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

Official PyTorch implementation of PhysioWave, accepted at NeurIPS 2025

A novel wavelet-based architecture for physiological signal processing that leverages adaptive multi-scale decomposition and frequency-guided masking to advance self-supervised learning

---

🌟 Key Features

✨ Learnable Wavelet Decomposition Adaptive multi-resolution analysis Soft gating mechanism for optimal wavelet selection 📊 Frequency-Guided Masking Novel masking strategy prioritizing high-energy components Superior to random masking for signal representation

🔗 Cross-Scale Feature Fusion Attention-based fusion across decomposition levels Hierarchical feature integration 🧠 Multi-Modal Support Unified framework for ECG and EMG signals Extensible to other physiological signals

📈 Large-Scale Pretraining: Models trained on 182GB of ECG and 823GB of EMG data

---

🏗️ Model Architecture

Pipeline Overview

The PhysioWave pretraining pipeline consists of five key stages:

1. Wavelet Initialization: Standard wavelet functions (e.g., 'db6', 'sym4') generate learnable low-pass and high-pass filters
2. Multi-Scale Decomposition: Adaptive wavelet decomposition produces multi-scale frequency-band representations
3. Patch Embedding: Decomposed features are processed into spatio-temporal patches with FFT-based importance scoring
4. Masked Encoding: High-scoring patches are masked and processed through Transformer layers with rotary position embeddings
5. Reconstruction: Lightweight decoder reconstructs masked patches for self-supervised learning

Core Components

Component	Description
🌊 Learnable Wavelet Decomposition	Adaptively selects optimal wavelet bases for input signals
📐 Multi-Scale Feature Reconstruction	Hierarchical decomposition with soft gating between scales
🎯 Frequency-Guided Masking	Identifies and masks high-energy patches for self-supervised learning
🔄 Transformer Encoder/Decoder	Processes masked patches with rotary position embeddings

---

📊 Performance Highlights

Benchmark Results

Task	Dataset	Metric	Performance
ECG Arrhythmia	PTB-XL	Accuracy	73.1%
ECG Multi-Label	CPSC 2018	F1-Micro	77.1%
ECG Multi-Label	Shaoxing	F1-Micro	94.6%
EMG Gesture	EPN-612	Accuracy	94.5%

Multi-Label Classification Detailed Metrics

CPSC 2018 Dataset (9-Class Multi-Label)

Metric	Micro-Average	Macro-Average
Precision	0.7389	0.6173
Recall	0.8059	0.6883
F1-Score	0.7709	0.6500
AUROC	0.9584	0.9280

Dataset Details:

• 9 official diagnostic classes (SNR, AF, IAVB, LBBB, RBBB, PAC, PVC, STD, STE)
• 12-lead ECG signals at 500 Hz
• Record-level split to prevent data leakage

Chapman-Shaoxing Dataset (4-Class Multi-Label)

Metric	Micro-Average	Macro-Average
Precision	0.9389	0.9361
Recall	0.9536	0.9470
F1-Score	0.9462	0.9413
AUROC	0.9949	0.9930

Dataset Details:

• 4 merged diagnostic classes (SB, AFIB, GSVT, SR)
• 12-lead ECG signals at 500 Hz
• Balanced multi-label distribution

---

💾 Pretrained Models

📥 Download Pretrained Models

Model	Parameters	Training Data	Description
ecg.pth	14M	182GB ECG	ECG pretrained model
emg.pth	5M	823GB EMG	EMG pretrained model

Usage:

# Load pretrained model
checkpoint = torch.load('ecg.pth')
model.load_state_dict(checkpoint['model_state_dict'])

---

🚀 Quick Start

Prerequisites

# Clone repository
git clone https://github.com/ForeverBlue816/PhysioWave.git
cd PhysioWave

# Create conda environment
conda create -n physiowave python=3.11
conda activate physiowave

# Install PyTorch (CUDA 12.1)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# Install requirements
pip install -r requirements.txt

📦 Data Preparation

Dataset Download Links

ECG Datasets

• PTB-XL Database - 21,837 clinical ECG records
• MIMIC-IV-ECG - 800K+ ECG recordings
• PhysioNet Challenge 2021 - Multi-database ECG
• CPSC 2018 - Arrhythmia classification challenge
• Chapman-Shaoxing - Large-scale 12-lead ECG

EMG Datasets

• EPN-612 Dataset - 612 hand gestures
• NinaPro Database DB6 - HD-sEMG recordings

Data Format Specifications

HDF5 Structure

# Single-label classification
{
    'data': (N, C, T),   # Signal data: float32
    'label': (N,)        # Labels: int64
}

# Multi-label classification
{
    'data': (N, C, T),   # Signal data: float32
    'label': (N, K)      # Multi-hot labels: float32
}

Dimensions:

• N = Number of samples
• C = Number of channels
• T = Time points
• K = Number of classes (multi-label only)

Signal Specifications

Signal	Channels	Length	Sampling Rate	Normalization
ECG	12	2048	500 Hz	MinMax [-1,1] or Z-score
EMG	8	1024	200-2000 Hz	Max-abs or Z-score

🔄 Preprocessing Examples

ECG Preprocessing (PTB-XL - Single-Label)

# Download PTB-XL dataset
wget -r -N -c -np https://physionet.org/files/ptb-xl/1.0.3/

# Preprocess for single-label classification
python ECG/ptbxl_finetune.py

Output files:

• train.h5 - Training data with shape (N, 12, 2048)
• val.h5 - Validation data
• test.h5 - Test data

Label format: (N,) with 5 superclasses (NORM, MI, STTC, CD, HYP)

ECG Preprocessing (CPSC 2018 - Multi-Label)

# Preprocess CPSC 2018 dataset
python ECG/cpsc_multilabel.py

Output files:

• cpsc_9class_train.h5 - Training data
• cpsc_9class_val.h5 - Validation data
• cpsc_9class_test.h5 - Test data
• cpsc_9class_info.json - Dataset metadata
• label_map.json - Class mappings
• record_splits.json - Record-level split info

Label format: (N, 9) with 9 official CPSC classes

ECG Preprocessing (Chapman-Shaoxing - Multi-Label)

# Preprocess Chapman-Shaoxing dataset
python ECG/shaoxing_multilabel.py

Output files:

• train.h5 - Training data
• val.h5 - Validation data
• test.h5 - Test data
• dataset_info.json - Metadata
• record_splits.json - Split information

Label format: (N, 4) with 4 merged classes (SB, AFIB, GSVT, SR)

EMG Preprocessing (EPN-612)

# Download from Zenodo and preprocess
python EMG/epn_finetune.py

Output files:

• epn612_train_set.h5 - Training set (N, 8, 1024)
• epn612_val_set.h5 - Validation set
• epn612_test_set.h5 - Test set

Label format: (N,) with 6 gesture classes

---

🎯 Training

Pretraining

ECG Pretraining

# Edit ECG/pretrain_ecg.sh to set data paths
bash ECG/pretrain_ecg.sh

Key parameters:

--mask_ratio 0.7                    # Mask 70% of patches
--masking_strategy frequency_guided # Use frequency-guided masking
--importance_ratio 0.7              # Balance importance vs randomness
--epochs 100                        # Pretraining epochs

EMG Pretraining

# Edit EMG/pretrain_emg.sh to set data paths
bash EMG/pretrain_emg.sh

Key parameters:

--mask_ratio 0.6                    # Mask 60% of patches
--in_channels 8                     # 8-channel EMG
--wave_kernel_size 16               # Smaller kernel for EMG

---

Fine-tuning

Single-Label Classification

Standard Fine-tuning (ECG/EMG)

# ECG fine-tuning (PTB-XL)
bash ECG/finetune_ecg.sh

# EMG fine-tuning (EPN-612)
bash EMG/finetune_emg.sh

Example command:

torchrun --nproc_per_node=4 finetune.py \
  --train_file path/to/train.h5 \
  --val_file path/to/val.h5 \
  --test_file path/to/test.h5 \
  --pretrained_path pretrained/ecg.pth \
  --task_type classification \
  --num_classes 5 \
  --batch_size 16 \
  --epochs 50 \
  --lr 1e-4

Multi-Label Classification

Multi-Label Fine-tuning (CPSC/Shaoxing)

This repository uses finetune_multilabel.py for multi-label classification tasks. First, prepare your data using the corresponding preprocessing scripts.

CPSC 2018 Example:

# Edit paths in ECG/cpsc_multilabel.sh
bash ECG/cpsc_multilabel.sh

Shaoxing Example:

# Edit paths in ECG/shaoxing_multilabel.sh
bash ECG/shaoxing_multilabel.sh

Manual command:

NUM_GPUS=4
torchrun --nproc_per_node=${NUM_GPUS} finetune_multilabel.py \
  --train_file "path/to/train.h5" \
  --val_file "path/to/val.h5" \
  --test_file "path/to/test.h5" \
  --pretrained_path "path/to/pretrained_ecg/best_model.pth" \
  \
  `# Task Configuration` \
  --task_type multilabel \
  --threshold 0.3 \
  \
  `# Model Architecture` \
  --in_channels 12 \
  --max_level 3 \
  --wave_kernel_size 24 \
  --wavelet_names db4 db6 sym4 coif2 \
  --use_separate_channel \
  --patch_size 64 \
  --embed_dim 384 \
  --depth 8 \
  --num_heads 12 \
  --mlp_ratio 4.0 \
  --dropout 0.1 \
  \
  `# Training Parameters` \
  --batch_size 16 \
  --epochs 50 \
  --lr 1e-4 \
  --weight_decay 1e-4 \
  --scheduler cosine \
  --warmup_epochs 5 \
  --grad_clip 1.0 \
  --use_amp \
  \
  `# Classification Head` \
  --pooling mean \
  --head_hidden_dim 512 \
  --head_dropout 0.2 \
  --label_smoothing 0.1 \
  \
  `# Output` \
  --seed 42 \
  --output_dir "./checkpoints_multilabel"

Key Parameters for Multi-Label:

• --task_type multilabel - Enable multi-label classification
• --threshold 0.3 - Decision threshold (adjust based on validation)
• --label_smoothing 0.1 - Regularization for better generalization

Zero-Shot Evaluation

Linear Probing

Evaluate pretrained representations by freezing the encoder and training only the classification head:

torchrun --nproc_per_node=4 finetune.py \
  --train_file path/to/train.h5 \
  --val_file path/to/val.h5 \
  --test_file path/to/test.h5 \
  --pretrained_path pretrained/ecg.pth \
  --freeze_encoder \
  --num_classes 5 \
  --epochs 10 \
  --lr 1e-3

---

🔧 Configuration Guide

Model Configuration

Architecture Parameters

Parameter	Description	Options	Recommendation
--in_channels	Input channels	12 (ECG), 8 (EMG)	Match your data
--max_level	Wavelet decomposition levels	2-4	3 (default)
--wave_kernel_size	Wavelet kernel size	16-32	24 (ECG), 16 (EMG)
--wavelet_names	Wavelet families	db, sym, coif, bior	See tips below
--embed_dim	Embedding dimension	128-768	256/384/512
--depth	Transformer layers	4-12	6/8/12
--num_heads	Attention heads	4-16	8/12
--patch_size	Temporal patch size	20-128	64 (ECG), 32 (EMG)

💡 Wavelet Selection Tips:

Signal Type	Recommended Wavelets	Rationale
ECG	db4 db6 sym4 coif2	Optimal for QRS complex detection
EMG	sym4 sym5 db6 coif3 bior4.4	Best for muscle activation patterns
Custom	Experiment with combinations	Domain-specific optimization

Training Configuration

Pretraining Parameters

Parameter	Description	ECG	EMG
--mask_ratio	Masking ratio	0.7	0.6
--masking_strategy	Masking type	frequency_guided	frequency_guided
--importance_ratio	Importance weight	0.7	0.6
--epochs	Training epochs	100	100
--lr	Learning rate	2e-5	5e-5

Fine-tuning Parameters

Parameter	Description	Default	Range
--batch_size	Batch size per GPU	16	8-64
--epochs	Training epochs	50	20-100
--lr	Learning rate	1e-4	1e-5 to 1e-3
--weight_decay	L2 regularization	1e-4	1e-5 to 1e-3
--scheduler	LR scheduler	cosine	cosine/step/plateau
--warmup_epochs	Warmup epochs	5	0-10
--grad_clip	Gradient clipping	1.0	0.5-2.0

Multi-Label Specific

Parameter	Description	Default	Notes
--threshold	Decision threshold	0.3-0.5	Tune on validation set
--label_smoothing	Label smoothing	0.1	0.0-0.2 for regularization
--use_class_weights	Class balancing	False	Enable for imbalanced data

Hardware and Performance

Performance Tips

# Enable mixed precision for 2x speedup
--use_amp

# Increase batch size with gradient accumulation
--batch_size 8 --grad_accumulation_steps 4  # Effective batch size: 32

# Multi-GPU training
torchrun --nproc_per_node=4 [script.py]

📖 Citation

If you find our work helpful, please cite:

@article{chen2025physiowave,
  title={PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation},
  author={Chen, Yanlong and Orlandi, Mattia and Rapa, Pierangelo Maria and Benatti, Simone and Benini, Luca and Li, Yawei},
  journal={arXiv preprint arXiv:2506.10351},
  year={2025}
}

---

🤝 Contact & Contributions

Lead Author: Yanlong Chen
Email: yanlchen@student.ethz.ch

We welcome contributions! Feel free to:

• 🐛 Report bugs
• 💡 Suggest enhancements
• 🔧 Submit Pull Requests
• ⭐ Star this repository if you find it useful!

---

🙏 Acknowledgments

We thank:

• The authors of PTB-XL, MIMIC-IV-ECG, CPSC 2018, Chapman-Shaoxing, and EPN-612 datasets
• The PyTorch team for their excellent framework
• The open-source community for inspiration and tools

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

_{Built with ❤️ for the physiological signal processing community}