Project / product name: Temporal Convolutional Network for AS detection
Link to the project: http://
Team leader: Armand Rafael Bálint
Challenge: no. 10: Aortic Stenosis Detection from electrocardiogram data
Problem: Our challenge was to unify demographic, admission and raw ECG data from multiple sources to predict aortic stenosis. This is usually a very difficult task when you have different formats and missing data pieces.
Solution: We decided to design both a new neural network architecture and a preprocessing framework to solve this problem. Our network processes demographic data and metrics obtained from the ECG such as QRS amplitude and duration using a traditional architecture. For the temporal data, though, we used a temporal convolution network design.
Impact: With a high accuracy and a robust preprocessing framework we can integrate multiple sources of clinical data to assess the risk of aortic stenosis for any given patient.
Feasibility & financials: The network was lightweight enough to train on a laptop, which means it can easily implemented into any clinical system and can be further trained with future data. To improve the performance of the network, only a server with a GPU is required.
What you have built at the hackathon - text explanation + code (e.g. GitHub link): We have created a preprocessing pipeline with an integrated TCN to process the data and predict the risk of aortic stenosis from ECG, demographic and admission report data. The project is available at: https://github.com/EHH2021-HvN3rQ/hack
What you had before the hackathon, please mention open source as well: We have built this project from scratch. We have used tensorflow for network training, and some widely available Python data processing libraries.
What comes next and what you wish to achieve: To improve the performance of the network, the preprocessing pipeline needs to be made more streamlined, and the network needs to be trained on a larger dataset. :)