PhysioWave

PhysioWave is a foundation model family for physiological time-series signals that replaces the fixed patch-embedding front end of standard signal transformers with a learnable, multi-scale wavelet decomposition. Physiological recordings—electrocardiography (ECG), electromyography (EMG), and electroencephalography (EEG)—are frequently corrupted by motion artifacts, baseline drift, and other low signal-to-noise disturbances, and they carry diagnostically important structure at very different time scales. Wavelet analysis is a classical tool for this kind of non-stationary, multi-resolution signal, and PhysioWave makes that decomposition trainable and couples it to a transformer encoder pretrained with masked reconstruction.

Developed by researchers at ETH Zurich, the University of Bologna, and the University of Modena and Reggio Emilia, the work was released as a preprint in June 2025 and accepted at NeurIPS 2025. The authors pretrain separate modality-specific backbones for ECG and EMG on large unlabeled corpora, then show that a single architecture transfers across a wide range of downstream biosignal tasks. They additionally demonstrate a multi-modal extension that fuses ECG, EMG, and EEG branches through a learnable weighted combination.

PhysioWave sits within the emerging class of biosignal foundation models, applying the self-supervised pretrain-then-adapt recipe—well established for language and protein models—to wearable and clinical physiological sensing, where raw recordings are abundant but task labels are scarce.

#Key Features

• Learnable wavelet decomposition: An adaptive wavelet selector chooses among candidate basis functions and performs multi-resolution analysis across several decomposition levels, so the time-frequency front end is optimized jointly with the network rather than fixed in advance.
• Frequency-guided masked pretraining: Masking is steered by FFT-based spectral energy to prioritize high-energy components, and patches are reconstructed with a Smooth-L1 loss, focusing self-supervision on the most informative parts of the signal.
• Multi-scale transformer encoder: A Cross-Scale Channel-Aggregation Feed-Forward Network (CAFFN) and rotary positional embeddings (RoPE) let the attention layers integrate features across resolutions and channels.
• Multi-modal fusion: Modality-specific ECG, EMG, and EEG branches are combined through a learnable weighted fusion, enabling joint inference when several signal types are recorded together.
• Open code and three model sizes: The implementation is released under the MIT license, with Small (5M), Base (15M), and Large (37M) parameter configurations per modality.

#Technical Details

The ECG backbone is pretrained on roughly 182 GB of data drawn from MIMIC-IV-ECG, MedalCare-XL, CODE-15%, the Norwegian Athlete ECG Database, and the Georgia cohort; the EMG backbone is pretrained on roughly 823 GB spanning NinaPro DB6/DB7/DB8, EMG2Pose, and EMG2Qwerty. Each modality is offered in Small (5M), Base (15M), and Large (37M) parameter variants. On downstream evaluation, the Large ECG model reaches 66.7% F1 and 94.6% AUROC on PTB-XL; EMG models reach 94.50% accuracy on EPN-612, 87.53% on NinaPro DB5, and 93.19% on UCI EMG. The multi-modal variant improves emotion recognition on DEAP to 85.2% valence (+6.1%) and 88.6% arousal (+7.3%) accuracy, and reaches 74.9% (+4.5%) on MPDB, indicating that the wavelet front end and fusion add value over prior baselines.

#Applications

PhysioWave targets researchers and engineers building models for wearable and clinical biosignals. The pretrained ECG backbone supports cardiac tasks such as arrhythmia detection and multi-label diagnostic classification; the EMG backbone supports gesture and movement decoding for prosthetics, human-computer interaction, and rehabilitation; and the multi-modal configuration supports affective-computing tasks such as emotion recognition from combined physiological streams. Because the backbones are pretrained on large unlabeled corpora, teams with limited labeled data can fine-tune for a specific downstream task instead of training a bespoke model from scratch.

#Impact

PhysioWave contributes a reusable, openly licensed architecture that brings learnable time-frequency analysis into the biosignal foundation-model toolkit, establishing new baselines across ECG, EMG, and multi-modal benchmarks and demonstrating that wavelet-based front ends can outperform fixed patch embeddings on non-stationary physiological data. Its acceptance at NeurIPS 2025 and release of training code with multiple model sizes lower the barrier for downstream adoption. Notable limitations include the absence of a published EEG-only pretrained backbone alongside the released ECG and EMG weights, distribution of weights via Google Drive rather than a versioned model hub, and the lack of formal model and data cards.