Lightweight representation bridges enable pretrained, task‑adapted EEG foundation models to be executed on other biosignal modalities without retraining their backbones. A single global bridge pathway, however, forces every instance through the same transfer route and ignores complementary errors across routes. We therefore cast 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.
During training, task labels are used to compute the reduction in negative log‑likelihood (gain) that each alternative route provides relative to a validation‑selected Single Best Solver (SBS). A lightweight MLP predicts these route‑specific gains from a target‑modality representation extracted by a frozen encoder. At inference time only the target modality is observed; no EEG, task labels, or predictions from all candidate routes are required. BridgeRouter executes the route with the largest predicted gain when that gain exceeds a validation‑calibrated threshold, otherwise it falls back to the SBS.
We evaluate seven EEG foundation‑model bridge pathways on two tasks: EMG‑based freezing‑of‑gait (FOG) detection and ECG‑based sleep classification (ISRUC). Across five random 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, indicating substantial remaining portfolio headroom.
These findings demonstrate that anchor‑relative gain prediction can convert complementary cross‑modal bridge behavior into deployable instance‑level routing improvements while retaining single‑route inference efficiency.
Review