13th EAI International Conference on Pervasive Computing Technologies for Healthcare - Demos and Posters

Research Article

WheezeD: Respiration Phase Based Wheeze Detection Using Acoustic Data From Pulmonary Patients Under Attack

Download1033 downloads
  • @INPROCEEDINGS{10.4108/eai.20-5-2019.2283516,
        author={Soujanya Chatterjee and Md Mahbubur Rahman and Ebrahim Nemati and Jilong Kuang},
        title={WheezeD: Respiration Phase Based Wheeze Detection Using Acoustic Data From Pulmonary Patients Under Attack},
        proceedings={13th EAI International Conference on Pervasive Computing Technologies for Healthcare - Demos and Posters},
        publisher={EAI},
        proceedings_a={PERVASIVEHEALTH - EAI},
        year={2019},
        month={6},
        keywords={wheeze asthma pulmonary condition wheezing severity mhealth},
        doi={10.4108/eai.20-5-2019.2283516}
    }
    
  • Soujanya Chatterjee
    Md Mahbubur Rahman
    Ebrahim Nemati
    Jilong Kuang
    Year: 2019
    WheezeD: Respiration Phase Based Wheeze Detection Using Acoustic Data From Pulmonary Patients Under Attack
    PERVASIVEHEALTH - EAI
    EAI
    DOI: 10.4108/eai.20-5-2019.2283516
Soujanya Chatterjee1,*, Md Mahbubur Rahman2, Ebrahim Nemati2, Jilong Kuang2
  • 1: University of Memphis
  • 2: Samsung Research America
*Contact email: schttrj1@memphis.edu

Abstract

Wheezing is one of the most prominent symptoms for pulmonary attack. Hence, wheezing detection has attracted a lot of attention in recent years. However, there is a dearth of a reliable method that can automatically detect wheezing events during each respiration phase in presence of several concurrent sounds such as cough, throat clearing, and nasal breathing. In this paper, we develop a model called WheezeD which, to the best of our knowledge, represents the first step towards developing a computational model for respiration phased based wheeze detection. WheezeD has two components, first, we develop an algorithm to detect respiration phase from audio data. We, then transform the audio into 2-D spectro-temporal image and develop a convolutional neural network (CNN) based wheeze detection model. We evaluate the model performance and compare them to conventional approaches. Experiments on a public dataset show that our model can identify wheezing event with an accuracy of 96.99%, specificity of 97.96%, and sensitivity of 96.08%, which is over 10% improvement in performance compared to the best accuracy reported in the literature by using traditional machine learning models on the same dataset. Moreover, we discuss how WheezeD may be used towards assessment and computation of patient severity.