The initial version of the seizure detector (or should it really be called an apnea monitor?) has been installed on test in our house for 2 weeks. This version raises only visual alarms, because I don't trust it enough yet to produce audible alarms in the middle of the night. Its function is to detect an abnormally low breating rate, which occurs after a major seizure.
The findings of this initial soak test are: * It basically works - the assessed breathing rate falls in deep sleep compared to lighter sleep with 'fidgeting', so it is promising. * It needs a good background image to work reliably - without this it may average the intensity of a larger area of the room (not just Benjamin), which reduces its sensitivity. I added a 'reset background image' to the raspberry pi viewers so that Sandie can easily reset the background herself when necessary. Another symptom of the background being incorrect is the system alarming if Benjamin is not in the room - it may monitor his pile of toys rather than saying it can not find Benjamin. * It gives some false alarms during deep sleep, so we need a slightly higher sensitivity for breath detection.
So it looks promising, but some development is necessary before I connect it to an audible alarm. The main development is to tune the peak detection parameters to reduce the false alarms. To achieve this, I have added more functionality to benFinder.py so that it records a short video of the raw data periodically, along with the assessed brightness timeseries, so I can see if I can find a better way of analysing the very small movement cases. It also saves a video and time series when it alarms.
I will give it another week on soak test collecting this data, then have a go at tuning the parameters to get the required sensitivity.
A later improvement will be to change the Benjamin detection from just being the largest object in the field of view, to doing some image processing to look for somethign that looks like him - this should reduce the requirement for having a good background image.