Dubey et al., International Journal on Emerging Technologies 11(3): 320-327(2020) 320 International Journal on Emerging Technologies 11(3): 320-327(2020) ISSN No. (Print): 0975-8364 ISSN No. (Online): 2249-3255 Comparative Analysis of Wavelet bi-spectrum and Power Spectrum Features for the Classification of Adventitious Lung Sound Rupesh Dubey 1,2 , Rajesh M. Bodade 3 and Divya Dubey 4 1 Associate Professor, Department of Electronics and Communication Engineering, Indore (Madhya Pradesh), India. 2 Research Scholar, Military College of Telecomm. Engineering (MCTE), MHOW, Ministry of Defense, Government of India, DAVV University, Indore (Madhya Pradesh), India. 3 Professor, Military College of Telecommuncation Engineering (MCTE), MHOW, Ministry of Defense, Government of India, Indore (Madhya Pradesh), India. 4 Assistant Professor, Department of Electronics and Communication Engineering, SDBCE Indore (Madhya Pradesh), India. (Corresponding author: Rupesh Dubey) (Received 15 January 2020, Revised 14 April 2020, Accepted 15 April 2020) (Published by Research Trend, Website: www.researchtrend.net) ABSTRACT: Lung sounds convey essential information about the health of the patient. Accurate detection, followed by the analysis of adventitious sounds, is still a significant challenge. Suitable techniques for the same shall empower to deal with the problem of scarcity of expert physicians. The conventional techniques suffer from the identification of suitable features that can provide higher accuracies of detection. Here features based on higher-order spectral for wavelet bi-spectrum (WBS) and Power Spectrum (PS) are introduced and analyzed for the classification of adventitious sounds of lungs, namely wheezes, crackle, and normal sound. Comparison is presented between higher-order spectral features based on wavelet bi- spectrum and power spectrum using classifiers like decision tree, SVM, k-NN, ensemble learner. Here results of accuracy are explored for the feature and sub classifier combination. Many feature classifier combination has yielded accuracy as high as 100%. Average accuracy achieved in the case of wavelet bi-spectrum outperforms that achieved for features based on the power spectrum. Keywords: Higher-order spectral, k-NN, SVM, classification tree, wheezes, crackle. Abbreviations: Global Peak GP, Local Peak LP, Support Vector Machine SVM, k-Nearest Neighbor k-NN, A- Asthma, Acy. Accuracy, AWS- Available with Source, BPF- Band Pass Filter, Bu.- Butter Worth C-Crackle, CA-Commercially Available, CAcc. Contact Accelerometer, CC- Coarse Crackle, DT- Decision Tree, DR- Dry rale, DS- Digital Stethoscope, EL- Ensemble Learner, ES- Electronics Stethoscope, Fpr- Friction pleural rub, HLS-Healthy Lung Sound, HPF- High Pass Filter, Lab. – Laboratory, LPF- Low Pass Filter, MR- Moist rale, N-Subjects/Signals, NA- Not Available, NM- Not Mentioned, NU- Not Used O- Order, S-Sensitivity, SF- Sampling Frequency, So- Stridor, SP- Specificity, SR- Sampling Rate, W- Wheeze, WCC- Wavelet Packet Transform, WBP-Wavelet bi-phase, WBS-Wavelet bi-spectrum, Yrs.- Years. I. INTRODUCTION As per WHO [1], over 80% of death in the case of obstructive pulmonary diseases like Asthma and Chronic Obstructive Pulmonary Diseases (COPD) occurs in low and lower-middle-income countries. Adventitious sounds like wheezes and crackle are heard during breathing cycles of patients suffering from obstructive pulmonary diseases [2]. Wheezes have a time length greater than 150 ms, whereas for crackle, it less than 20 ms [2]. Wheezes exhibit continuous waveform, while for crackle, the waveforms are discontinuous. Experienced human ears can easily recognize adventitious sounds. The traditional stethoscope was invented in the year 1821. These suffer limitations of lower frequency (<120Hz). Previous signal processing techniques for the identification of wheezes were based on time-expanded waveforms [3]. Further works followed to focus on peaks detection in the spectrum, their amplitude, and pitch range [4, 5]. Researchers employed various classification techniques for the labeling of features like neural networks [6]. Shi et al ., (2019) also used BPNN as a classification method with the WCC feature but achieved 92.5% accuracy [7]. CNN, which was initially introduced for image processing, is being employed for lung sound analysis with a lower number of data set achieved 97% accuracy [8]. The accuracy of algorithms that are based on the above criteria depends upon the acoustics amplitude of the signals. There remains a need to focus on techniques independent of amplitudes, so that accuracy of detection remains unaffected by location and device of sound capture. The obstruction in airways results in changes in non-linear harmonic peaks interaction. They have non-stationary characteristics; contain non-linearity in their harmonic interactions. Wheezes show a phase relationship with quadratic phase coupling. Some researchers have employed bi- coherence [9] and phase spacing features for the analysis of lung sound [10]. Also, works are clubbing continuous wavelet transform with third-order spectra [11]. The severity of asthma is also identified in some of the approaches by using integrated power [12]. e t