the two token sets that gives each EEG band its own fMRI target. In subject-grouped cross-validation on the PsiConnect psilocybin data set, Band2Net reaches 0.801 ROC-AUC for EEG-only session-condition decoding, 5.0 AUC points (95% CI 1.8-8.2) above an EEG-only student. Although its margins over a single shared plan, a KL-relaxed plan and label-free network descriptors remain unresolved, secondary comparisons favour band-specific transported targets over logit, global-feature, contrastive and pooled-target distillation, though not after Holm correction.
BibTeX
@misc{paul2026band,
  author = {Nishi Kanta Paul and Md Shihabul Islam Shovo and Dwarikanath Mahapatra},
  title = {{Band2Net: Frequency-to-Network Transport Distillation for fMRI-Privileged, EEG-Only Session-Condition Decoding}},
  year = {2026},
  eprint = {2611.01548},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.01548},
}
Figure