Research map / Vision-language and multimodal / Cross-modal learning
PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PhysioWave, a learnable wavelet-decomposition Transformer with frequency-guided masking, pretrains EMG and ECG foundation models, and fuses them with frozen EEG encoders for multimodal physiological classification.
Problem
Physiological signals (EMG, ECG, EEG) are non-stationary, noisy, and heterogeneous across subjects and devices, and there are few pretrained foundation models for EMG and ECG; standard time-domain/Fourier front-ends and random masking are ill-suited.
Approach
A learnable wavelet front-end (adaptive selection over multiple candidate wavelet bases, soft gating, and cross-scale attention fusion) feeds a masked-patch Transformer; Frequency-guided Masking masks high spectral-energy patches (blended with noise) rather than random patches.
Key results
- PhysioWave-ecg Large (37M) reaches 66.7% F1 / 94.6% AUROC on PTB-XL, surpassing the prior best reported in its table.
- PhysioWave-emg Large (37M) reaches 94.50% accuracy on EPN-612, +3.6% over Moment, using under 1/10 of Moment's parameters.
- Multimodal fusion with frozen EEGPT and PhysioWave ECG/EMG encoders raises DEAP valence accuracy from 79.1% to 85.2% and arousal from 81.3% to 88.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase model capacity from Small to Large | PhysioWave-ecg · PTB-XL | F1: 65.8 -> 66.7 | Helped |
| add PhysioWave ECG and EMG branches to EEGPT | EEGPT on DEAP · DEAP valence | accuracy: 79.1% to 85.2% (+6.1%) | Helped |
| add PhysioWave ECG and EMG branches to EEGPT | EEGPT on DEAP · DEAP arousal | accuracy: 81.3% to 88.6% (+7.3%) | Helped |
| add PhysioWave EMG and ECG branches to LaBraM | LaBraM on MPDB · MPDB | accuracy: 70.4% to 74.9% (+4.5%) | Helped |
| add Frequency-guided Masking (ablation) | PhysioWave-emg · EPN-612 | F1: 92.85 -> 93.67; 0.8% | Helped |
| add large-scale self-supervised pretraining (ablation) | PhysioWave-emg · EPN-612 | F1: 91.57 -> 93.67; 2.1% | Helped |
| add Soft Gate, CAFFN, and Adaptive Wavelet together (ablation) | PhysioWave · EPN612 | Accuracy: 89.88 -> 93.85 | Helped |
| use Adaptive Wavelet instead of the fixed-wavelet baseline (ablation) | PhysioWave · EPN612 | Accuracy: +2.35% | Helped |
| use Soft Gate instead of Hard Gate (threshold 0.5) (ablation) | PhysioWave · EPN612 | Accuracy: 92.25 -> 93.85; ↓1.60% | Helped |
| use CrossScale CAFFN instead of Standard FFN (ablation) | PhysioWave · EPN612 | Accuracy: 91.95 -> 93.85; ↓1.90% | Helped |
| use multi-wavelet combination db4+db6+sym4 instead of Random Xavier (ablation) | PhysioWave · EPN612 EMG | Accuracy: 92.25 -> 93.90; ↑1.65% | Helped |
| use multi-wavelet combination db6+coif4+sym8 instead of Random Xavier (ablation) | PhysioWave · Georgia ECG | Accuracy: 62.38 -> 64.38; ↑2.00% | Helped |
Surprising findings
- The multi-wavelet combinations (e.g., db4+db6+sym4) outperform single wavelets, and the optimal wavelet combination differs between EMG and ECG.
Limitations
- Multimodal results rely on frozen encoders with linear probing; end-to-end fusion is not explored.
- Several appendix experiments (EEG, PPG) are reported with unclear or inconsistent details and comparisons to baselines in different settings.
- Reported ablation numbers are inconsistent across tables (e.g., the full model is 93.12% in one table and 93.85% in another).
Experimental setup
Datasets: MIMIC-IV-ECG NinaPro DB5 EPN-612 UCI EMG PTB-XL CPSC 2018 Chapman-Shaoxing DEAP MPDB Sleep-EDF PPG-DaLiA Georgia ECG
Benchmarks: PTB-XL CPSC 2018 Chapman-Shaoxing NinaPro DB5 EPN-612 UCI EMG DEAP MPDB Sleep-EDF PPG-DaLiA Georgia ECG
Metrics: F1 AUROC accuracy balanced accuracy Cohen's kappa MAE MSE MAPE
Baselines: ECG-Chat MERL MaeFE OpenECG-SimCLR OpenECG-BYOL OpenECG-MAE Moment OTiS EEGPT LaBraM Pulse-PPG BENDR
Builds on: masked autoencoder SpecAugment HuBERT rotary positional embeddings PyWavelets
Closest papers by meaning
- FEEL: Quantifying Heterogeneity in Physiological Signals for Generalizable Emotion Recognition (NeurIPS 2025 · similarity 0.83)
- PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement (AAAI 2025 · similarity 0.82)
- Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (ICML 2025 · similarity 0.82)
- Cylindrical Geodesic Flow Matching for Quasiperiodic Physiological Signal Transformation (NeurIPS 2026 · similarity 0.82)
- Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation (AAAI 2024 · similarity 0.82)
- PhysDrive: A Multimodal Remote Physiological Measurement Dataset for In-vehicle Driver Monitoring (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.