PT  - JOURNAL ARTICLE
AU  - Konstantinos Tatsis
AU  - Niels Agerskov
AU  - Daniela Savi
AU  - Agni Sioutkou
AU  - Sofia Peristeri
AU  - Evgenia Salla
AU  - Christos Kyriakopoulos
AU  - Athena Gogali
AU  - Konstantinos Kostikas
TI  - Automated detection of cough events with a Smartphone
AID  - 10.1183/13993003.congress-2021.PA1938
DP  - 2021 Sep 05
TA  - European Respiratory Journal
PG  - PA1938
VI  - 58
IP  - suppl 65
4099  - http://erj.ersjournals.com/content/58/suppl_65/PA1938.short
4100  - http://erj.ersjournals.com/content/58/suppl_65/PA1938.full
SO  - Eur Respir J2021 Sep 05; 58
AB  - Rationale: A common symptom of worsening respiratory disease is an increase in cough. Patients currently self-report changes in cough as part of routine follow-up. However, self-reports are inaccurate. This project evaluated an automated process for quantifying cough during sleep.Methods: A proprietary deep learning model was developed to detect “cough-events” from audio collected with a smartphone. The model was trained using data from 169 subjects from publicly available sources and the University of Ioannina, Greece. Data consisted of coughs from healthy controls and patients with asthma and COPD. The trained model was evaluated in a quiet ‘nighttime like’ environment at the University of Ioannina, Greece with a new cohort of 40 subjects (15 male, age 54.6±12.2, 21 healthy controls, 5 asthmatics, 11 COPD and 3 other respiratory diseases). A total of 5 coughs were recorded from each subject at intervals of 30 seconds. Recordings were annotated by a respiratory physician and analyzed offline using the trained model.Results: In total 218 events were counted by the physician and 221 detected by the model. The model showed good agreement with the physician annotations, with a true positive rate (precision) of 84% and recall rate of 85%. On average, there was a 1.1±1.45 difference between physician counted events and the model. Comparing healthy and unhealthy subjects there was an average difference between annotator and system of 1.28±1.7 for healthy patients, and 0.95±1.2 for unhealthy patients Conclusion: The model shows good agreement with annotations from a respiratory physician. The recall and precision of the model was sufficiently high to give an accurate measurement of a patient’s nocturnal cough trend.FootnotesCite this article as: European Respiratory Journal 2021; 58: Suppl. 65, PA1938.This abstract was presented at the 2021 ERS International Congress, in session “Prediction of exacerbations in patients with COPD”.This is an ERS International Congress abstract. No full-text version is available. Further material to accompany this abstract may be available at www.ers-education.org (ERS member access only).