A step-counting model based on self-supervised learning for wrist-worn accelerometer data.

The SSL model was pre-trained using the large-scale UK Biobank Accelerometer Dataset, and fine-tuned on the OxWalk Dataset.

The command-line tool can process Axivity AX3 files (UK Biobank, China Kadoorie Biobank) directly. For consumer devices like Fitbit and Apple Watch, convert them to raw CSV first.

Available models:

• Self-supervised learning model of Hang et al. (default): https://www.nature.com/articles/s41746-024-01062-3
• Random forest (enable with the flag -t rf)

Minimum requirements: Python>=3.8, Java 8 (1.8)

The following instructions make use of Anaconda to meet the minimum requirements:

1. Download & install Miniconda (light-weight version of Anaconda).
2. (Windows) Once installed, launch the Anaconda Prompt.
3. Create a virtual environment:

   $ conda create -n stepcount python=3.9 openjdk pip

   This creates a virtual environment called stepcount with Python version 3.9, OpenJDK, and Pip.
4. Activate the environment:

   $ conda activate stepcount

   You should now see (stepcount) written in front of your prompt.
5. Install stepcount:

   $ pip install stepcount

You are all set! The next time that you want to use stepcount, open the Anaconda Prompt and activate the environment (step 4). If you see (stepcount) in front of your prompt, you are ready to go!

Check out the 5-minute video tutorial to get started: https://www.youtube.com/watch?v=FPb7H-jyRVQ.

# Process an AX3 file
$ stepcount sample.cwa

# Or an ActiGraph file
$ stepcount sample.gt3x

# Or a GENEActiv file
$ stepcount sample.bin

# Or a CSV file (see data format below)
$ stepcount sample.csv

Output:

Summary
-------
{
    "Filename": "sample.cwa",
    "Filesize(MB)": 65.1,
    "Device": "Axivity",
    "DeviceID": 2278,
    "ReadErrors": 0,
    "SampleRate": 100.0,
    "ReadOK": 1,
    "StartTime": "2013-10-21 10:00:07",
    "EndTime": "2013-10-28 10:00:01",
    "TotalWalking(min)": 655.75,
    "TotalSteps": 43132,
    ...
}

Estimated Daily Steps
---------------------
              steps
time
2013-10-21     5368
2013-10-22     7634
2013-10-23    10009
...

Output: outputs/sample/

Refer to the GLOSSARY.md for a comprehensive list of outputs.

Some systems may face issues with Java when running the script. If this is your case, try fixing OpenJDK to version 8:

$ conda install -n stepcount openjdk=8

By default, output files will be stored in a folder named after the input file, outputs/{filename}/, created in the current working directory. You can change the output path with the -o flag:

$ stepcount sample.cwa -o /path/to/some/folder/

The following output files will be generated:

• Info.json Summary info, as shown above.
• Steps.csv Raw time-series of step counts
• Minutely.csv Minutely summaries
• Hourly.csv Hourly summaries
• Daily.csv Daily summaries

By default, the stepcount tool employs a self-supervised Resnet18 model to detect walking periods. However, it is possible to switch to a random forest model, by using the -t flag:

$ stepcount sample.cwa -t rf

When using the random forest model, a set of signal features is extracted from the accelerometer data. These features are subsequently used as inputs for the model's classification process. For a comprehensive list of the extracted features, see the glossary.

Adjusted estimates are provided that account for missing data. Missing values in the time-series are imputed with the mean of the same timepoint of other available days. For adjusted totals and daily statistics, 24h multiples are needed and will be imputed if necessary. Estimates will be NaN where data is still missing after imputation.

If a CSV file is provided, the following header is expected: time, x, y, z.

Example:

time,x,y,z
2013-10-21 10:00:08.000,-0.078923,0.396706,0.917759
2013-10-21 10:00:08.010,-0.094370,0.381479,0.933580
2013-10-21 10:00:08.020,-0.094370,0.366252,0.901938
2013-10-21 10:00:08.030,-0.078923,0.411933,0.901938
...

If the CSV file has a different header, use the option --txyz to specify the time and x-y-z columns, in that order. For example:

HEADER_TIMESTAMP,X,Y,Z
2013-10-21 10:00:08.000,-0.078923,0.396706,0.917759
2013-10-21 10:00:08.010,-0.094370,0.381479,0.933580
2013-10-21 10:00:08.020,-0.094370,0.366252,0.901938
2013-10-21 10:00:08.030,-0.078923,0.411933,0.901938
...

then use:

$ stepcount my-file.csv --txyz HEADER_TIMESTAMP,X,Y,Z

To process multiple files you can create a text file in Notepad which includes one line for each file you wish to process, as shown below for file1.cwa, file2.cwa, and file2.cwa.

Example text file commands.txt:

stepcount file1.cwa &
stepcount file2.cwa &
stepcount file3.cwa 
:END

Once this file is created, run cmd < commands.txt from the terminal.

Create a file command.sh with:

stepcount file1.cwa
stepcount file2.cwa
stepcount file3.cwa

Then, run bash command.sh from the terminal.

A utility script is provided to collate outputs from multiple runs:

$ stepcount-collate-outputs outputs/

This will collate all *-Info.json files found in outputs/ and generate a CSV file.

Validation for this algorithm is presented in a preprint on medRxiv at: https://www.medrxiv.org/content/10.1101/2023.02.20.23285750v1.

If you would like to contribute to this repository, please check out CONTRIBUTING.md. We welcome contributions in the form of bug reports, feature requests, and pull requests.

When using this tool, please consider citing the works listed in CITATION.md.

See LICENSE.md.

We would like to thank all our code contributors, manuscript co-authors, and research participants for their help in making this work possible.