Can your voice, cough or breath help diagnose COVID-19?
Living already two years with COVID-19 pandemic, it has been clear that enormous research resources have been mobilised to help this fight. Especially, at the front of diagnosis, where nasopharyngeal tests have been a crucial advancement, that also facilitated the continuity of everyday life. However, they remain an invasive procedure and not at all pain free. What if there is a way to detect COVID-19 by simply recording your voice, cough or breath? In fact, Artificial Intelligence (AI) can make this possible!
EURECOM’s professors Massimiliano Todisco, expert in speech and audio signal processing and Maria A. Zuluaga, expert in AI for health, joined forces and created a voice and respiratory sound-based AI system for COVID -19 diagnosis.
“The human voice conveys much more information than we imagine. For example, in addition to the meaning of the discourse, it is often possible to determine the speaker’s identity and his or her emotional state. Other information remains somewhat more difficult to reveal, such as information about pathological conditions.”, explains Prof. Todisco.
Doctors are trained to listen to patients’ voice, breath or cough and they can distinguish different respiratory diseases. In fact, the entire human phonatory apparatus, from the vocal cords to the thoracic cage, trachea, nose and mouth, acts as an acoustic space, where sound is distorted by possible infections. Unfortunately, these distortions can be very weak to detect. This is where AI intervenes, in order to detect different patterns of such characterisation of distinct sounds in normal conditions and in pathology. These machine learning (ML) tools try to quantify doctors’ estimations and intuitions and validate their diagnosis. One of the challenges, however, is to acquire sufficient data of good quality, as the presence of large amounts of noise is a typical problem to be faced.
“In our project, we tried different ML methods together with acoustic features that are more related to the auditory system and the way humans perceive sounds”. Prof. Todisco says.
“What I really enjoy in this project is that we tried demystifying machine learning and AI. There is a common conception that we can put any data through an AI system and magic will come out. We participated in an international challenge and achieved almost 90% accuracy with our method. This result was really the product of collaborating and combining our two expertises; in ML and audio processing.”, says Prof. Zuluaga.
“This technology can address a lot of respiratory diseases and patients with reduced mobility could benefit from online diagnostic tools like that. But in the end, it is all about data, in order to train AI models correctly and achieve high accuracy results. For that, synergy and close collaborations between research labs and hospitals is crucial.”, explains Prof. Zuluaga.
In the future, it would be great to have an application with this algorithm implemented providing a non-invasive solution for COVID-19 detection or for other respiratory conditions, with the help of industrial collaboration.
by Dora Matzakou for EURECOM
