Design of a Cough Detection System Based on Vibration and Audio Signals for Post- Thoracic Surgery Patients

Cough is a vital rehabilitation indicator which can exhibit many complications after thoracic surgical procedures. In this paper, a portable bimodal cough detector that can achieve consecutive detection is designed to recognize postoperative cough. The detector collects vibration and audio signals of cough and sends the raw data to the base station for data processing. The time domain and frequency domain features of vibration data are extracted. Meanwhile, 13 dimensions of Mel-frequency Cepstral Coefficients (MFCC) features and two dimensions of physical features are extracted from audio data. The threshold distinguish method and the bagged tree model are applied to classify the vibration and audio signals features respectively. The fusion discrimination method determines the result of the two signals and then gives the final judgment based on the two classifiers. The bagged tree model training results show that the audio classifier achieved an accuracy of 94.20% with an Area Under the Receiver Operating Characteristics Curve (AUC) of 0.984. We evaluate the performance of the fusion discrimination in actual use scenarios. The experiment results show that the detector achieved a sensitivity of 85.54% and a specificity of 98.76% in the silent environment. The detector in the interference environment can achieve a sensitivity of 83.44% and a specificity of 98.50% which exhibits a strong interference resisting ability.