Stefan Jokic

Stefan Jokic

Auto-generated profile, based on publicly available information

Stefan Jokic

Department of Computer Science, ETH Zurich, Zurich, Switzerland

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
23 February 2022 in IEEE Journal of Biomedical and Health Informatics
Verified
Journal Article
Open access
SJStefan Jokic
Stefan Jokic
DCDavid Cleres
David Cleres
FRFrank Rassouli
Frank Rassouli
CSClaudia Steurer-Stey
Claudia Steurer-Stey
MPMilo A. Puhan
Milo A. Puhan
FBFilipe Barata
Filipe Barata
Cough, a symptom associated with many prevalent respiratory diseases, can serve as a potential biomarker for diagnosis and disease progression. Consequently, the development of cough monitoring systems and, in particular, automatic cough detection algorithms have been studied since the early 2000s. Recently, there has been an increased focus on the efficiency of such algorithms, as implementation on consumer-centric devices such as smartphones would provide a scalable and affordable solution for monitoring cough with contact-free sensors. Current algorithms, however, are incapable of discerning between coughs of different individuals and, thus, cannot function reliably in situations where potentially multiple individuals have to be monitored in shared environments. Therefore, we propose a weakly supervised metric learning approach for cougher recognition based on smartphone audio recordings of coughs. Our approach involves a triplet network architecture, which employs convolutional neural networks (CNNs). The CNNs of the triplet network learn an embedding function, which maps Mel spectrograms of cough recordings to an embedding space where they are more easily distinguishable. Using audio recordings of nocturnal coughs from asthmatic patients captured with a smartphone, our approach achieved a mean accuracy of 88% (10% SD) on two-way identification tests with 12 enrollment samples and accuracy of 80% and an equal error rate (EER) of 20% on verification tests. Furthermore, our approach outperformed human raters with regard to verification tests on average by 8% in accuracy, 4% in false acceptance rate (FAR), and 12% in false rejection rate (FRR). Our code and models are publicly available.