Private Audio-Based Cough Sensing for In-Home Pulmonary Assessment Using Mobile Devices

With the prevalence of smartphones and smartwatches during the last decade, ambulatory audio-based monitoring of lung function has become more common. A symptom that provides significant information for determining the pulmonary patients’ state is cough. This work proposes a novel algorithm to detect the cough features such as frequency, type, and intensity, in a privacy preserving manner using readily-available mobile devices for daily monitoring. The cough classification algorithm consists of three main modules. Audio events will be detected in the first layer after pre-processing of raw audio data. Then, a random forest algorithm is applied to classify the event as cough, speech, or none for each of the time frames. Lastly, a majority voting scheme is applied on the output of the second layer for each time window (which consists of multiple time frames) to determine the final label for that window. The implemented sound event detector has detection error of less than 3% in an indoor situation and cough/speech/none classification reaches 96% accuracy using random forest algorithm. The system benefits from a privacy preservation algorithm (proposed in our previous work) to make the speech unintelligible while keeping the important cough features recognizable. The privacy preservation algorithm reduces the accuracy of classification by 11% (8% for cough and speech labels) which is not significant comparing to the benefit it delivers.