I have spent some time over the Christmas break using the data that users have contributed via our Data Sharing (the Open Seizure Database) system.
It looks like I have a model for a single user which detects seizures better than our deterministic algorithm and has a much lower (4.7x) false alarm rate, which is what I was hoping to achieve with a Machine Learning model.
There is a very technical description of the progress for those who are interested here: https://github.com/OpenSeizureDetector/OpenSeizureDatabase/blob/main/Publications/User_Specific_Training_Report_1A.pdf
The next steps are to update the Android App so that it can download different models from our web site and publish that for beta testing.
I will then test the new model to make sure that we have not introduced a lot of other possible false alarm initiators before I train and test it against the whole Open Seizure Database. If that is testing I will make the new version of the app and this model available for all users for testing to see how it performs in the real world.
Graham
(graham@openseizuredetector.org.uk)
