BRIDGE-EEG

A clinical EEG cap with 64 electrodes and a three-electrode consumer headband record the same cortical rhythms, but almost nothing reads both. Montages, channel counts and sampling rates differ per device, so most EEG pipelines stay welded to the one dataset they were trained on. The models that do generalize — EEG foundation models such as LaBraM, CBraMod and NeuroLM — buy that generality with size and compute that a battery-powered wearable cannot supply.

BRIDGE-EEG, from Purdue University and Case Western Reserve University, asks how much of a pretrained EEG representation survives compression to edge scale. Its answer has three parts. First, unified preprocessing projects every recording onto the international 10–10 62-channel layout, zero-filling absent electrodes rather than interpolating them so that no synthetic spatial correlations are introduced; squeeze-and-excitation blocks then learn to down-weight those empty channels. Each 4-second window becomes a log-scaled short-time Fourier spectrogram, one 128×128 time–frequency image per channel. Second, an SE-ResNet18 encoder is pretrained once with SimCLR on pooled, label-stripped EEG from five datasets. Third, that teacher is distilled into SE-ResNet8 and SE-ResNet4 students by two routes: task-agnostic distillation on the same unlabeled data, producing a shared backbone, and task-specific distillation from per-task fine-tuned teachers.

Released as a preprint in September 2026, the work extends the same group's SSL-SE-EEG (EMBC 2025), which introduced the squeeze-and-excitation encoder and the 2D EEG image representation, by adding the distillation stage, a six-benchmark evaluation and measured on-device energy.

#Key Features

• Device-agnostic input representation: A single 62×128×128 spectrogram tensor accepts recordings from 3 to 64 channels at 128–512 Hz, letting one backbone serve datasets acquired on different hardware.
• Pretraining disjoint from evaluation: The pretraining pool (SEED-IV, DEAP, PhysioNet MI, TUH-Sz, CHB-MIT) shares no dataset with the six evaluation benchmarks, so downstream scores measure transfer to unseen cohorts and devices.
• Two distillation strategies: Strategy A yields a reusable task-agnostic SE-ResNet8 backbone; Strategy B yields per-task students that trade reusability for accuracy.
• Energy measured, not estimated: Latency and power are profiled on an NVIDIA L4 GPU, an Intel i7-8700 CPU and an NVIDIA Jetson Orin Nano, rather than reported as FLOPs.
• Channel recalibration under missing electrodes: SE blocks after each residual stage reweight channels, the mechanism that makes heavy zero-filling tolerable.

#Technical Details

The SE-ResNet18 teacher has 11.84 M parameters and maps spectrograms to 512-dimensional embeddings. Contrastive pretraining runs 50 epochs on four NVIDIA L40S GPUs at batch size 1024 with SGD and an NT-Xent loss over views generated by channel dropout, time- and frequency-bin masking, and additive Gaussian noise. Distillation produces a 1.56 M-parameter SE-ResNet8 and a 0.48 M-parameter SE-ResNet4; Strategy A combines NT-Xent with an MSE term anchoring the student to the teacher's embedding space, Strategy B uses the standard Hinton loss. On TUAB abnormality detection the SE-ResNet8 student reaches 90.35% accuracy against 82.58% for LaBraM-Huge (369 M) and 83.15% for REVE (69 M); on SIENA it reaches 99.54%. On SEED emotion recognition it scores 80.0% against 73.18% for LaBraM-Base. Motor imagery is the exception: the best BCI-IV-2a result is 38.66% against a 25% chance level, below Neuro-GPT at 58.60%, which the authors attribute to a single motor-imagery dataset in the pretraining pool and to the 22- and 3-channel recordings being mostly zero-filled. On TUAB the students also score higher accuracy than the teacher but lower macro-F1 (0.83 versus 0.87), favoring the majority class. On the Jetson Orin Nano, SE-ResNet8 runs at 5.73 ms and 15.64 mJ of dynamic energy per inference against 12.93 ms and 46.67 mJ for the teacher.

#Applications

The compact students target continuous and ambulatory EEG, where recordings must be screened on the device rather than shipped to a server: abnormality screening from clinical montages, emotion decoding, and brain-computer interface tasks on low-density consumer headsets. The 0.48 M-parameter SE-ResNet4 occupies roughly 1.9 MB at 32-bit precision and about 0.48 MB when quantized to 8 bits, which the authors present as a candidate footprint for Cortex-M-class microcontroller deployment, with runtime activations still to be fitted.

#Impact

BRIDGE-EEG is evidence that EEG foundation-model accuracy on binary clinical tasks is not intrinsically tied to parameter count, and that measured energy is a more useful deployment metric than FLOPs — a comparison it makes explicitly against FEMBA, whose smallest variant is an order of magnitude larger. The motor-imagery result is the honest counterweight: the distilled students track the teacher closely there, so the gap lies in the pretrained representation rather than in compression, and the paper reads it as a limit on pretraining diversity. The work is a preprint awaiting peer review, and it carries no code or data availability statement; no public repository or released checkpoint accompanies it, so the results are not yet independently reproducible.