/new_mgsiameseTS

Primary LanguageJupyter Notebook

Tutorial for development system

For processed data, please refer the 'RRT_processed_data.zip' files

1. Install requirement package base on the tutorial in 'requirement.txt'

2. Training

  • Refer 'Detection_train.ipynb'.

  • Change the directory based on your specific data directory.

  • Train process includes:

    (a) Window Interval Processing (ts_dataloader) (refer 'dataloader.py')

      +  For 'ts_dataloader' function, you can select your specific input features ('lab_list', 'sign_list', 'dem_list').

  + You can also select your specific task corresponding each label column.(Tasks colum's name includes: 'label' (normal or abnormal), 'is_detection', 'is_event', 'ev_w_dec').

  + You can also select 'window_len' (history data using for prediction); 'stride' (Number of future timesteps we want to predict).

  + Finally, we have sequences process data for training DL models.

  Specifically for SiameseTS_model, It should includes 2 type of input features: 'x_t' (time based features); 'x_d' (non-time based features).

    (b) Training

      + Change your specific training folder to save the trained models.

  + Select model (refer 'model.py')

  Example: 'model = SiamseTS_model(x_t.shape[1:], x_d.shape[1:], num_classes=num_classes)'

    You can design your training strategy to improve the model's performance.

3. Evaluation

  • Refer 'Detection_evaluation.ipynb'.

  • Change the directory based on your specific data directory.

  • Using ts_dataloader for test data. Input features list and task should be similar with the trained model.

  • Get prediction and evaluation (refer 'eval.py')

4. Publication:

Nguyen, T. N., Kim, S. H., Kho, B. G., & Yang, H. J. (2024). Multi-Gradient Siamese Temporal Model for the Prediction of Clinical Events in Rapid Response Systems. IEEE Intelligent Systems. DOI: 10.1109/MIS.2024.3408290