Concepts of machine learning are potentially useful tools in reducing human effort and time. In rural India, there is a dearth in accessibility and affordability of excellent and sound medical diagnostic facilities. Implementing machine learning concepts to predict the presence of an illness as a part of an automated diagnostic system can go a long way in bridging the gap. As a part of developing one such system to diagnose respiratory illness using respiratory sound, an attempt has been made to analyze the performance of a set of six classifiers that include the nearest neighbor, the parzen window, the support vector machine, the relaxation batch margin, the relaxation single sample margin and the least square classifier with respect to their ability to classify the healthy and the non-healthy subjects. Four sets of features have been used, namely the statistical feature set, feature set based on the Gray Level Co-occurrence Matrix (GLCM) obtained from the spectrogram of the sound, the Mel- Frequency Cepstral Coefficients (MFCC) and Wavelet Packet Decomposition Coefficients. These features have been employed individually and in combinations to train and test the classifier performance. The performance has been interpreted by obtaining the confusion matrix and parameters such as sensitivity, specificity, precision, negative predictive value and accuracy and also by plotting the Receiver Operating Characteristic (ROC) curve for each classifier. Based on sensitivity that measures the ability of a classifier to identify the true class correctly and the accuracy that measures the correctness of the predicted class, it is inferred that the wavelet packet decomposition coefficients and MFCC are good features in characterizing respiratory sound. Further, in terms of the classifier, the relaxation classifiers, both batch margin and single sample margin and the support vector machine that classifies using a hyper plane yielded appreciable results with a maximum accuracy of 0. 83 in clear contrast to that of Parzen and Nearest Neighbor. Further the result demonstrates that the classifiers used in this work will assist the physician in diagnosing the abnormal nature of respiratory sound and the system can be used as a mass screening tool.

1. Sovijarvi ARA, Maimberg LP, Charbonneau G, Vanderschoot J, Dalmasso F, Sacco C, et al. 2000. Characteristics of breath sounds and the adventitious respiratory sounds. European Respiratory Review, 10: 591–596.
2. Forgacs P. 1967. Crackles and wheezes. Lancet, 2: 203–205.
3. H. Pasterkamp. 1990. "RALE: A computer-assisted instructional package," Respiratory Care, 35: 1006.
4. Respiratory sounds from www. littmann. com/wps/portal/3M/en_US/3M-Littmann/stethoscope/littmann-learning-institute/heart-lung-sounds/
5. Pasterkamp H, Fenton R, Tal A, Chernick V. 1984. Tracheal vs lung sounds in acute asthma. Am Rev Respir Dis, 129: 256A.
6. R Palaniappan and K Sundaraj. 2013. 'Respiratory Sound Classification using Cepstral Features and Support Vector machine'. Proceedings of 2nd Conference on Recent advances in Intelligent Computational Systems, IEEE; Kerala, India.
7. Kandaswamy, A and Kumar, CS and Ramanathan, RP and Jayaraman, S and Malmurugan, N. 2004. 'Neural classification of lung sounds using wavelet coefficients'. Computers in Biology and Medicine, 34(6): pp. 523-37.
8. Kahya, YP and Bayatli, E and Yeginer , M and Ciftci, K and Kilinc, G. 2003. 'Comparison of different feature sets for respiratory sound classifiers'. Proceedings of the 25th Annual International Conference of the IEEE EMBS; Cancun, Mexico: pp. 28532856.

Classifier Roc Curve Support Vector Machine Respiration diagnostic Tool