BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Improved COVID-19 classification of cough audio
DTSTART;VALUE=DATE-TIME:20211203T131500Z
DTEND;VALUE=DATE-TIME:20211203T134500Z
DTSTAMP;VALUE=DATE-TIME:20260711T092337Z
UID:indico-contribution-1517@events.chpc.ac.za
DESCRIPTION:Speakers: Thomas Niesler (University of Stellenbosch)\nWe repo
 rt progress in our research aiming to detect COVID-19  from smartphone aud
 io recordings. In our previous work we reported that  it is  possible to d
 iscriminate between recordings of COVID-19 positive coughs  and coughs by 
 COVID-19 negative or healthy individuals using machine  learning algorithm
 s. Since the available datasets of COVID-19 coughs are  small\, the classi
 fiers exhibited a fairly high variance. In subsequent  work we have invest
 igated the effectiveness of transfer learning and  bottleneck feature extr
 action for audio COVID-19 classification\, in this  case performing experi
 ments for three sound classes: cough\, breath and  speech. For pre-trainin
 g\, we use datasets that contain recordings of  coughing\, sneezing\, spee
 ch and other noises\, but do not contain COVID-19  labels.  Convolutional 
 neural network (CNN)\, long short term memory (LSTM) and  Resnet50 archite
 ctures were considered. The pre-trained networks are subsequently either f
 ine-tuned using smaller datasets of coughing with COVID-19 labels in the p
 rocess of transfer learning\, or are used as bottleneck feature extractors
 . Results show that a Resnet50 classifier  trained by this transfer learni
 ng process delivers optimal or near-optimal performance across all dataset
 s achieving areas under the \nreceiver operating characteristic (ROC AUC) 
 of 0.98\, 0.94 and 0.92  respectively for the three sound classes (coughs\
 , breaths and speech).  This indicates that coughs carry the strongest COV
 ID-19 signature\,  followed by breath and speech. Our results also show th
 at applying  transfer learning to capitalise on the larger datasets withou
 t COVID-19  labels leads not only to improved performance\, but also stron
 gly reduces  the standard deviation of the classifier AUCs measured on the
  test sets  during cross-validation\, indicating better generalisation. We
  conclude  that transfer learning and bottleneck feature extraction can im
 prove COVID-19 cough\, breath and speech audio classification\, yielding  
 automatic classifiers with higher accuracy. Since audio classification is 
 non-contact\, does not require specialist medical expertise or  laboratory
  facilities and can be deployed on inexpensive consumer  hardware\, it rep
 resents an attractive method of screening.\n\nhttps://events.chpc.ac.za/ev
 ent/98/contributions/1517/
LOCATION:
URL:https://events.chpc.ac.za/event/98/contributions/1517/
END:VEVENT
END:VCALENDAR