Inspiration

Helping a quick, automated diagnostic for early stages of Parkinson and removing workload from doctors around the world.

What it does

The patient calls the number shown in the screen, where it has to talk for about 30 sec. Quickly after, they receive a SMS telling the diagnostic.

How we built it

We worked with a public databased of diagnosed Parkinson patients which contains some speak parameters. Through Machine Learning, we are able to predict, with around 70% of accuracy, if a new sample belongs to an affected person or not. In order to process the call, we use Nexmo Voice API and run it against the database. The SMS is also sent thanks to Nexmo SMS API.

Challenges we ran into

The database used was quite small, containing around 200 samples, although from just 23 different patients. The predictions could be easily improved with a larger database.

The Nexmo Voice API was really a challenge, having problems to receive correctly the information needed. Fortunately, we managed to solve this in time (thanks to the help of one of the developers there) and we are able to retrieve audio files from a phone call.

As for the treatment of the voice signal, it was a challenge on its own. We tried, really hard, to use Praat phonetics program (http://www.fon.hum.uva.nl/praat/) and implement a script via Python. After a lot of try and error, we coded a functional Praat script that returned the data we needed from a voice sample to run it agains the database. However, due to the large amount of time spent doing this, we haven't got enough time to merge it with Python.

All in all, the main problem of our project was time and, because of it, not being able to fit all the pieces of the puzzle together.

Accomplishments that we're proud of

  • A deeper understanding of machine learning and deep learning

  • Using random APIs

  • Understanding a little bit more flask and HTML

  • Uploading our website to a .org domain

  • Discovering that Parkinson can be diagnosed with your speach

What we learned

24h to code is shorter than we thought

Citations

'Exploiting Nonlinear Recurrence and Fractal Scaling Properties for Voice Disorder Detection', Little MA, McSharry PE, Roberts SJ, Costello DAE, Moroz IM. BioMedical Engineering OnLine 2007, 6:23 (26 June 2007)

Oxford Parkinson's Disease Detection Dataset: https://archive.ics.uci.edu/ml/datasets/Parkinsons

Share this project:

Updates