1. Tonal Index in Digital Recognition of Lung Auscultation Marcin Wiśniewski,Tomasz Zieliński 2016/7/12 Signal Processing Algorithms, Architectures,Arrangements, and Applications Conference Proceedings (SPA), 2011 Presenter : Kun-Han Jhan Advisor : Dr. Chun-Ju Hou Date : 2012.12.26

2. Outline Introduction Lung sounds Wheezes Descriptors Testing methodology Experiments Conclusions References 2016/7/12

3. Introduction Asthma ◦ Secretion or mucus  ◦ Muscle contraction ◦ Main indicator  An appearance of wheezes in breath cycle 2016/7/12

4. Introduction Lungs auscultation ◦ A non-invasive test in asthma ◦ Evaluate the stage of the disease ◦ Evaluate the level of wheeze appearance The main problem of lungs auscultation ◦ Depend on doctor’s experience ◦ Subjective 2016/7/12

5. Introduction The advantage of digital lung auscultation ◦ Objective and unambiguous ◦ Telemedicine  Doctor can see the results without necessity of direct meeting  Patients not have to go to the hospital  Increase the comfort of patient’s life  Decrease stress in direct meetings with doctors 2016/7/12

6. Lung sound Location ◦ Trachea ◦ Bronchus ◦ Alveoli Characteristics ◦ Like a noise ◦ Frequency rang : 20 ~ 1.6kHz 2016/7/12 Lung sound

7. Wheeze ◦ A single or multi tone sound  Duration > 80 ms  Frequency: 100 ~ 1k Hz ◦ Mixed normal lung sounds with wheezes 2016/7/12

8. Wheezes Descriptors Features ◦ Kurtosis (K) ◦ Spectral Peaks Entropy (SPE) ◦ Frequency Ratio (FR) ◦ Spectral Flatness (SF) ◦ Tonal index 2016/7/12

9. Wheezes Descriptors Kurtosis (K) ◦ Measure a level of peakedness of a probability distribution in time domain ◦ k = 3 (noisy signal with normal or sub-gaussian distribution) ◦ k > 3 (the signal with wheezes) 2016/7/12 x : input signal μ : mean σ 2 : variance 眾數眾數 中位數中位數 μ

10. Wheezes Descriptors 2016/7/12 Spectral peaks entropy (SPE) ◦ Frequency domain Cn : peak value of frequency spectrum : total sum of these peaks Entropy:

11. Wheezes Descriptors 2016/7/12 Frequency ratio ◦ Frequency feature ◦ The signal with wheezes has higher values of this ratio than normal lung sounds ◦ The area under the power spectral density of ROI ◦ The area of total power spectral density ◦ FR descriptor was modified and tested once again as a Energy Ratio (ER) descriptor

12. Spectral flatness ◦ A signal feature defined in frequency domain ◦ A ratio of geometrical and arithmetical mean values Wheezes Descriptors 2016/7/12 : geometrical mean value : arithmetical mean value

13. Tonal index ◦ A spectral feature ◦ MPEG psychoacoustic model ◦ FFT module and phase Wheezes Descriptors 2016/7/12

14. Testing methodology 2016/7/12 Tonal signals simulation  Artificial wheezes: multi-frequency signals with random three frequencies (100~1200Hz)  Normal breathing signals Features testing  Signal samples: 1024 points  Add white Gaussian noise with different SNR scale  Sampling frequency: 8 KHz Recognition  SVM(Support Vector Machine)

15. Experiments The modeled wheezes ◦ Artificial noise with normal ◦ Training signals with different gains To recognition process ◦ 100 samples 2016/7/12

16. Experiments Artificial signals

17. Experiments 2016/7/12 Artificial signals

18. Experiments Hybrid data ◦ Artificial wheezes added to the normal lung sounds taken from chest auscultation. ◦ 8 kHz/16-bit recorder ◦ Panasonic WM-61 microphones To recognition process ◦ 28 samples 2016/7/12

19. Experiments Hybrid signals

20. Experiments 2016/7/12 Hybrid signals

21. Experiments 2016/7/12 Artificial data ◦ TI ◦ SPE. Hybrid data ◦ TI In both study the FR shows the worst effectiveness.

22. Conclusions The tonal index is more sensitive to tonality in noisy signals reaches full efficiency in lower SNR as well. Increasing number of features in algorithm not necessarily improves effectiveness of recognition. The best result reaches the algorithm with 2 features{TI,ER}– 94.2% and {K,TI}– 94.6% effectiveness. 2016/7/12

23. References [1] Aydore, S.; Sen, I.; Kahya, Y.P.; Mihcak, M.K., Classification of respiratory signals by linear analysis, Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE Publication Year: 2009, pp. 2617 - 2620 [2] Jianmin Zhang; Wee Ser; Jufeng Yu; Zhang, T.T.; A Novel Wheeze Detection Method for WearableMonitoring Systems Intelligent Ubiquitous Computing and Education, 2009 International Symposium on Publication Year: 2009 pp. 331 - 334 [3] A.H. Gray, J.D. Markel, A spectral-flatness measure for studying the autocorrelation method of linear prediction of speech analysis, IEEE Trans. Acoust. Speech Signal Process., 1974, 22, pp. 207–217 [4] H. Pasterkamp, S.S. Kraman, G. R. Wodicka, Respiratory Sounds. Advances Beyond the Stethoscope, Am. J. Respir. Crit. Care Med., Volume 156, Number 3, pp. 974-987, September 1997 2016/7/12