TriRouter: Sparse-to-Dense Electrode-Space Routing for Robust EEG Motor Imagery Decoding

• Motor imagery decodes user intention from sensorimotor rhythm modulation, and EEG is a portable, low-cost, temporally precise noninvasive BCI modality.Source: §1 · p. 1
• Electrode detachment, abnormal impedance, artifact rejection, and low-density replacement can make test-time channels differ from training and degrade complete-channel models.Source: §1 · p. 1
• The paths are fused using a validation-calibrated missing-severity router, preserving mild-missingness discrimination while strengthening recovery under severe missingness.Source: §1 · p. 1
• The framework treats channel availability as an acquisition state and routes among discriminative-preservation, mask-robustness, and latent-compensation paths.Source: §1 · p. 1
• TriRouter shows consistent gains over public backbones and a corrupted-EEG baseline under random, regional, and low-channel missingness.Source: §1 · p. 1

• Early MI-EEG decoding used common spatial patterns, filter banks, and linear classifiers, while BCI Competition IV and MOABB established widely used benchmarks and protocols.Source: §2 · p. 1
• Deep MI-EEG models include ShallowConvNet, EEGNet, FBCNet, attention-based architectures, EEG Conformer, and CTNet, but most methods still assume consistent training and test channel sets.Source: §2 · p. 2
• Completion, inpainting, dynamic filtering, channel dropout, and few-channel prediction address corrupted channels, but usually apply one strategy across missingness levels.Source: §2 · p. 2
• The reviewed work leaves underexplored specialization that preserves spatial discrimination under mild missingness while strengthening compensation under severe loss.Source: §2 · p. 2

3. METHOD

• TriRouter assigns discrimination, robustness, and compensation to specialized paths rather than forcing one encoder to serve all three objectives.Source: §3 · p. 2

3.1. Three-Path Decoding for Discriminative Preservation, Mask Robustness, and Latent Compensation

• The discriminative path learns the complete-observation decision boundary, while inference supplies all paths with zero-filled inputs and reduces the discriminative path's contribution as missingness rises.Source: §3.1 · p. 2
• The mask-robust path trains on uniformly sampled random channel subsets, while the latent-compensation path completes electrode-space representations using a mask-conditioned MLP.Source: §3.1 · p. 2
• The latent-compensation objective uses complete-trial channel-wise log-bandpower targets across θ, µ, β, and low γ bands; experiments fix the compensation weight at λ = 0.1.Source: §3.1 · p. 2

3.2. Validation-Calibrated Missing-Severity Routing and Path Fusion

• For each missing-rate anchor, the router combines validation accuracy with a functional prior to calibrate path weights; calibration settings are selected on validation and held fixed across reported tests.Source: §3.2 · pp. 2–3
• The router linearly interpolates weights between anchors and clips rates beyond their range to the nearest endpoint; low, intermediate, and high missingness favor Ddis, Drob, and Dlat, respectively.Source: §3.2 · p. 3

4.1. Experimental Setup

• Evaluation follows the official subject-specific train-test protocol, with nine subjects, three random seeds, and validation-only model selection and routing calibration.BNCI2014-001 uses 22 channels and nominal 0%, 25%, 50%, and 75% missing settings; BNCI2014-004 uses three channels.Source: §4.1 · p. 3

4.2. Main Results on BNCI2014-001

• TriRouter achieves the best accuracy from complete inputs to severe missingness, with its margin over DSF-CTNet growing as channels are removed.The reported comparison includes nominal random missing-channel settings on BNCI2014-001.Source: Table 1 · p. 3, §4.2 · p. 3
• Paired subject-level gains over EEGNet at 50% and 75% missingness favor TriRouter for most subjects, and advantages persist across most non-anchor missing rates.The subject-level comparisons use a fixed EEGNet reference; the continuous evaluation spans 0% to 90% missing channels.Source: §4.2 · p. 3, Figure 2 · p. 3, Figure 3 · p. 4

4.3. Mechanism Analysis

• Path reliability shifts with missing severity: the discriminative path works best with intact spatial layout, latent compensation dominates severe loss, and mask robustness stabilizes the intermediate range.Ablations associate removing latent compensation with weaker high-missingness robustness, removing discrimination with worse complete and mild inputs, and removing mask robustness with reduced low-to-medium stability.Source: §4.3 · pp. 3–4, Figure 4 · p. 4
• Validation-only calibration helps at severe missingness but is not uniformly better at lower rates; adding the functional prior yields a more stable overall allocation, while compensation reduces missing-channel log-bandpower error versus zero filling.The reported Pearson correlation compares reconstructed missing-channel bandpower vectors with complete-trial targets.Source: §4.3 · pp. 3–4, Table 3 · p. 4

4.4. Structured Regional Missingness and Computational Complexity

• Under structured regional missingness, TriRouter consistently outperforms baselines, with an increasing margin at higher missing rates; it has 4.039M parameters and 0.096 ms single-trial latency.TriRouter is smaller than ShallowConvNet and faster than FBCNet; the regional setting samples 70% of missing channels from the sensorimotor set.Source: Table 4 · p. 4, §4.4 · p. 4

4.5. Additional-Dataset Evaluation on BNCI2014-004

• On three-channel BNCI2014-004, TriRouter achieves the best accuracy with zero or one missing channel and remains competitive with two missing channels.Each method uses the dataset’s own subject-specific protocol; results suggest the routing principle remains useful beyond the 22-channel setting.Source: Table 5 · p. 4, §4.5 · p. 4

• TriRouter combines three paths—discriminative preservation, mask robustness, and latent compensation—through a validation-calibrated missing-severity router.Source: §5 · p. 4
• Experiments show consistent gains over public backbones and a corrupted-EEG baseline under random and regional missingness.Source: §5 · p. 4
• Latent recovery analysis confirms that compensation restores meaningful bandpower structure, suggesting missing severity is informative for adaptive EEG decoding.Source: §5 · p. 4