Inspiration
Read more- Testing and screening of SARS-CoV-2 requires time, money, a good health care system, trained staff. It is still not sufficient in many countries.
- We want that everybody in need can get screened at zero costs from anywhere.
As cough is one of the main symptoms of COVID-19 and based on the idea of cough analysis via AI, our project aims to find the unique biosignals of the COVID19 cough and calculate the probability of a SARS-CoV-2 infection. The whole process is available through a web application, making it easy, convenient and accessible even in low-resource areas.
DISCLAIMER: The diagnostics function is not live yet, as this will require a medical trial first
Challenges- It needs to be clear that this can not be used to replace a medical examination but instead shall assist the national health care systems in the screening , outbreak prevention and outbreak control process.
- We are awaiting verified data from COVID-19 positive tested people. This has to be done in a clinical trial as to protect data privacy.
- We are a group of machine learning experts, doctors, and entrepreneurs from Switzerland, Egypt, Germany, Greece, Spain, China, Ukraine, India, Pakistan
- We initially found together through a Slack group during the #CodevsCOVID19 challenge and are working completely remotely
- Web App: Frontend is built using bootstrap, Jquery and integrated with the backend via SSL secure http protocol to the flask server. WebRCT combined with native JS API are used to cover a broad range of devices for recording media. Web app is deployed on scalable Azure webapp with minor hacks including kubernetes runtime initialization to include custom libraries for AI.
- Data Collection: As of June 2020 we have collected 600 cough samples through crowdsourcing. Now we are in the process of collecting confidential verified data in a clinical setting.
- Signal Processing: Considering that we want to CLEAN a drum recording (highly explosive sounds), we removed low frequencies below 40 Hz and high frequencies above 15 kHz - 18 kHz because of microphone limitations. We also used a gate to remove unwanted noise between the coughs and make sure that you do not remove the "silent coughs" or "heavy breathing" because the probability for it to be an important features for the model is high. Finally normalizing the sound, so that we do not have different loudness/ amplitudes among the files.
- Check Signal Similarity: We used cross correlation (or correlation coefficient as normalized measure). Also as an alternative correlation approach we will consider rank based correlation. Moreover similarity of signals can be accessed in the frequency domain. So, we look for "coherence" to find more information.
- ML Model: we apply CNN as a binary classifier. This step is to classify cough sounds into two main categories; COVID19 cough (dry cough) and None COVID19 cough based on different cough patterns.
- Working on a common mission with a team of machine learning engineers, doctors, and entrepreneurs. We are non profit, work for the social good and are humanitarian. We plan to stay this way.
- Having built a functional prototype over a weekend: https://www.detect-now.org/
- We got listed on Forbes, CNN, Live HUM TV interview and we are leading efforts for an umbrella organization with top research universities.
- Interdisciplinary learning and working is highly efficient and can achieve goals quicker
- Information privacy is very important no matter how limiting it is in innovation and progress in the health care field.
- Collecting verified, qualitative data , therefore making it an ethical and responsible application to a philanthropic cause.
- Make it an official medical application . It shall stay easy, convenient and accessible even in low-resource areas, for efficient outbreak control during epidemics.
- This approach to cough analysis might provide a foundation towards further clinical research with AI on pulmonary diseases.
- Share knowledge and promote research by creating an umbrella organization with top research organizations (MIT, Stanford, CMU, Cambridge, EPFL, Bill & Melinda gates foundation)
Muhammad Zeeshan Karamat
1 project • 4 followers
Software engineer with extensive experience in leading successful large scale machine learning solutions.
Thanks to Melia Flieshmann, Gabriel Brandel, Pia Eggert, Patrick Betz, Shresth Agrawal, Vlada Petrusenko, Mohammed Amr, Shizhe He, Alexandre Micheloud, Stephana Muller, Simon Hofer , and Muhammad Zeeshan Karamat.
Comments