An accelerated COVID-19 diagnosis tool using interpretable deep learning

This research aims to develop an efficient machine learning model to detect COVID-19 patients by using their cough signals. The model is trained using thousands of audio recordings of cough signals from different subjects. The audio signals can be converted into spectrogram images that can be visually inspected to determine relevant regions of interest. There are two challenges in building the prediction models, including 1) complexity of finding the best architecture of the machine learning model and 2) understanding the reasons behind specific predictions. Since we model the problem as a set of images, we can exploit previously published state-of-the-art deep learning models under certain modifications. Further, we address interpretable approaches that can identify different conditions and highlight areas of interest while predicting positive or negative COVID-19 cases. Domain experts can then look at the outcome and confirm if the right parts of the cough spectrum are considered.

Faculty Supervisor:

Othman Soufan

Student:

Partner:

Zensark, Inc.

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

St. Francis Xavier University

Program: