Diagnosis per voice: AI for coughs and colds

More than 300 million audio snippets with voices can be easily found on YouTube. And every uploaded video can be used and analyzed by the parent company Alphabet, i.e. Google.

This means there is a treasure trove of data that, in contrast to some current controversies surrounding the training of various AI systems, can be analyzed legally and without any legal issues.

And after all, researchers from Google and various other research institutions have used it to train a rather exciting AI that is to be used as a diagnostic tool for respiratory diseases. It is called HeAR, which stands for "Health Acoustic Representation".

Such analyses have existed in medical research for a long time; after all, it is usually very easy to hear when someone else has a cold. Pre-examined data was and is generally used for this purpose. It is therefore known that the voice in question resonates a specific illness or that it is healthy.

The problem, however, is that the amount of this data is very limited. Not so with the training of HeAR: the system was trained with the really huge data set of 300 million completely unlabeled audio files.

A few additional voices with identified diseases were then added. In this case, these were tuberculosis, COVID-19 and the fact that someone smokes. For a newly developed system, the results of 74% correct detection of tuberculosis and 71% for COVID-19 seem quite impressive.

The next step is to continue with controlled clinical trials. This would be the first opportunity to obtain an initial diagnosis using only the voice. Even if there is still room for improvement in the results, this is a test that is based solely on the spoken word - neither invasive nor costly.