Diagnosing the Stage of COVID-19 using Machine Learning on Breath Sounds

Abstract

With rapidly increasing COVID-19 cases, patients with mild
and moderate symptoms are being asked to home-isolate themselves
to save hospital resources for more severe patients.
Such patients have been asked to self-monitor themselves and
seek medical attention if their condition worsens. COVID-
19 affects the respiratory system and home-isolated patients
must monitor their lung condition continuously before it quickly
deteriorates. But this is difficult to monitor by oneself, and the
patient may not notice his worsening lung condition before it
is too late. A machine-learning based approach is proposed
to monitor lung condition by analyzing the breath sounds of
a patient for respiratory sounds like wheezes, crackles and
tachypnea, which in turn can identify the stage of COVID-
19. Data from a respiratory sound database with recordings
from 226 patients was split into 6898 respiratory cycles and
pre-processed. In this paper, two approaches are evaluated.
The first approach is demonstrated using Google Cloud AutoML
with the recordings of respiratory cycles which were
converted to spectrograms to train the model. In the second
approach, Log Mel filter-bank features were extracted from
the breath sounds and used to train multiple CNN models to
hierarchically classify breath sounds. This ensemble-learning
with hierarchical-model approach achieved a better accuracy
of 78.12%. This model can be integrated with a mobile application
to record and analyze breath-sounds. This will enable
the patient to admit himself sooner if he is progressing to a
severe stage of COVID-19.