As a first approach, we have implemented a cough classification model based on cough sound samples that were previously labelled wet/dry by one of the doctors in our team.
Over the past month, we achieved 90% accuracy on the cough classification model.
Currently, we are collecting data in pilot studies with partnerships with hospitals and research centres.
Our pilot study includes confirmed COVID-19 patients, patients diagnosed with other breathing related diseases (and who tested negative for COVID-19) and healthy people, as a control. Therefore, the audio data is labelled according to the confirmed infection status of the patients and acquired by trained healthcare professionals in a controlled setting. Additionally, we will also rely on real-world-voice-data acquired through the web app and in partnership with real-world COVID-19 monitoring solutions (for example, the geoHealthApp from Germany).
This data will allow us to train an artificial intelligence to estimate the probability of someone being affected by COVID-19 from the analysis of their vocal biomarkers and accompanying symptoms. Our product builds on medical research that has provided proof of concept for the detection, with a high level of accuracy, of diseases, including COVID-19, through cough and other vocal biomarkers (see medical research file in attachment).
The interface for data collection is implemented in the form of a web app and an automated phone call. In particular, the phone line consists of a highly scalable (million simultaneous users) Interactive Voice Response (IVR) system that collects cough, voice and breath audio data. Alternatively, our software is also available through an API.