Missing Most of the Functionality In The Paper

[mrucker]

After reviewing your code more it appears that most of the functionality claimed in the paper is missing.

https://www.sciencedirect.com/science/article/pii/S0169260721005356

Are there plans to fix this?

[enniin]

I noticed the same issue. I find this a bit misleading. Please at least inform us of this on the readthedocs page of the project -- not everyone is capable of examining the code in such detail as to notice the lack of e.g. preprocessing steps for the EDA pipeline.

That said, it would also be marvellous to have the full functionality. Please consider continuing with this!

[mrucker]

@enniin, If you really need the functionality it all exists in various feature branches that have never been merged into master (or published on pypi). If you are using more than EDA, also note that there are problems with the ACC preprocessing and HRV preprocessing as well. For our work, my team ended up going with the Neurokit2 package.

[hidaIgo]

@enniin, If you really need the functionality it all exists in various feature branches that have never been merged into master (or published on pypi). If you are using more than EDA, also note that there are problems with the ACC preprocessing and HRV preprocessing as well. For our work, my team ended up going with the Neurokit2 package.

Thanks for the clarification, I am in the same situation. Neurokit2 works with empatica E4 wristband?

[mrucker]

You have to load the data into numpy arrays before passing it to NeuroKit2. But yes, it does work with E4 wristband. That's what my lab uses.

[enniin]

@enniin, If you really need the functionality it all exists in various feature branches that have never been merged into master (or published on pypi). If you are using more than EDA, also note that there are problems with the ACC preprocessing and HRV preprocessing as well. For our work, my team ended up going with the Neurokit2 package.

Thank you!! I think for this situation I will try to use the EDA-branch, though it does seem unmaintained -- there are some problems with getting it installed. (I guess I could make a pull request of it, as I managed to solve them, even if there's no maintainer to take them on board...) Thank you also for the tip on Neurokit2! It will be my second option.

[hidaIgo]

You have to load the data into numpy arrays before passing it to NeuroKit2. But yes, it does work with E4 wristband. That's what my lab uses.

Hi again, after a few days testing with neurokit2 I can't get features for the data that empatica E4 offers, if it's no bother, would you have any example that I can see? I am a student, it is not for professional purposes.

My purpose is to introduce to a ML model, different events of about 10 minutes labeled together with relevant features of the signals.

[mrucker]

@hidaIgo Yeah, totally, What E4 sensor are you interested in? I assume EDA and BVP?

[hidaIgo]

@hidaIgo Yeah, totally, What E4 sensor are you interested in? I assume EDA and BVP?

For the moment yes, later I would like to try them all, but to start with EDA and BVP is fine. Thank you very much!

[mrucker]

For EDA it might look something like this:

[mrucker]

For HRV from BVP it looks something like this:

[hidaIgo]

For EDA it might look something like this:

Thank you very much! And how could I do this EDA feature extraction by specifying a window size and the window step size just like with flirt? What values for these should I test?

You are being a great help, sorry for the inconvenience.

[mrucker]

You just have to split your EDA.csv file down into the windows/steps you want to calculate features for.

For example, the EDA signal on the E4 samples at 4hz.

So if you wanted features for 10 minute windows with a 5 minute step size then you'd do something like:

stp_size = 4*60*5 #5 minutes
win_size = 4*60*10 #10 minutes
n_windows = int(len(signal)>win_size) + (len(signal)-stp_size)/win_size
windows = [ signal[i*step_size: i*step_size+win_size] for i in range(n_windows) ]

for window in windows:
   # Calculate the EDA features using this signal window.

The most appropriate window size and step size really depends on your use case and what you're trying to predict. In general I'd say having a step size that is half of your window size is a good heuristic. That will let you stretch your data a little bit further so long as you're careful that overlapping windows don't get split between training and test. To figure out window size you'll just need to test different sizes.

[hidaIgo]

You just have to split your EDA.csv file down into the windows/steps you want to calculate features for.

For example, the EDA signal on the E4 samples at 4hz.

So if you wanted features for 10 minute windows with a 5 minute step size then you'd do something like:

stp_size = 4*60*5 #5 minutes
win_size = 4*60*10 #10 minutes
n_windows = int(len(signal)>win_size) + (len(signal)-stp_size)/win_size
windows = [ signal[i*step_size: i*step_size+win_size] for i in range(n_windows) ]

for window in windows:
   # Calculate the EDA features using this signal window.

The most appropriate window size and step size really depends on your use case and what you're trying to predict. In general I'd say having a step size that is half of your window size is a good heuristic. That will let you stretch your data a little bit further so long as you're careful that overlapping windows don't get split between training and test. To figure out window size you'll just need to test different sizes.

Thanks, that's exactly what I meant!

I don't quite understand how to make the overlapping windows not split between training and testing. I may have misunderstood the construction of the dataset.

An example: I have 2 EDA.csv files corresponding to two events (each 1 hour long), I process them separately as you said (with 10 minutes of window and a 5 minutes jump). This way I will have 7 windows for each file/event.

At the same time as I calculate the features for each window, I add the 'Mood' column to predict, so I have 7 rows with the same value of the target column, e.g. 'Happy'. On the other hand, I have another 7 rows coming from the other event with the column 'Mood' to 'Sad'.

Is this approach correct? If so, how would it be done so as not to split the overlapping windows between training and test as you say?

If not, as long as it is not a bother, could I contact you by mail so as not to fill this Issue with so many offtopic messages?

Thank you very much, I really appreciate this help.

[mrucker]

Yeah, good idea. Why don't you email me.

I believe you can get my email address from my GitHub profile. If this isn't the case let me know.