A novel method for automatic separation of pulmonary crackles from normal breath sounds

Pulmonary crackles are an important physiological parameter for evaluating lung condition of an individual and usually determined at auscultation by conventional stethoscope. The presence of crackles is generally an early indication of the disease and their number per breath cycle can indicate the severity of the disease. A conventional stethoscope placed on the chest wall can identify the presence of crackles, but this approach is subjective and the accurate detection of crackles and the identification of their type (fine or coarse) is highly dependent on clinician hearing ability and expertise. The misinterpretation of crackles may lead to inappropriate treatment of the patient. Computer aided lung sound analysis (CALSA) using advanced signal processing techniques can provide an objective way of analysing recorded lung sounds and hence can play important role in diagnosing or monitoring pulmonary diseases. In this study, a novel crackle separation technique: iterative envelope mean fractal dimension (IEM-FD) filter is developed for automatically separating crackles from normal breath sounds. The separation of crackles from normal breath sounds is an initial processing stage which can lead to better estimation of crackle features such as number of crackles and two-cycle deflection. To test the crackle separation ability of the IEM-FD filter, a dataset was generated. The performance of the IEM-FD filter was compared with the selected previously published crackle separation techniques using the developed dataset. The experimental results show the proposed method can achieve high accuracy for the number of crackles identified with low computational cost, better quality of crackle separation (less over or underestimation), and good preservation of crackle morphology and hence it may be useful in a clinical setting for determining number of crackles and characteristics of crackles in a recorded lung sound. The proposed IEM-FD filter is applied to two different datasets: (a) longitudinal dataset recorded from 19 idiopathic pulmonary fibrosis (IPF) patients in 7 visits (every visit in 2 months) over a 1 year time period and (b) Cross-sectional dataset recorded from 55 subjects who were referred for a high-resolution computed tomography (HRCT) scan of the chest for various clinical indications. In the longitudinal study application of the IEM-FD filter prior to counting the number of crackles present, allowed evaluation of the association between number of crackles per breath cycle (NOC/BC) and reproducible acoustic features directly generated from the original signal. In this study, it was found that some of these acoustic features were directly associated with NOC/BC therefore might be useful for monitoring progression of IPF. In the cross-sectional study, the IEM-FD filter was applied as a first stage of an automatic crackle counting system which can be used for differentiating idiopathic pulmonary fibrosis patients from patients with other types of pulmonary pathology based on the average NOC/BC. The diagnosis given by two radiologists using the HRCT scan was used as a gold standard for classifying IPF and non-IPF groups. The ability of the automatic system to differentiate IPF patients from non-IPF patients was compared with the individual and average assessment by two respiratory physicians based on listening for the presence of Velcro crackles. Velcro crackles are generally considered as an early clue to the presence of fibrosis. The results show that the automatic system can perform as well as the expert physicians’ assessments and hence could support the auscultatory findings of lung sounds in less specialist clinics. In both the longitudinal and cross-sectional studies, in each recorded lung sound file the number of breathing cycles was audio-visually marked by the Author with the help of open access Audacity software. Audio-visual marking is a highly time-consuming process, therefore a new automatic breath cycle detection method based on the estimation of breathing phases was developed. The performance of the method was tested using both the longitudinal and crosssectional datasets, and a dataset recorded from 10 healthy subjects in 7 visits (each visit in 2 months) over a period of 1 year. The promising results show the possibility of the developed algorithm as an automatic method for breath cycle detection in lung sound recordings