PaperlayerView exact revision on arXiv

Source-linked AI summary

BridgeRouter: Gain-Aware Instance Routing over Cross-Modal Bridge Portfolios

Ruoxi Piao, Alicia Lu, Yunzhen Zhang, Mustafa Mısır, Hasan Onur Keleş

arXiv:2610.38152v1cs.AI

TL;DR

A global bridge choice ignores the complementary errors of different transfer pathways. BridgeRouter predicts anchor-relative gains from target-modality representations and routes each instance to one pathway. It improves over the validation-selected SBS on both evaluated datasets, while gains are smaller and more variable under the subject-disjoint ISRUC setting.

  • Problem

    A globally selected transfer pathway can miss complementary pathway strengths on individual instances, while the source modality and task labels are unavailable at deployment.

    Source: §1 · p. 1

  • Method

    BridgeRouter learns anchor-relative pathway gains from training-time transfer outcomes and selects one route at inference using only the target-modality representation.

    Source: §1 · p. 2

  • Results

    BridgeRouter improves balanced accuracy over the validation-selected single best solver by 11.49 percentage points on FOG and 3.83 points on ISRUC.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The results show substantial complementarity among frozen EEG-to-target-modality bridges, with remaining oracle gaps indicating portfolio headroom.

    Source: §5 · p. 9

    The subject-disjoint ISRUC setting has smaller and more variable routing gains, and its validation-selected SBS is not the strongest individual test pathway.

    Source: §4.4 · p. 9, §5 · p. 9

Abstract

from arXiv · show

Lightweight representation bridges make pretrained, task-adapted EEG foundation models executable on other biosignal modalities without retraining their backbones. However, selecting a single bridge pathway globally assigns every instance to the same transfer route and ignores complementary errors across routes. We therefore formulate multi-bridge transfer as an instance-level algorithm-selection problem and introduce BridgeRouter, a gain-aware router over a portfolio of frozen cross-modal pathways. A validation-selected Single Best Solver (SBS) serves as a protected default route. During training, task labels are used to compute the reduction in negative log-likelihood provided by each alternative route relative to the SBS, and a lightweight MLP predicts these route-specific gains from a target-modality representation extracted by a frozen encoder. At inference, BridgeRouter observes only the target modality, requires neither EEG, task labels, nor predictions from all candidate routes, and executes exactly one bridge pathway: the route with the largest predicted gain when that gain exceeds a validation-calibrated threshold, and the SBS otherwise. We evaluate seven EEG foundation-model bridge pathways for EMG-based freezing-of-gait detection on FOG and ECG-based sleep classification on ISRUC. Across five seeds, BridgeRouter achieves $81.38\pm1.11\%$ balanced accuracy on FOG, improving over the BIOT SBS by $11.49$ percentage points, and $62.92\pm2.17\%$ on ISRUC, improving over the CBraMod SBS by $3.83$ points. Per-instance bridge oracles reach $99.96\%$ and $78.61\%$, respectively, revealing substantial remaining portfolio headroom. These results show that anchor-relative gain prediction can convert complementary cross-modal bridge behavior into deployable instance-level routing improvements while retaining single-route inference.

1 INTRODUCTION

A single globally selected bridge can miss complementary pathway strengths on individual instances. BridgeRouter learns to predict anchor-relative gains from target-modality representations and selects a route using only the target modality at inference.

1 INTRODUCTION

  • Different pathways can fail on different instances, so a weaker average pathway may correctly predict examples missed by a stronger one.This makes pathway complementarity, rather than standalone average performance alone, relevant to transfer portfolios.Source: §1 · p. 1
  • The routing problem is to identify whether another transfer pathway can do better on an instance, although the information revealing pathway success is unavailable at inference.Paired EEG–target recordings and task labels reveal pathway outcomes during training, while deployment provides only the target modality.Source: §1 · pp. 1–2
  • The paper formulates multi-model cross-modal transfer as instance-level pathway selection, emphasizing complementary errors rather than a single globally preferred route.BridgeRouter uses a validation-selected default and switches to an alternative only when its predicted gain exceeds a validation-calibrated threshold.Source: §1 · p. 2
  • BridgeRouter learns task-label-derived, anchor-relative negative-log-likelihood gains from target-modality representations and makes target-only, single-route selections at inference.It evaluates transfer on EEG-to-EMG and EEG-to-ECG tasks, examining fixed-route performance, oracle headroom, and rescue-harm behavior.Source: §1 · p. 2

2 RELATED WORK

Prior work supports cross-modal reuse of pretrained biosignal representations and adaptive model selection, but typically evaluates transfer routes independently. BridgeRouter connects these ideas by predicting anchor-relative gains among cross-modal pathways.

2.1 EEG FOUNDATION MODELS AND CROSS-MODAL REPRESENTATION TRANSFER

  • Cross-modal bridges make frozen pretrained biosignal representations reusable, but existing methods generally select a fixed pathway by aggregate performance without modeling instance-level complementary predictions.EEG foundation models also vary in downstream performance across datasets and tasks.Source: §2.1 · p. 2

2.2 MODEL ROUTING AND PER-INSTANCE MODEL SELECTION

  • Conditional model selection routes different inputs to different models, while cross-modal routing must infer pathway usefulness from target-side information when the source modality is unavailable.Executing every source pathway at deployment can undermine lightweight transfer.Source: §2.2 · p. 3

2.3 ALGORITHM SELECTION AND COMPLEMENTARY MODEL PORTFOLIOS

  • Algorithm selection exploits differences in algorithm performance across instances, and EEG-AS applies this idea to within-modality EEG model choice.BridgeRouter instead selects among cross-modal pathways combining source models and target-side bridges, predicting relative gains over an anchor rather than absolute candidate performance.Source: §2.3 · p. 3

3 METHODS

The methods build multiple target-executable pathways from frozen EEG foundation models and lightweight bridges, then learn instance-level route selection from training-time outcomes. At inference, routing uses target-modality information and executes one pathway.

3.1 PROBLEM SETUP AND NOTATION

  • Training uses paired EEG and target-modality samples, but inference must produce predictions from the target modality alone.The portfolio comprises pathways built from different EEG foundation models and corresponding cross-modal bridges, with the validation-balanced-accuracy SBS frozen as the default anchor.Source: §3.1 · pp. 3–4

3.2 CROSS-MODAL PATHWAYS VIA REPRESENTATION BRIDGING

  • Each lightweight bridge maps a target-modality representation into a frozen EEG foundation model’s representation space, using paired EEG observations as privileged alignment information during training.After bridge learning, the bridged representation passes through the remaining frozen model layers, so the pathway runs from target input without EEG.Source: §3.2 · p. 4

3.3 PORTFOLIO VIEW OF MULTI-BRIDGE TRANSFER

  • Pathways may make different predictions on the same sample, so average accuracy can miss a weaker pathway’s correct predictions where the anchor fails.The portfolio oracle marks whether at least one pathway solves an instance; its gap from a fixed route exposes routing opportunity.Source: §3.3 · pp. 4–5

3.4 BRIDGEROUTER: TRAINING AND INFERENCE

  • The router learns each alternative’s loss reduction relative to the anchor from training-time pathway outcomes, then predicts gains from the target-modality representation.It selects the most promising alternative only when its predicted gain clears a validation-selected protection threshold; otherwise, the anchor remains the default, and exactly one pathway executes.Source: §3.4 · pp. 5–6, Figure 1 · p. 5

4 EXPERIMENTS AND DISCUSSION

Across FOG and ISRUC, BridgeRouter improves on the validation-selected SBS by routing instances among frozen cross-modal pathways, while executing only one route per instance. The experiments also reveal substantial pathway complementarity and a remaining challenge in reliable route selection under cross-subject shift.

4.1 EXPERIMENTAL SETUP

  • FOG uses synchronized EEG–EMG windows and a sample-wise split, whereas ISRUC uses synchronized EEG–ECG epochs and a subject-disjoint split.Seven frozen EEG foundation models define the candidate routes; the SBS is selected by validation balanced accuracy, yielding BIOT on FOG and CBraMod on ISRUC.Source: §4.1 · pp. 6–7

4.3 GAIN-AWARE BRIDGE ROUTING

  • BridgeRouter predicts route-specific gains relative to the SBS, switches only when the largest predicted gain clears a validation threshold, and executes one pathway per instance.On ISRUC, the improvement is smaller and more variable, and the router mean remains below EEGPT and BENDR test balanced accuracy.Source: §4.1 · p. 7, §4.3 · p. 8

4.2 FROZEN BRIDGE PERFORMANCE

  • On FOG, BIOT is the strongest test pathway at 69.89% balanced accuracy; on ISRUC, validation-selected CBraMod scores 59.09%, below EEGPT at 65.17% and BENDR at 64.45%.Selecting a pathway based on test labels would be test-set model selection, so the SBS remains the validation-selected route.Source: §4.2 · p. 7, Table 1 · p. 8
  • The seven-route oracle reaches 99.96% balanced accuracy on FOG and 78.61% on ISRUC, with the FOG compact-four oracle still reaching 93.69%.Oracles are non-deployable; the compact-four result shows complementarity remains after removing collapsed pathways.Source: §4.2 · pp. 7–8, Table 2 · p. 8

4.4 CROSS-DATASET FINDINGS

  • Across datasets, pathway errors are heterogeneous, relative-loss routing improves over the SBS without evaluating every route at inference, and cross-subject route selection remains unresolved.FOG is sample-wise, whereas ISRUC is subject-disjoint and has a validation-to-test pathway-ranking reversal.Source: §4.4 · pp. 8–9

4.3 GAIN-AWARE BRIDGE ROUTING

  • 81.38 ± 1.11% balanced accuracy on FOG (11.49 points above BIOT SBS) and 62.92 ± 2.17% on ISRUC (3.83 points above CBraMod SBS) are BridgeRouter’s five-seed results.Every ISRUC seed improves over CBraMod in balanced accuracy.Source: §4.3 · p. 8, Table 3 · p. 9, §4.4 · p. 9

5 CONCLUSION

BridgeRouter selects among cross-modal bridge routes by predicting their task-loss gains over an anchor, while executing only one route per instance. Across FOG and ISRUC, it improves on the validation-selected solver, though gains are smaller and more variable under subject-disjoint evaluation.

5 CONCLUSION

  • The router predicts candidate bridges’ relative task-loss gains and switches from the anchor only when the predicted gain is sufficiently large.Inference executes only one bridge pathway.Source: §5 · p. 9
  • BridgeRouter improves balanced accuracy over the validation-selected single best solver by 11.49 points on FOG and 3.83 points on ISRUC.The smaller and more variable ISRUC gains occur under its subject-disjoint setting.Source: §5 · p. 9
  • Future work will investigate more robust routing under cross-subject and cross-domain shifts and richer signals for estimating transfer gains.Source: §5 · p. 9
  • The study evaluates existing physiological datasets for methodological research and makes no clinical decisions or medical recommendations.Source: ETHICS STATEMENT · pp. 9–10

A.1 EXPERIMENTAL SETUP

The experiments evaluate EEG-to-EMG freezing-of-gait detection and EEG-to-ECG sleep classification using frozen EEG foundation models and modality-specific encoders. Dataset partitions support bridge learning and model selection, but FOG uses a sample-wise split rather than an unseen-subject evaluation.

A.1 EXPERIMENTAL SETUP

  • FOG contains 37,200 paired EEG–EMG windows from 12 participants for binary freezing-of-gait classification.Recordings are segmented into three-second windows; Table 4 summarizes the partitioning, with all subsets containing the same participants.Source: §A.1 · p. 11, Table 4 · p. 12
  • ISRUC Subgroup I provides synchronized EEG and ECG overnight recordings for binary Wake-versus-Sleep classification using 30-second epochs.The original five sleep stages are collapsed into Wake and grouped sleep; the dataset follows BioX-Bridge Li et al. (2026).Source: §A.1 · p. 11
  • EEG is the source modality, while EMG and ECG are the target modalities for FOG and ISRUC, respectively.Source: §A.1 · p. 11
  • Each dataset is divided into Dold, Dpair, Dval, and Dnew for foundation-model head training, bridge learning, model selection, and final evaluation, respectively.ISRUC partitions are subject-level; FOG follows a sample-wise protocol and is not an unseen-subject evaluation.Source: §A.1 · p. 12
  • Seven frozen EEG foundation models each receive a task-specific linear head, with one bridge route per model and a frozen target-modality encoder.NormWear encodes FOG EMG and HuBERT-ECG encodes ISRUC ECG; the evaluated EEG models use publicly released pretrained weights.Source: §A.1 · p. 12

A.2 TASK-LABEL-FREE TRANSFER-AWARE ROUTING

The task-label-free experiment tests whether target-side transfer signals can guide pathway selection without label-derived gain targets. It improves over the BIOT solver on FOG, but no tested label-free variant exceeds the CBraMod solver on ISRUC.

A.2 TASK-LABEL-FREE TRANSFER-AWARE ROUTING

  • The label-free router predicts teacher competence, normalized teacher-to-bridge Jensen-Shannon divergence, and prediction-flip probability from target-modality representations.The experiment also tests competence-only and transfer-only variants; FOG uses a compact-four portfolio, whereas ISRUC uses all seven routes.Source: §A.2 · p. 12
  • 2.02 ± 0.43 balanced-accuracy points: transfer-aware routing improves over the BIOT SBS on FOG.Combining competence and transfer risk yields 580.4 ± 69.9 rescues and 307.4 ± 80.1 harmful switches, for a net gain of 273.0 ± 26.5 correct predictions.Source: §A.2 · p. 12
  • No label-free variant exceeds the CBraMod SBS on ISRUC; transfer-only reaches 64.54 ± 0.07% balanced accuracy, 1.38 points below CBraMod.The transfer-aware combination reaches 61.31 ± 0.25%. These ISRUC results are reported on VAL, while the FOG experiment is reported on TEST, and the extension is exploratory.Source: §A.2 · p. 13, Table 5 · p. 13
  • Transfer-derived signals may inform utility within one pathway without being calibrated across heterogeneous foundation models.The authors note that teacher confidence, divergence, and disagreement can have different numerical meanings across receivers, limiting the reliability of a fixed transfer-aware score.Source: §A.2 · p. 13

A.3 FOG DIRECT-EEG PERFORMANCE

FOG direct-EEG performance does not determine bridged-EMG performance, motivating selection of the bridge pathway rather than the receiver’s native-EEG ranking.

A.3 FOG DIRECT-EEG PERFORMANCE

  • CSBrain and BENDR improve after bridging, while LaBraM, EEGPT, and REVE substantially degrade and produce strongly class-biased predictions.The comparison concerns balanced accuracy on the final FOG test split.Source: §A.3 · p. 13, Table 6 · p. 13
  • Direct-EEG ranking is therefore not a reliable basis for selecting a cross-modal bridge route.Source: §A.3 · p. 13

A.4 FOG OUTPUT-LEVEL TRANSFER LOSS AND REPRESENTATION DISTANCE

The FOG diagnostics compare bridged and direct-EEG outputs and examine whether representation discrepancies indicate bridge failures. Their diagnostic value varies by receiver, while BIOT’s remaining layers contract representation discrepancy without eliminating output discrepancy.

A.4 FOG OUTPUT-LEVEL TRANSFER LOSS AND REPRESENTATION DISTANCE

  • The FOG test diagnostics compare bridged and direct-EEG class distributions using JS divergence, prediction-flip rate, and conditional failure-detection AUC.The comparison is made for paired instances and receivers, with failure detection conditioned on a correct EEG teacher.Source: §A.4 · p. 13, Table 7 · p. 14
  • JS divergence is especially diagnostic for EEGPT and REVE, but less discriminative for CSBrain and BENDR, limiting the transferability of one uncalibrated threshold.The reported output discrepancy is highly receiver-dependent.Source: §A.4 · p. 14
  • For BIOT, test cosine distance falls from 0.226 at the bridge destination to 0.134 at the final representation, while mean test JS divergence remains 0.055.Destination-layer discrepancy is nearly identical across paired, validation, and test splits; the remaining BIOT layers contract rather than amplify it.Source: §A.4 · p. 14, Table 8 · p. 14

A.5 COLLAPSED ROUTES AND COMPACT-PORTFOLIO HEADROOM

The section audits collapsed FOG routes and shows that removing them reduces, but does not eliminate, the portfolio’s oracle headroom over the BIOT SBS.

A.5 COLLAPSED ROUTES AND COMPACT-PORTFOLIO HEADROOM

  • A route is classified as collapsed using validation data when one predicted class accounts for at least 95% of outputs and balanced accuracy is below 55%.This validation-only rule removes LaBraM, EEGPT, and REVE, retaining BIOT, CBraMod, CSBrain, and BENDR.Source: §A.5 · p. 14, Table 9 · p. 14
  • 99.96% test balanced accuracy is reached by the seven-route oracle, but removing three collapsed routes reduces it to 93.69%.The full oracle misses three instances; the compact portfolio oracle misses 514 jointly.Source: §A.5 · p. 14
  • The compact-four oracle remains 23.80 points above the BIOT SBS, so collapsed routes do not explain away portfolio complementarity.With the target-only MLP as default predictor, the compact-four residual oracle reaches 97.07 ± 0.08% balanced accuracy across five MLP seeds.Source: §A.5 · p. 14

A.6 FOG SUPERVISED CONTROLS AND HIGHER-CAPACITY ROUTING

These supplementary analyses delimit the routing result: target-only prediction is a strong control, sensitivity findings are development-stage only, and inference still executes one selected bridge route. Rescue-harm behavior and implementation safeguards further characterize the results and their scope.

A.6 FOG SUPERVISED CONTROLS AND HIGHER-CAPACITY ROUTING

  • The target-only MLP is the strongest deployable method in the FOG controls, narrowing BridgeRouter’s result to improving the predefined SBS while retaining single-route execution.The comparison also includes a higher-capacity strict EEG-AS adaptation.Source: §A.6 · p. 15, Table 10 · p. 15

A.7 ISRUC DEVELOPMENT-STAGE SENSITIVITY ANALYSES

  • Within the earlier H4/K7 ISRUC development configuration, NLL-gain routing with validation-selected protection gives the highest mean validation balanced accuracy among the tested changes.Correctness-gain thresholds suppress switching, while forcing alternatives with NLL-gain scores slightly reduces performance relative to retaining the anchor prediction; these are validation analyses, not test results.Source: §A.7 · p. 15, Table 11 · p. 15
  • ISRUC oracle balanced accuracy rises monotonically as routes are added, but deployable routing is nonmonotonic and its strongest exploratory result is at K4.The K4 result and nested ordering are development-stage sensitivity analyses and do not replace the final canonical K7 portfolio.Source: §A.7 · p. 16, Table 12 · p. 16
  • The ISRUC source-representation analysis finds a broad H3–H5 plateau, with only 0.0491 balanced-accuracy points between H4 and H5, below observed seed dispersion.Concatenating all eight layers expands the representation from 512 to 4096 dimensions without improving on the single-layer plateau.Source: §A.7 · p. 16, Table 13 · p. 16

A.8 RESCUE-HARM DIAGNOSTICS

  • On FOG, the router rescues many more BIOT errors than it destroys BIOT successes; on ISRUC, seed 46’s 42.2% rescue precision explains much of aggregate-result variance.Seeds 42–45 have ISRUC rescue precision between 75.6% and 94.0%.Source: §A.8 · p. 16

A.9 ROUTER OPTIMIZATION AND REPRODUCIBILITY DETAILS

  • The validation-selected threshold uses the more conservative policy when balanced-accuracy checkpoints tie, and no threshold or checkpoint is reselected on test data.Gain targets are clipped to training-set percentiles, standardized, and fit with smooth-L1 regression; instances are weighted inversely to training class frequency.Source: §A.9 · p. 17
  • The evaluation selects the SBS by validation balanced accuracy and freezes it before testing; the ISRUC split is subject-disjoint, whereas FOG uses a sample-wise split with the same 12 participants in all subsets.The portfolio oracle uses ground-truth labels and is reported only as a non-deployable upper bound.Source: §A.9 · p. 17

A.10 INFERENCE REQUIREMENTS AND COMPUTATIONAL COST

  • At inference, the router uses the frozen target encoder once, predicts six gains for the seven-route portfolio, and executes exactly one selected bridge pathway.It requires neither task labels nor paired EEG, and its adaptive routing does not multiply bridge-pathway inference cost by portfolio size.Source: §A.10 · pp. 17–18, Table 17 · p. 18

Related papers

Loading 2610.38152v1…