PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation

Conference: NeurIPS 2025 arXiv: 2506.10351 Code: Available Area: Medical Imaging / Biosignal Processing Keywords: Wavelet transform, physiological signals, self-supervised learning, multimodal fusion, foundation model

TL;DR

This paper proposes PhysioWave, a multi-scale Transformer architecture based on learnable wavelet decomposition and frequency-guided masking. It establishes, for the first time, large-scale pretrained foundation models for EMG and ECG, and achieves state-of-the-art performance on both unimodal and multimodal physiological signal tasks through a multimodal fusion framework.

Background & Motivation

Physiological signals (EEG, EMG, ECG) are core data sources for health monitoring, clinical diagnosis, and brain-computer interfaces, yet they present three major challenges: (1) low signal-to-noise ratio: severe interference from motion artifacts, baseline drift, etc.; (2) strong non-stationarity: signals contain spikes and abrupt transitions that conventional time-domain or fixed-window Fourier methods cannot effectively capture; (3) cross-modal heterogeneity: significant differences in sampling rates and dimensionality across modalities.

Although pretrained models such as LaBraM and EEGPT exist for EEG, foundation models for EMG and ECG remain absent. Existing NLP-inspired self-supervised methods (e.g., random token masking) are ill-suited for physiological signals—raw signal segments do not correspond to meaningful units like words, and random dropping may remove critical events or mask redundant regions.

Method

Overall Architecture

The pretraining pipeline of PhysioWave consists of four stages: (1) learnable wavelet decomposition decomposes raw multi-channel signals into multi-scale frequency-band representations; (2) frequency-guided masking (FgM) selectively masks high-information patches based on FFT energy; (3) a Transformer encoder processes the token sequence; (4) a lightweight decoder reconstructs the masked patches.

Key Designs

1. Adaptive Wavelet Selector: Maintains \(M\) candidate wavelet bases \(\{(k_w^{\text{low}}, k_w^{\text{high}})\}_{w=1}^M\) and computes selection weights via MLP + Softmax: \(\alpha = \text{Softmax}(\text{MLP}(\text{AvgPool}(x)))\), adaptively combining filters as \(k^{\text{low}} = \sum_w \alpha_w k_w^{\text{low}}\). The design motivation is that different signal characteristics require different wavelet bases, and manual selection cannot accommodate diverse signals.

2. Soft-Gated Multi-Resolution Analysis: After each decomposition layer, the signal is upsampled back to the original length, and multi-head attention estimates an adaptive gate \(G_c^{(\ell)} \in [0,1]\) that dynamically weights the current-layer signal and the upsampled signal: \(\hat{a}_c^{(\ell)}[n] = G_c^{(\ell)} a_c^{(\ell)} + (1-G_c^{(\ell)}) \tilde{a}_c^{(\ell+1)}\). Compared to U-Net's hard skip connections, soft gating enables per-channel adjustment of frequency content emphasis, reducing aliasing and ringing artifacts.

3. Cross-Scale Channel-Aggregating Feed-Forward Network (Cross-Scale CAFFN): Wavelet features at each layer are fused across scales via channel aggregation and multi-head attention, where the current-layer features serve as queries and shallower features serve as keys/values: \(Y^{(\ell)} = U^{(\ell)} + \beta \cdot \text{Attention}(U^{(\ell)}, \{Y^{(i)}\}_{i<\ell})\), enabling fine-grained subband features to incorporate long-range patterns from coarser resolutions.

4. Frequency-guided Masking (FgM): The FFT spectral energy of each patch is computed and mixed with random noise to yield an importance score \(s_n = \alpha \cdot e_n + (1-\alpha) \cdot z_n\), prioritizing high-energy patches for masking. This forces the model to infer critical information from context, producing richer discriminative features compared to random time-domain masking.

Loss & Training

• Pretraining loss: Smooth-L1 reconstruction loss computed only on masked patches: \(\mathcal{L} = \frac{1}{|\mathcal{M}|} \sum_{n \in \mathcal{M}} \text{SmoothL1}(\hat{p}_n, p_n)\)
• Unimodal downstream: End-to-end fine-tuning; mean pooling followed by a two-layer MLP classifier.
• Multimodal downstream: Pretrained encoders for each modality are frozen; only the classification head and softmax-constrained fusion weights are trained: \(z_{\text{fused}} = \sum_{m \in \mathcal{M}} \alpha_m z_m\)

Key Experimental Results

Main Results

ECG Rhythm Classification (PTB-XL)

Method	Params	F1 (%)	AUROC (%)
ECG-Chat (2024)	13B	55.9	94.1
MaeFE (2023)	9M	64.7	88.6
PhysioWave-Large	37M	66.7	94.6

EMG Gesture Recognition (EPN-612)

Method	Params	Acc (%)	F1 (%)
Moment (2024)	385M	90.87	90.16
OTiS (2024)	45M	87.55	88.03
PhysioWave-Large	37M	94.50	94.56

Ablation Study

Configuration	Training Loss	Acc (%)	F1 (%)
w/o FgM (random masking)	0.24	92.48	92.85
w/o pretraining	0.27	91.67	91.57
Full model	0.22	93.12	93.67

Key Findings

• PhysioWave-Large achieves 66.7% F1 on PTB-XL, surpassing the 13B-parameter ECG-Chat.
• The EMG model outperforms Moment across three benchmarks using fewer than one-tenth of its parameters.
• Multimodal fusion yields a +7.3% accuracy improvement on DEAP emotion recognition (81.3%→88.6%).
• FgM improves F1 by 0.8% over random masking; pretraining contributes an additional 2.1% F1 gain.

Highlights & Insights

• The first large-scale pretrained foundation models for EMG and ECG, trained on 823 GB of EMG and 182 GB of ECG data respectively, filling a critical gap in the field.
• Learnable wavelet decomposition elegantly addresses the multi-scale and non-stationary characteristics of physiological signals, proving highly effective as a general-purpose front-end.
• The FgM strategy upgrades masking from "random dropout" to "information-guided selection," aligning well with the non-uniform information distribution inherent in physiological signals.
• The multimodal framework adopts a frozen-encoder plus lightweight-fusion design, mitigating overfitting risks on small downstream datasets.

Limitations & Future Work

• Other modalities such as optical biosensors are not yet supported.
• The multimodal fusion relies on simple weighted summation; more sophisticated attention-based fusion may yield further improvements.
• Pretraining data are biased toward Western populations, and generalization across different ethnic and age groups remains to be validated.
• Real-time inference efficiency is not thoroughly discussed, which may pose challenges for edge device deployment.

Related Work & Insights

• Relation to LaBraM / EEGPT: PhysioWave extends the pretraining paradigm from EEG to EMG and ECG, providing a unified multimodal framework.
• Comparison with Moment / OTiS: General-purpose time-series foundation models underperform domain-specific models on physiological signals, underscoring the importance of domain-specific design choices such as the wavelet front-end.
• Insight: The combination of learnable wavelet decomposition and frequency-guided masking may generalize to other non-stationary signal analysis tasks, such as seismic or acoustic signal processing.

Rating

• Novelty: ⭐⭐⭐⭐ (The combination of learnable wavelets and FgM is novel, though individual components have prior precedents.)
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Three modalities, 6+ datasets, comprehensive ablations.)
• Writing Quality: ⭐⭐⭐⭐ (Clear structure with detailed formulations.)
• Value: ⭐⭐⭐⭐⭐ (Fills a critical gap in EMG/ECG foundation models with high practical applicability.)

Related Papers

• [NeurIPS 2025] Unpaired Image-to-Image Translation for Segmentation and Signal Unmixing
• [NeurIPS 2025] STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology
• [NeurIPS 2025] LoMix: Learnable Weighted Multi-Scale Logits Mixing for Medical Image Segmentation
• [CVPR 2026] MUST: Modality-Specific Representation-Aware Transformer for Diffusion-Enhanced Survival Prediction with Missing Modality
• [NeurIPS 2025] Domain-Adaptive Transformer for Data-Efficient Glioma Segmentation in Sub-Saharan MRI