此研究探索ＭＦＣＣ（梅爾倒頻譜係數）為基礎的診斷與以卷積神經網路及徑向基函數網路分析藍牙診器紀錄的肺音。在預處理的部分使用了滑動窗以及時間域從呼吸肺音循環採樣。每一個樣本都是連續Ｋ個值的框訊號，其本身為Ｋ個d 組合成的ＭＦＣＣ向量，其d值為12，用Ｋxd的特徵圖來進行神經診斷及分析。 此研究有效地分類哮喘音循環對正常音循環、爆裂音循環對正常音循環、哮喘音循環對爆裂音循環，其用的是監督式學習中的CNN以及ＲＢＦ網路來進行訓練。這份研究開拓了新的方法其以Ｋxd的特徵圖作為輸入使得深層卷積神經網路得以使用二維的濾波器在隱藏層中，並在數值實驗得到了激勵人心的結果，正確率在哮喘音循環對正常音循環以及爆裂音循環對正常音循環分別為0.77以及0.81。

This work explores MFCC(Mel-frequency cepstrum coefficient) based diagnosis and analysis of bluetooth stethoscope recorded lung sounds using deep convolutional neural network (CNN) and RBF (radial basis function) neural networks. The preprocess applies a sliding window to have time domain samples from a recorded respiratory lung sound cycle. Each sample contains K consecutive frames of signals, which are translated to K d-component MFCC vectors, where d=12, and constitute a K × d feature pattern for neural diagnosis and analysis. This work presents effective classification of wheeze cycles versus normal cycles and crackles cycles versus normal cycle is based on the classifier derived by training a convolutional neural network or an RBF neural network subject to labeled feature patterns for supervised learning. This work pioneers deep convolutional neural diagnosis of lung sounds of respiratory cycles with two-dimensional filters in the first hidden layer for the process of network inputs of K × d feature patterns. Numerical results show encouraging results. The testing accuracy is 0.77 and 0.81 respectively for classification of wheeze cycles versus normal cycles and crackles cycles versus normal cycles.

1 Introduction 1
2 Lung sound preprocess using MFCCs 3
2.1 Mel-scale Frequency cepstral Coefficients(MFCCs) 3
2.2 Data preparation of a single respiratory lung sound cycle 5
2.3 Pairedtrainingandtestingdata 7
3 Method 9
3.1 RadialBasisFunctionNetwork(RBF) 9
3.2 ConvolutionalNeuralNetworks(CNN) 12
4 Experiment 18
4.1 EX-1 18
4.2 The Influence of C 23
4.3 OtherComparison 25
4.4 CNN 32
4.5 LungSoundcycleClassifier 41
5 Conclusion 46

[1] ICBHI 2017 Challenge Respiratory Sound Database. https:// bhichallenge.med.auth.gr/ICBHI_2017_Challenge.
[2] Large Scale data clustering Models and codes. http://134.208.26.59/ AdvancedNA2018/large-scaled-clustering.pdf.
[3] LM learning of MLP and RBF neural networks. http://134.208.26. 59/AdvancedNA/Levenberg-Marquardt_Learning.pdf.
[4] D. Bardou. Lung sounds classification using convolutional neural networks. Artificial Intelligence in Medicine, 88(58–69), 23 April 2018.
[5] B.-X. Chiou. Deep perceptrons and supervised levenberg-marquardt learning with applications to key speaker verification. July 2016.
[6] B. Flietstra, N. Markuzon, A. Vyshedskiy, and R. Murphy. Automated analysis of crackles in patients with interstitial pulmonary fibrosis. Pulm. Med., 10(77):1–7, 2011.
[7] M. Grønnesby, J. C. A. Solis, H. M. Einar Holsbø, and L. A. Bongo. Feature extraction for machine learning based crackle detection in lung sounds from a health survey, 2017.
[8] X. Huang, A. Acero, and H. Hon. Spoken Language Processing: A guide to theory, algorithm, and system development. Prentice Hall, 2001.
[9] N. Jakovljevic. Hidden markov model based respiratory sound classification. Precision Medicine Powered by Health and Connected Health , (39-43), 17 November 2017.
[10] Y. LeCun, Y. Bengio, and G. Hinton. Deeping learning. Nature, 25 February 2015.
[11] S. S. Li. Classification of bluetooth stethoscopic heart sounds using convolution neural network. 2017.
[12] H. Pasterkamp, S. Kraman, and G. Wodicka. Respiratory sounds, advances beyond the stethoscope. Am. J. Respir. Crit. Care Med., 156:974–987, 1997.
[13] G. Serbes, S. Ulukaya, and Y. P. Kahya. An automated lung sound preprocessing and classification system based onspectral analysis methods. In Precision Medicine Powered by pHealth and Connected Health, January 2018.
[14] I. Sutskever and G. Hinton. Imagenet classification with deep convolutional neural networks. In NPS, 2012.
[15] R. Xaviero, A. Pramono, S. Bowyer, and E. Rodriguez-Villegas. Au- tomatic adventitious respiratory sound analysis: A systematic re- view. https://journals.plos.org/plosone/article?id=10.1371/ journal.pone.0177926, May 26 2017.
[16] 李伯璋, 張禹斌, 林子量, 黃信忠, 詹建富, 龐一鳴, and 蔡淑鈴. 從健保 大數據分析，邁向健保改革之路. 臺灣醫界, 60卷(P8 – 13), 2017.