PRiSE-EEG

Sleep staging wants slow whole-night rhythms; a visual-oddball task wants a few hundred milliseconds over occipital electrodes; emotion recognition leans on frontal asymmetry. Train one dense Transformer on all of them and those objectives pull the shared weights apart. PRiSE-EEG measures that conflict rather than assuming it: pairwise cosine similarity of fine-tuning gradients across six datasets, computed for LaBraM and CSBrain, shows consistently low agreement between paradigms. Centered Kernel Alignment on mixed-paradigm batches of a dense pilot model then shows the conflict is not uniform with depth — shallow blocks keep high cross-paradigm similarity, deeper blocks drift apart.

The model turns both measurements into architecture. PRiSE-EEG is a Mixture-of-Experts Transformer in which the split between shared and specialized experts is set per block by that block's CKA sharedness, mapped through a sigmoid: early blocks get mostly shared experts, later blocks get mostly specialized ones. The allocation is computed once and held fixed for training, so only token-level routing stays adaptive. Feeding it is a tokenizer that forms continuous multi-channel patches using weak sensor-space priors — coarse anatomical regions and functional networks over the 10-10 montage — instead of the per-channel discrete codes used by CodeBrain or the topology hierarchy of MMM. It was released as a preprint in May 2026 by a group at Tongji University, with a co-author at Shanghai Yangzhi Rehabilitation Hospital, and revised that September.

#Key Features

• CKA-calibrated depth-stratified experts: Each MoE block carries nine experts — one permanently shared, eight apportioned by the sigmoid rule — with the top two specialized experts routed per token. Reversing the allocation order costs 7.93–21.02 points of balanced accuracy, the sharpest evidence that the depth ordering, not merely the use of experts, is doing the work.
• Prior-guided continuous tokenization: Static group queries carry attention biases from cortical regions and eight functional networks, plus a learnable relaxation term so the grouping can bend to a dataset's montage. Dynamic groups add lag-limited channel affinities computed over micro-frames inside each patch.
• Three pretraining objectives: The reconstructor is trained jointly on masked temporal-patch reconstruction, masked-channel reconstruction and next-patch prediction, against both waveform and spectral targets, plus a router load-balancing loss.
• Unified multi-dataset adaptation: One backbone is fine-tuned across all downstream datasets at once with dual-stream heads that pool separately over time and over groups, cutting head parameters by more than 90% against a flattened MLP.

#Technical Details

Five configurations are published, spanning 9.41M parameters (PRiSE-EEG-S) to 125.58M (-XL). Reported results use PRiSE-EEG-B: 41.06M total parameters with 18.05M activated per token, model dimension 384, eight attention heads, and a 12-block stack of two tokenizer cross-attention blocks, four group MoE blocks, four temporal MoE blocks and two decoder blocks. Pretraining draws on 17 public datasets totaling over 28,000 hours, dominated by TUEG and the Healthy Brain Network corpus and reaching into emotion, motor-imagery, ERP, resting-state and imagined-speech recordings; TUEG subjects overlapping the TUAB, TUEV or TUSL evaluation sets are removed. An MNE-based pipeline resamples to 256 Hz, band-passes 1–100 Hz, notch-filters line noise, and segments one-second patches on a standardized 10-10 channel superset, with rotary positional encoding over time.

Evaluation covers 12 datasets across 10 task categories, averaged over 10 runs. The largest gains fall on disease-related and cognitive tasks — roughly 6–8 points on ADFTD, SEED-V and mental workload over prior EEG foundation models — with 2–5 point gains on TUEV, TUSL, PhysioMI, SEED and HMC, and a >9 point AUROC gain on Things-EEG-2. Replacing the MoE layers with dense feed-forward blocks costs 2.21–5.94 points. The headline tables set per-task-trained baselines, including CBraMod and BENDR, against PRiSE-EEG under unified multi-task fine-tuning; a separate single-task table reports the same weights transferring alone.

#Applications

The downstream suite is the clinical and brain-computer-interface workload: seizure detection on Siena, abnormality and event screening on TUAB and TUEV, sleep staging on HMC, Alzheimer's and frontotemporal dementia identification on ADFTD, mental workload, motor imagery, emotion recognition and visual target detection. Because one backbone serves all of them, a lab with a small labelled cohort can attach a head rather than train a decoder from scratch. The authors scope this explicitly: PRiSE-EEG is intended for research use, and clinical or other high-stakes deployment requires independent validation and professional oversight.

#Impact

PRiSE-EEG's contribution is methodological more than numerical — it treats representation similarity as a design input that fixes the architecture before training rather than as a diagnostic plotted afterwards, and the reversed-allocation ablation gives that idea a falsifiable test. The caveats are real: the training and preprocessing code is public under GPL-3.0, but no pretrained checkpoints have been released, so reproducing it means re-running pretraining over a multi-terabyte corpus drawn from 17 dataset agreements. Evaluation is retrospective on public benchmarks, with no prospective or cross-site validation, and the work remains a preprint.