With their Android app, called CloudUPDRS, Roussos and his colleagues want to make it easier to track symptoms and flag potential problems earlier. Similar to how a clinician would conduct a Parkinson’s severity test, the app includes both self-assessment questions and physical tests using a smartphone’s sensors.
For example, one test measures tremors by asking the user to hold the phone flat in their hand. Another measures gait by getting the user to walk 5 metres in a straight line and back with the phone in their pocket.
The first version of the app directly mimicked the role of a clinician, so the assessment took around 25 minutes. “The reason it takes so long is because it’s hard to make sure that you get enough good data to make the tests reliable. So you have to overcompensate by performing each test for longer than necessary,” says team member Cosmin Stamate.
Stamate added a deep learning feature so that subsequent versions of the app can distinguish between good data, like a measurement of tremors, and bad data, like the smartphone being knocked. If someone performs the wrong action or the smartphone sensor picks up meaningless vibrations, the app simply ignores it.
Having been trained to recognise these differences using data labelled by experts, the system discards bad data with an accuracy of 92.5 per cent. “Then as soon as we’ve registered enough good data, the user is told to stop,” says Stamate.
The app can also personalise assessments to provide a “quick test” option. This measures only three symptoms that are most indicative of an individual’s overall performance and could reduce assessment times to less than 4 minutes. The team will present the work later this month at the International Conference on Pervasive Computing and Communications in Hawaii.