Hi everybody,
I have time series heart rate data of patients that are sick or healthy. At each point in time is a measurement of their heart pumping ability so it is univariate. I want to try the CNN approach on this: Image this time series and then train ResNet to classify sick or healthy.
How would I go about imaging this time series? I have tried time series AI and it worked but I was only getting ~72% validation accuracy. I was wondering if there any other state of the art imaging techniques? Also, if you have any ideas on how to get state of the art results, that would be appreciated.
Thanks
1 Like
HI Varun,
Depending on whether there are relationship between your different dataset, you might consider stacking your data. Say, your time series lies on the horizontal, with perhaps 100 timesteps or something, and you want to create an image 100x100, so you can stack 100 time series data on top of each other, and that would create the image. Similarly for the output it would now a multi-output (output 100 value). Furthermore, you could also mix them up and stack horizontally say having 150x150 so some of the time series will be split into two, and this will require experimentation. Try to visualize with a heatmap or plot with cmap=“coolwarm” or something that is clear to visualize and see if there are any patterns.
However, I only tried it on regression output, and since you want classification output, I am unsure of how you might want to stack this. Perhaps stacking only those with one group together instead of mixing them together might be a good idea? You might need to experiment with this more.
Have fun experimenting.
1 Like