Development Priorities

[jones139]

There are plenty of opportunities to improve the various tools in the OpenSeizureDetector system. The idea of this discussion is to help to decide where our efforts are best used to improve the system. I'll do a list in the next post for anyone to comment on.

[jones139]

My current (and recent) priorities:

• Develop Machine Learning (ML) Algorithm to reduce false alarm rate (First go is done - included in V5 of OSD App, but this can definitely be improved.)
• Update Android App to use Machine Learning Algorithm, and improve user experience with easier set up - Done - V5 merged in to msater branch and published on Play Store.
• Improve OpenSeizureDatabase publishing to remove duplicates automatically to make it easier to do more regular releases - In progress 05/07/2026
• Implement a 'Frequency Augmentation' feature in the nnTraining toolchain that we use for the ML models to recognise that different size people are likely to shake at different frequencies, so re-sample to a different timebase to augment the data to improve generalisation.
• Re-train ML Algorithm on updated OpenSeizureDatabase to try to address false alarms.
• Automate re-training of the ML algorithm when a new version of OSDB is published.
• Implement translations into V5 of OSD Android App, and include feature requests from users.
• Investigate user-specific ML models
• Improve the local http web page (Update Monitoring Web Page #259 )
• Implement a Group Monitoring web site (Support for group of kids #256 )

[pmithrandir]

I think we can add the following:

• simplify the seizure type selection. the text based is not that easy and on my side I select always the same option. (I miss one line: standing, which is enough to start OSD for my son)
• improve / remove local http page
• FE group monitoring website.

[jones139]

Thanks - have added the second two to the list above. The first one is very easy to do - the seizure and false alarm types are defined here: https://osdapi.org.uk/static/eventTypes.json and it is downloaded dynamically to create the selection UI.
Let me know what you would like to change (I should really have this in a repository, but haven't thought which one....)

[pmithrandir]

For the value, one would be doing small mouvements (uncontrolled, exitation, etc...) and standing

For the interface, it's a bit unfriendly I think.

And I'm even wondering if these values are really useful for a ML training, or if SEIZURE(maybe with type), FALSE ALARM and test would not be enough.

The easier to fill, the more report we will have.