Sber: The algorithm determining COVID-19 disease probability

Main articles:

• Tests for COVID-19 coronavirus
• Neuronets (neural networks)
• Artificial intelligence in medicine

2021: Representation of an algorithm

The laboratory of artificial intelligence Sbera developed an algorithm which within 60 seconds can determine the probability of a disease of a koronavirusny infection of COVID-19. Presence of a virus is defined on the basis of results of short poll on symptomatology and three sound models — voices, breath and cough. Sber reported about it on January 12, 2021.

Sound files turn into the spectrogram showing sound energy at different frequencies and further are analyzed using deep convolution neural network. For its training only open data were used - it is more than one thousand samples of sounds of breath and cough collected from the diagnosed patients in the Russian clinics.

at the beginning of November, 2020 we declared determination to create such algorithm and since then in vain did not waste time. Of course, our model does not reach the accuracy of biological PCR, which is not surprising, yet but for January, 2021 has comparable characteristics. At the same time it gives the chance to make the configured sensitivity, it is much easier to use, more convenient and that is important, cheaper. It is not the medical diagnostic tool, but rather personal daily cheker — test passing and obtaining result take only 60 seconds. In the nearest future we are going to create a special application which will become available in App Store and Google Play. It will allow to configure even better model accuracy,

Average ROC AUC (the area under "error curve") the model created by Sber for January, 2021 godp is equal to 0.8. Further improvement of quality of model at increase in amount of data, including collected using mobile application is expected .