Use LSTM on fmri activity vector to predict stimuli label

As mentioned in How to use RSA on fmri data,
I have a dataset of n subjects, with 347 stimuli and activity in 8 regions.
Each stimuli have a specific label value, ranging 0-1, that will be binned to 4 labels (Low, mid-low, mid-high, High).
The stimuli have temporal dependency.
I want to build a LSTM NN to predict the label based on activity.
So the input at each point will be the activity (convolved with HRF) in the last, let’s say, 4 seconds, and the output will be the average label at this segment. (so if 3 stimuli were presented with the label values of 0. 0.3. 0.1 - the mean value will be 0.1333 - so label category is low).
Does this make sense? I couldn’t find any similar studies or related repos in github - Any input will be appreciated :slight_smile:
Thanks!