Efficient Online Cough Detection with a Minimal Feature Set Using Smartphones for Automated Assessment of Pulmonary Patients

摘要

An automated monitoring of chronic diseases may help in the early identification of exacerbation, reduction of healthcare expenditure, as well as improve patient's health-related quality of life. Cough monitoring provides valuable information in the assessment of asthma and Chronic Obstructive Pulmonary Disease (COPD). In this multi-cohort study, we have used every-day wearables such as smartphone and smartwatch to collect cough instances from 131 subjects including 69 asthma patients, 9 COPD patients, 13 patients with a co-morbidity of asthma and COPD and 40 healthy controls. For online cough detection we have identified the audio features suitable for resource-constrained platforms (e.g., smartphone), ranked the features and identified top 9 features to obtain an F-1 score of 99.8% in the offline classification of 23,884 cough instances from non-cough (speech/silence, etc.) events using Random Forest classifier. Finally, a power and time-efficient scheme for continuous online cough detection from the audio stream has been illustrated. The proposed model has an online cough detection sensitivity of 93.3%, specificity of 98.8% and accuracy of 98.8%. In addition, a good improvement in reducing the on-device execution (feature extraction and classification) time and power consumption has been achieved compared to the current state of the art algorithms. The proposed on-device cough detector has been implemented to meet the criteria for integration in the passive monitoring and online assessment of asthma/COPD patients.