An attempt to process audio signals, applying various feature extraction, augmentation techniques and develop a CNN and RNN model to detect abnormal heart noises to signify some undetected ailment in the patient.
Collecting the data
We used a published dataset from Russian University providing us with various classes of Heart-beat audio noises collected using digital Stethoscopes.
Approach
A wide range of processing, such as downsampling, appplying Chroma CENS filters, vltp, feature extration and data augmentation was d-none on the audio data. Then a multi-layer cnn was trained upon it to perform detection of anomalous heart behaviour.
Deploying and Publishing
The model was containerized and deployed on cloud, and connected to FAST api such that anyone can go to the site and classify their noise files. The work is under peer review and is about to be published by Feb 2021.
Technologies used and skills learned were:
- Python 3
- Tensorflow
- Data Pre processing
- Digital Signal Processing
- Azure cloud deployment
- Feature Extraction
- Data augmentation
- Deep Learning