SEAR: Speed estimation algorithm for running

Estimate running speed from raw wrist or waist-worn accelerometer data

Please cite:

Davis, J., Oeding, B., and Gruber, A., 2022. Estimating Running Speed From Wrist- or Waist-Worn Wearable Accelerometer Data: A Machine Learning Approach. Journal for the Measurement of Physical Behaviour, ahead of print. DOI: 10.1123/jmpb.2022-0011

Downloading the data

To reproduce the paper results you will need to download the data from FigShare DOI: 10.6084/m9.figshare.21507180

The data should be unzipped to the /raw data/ folder.

Reproducing the paper

First: run extract_features.py to extract features from raw 10sec windows. These are saved in /feature exports/

Second: run main.py to run nested subject-wise cross-validation. This writes result file to /cross-validation results and prediction results.

For speed, main.py is set up to reproduce the ridge model results with fewer random searches of the hyperparameter space than the paper (here, 200; paper is 1000). Change "ridge" to "xgb" in the code, and n_search_list to be [50]*2 to reproduce the xgboost model results (warning, this could take ~12hrs or more on a fast desktop computer).

Using on new data

(Demo forthcoming)

1. Resample to 100 Hz if necessary
2. Calculate resultant acceleration, in g-units.
3. Extract bouts of running (perhaps using a running recognition algorithm)
4. Window data into non-overlapping 10 second, 100 Hz windows and reshape into an n x 1000 numpy array. The helper function in sear_features.py might be useful here.
5. Extract matrix of features using sear_features.extract_features()
6. Feed into pre-trained algorithm using `.predict()`` method