Weighted-ShapeDBA-4-Rehab

This is the code of our paper "Weighted Average of Human Motion Sequences for Improving Rehabilitation Assessment" [pdf] accepted at the 9th International Workshop on Advanced Analytics and Learning on Temporal Data, at ECML-PKDD 2024
This work was done by Ali Ismail-Fawaz, Maxime Devanne, Stefano Berretti, Jonathan Weber and Germain Forestier.

Requirements

tensorflow
numpy
hydra-core
aeon
scikit-learn
matplotlib

Usage

Create the extended dataset

This code utilizes the hydra configuration setup, simply edit the parameters of the configuration file ('config/config_hydra.yaml`) and run the following:

python main_data_extension.py

It will create the base dataset and extend it using a noisy version (baseline) as well as using the weighted shape DBA for various number of neighbors. The different parameters include:

• original_dir: The directory where the original Kimore dataset is stored
• res_directory: The output directory where the results are stored
• exercise: The exercise to consider
• num_folds: The number of folds per exercise to consider
• num_neighbors_wsdba: The list of number of neighbors used for weighted shape DBA

Train and evaluate a Regressor

This code utilizes the hydra configuration setup, simply edit the parameters of the configuration file ('config/config_hydra.yaml`) and run the following:

python main_regression.py

It creates and trains and avaluates a regressor for all version of the dataset (original + extended version). For now, only the FCN regressor is available. The different parameters include:

• res_directory: The output directory where the results are stored
• exercise: The exercise to consider
• num_folds: The number of folds per exercise to consider
• num_neighbors_wsdba: The list of number of neighbors used for weighted shape DBA
• nb_epochs: Number of epochs to train the regressor
• batch_size: The batch size to use in the regressor

Acknowledgments

This work was supported by the ANR DELEGATION project (grant ANR-21-CE23-0014) of the French Agence Nationale de la Recherche. The authors would like to acknowledge the High Performance Computing Center of the University of Strasbourg for supporting this work by providing scientific support and access to computing resources. Part of the computing resources were funded by the Equipex Equip@Meso project (Programme Investissements d’Avenir) and the CPER Alsacalcul/Big Data. The authors would also like to thank the creators and providers of the Kimore dataset. We would also like to thank the aeon python package for providing an implementation of ShapeDBA that we utilize for our experiments.