Hello!
I am a student at a tertiary Epilepsy Center. Our team is building an open source seizure alert app for the apple watch that will undergo a clinical trial. A well conducted clinical trial is crucial for convincing insurance to pay for apple watches, which correlates with tech adoption. This imposes some restrictions on data collection.
Our use case requires that the app run continuously as seizures are not predictable. The documentation on extended runtime sessions recommended the smart alarm mode which lasts for 30 minutes. Individuals would not use an alarm system that only works 30 minutes at a time. Is there any way we could deploy the alarm continuously?
The requirements for the background execution are two fold:
-
During model development phase the app would need to continuously collect data from accelerometer, gyro-meter, magnetometer, blood oxygen and EKG. I understand these are very demanding tasks so even one sensor would be good to start with. The app would then upload this to a server over the web (perhaps using CloudKit). If continuous data collection is not possible then what is the longest guaranteed latency between samples? What amount of CPU usage would be permissible and how would it impact sampling rates?
-
During deployment the app would need to run an ML model in the background continuously. We have chosen a light-weight model[1] that has been optimized for mobile devices. It has ~ 1.3 million parameters. What would be the maximum guaranteed latency for running a ML model in the background?
Please let me know if you need any more information. Thank you.
Reference:
- Lightweight Transformers for Human Activity Recognition on Mobile Devices, Sannara EK, François Portet, Philippe Lalanda. arXiv:2209.11750
