Transformer Encoder with Multiscale Deep Learning for Pain Classification Using Physiological Signals

Our new paper just released on Frontiers in Physiology: https://www.frontiersin.org/articles/10.3389/fphys.2023.1294577/full

Abstract

Pain is a global health problem impacting a significant proportion of the population. Accurate assessments are crucial for effective pain management and treatment. Current methods, \eg self-report scales, can be biased and inconsistent. Therefore, there is a need for objective and automatic pain intensity evaluation. In this paper, we propose PainAttnNet, a novel transformer-encoder deep learning framework for classifying pain intensities using physiological signals. PainAttnNet integrates multiscale convolutional networks to capture short- and long-window sequential features, a squeeze-and-excitation residual network emphasizing interdependencies among relevant features, and a transformer encoder block for optimal feature extraction and temporal dependency analysis. Utilizing the publicly available BioVid heat pain database, PainAttnNet outperforms state-of-the-art models, demonstrating its potential for automated pain intensity classification. Our results suggest that PainAttnNet can enhance pain assessment using physiological signals benefitting further pain management and treatment.

Directory Structure

PainAttnNet
|   environment.yml # Requirements for conda environment
|   LICENSE
|   README.md
|   requirments.txt # Requirements for pip environment
|          
+---jq
|       jq-win64.exe
|       
\---src
    |   batch_train.sh # Training script
    |   config.json # Training configurations
    |   logger_config.json # Logger configurations
    |   parser.py # Parser for training configurations
    |   train_kfold_cv.py # Main training script
    |  __init__.py
    |   
    +---models
    |   |   main_painAttnNet.py # Main model wrapper
    |   |   module_mscn.py* # Multiscale convolutional network
    |   |   module_se_resnet.py # Squeeze-and-excitation residual network
    |   |   module_transformer_encoder.py # Transformer encoder block
    |   \   __init__.py
    |           
    +---tests # Unit tests
    |   |   test_generate_kfolds_index.py
    |   |   test_mscn.py
    |   |   test_PainAttnNet_output.py
    |   |   test_process_bioVid.py
    |   |   test_se_resnet.py
    |   |   test_transformer_encoder.py
    |   \   __init__.py
    |           
    +---trainers # Training modules
    |   |   checkpoint_handler.py # Checkpoint handler
    |   |   device_prep.py # Device preparation, CPU or GPU
    |   |   main_trainer.py # Main trainer scripts
    |   |   metrics_manager.py # Metrics manager and other metrics functions
    |   \   __init__.py
    |           
    \---utils
        |   process_bioVid.py # Data processing for BioVid
        |   utils.py # Other utility functions
        \   __init__.py

Updates

update 12/11/2023 Paper published on Frontiers in Physiology.

update 4/21/2023 Full code is released.

update 4/20/2023 The strutrue of MSCN currently in the code has three convolutional layers for three scales (slightly different from the paper), but this can be easily changed by modifying the code.

Get Started

torchaudio==0.13.0
python==3.10.8
pytorch-cuda==11.7
pytorch==1.13.0
torchvision==0.14.0
scikit-learn==1.0.1
pandas
matplotlib
openpyxl

For Linux users, install jq package via conda or pip.

For Windows users, install jq package from here, and put the jq.exe file in the local directory.

Training

Training k-fold cross validation with script

sh batch_train.sh

Training individual fold in terminal

python train_kfold_cv.py --fold_i {fold index}

You can change settings at main_painAttnNet.py for tuning model structure, config.py for training configurations and train_kfold_cv.py for others.

Dataset

The dataset is available at BioVid Heat Pain Database.

Reference

If you find our study useful, please cite our paper on Frontiers

@article{10.3389/fphys.2023.1294577,
AUTHOR={Lu, Zhenyuan and Ozek, Burcu and Kamarthi, Sagar},   
TITLE={Transformer encoder with multiscale deep learning for pain classification using physiological signals},      
JOURNAL={Frontiers in Physiology},      
VOLUME={14},           
YEAR={2023},      
URL={https://www.frontiersin.org/articles/10.3389/fphys.2023.1294577},       
DOI={10.3389/fphys.2023.1294577},      
ISSN={1664-042X},
}

Contact

Zhenyuan Lu
Northeastern University, Boston
Email: lu.zhenyua[at]northeastern[dot]edu