Crackles, caused by the abrupt opening of previously closed airways, are often heard in patients with pulmonary diseases, such as pneumonia, bronchiectasis or pulmonary fibrosis. They are transient, discontinuous, explosive and non-musical lung sounds which are characterized by their duration and pitch.
However, auscultating is subjective and depends on the physician’s experience. Researchers from the Institute of Acoustics of the Chinese Academy of Sciences have recently found an objective and noninvasive way to detect crackles, which has been proved rapid and efficient for crackles detection.
The improved noninvasive way consists of preprocessing which comprises filtering and decomposition by short-time Fourier transform (STFT), features extracting and crackles detecting based on support vector machines (SVM). From Fig.1, its block diagram can be depicted.
Fig. 1 The steps in block diagram of the algorithm (Image by LI Jiarui et al.)
In this new way, a band pass filter that keeps frequencies in range from 150Hz to 1800Hz is applied to the lung sound.
In the following, three features, which are the ratio of fmin to fmax of the frequency limbic signal, the standard deviation of the time limbic signal and the smoothing time limbic signal, are extracted respectively. Herein, the frequency limbic signal is equal to the power spectral density, which is calculated by integrating the power spectrogram signals along time axes, while the time limbic signal is computed by integrating the power spectrogram signals along frequency axes. Fmin/fmax is the frequency ratio of the frequency limbic signal.
Finally, a support vector machine, which is a supervised learning method, is used here to classify crackles and normal lung sounds. In addition, the parameters of the support vector machines are optimized to get a better performance.
The simulation result proves that the sensibility (SE), specificity (SP) and accuracy (AC) of the first testing dataset are 94.29%, 100% and 97.14%, respectively, while that of the second testing dataset are 100%, 100% and 100%, respectively.
As a result, the accuracy of the new way is 98.57%. Compared to the past method which only uses fmin/fmax as the feature to detect crackles, the accuracy of the new way is increased by 10.71%.
In particular, the proposed features in time domain are key innovations of this research which improves the accuracy of crackles detecting efficiently.
LI Jiarui, HONG Ying. Crackles Detection Method Based on Time-Frequency Features Analysis and SVM. The 13th IEEE International Conference on Signal Processing (November 6th to 9th, 2016, Chengdu, China).
Institute of Acoustics, Chinese Academy of Sciences, 100190 Beijing, China