Context
For this project I have been contacted by a Surgeon through a friend working in Lumos, a junior company in Vienna. The pitch was to use a specific type of machine learning model (Graph Neural Network) to detect a disease. As I was already using GNNs in another project (see OMR), and I was interested by working with real world data and conducting a small project beside my master thesis I signed in!
Project
Synkinesis is a neurological symptom in which a voluntary muscle movement causes the simultaneous involuntary contraction of other muscles. The goal was to detect this issue based on patient portrait.
I really appreciate this project because it involved a lot of discussion with the surgeons. The idea they had at first and the misunderstanding of data science they had almost lead to cancelling the project. The initial idea was to detect synkinesis in picture of patient with facial palsy but they could provided enough data. The project turned into recognizing patients with synkinesis among healthy people. Well, no need to be a doctor or a futuristic computer to perform this, anyone can do it! However, they maintained the project and I build a small app that included a model able to predict the disease. You can see a demo in the following video.
Another important aspect of this project was to temper the conclusion as the model abilities are very limited and the data clearly show some limit. This project also led me to my first paper (still under review at the Laryngoscope)