Towards Device-Agnostic Mobile Cough Detection with Convolutional Neural Networks

Abstract

Ubiquitous mobile devices have the potential to reduce the financial burden of healthcare systems by providing scalable and cost-efficient health monitoring applications. Coughing is a symptom associated with prevalent pulmonary diseases, and bears great potential for being exploited by monitoring applications. Prior research has shown the feasibility of cough detection by smartphone-based audio recordings, but it is still open as to whether current detection models generalize well to a variety of mobile devices to ensure scalability. We first conducted a lab study with 43 subjects and recorded 6737 cough samples and 8854 control sounds by 5 different recording devices. We then reimplemented two approaches from prior work and investigated their performance in two different scenarios across devices. We propose an efficient convolutional neural network architecture and an ensemble based classifier to reduce the cross-device discrepancy. Our approach produced mean accuracies in the range [85.9%, 90.9%], showing consistency across devices (SD = [1.5%, 2.7%]) and outperforming prior learning algorithms. Thus, our proposal is a step towards cost-efficient, ubiquitous, scalable and device-agnostic cough detection.

Automatic Recognition, Segmentation, and Sex Assignment of Nocturnal Asthmatic Coughs and Cough Epochs in Smartphone Audio Recordings: Observational Field Study

“…This work was built upon a convolutional neural network architecture for cough recognition that we introduced in previous work [ 32 ] which recognized coughs in Mel spectrograms. Mel-scaled spectrograms are visual representations of audio signals with respect to time and frequency.…”

“…The detailed calculation of the Mel spectrograms used can be found in Multimedia Appendix 1 . We evaluated this approach against different approaches for smartphone-based cough recognition in previous work on voluntary coughs and found that it performed best [ 32 ]. Our approach produced stable results across recordings of five different devices with different hardware and service life duration.…”

“…In addition, they can be used to passively monitor the health status of patients without an additional task of activating a monitor. Recent advances have used only a smartphone and its built-in microphone for cough monitoring [ 30 - 32 ].…”

“…Further, we aimed to use the classifier to segment and count cough events over the course of the night. Building upon previous work [ 32 ], we adopted a convolutional neural network architecture , which performed best in comparison to that of other machine learning approaches when using voluntary cough data from different smartphone recordings. We altered the learning part of the algorithm by combining three different techniques from literature to combat the high class imbalance encountered in this real-life data set—ensemble learning [ 38 ], minibatch balancing [ 39 ], and decision thresholding [ 40 ].…”

Breeze

“…To maximize the information that can be derived from a one-second sound, we propose a Convolutional Neural Network (CNN). CNNs have already been successfully applied to different audio event detection applications [1,5,63,83]. For BEM to identify breath sound, we employ a single layer CNN to extract information from the entire one-second sound features.…”