Attach your data. What if you just threshold? Why do you refuse to use a threshold? It looks like it would work well.
This will be true if the signal is more than 1 and false if it's less than 1.
If you have the Image Processing Toolbox, you can label each breath and then use regionprops(), or ismember() and find(), to find the starting and stopping index of each breath.
We've seen this before with audio signals and people wanted to determine the voiced and silent parts of the audio recording so you might search for silent, silence, voice, words, and things like that.