5.0.9 doesn't triggered on ML alg during seizure

[pmithrandir]

Hi,

I installed 5.09 beta version 2 days ago.

6h15 : standing and walking, one warning
19h45 : standing and walking one warning + alarm
4h50 : real seizure, about 3-5 minutes, not even a warning (seizure prob around 10%)
9h15 : real seizure, same length, no warning (seizure prob around 20%)

Usually, I get at least a warning, now, nothing.

Was there any change in 5.0.9 that could explain that poor performance ?

[jones139]

That is very odd, because we have been using it and haven't seen any issue - the only thing that changed in V5.0.9 is to do with SMS alerts, system re-starting if you change BLE device (PineTime), and general system shutdown and re-start process. So nothing that would affect ML model performance.
I think your data will be in the Data Sharing database, so I'll have a look at your data this evening to check.

Thanks for reporting this.

Graham.

[jones139]

I can't see this data in the database - I can see events at 0622 and 1944 on 31/07/2026, plus rather a lot of 'fault' events too - Maybe the server was running slow and the connection has dropped out - when you get chance, please will you try restarting the osd phone app to see if it uploads the extra data. Were you getting a lot of fault 'pip' warnings suggesting something was not working right?

[jones139]

@pmithrandir I had an unusual incident last night - I have both of our 'Network Data Source' devices go into Fault, but when I looked at our Son's phone I could not see anything wrong with it.
It was late so I did not investigate properly but re-started the phone app and re-booted his watch, and the fault went away.

Now it was probably some sort of local wifi network problem that I didn't detect, but it seemed to go away when I restarted everything, but it is possible that the main phone app was failing to report a fault condition that was visible over the Network Data Source. Is it possible that your system was in a similar condition when it failed to alarm for the seizure?

I think you have contributed quite a bit of data to the database, so we could re-train the ML model, weighted towards your data to improve your detection reliability - I'll look at this once I have fixed my database updating process and can do a new release of the database.

[pmithrandir]

Hi,

I don't think it was the case, both phone were quite ok.
Just no seizure detected while shaking was happenning.

As your suggest, it might be that the ML algo is less efficient to detect some seizure for Arthur.

If we can try a custom version, that would be quite nice.

Thanks,
Pierre