BIOT: Cross-data Biosignal Learning in the Wild

[Model] We first resample biosignals into the same frequency, such as 200Hz. Then, we transform biosignals of different channels, variable lengths, and with missing values into the consistent sentence structures. Next, we apply a transformer model to encode the sentence embeddings.

[Applications] Our model can be used for standard supervised learning, supervised learning with missing values, (supervised and unsupervised) pre-training on multiple data sources with different formats and fine-tuning on similar new datasets.

0. Quick Start

• Understand the usage of each model (input and output)

  # python run_example.py SPaRCNet
  # python run_example.py ContraWR
  # python run_example.py CNNTransformer
  # python run_example.py FFCL
  # python run_example.py STTransformer
  python run_example.py BIOT
  # python run_example.py BIOT-pretrain-PREST
  # python run_example.py BIOT-pretrain-SHHS+PREST
  # python run_example.py BIOT-pretrain-six-datasets 

1. Folder Structures

• datasets/: contains the dataset processing scripts for

  • The TUH Abnormal EEG Corpus (TUAB): 400K sample, 256Hz, 10 seconds per sample, 16 montages. https://isip.piconepress.com/projects/tuh_eeg/html/downloads.shtml
  • The TUH EEG Events Corpus (TUEV): 110K samples, 256Hz, 10 seconds per sample, 16 montages. https://isip.piconepress.com/projects/tuh_eeg/html/downloads.shtml
  • CHB-MIT: 326K samples, 256Hz, 10 seconds per sample, 16 montages. https://physionet.org/content/chbmit/1.0.0/
  • Sleep Heart Health Study (SHHS): 5M samples, 125Hz, 30 seconds per sample. https://sleepdata.org/datasets/shhs

• models/: contains the model scripts for

  • SPaRCNet [1]: can refer to the implementation in https://pyhealth.readthedocs.io/en/latest/api/models/pyhealth.models.SparcNet.html
  • ContraWR [2]: can refer to the implementation in https://pyhealth.readthedocs.io/en/latest/api/models/pyhealth.models.ContraWR.html or official github in https://github.com/ycq091044/ContraWR
  • CNNTransformer
  • FFCL
  • STTransformer [3]: can refer to https://github.com/eeyhsong/EEG-Transformer
  • Our BIOT Model: BIOTEncoder, BIOTClassifier, UnsupervisedPretrain, SupervisedPretrain

• pretrained-models/: contains the pretrained EEG models

  [1] Jing, J., Ge, W., Hong, S., Fernandes, M. B., Lin, Z., Yang, C., An, S., Struck, A. F., Herlopian, A., Karakis, I., et al. (2023). Development of expert-level classification of seizures and rhythmic and periodic patterns during eeg interpretation. Neurology.

[2] Yang, C., Xiao, D., Westover, M. B., and Sun, J. (2021). Self-supervised eeg representation learning for automatic sleep staging. arXiv preprint arXiv:2110.15278.

[3] Song, Y., Jia, X., Yang, L., and Xie, L. (2021). Transformer-based spatial-temporal feature learning for eeg decoding. arXiv preprint arXiv:2106.11170.

2. EEG Pre-trained models (all re-sampled to 200Hz)

• EEG-PREST-16-channels.ckpt: pretrained model on 5 millions of resting EEG samples from Massachusetts General Hospital (MGH) EEG corpus. The sample size is 16 montages x 2000 time points. The 16 channels are "FP1-F7", "F7-T7", "T7-P7", "P7-O1", "FP2-F8", "F8-T8", "T8-P8", "P8-O2", "FP1-F3", "F3-C3", "C3-P3", "P3-O1", "FP2-F4", "F4-C4", "C4-P4", "P4-O2".
• EEG-SHHS+PREST-18-channels.ckpt: pretrained model on 5 millions samples from MGH EEG corpus and 5 millions of sleep EEG from SHHS. The SHHS sample size is 2 channels x 6000 time points. The 18 channels are the above 16 channels plus "C3-A2" and "C4-A1".
• EEG-six-datasets-18-channels.ckpt: pretrained on 5M MGH EEG samples, 5M SHHS, and the training sets of TUAB, TUEV, CHB-MIT, and IIIC Seizure (requested from [1]). The same 18 channels as above.

How to use the pretrained models:

• start wtih run_example.py to understand how to use these DL models

To run your own pretrained models:

• use run_supervised_pretrain.py and run_unsupervised_pretrain.py for running the pre-training.

3. Performances on TUAB

• The first six models are trained from scratch. The last three models used the pre-trained BIOT.

Reference Runs

python run_binary_supervised.py --dataset TUAB --in_channels 16 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 10 --batch_size 512 --model BIOT
python run_binary_supervised.py --dataset TUAB --in_channels 16 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 10 --batch_size 512 --model BIOT --pretrain_model_path pretrained-models/EEG-PREST-16-channels.ckpt
python run_binary_supervised.py --dataset TUAB --in_channels 18 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 10 --batch_size 512 --model BIOT --pretrain_model_path pretrained-models/EEG-SHHS+PREST-18-channels.ckpt
python run_binary_supervised.py --dataset TUAB --in_channels 18 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 10 --batch_size 512 --model BIOT --pretrain_model_path pretrained-models/EEG-six-datasets-18-channels.ckpt

4. Performance on TUEV

• The first six models are trained from scratch. The last three models used the pre-trained BIOT.

Reference Runs

python run_multiclass_supervised.py --dataset TUEV --in_channels 16 --n_classes 6 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 5 --batch_size 128 --model BIOT
python run_multiclass_supervised.py --dataset TUEV --in_channels 16 --n_classes 6 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 5 --batch_size 128 --model BIOT --pretrain_model_path pretrained-models/EEG-PREST-16-channels.ckpt
python run_multiclass_supervised.py --dataset TUEV --in_channels 18 --n_classes 6 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 5 --batch_size 128 --model BIOT --pretrain_model_path pretrained-models/EEG-SHHS+PREST-18-channels.ckpt
python run_multiclass_supervised.py --dataset TUEV --in_channels 18 --n_classes 6 --sampling_rate 200 --token_size 200 --hop_length 100 --sample_length 5 --batch_size 128 --model BIOT --pretrain_model_path pretrained-models/EEG-six-datasets-18-channels.ckpt

5. Citations

@inproceedings{yang2023biot,
    title={BIOT: Biosignal Transformer for Cross-data Learning in the Wild},
    author={Yang, Chaoqi and Westover, M Brandon and Sun, Jimeng},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
    year={2023},
    url={https://openreview.net/forum?id=c2LZyTyddi}
}
@article{yang2023biot,
  title={BIOT: Cross-data Biosignal Learning in the Wild},
  author={Yang, Chaoqi and Westover, M Brandon and Sun, Jimeng},
  journal={arXiv preprint arXiv:2305.10351},
  year={2023}
}