This repository contains de-identified IMU walking data from seven subjects in in-clinic experiments. It also contains code to pre-process, train, and evaluate convolutional neural networks that identify freezing of gait in continuous inertial measurement unit data. The intent of this repository is to provide other researchers tools to develop and train their own FOG detection models.

Currently, this repo can only be built from source. To do so,

1. Clone the repo:

   git clone https://github.com/stanfordnmbl/imu-fog-detection
   cd imu-fog-detection  
   

2. Download the prerequisite packages:

   pip install -r requirements.txt
   

All raw IMU data is formatted in .xslx files (see data/raw). One file exists for each walking trial. The columns in each data file are:

Accelerometer (ax, ay, az) units are m/s², and gyroscope (gx, gy, gz) units are rad/s. A configuration file is necessary to indicate sensor configurations (see data/imus6_subjects7_configs.xlsx). Column names in the configuration file must be substrings of column names in the raw IMU data files. If the code base will be used, unaltered, .xlsx files will need to be formatted as stated above.

This script creates windowed examples and trains models across all sensor sets specified in the configuration file. Models, weights, and classification metrics are saved to a results directory. Note, on a standard laptop, this could take several hours per sensor set. Run with the commands:

cd code
python experiments.py

This script aggregates classification metrics (area under the receiver operating characteristic and average precision) across sensor sets and generates a summary figure. Top-ranked participant sets, as identified through the survey described in the accompanying paper, are also indicated. Run with the commands:

cd code
python plots.py

Example of summary figure of classification metrics from [experiments.py](code/experiments.py).

This script computes clinical metrics (percent time FOG and number of FOG events) from specified models. It then computes the intraclass correlation coefficient between the ground truth (human) and model-based clinical metrics. A summary figure is generated, and results are saved to the appropriate results directory. Note, on a standard laptop, this could take up to about 30 minutes per sensor set. Run with the commands:

cd code
python clinical_metrics.py

Example of summary figure of clinical metrics from human and model generated by [clinical_metrics.py](code/clinical_metrics.py).

The following constants are found across our code and may be important to consider in your own use:

• In datapreprocessing.py:

  • FREQ_SAMPLED and FREQ_DESIRED, for resampling purposes.

  • N_AUGMENTATION, the number of times to augment each example.

  • N_AXES and N_DATA_STREAMS, dependent on the data streams included from the IMUs.

• In experiments.py:

  • WINDOW_OVERLAPS, a parameter that can be increased to generate more examples or decreased to generate fewer examples.

Please cite our paper in your publications if our repository helps your research.

O’Day, J., Lee, M., Seagers, K. et al. Assessing inertial measurement unit locations for freezing of gait detection and patient preference. J NeuroEngineering Rehabil 19, 20 (2022). https://doi.org/10.1186/s12984-022-00992-x

• Johanna O’Day ([email protected])

• Marissa Lee ([email protected])

• Kirsten Seagers ([email protected])

Distributed under the BSD 3-clause License. See LICENSE for more information.