Sounds great. I remember there has been some prior research done, documented here:
A simple heart rate limit to decide the closeness was deemed too slow/inaccurate. However, adding AI into the equation might be what it takes to achieve a breakthrough in that regard. If AI is good at anything, it’s pattern recognition, predicting if a value is about to go somewhere (rather than relying on a fixed threshold), based on previous measurements/observations, tailored to the user. Also including the reference to what’s on screen sounds very interesting.
Definitely following this thread 