Open gcosne opened 3 years ago
Hey Gautier,
Thank you for your interest. The classification model was trained on several internal (not open-source) datasets of healthy volunteers and Parkinson's disease patients (approximately 140 total subjects). Positive labels included an in-lab walking task and negative labels included a variety of other activities of daily living. The model was trained using data from a lumbar-mounted accelerometer so I would urge caution trusting it on data not acquired at this location. Based on the dataset you are interested in, you may not need to use the classification function if the data is labeled appropriately. However, the extract_features function was also developed and tuned on data from a lumbar-mounted accelerometer so, again, I would urge caution when interpreting results from other sensor locations. I hope that helps and best of luck on your analysis.
Best, Matt
Thank you for your clarification, as it is always complicated to agree to use a model without understanding how it was trained, I would be careful if I use your classification method. On the other hand, I would like to use your feature extraction method on vertical acceleration without gravity, a form of user acceleration vs. your raw vertical acceleration. Do you think the wavelet transformation and the parameter are robust to this change? (acceleration would henceforth be centered in 0)
Best, Gautier
You may run into issues if your data is centered at 0 because line 145-150 of gait.py checks whether the mean of your data is below 0 to verify orientation. You may consider converting your data to acceleration with gravity, or modifying gait.py to fit your needs.
Best, Matt
Hi Matt,
I would like to use your library on Walk detection and step counting on unconstrained smartphones dataset I am interested in your gait classification algorithm and would love to better understand the training process. Would you be willing to share information about the type of data (number of sample of walking) used to make the classification? Did you use an open-source data set? Also would you agree to share the parameters of the RandomForestClassifier ?
Best,
Gautier