本論文發展了基於脈搏聲音訊號及時頻轉換之心律不整與器質性心臟病之人工智慧/機械學習分類辨識演算法。在心律不整方面，本研究探討了心房顫動、心房撲動、心室早期收縮、心房早期收縮；在器質性心臟病方面，本研究探討了主動脈瓣反流、主動脈瓣狹窄、肥厚型心肌症、鬱血性心衰竭。在訊號感測部分，本論文透過微機電麥克風，量測手腕橈動脈在脈搏跳動時所產生的聲音，本研究稱此訊號為脈音訊號(Pulse AudioGram, PAG)。根據血液動力學的理論，血液循環時在橈動脈上會產生壓力波，本研究將此壓力波變化以聲音的方式記錄。在訊號處理的過程，脈音的類比訊號首先會轉換為數位訊號，接著再透過連續小波時頻轉換將時域的脈音訊號轉換為時頻圖，接著將脈音訊號的時頻圖作為人工智慧辨識器（卷積神經網路）/機械學習分類辨識器（決策樹、支援向量機、最近鄰居法）的輸入。使用卷積神經網路辨識器時，本論文直接以時頻圖作為系統輸入，使用決策樹、支援向量機、最近鄰居法辨識器時，本論文搭配圖像檢索算法對時頻圖進行特徵提取，再以機械學習辨識器進行辨識。最後以k-fold cross validation (k = 5)以及Leave one subject out cross validation方法進行交叉驗證。在k-fold cross validation的交叉驗證中：人工智慧辨識器（卷積神經網路）辨識心律不整之最佳正確率、敏感度、特異性分別為：91.61%、92.31%、85.57%；人工智慧辨識器（卷積神經網路）在辨識器質性心臟病之最佳正確率、敏感度、特異性分別為：97.13%、 98.02%、 91.84%。機械學習辨識器（支援向量機）辨識心律不整之最佳正確率、敏感度、特異性分別為：91.34%、92.55%、80.95%；機械學習辨識器（支援向量機）辨識器質性心臟病之最佳正確率、敏感度、特異性分別為：99.01%、99.18%、96.97%。在Leave one subject out cross validation的交叉驗證中：人工智慧辨識器（卷積神經網路）在辨識器質性心臟病之最佳正確率、敏感度、特異性分別為：74.46%、70.93%、82.14%。機械學習辨識器（支援向量機）辨識器質性心臟病之最佳正確率、敏感度、特異性分別為：67.14%、65.73%、64.29%。在討論中，本研究發現機器學習抓取之特徵與肉眼觀察之特徵相似，且可能可以對應到疾病與其表現之生理現象，在竇性心律的時頻圖中，機器學習可抓取到能量分布的中心且形成一規律的心跳；在主動脈瓣反流的時頻圖中，能量分布的頻帶較大，可能與血液逆流時造成的心雜音有關；在主動脈瓣狹窄的時頻圖中，在3 Hz的地方有一連續能量訊號，可能為射血時間延長造成；在肥厚型心肌症的時頻圖中，機器學習抓取的特徵在心跳間格且無能量處，可能是流出道與主動脈無壓力差之瞬間；在鬱血性心衰竭中，機器學習抓取到的特徵大小不同，可能為心臟收縮力不足，造成心跳忽強忽弱的現象。本研究搭配時頻轉換以及人工智慧/機械學習演算法，成功的辨別出手腕橈動脈壓力波在不同心律不整/器質性心臟病的差異。在未來有潛力作為快速篩檢心臟疾病的辨識演算法。

This thesis develops an artificial intelligence/machine learning classification algorithm on arrhythmia and structural heart disease, based on pulse audiogram and time frequency transformation. In arrhythmia, this study explores atrial fibrillation, atrial flutter, ventricular premature contraction, and atrial premature contraction; in structural disease, this study explores aortic regurgitation, aortic stenosis, hypertrophic cardiomyopathy, and congestive heart failure. In the part of signal collecting, this thesis utilizes MEMS microphone to measure the sound of pulse beat on the radial artery of the wrist. The sound is named Pulse AudioGram (PAG) by this study. According to the theory of hemodynamics, the circulation of blood on the radial artery results in the pressure wave. The variation of the pressure is recorded by sound in this study. In signal processing, the analog signal of the PAG signal is converted into the digital signal at first; then the digital signal is transformed into to time frequency representation by a continuous wavelet transform. Time frequency representation is the input of artificial intelligence (convolutional neural network)/machine learning (decision tree, support vector machine, k-nearest neighbors). While using the convolutional neural network as the classifier, this thesis directly utilizes time frequency representation as the input. While using a decision tree, support vector machine, and k-nearest neighbors as the classifier, this thesis utilizes bag-of-feature which is a kind of feature extraction to produce features for the input. Last, this thesis utilizes k-fold cross-validation (k=5) and leave one subject out (LOSO) cross-validation for cross validation. In k-fold cross-validation, while using an artificial intelligence (convolutional neural network) classifier for arrhythmia classification, the best accuracy, sensitivity, and specificity are 91.61%, 92.31%, 85.57%, respectively; while using an artificial intelligence (convolutional neural network) classifier for structural heart disease classification, the best accuracy, sensitivity, and specificity is 97.13%, 98.02%, 91.84%, respectively. While using a machine learning (support vector machine) classifier for arrhythmia classification, the best accuracy, sensitivity, and specificity is 91.34%, 92.55%, 80.95%, respectively; while using a machine learning (support vector machine) classifier for structural heart disease classification, the best accuracy, sensitivity, and specificity is 99.01%, 99.18%, 96.97%, respectively. In leave one subject out cross-validation, while using artificial intelligence classifier (convolutional neural network) for structural heart disease classification, the best accuracy, sensitivity, and specificity is 74.46%, 80.93%, 82.14%, respectively. While using a machine learning (support vector machine) classifier for structural heart disease classification, the best accuracy, sensitivity, and specificity is 67.14%, 65.73%, 64.29%, respectively. In the discussion, this study found that the characteristics captured by machine learning are similar to those of observation of the naked eye, and may correspond to the physiological phenomena of the heart disease. In the spectrogram of sinus rhythm, machine learning can capture the center of energy distribution and look like regular heartbeats; in the spectrogram of aortic regurgitation, the frequency band of energy distribution is wide, which may be related to the heart murmur caused by blood reflux; in the time spectrogram of aortic stenosis, there is a continuous energy signal at 3 Hz, which may be caused by prolonged ejection time; in the spectrogram of hypertrophic cardiomyopathy, machine learning caught the characteristics of the heartbeat intervals with no energy, which may relate to no pressure difference between the outflow and the aortic artery; in the spectrogram of congestive heart failure, the size of features are different, which may be due to insufficient contraction of the heart. This study using time frequency transformation and artificial intelligence/machine learning algorithm successfully recognizes the difference between different arrhythmia/structural heart disease. The classification algorithm is potential for rapid screening of heart disease in the future.

[1] https://goo.gl/eVmaUb
[2] https://goo.gl/yd34XF
[3] Myerburg RJ, Interian A Jr, Mitrani RM, Kessler KM, Castellanos A., “Frequency of sudden cardiac death and profiles of risk,” Am J Cardiol, 80:10F-19F, 1997.
[4] R.M. Rangayyan, “Biomedical Signal Analysis: A Case-Study Approach,” Wiley–Interscience, New York, 2001.
[5] http://www.hemodynamicsociety.org/hemodyn.html
[6] Sowton E., “Hemodynamic consequences of arrhythmias,” J Cardiovasc Pharmacol, 17 (suppl. 6): S20–S23, 1991.
[7] https://www.hrsonline.org/Patient-Resources/Heart-Diseases-Disorders
[8] HA Guvenir, B. Acar, G. Demiroz, and A. Cekin, “A supervised machine learning algorithm for arrhythmia analysis,” Computers in Cardiology, pp. 433-436, 1997.
[9] Martis R. J., Krishnan M. M., Chakraborty C., Pal S., Sarkar D., Mandana K. M., “Automated screening of arrhythmia using wavelet based machine learning techniques,” Journal of Medical Systems, 2010.
[10] K. Polat, S. Sahan, S. Gnes, “A new method to medical diagnosis: Artificial immune recognition system (AIRS) with fuzzy weighted pre-processing and application to ECG arrhythmia,”Exp. Syst. Appl., vol. 31, no. 2, pp. 264-269, Aug. 2006.
[11] B. M. Asl, S. K. Setarehdan, M. Mohebbi, “Support vector machine-based arrhythmia classification using reduced features of heart rate variability signal,” Artificial Int. in Medicine, vol. 44, pp. 51-64, 2008.
[12] L. Qiao, “Ventricular fibrillation and tachycardia classification using a machine learning approach,” IEEE Trans. Biomed. Eng., vol. 61, no. 6, pp. 1607-1613, Jun. 2014.
[13] S. Osowski, L. T. Hoai, T. Markiewicz, “Support vector machine-based expert system for reliable heartbeat recognition,”IEEE Trans. on Biomed. Eng., vol. 51, no. 4, pp. 582-589, Apr. 2004.
[14] LM Eerik, J Vanschoren, MJ Rooijakkers, R Vullings, RM. Aarts, “Decreasing the false alarm rate of arrhythmias in intensive care using a machine learning approach,” Computing in Cardiology, vol. 42, pp. 293-296, 2015.
[15] V. Kalidas et al., “Enhancing accuracy of arrhythmia classification by combining logical and machine learning techniques,” 2015 Computing in Cardiology Conference (CinC), IEEE, pp. 733-736, 2015.
[16] J. J. Oresko, et al., “A Wearable Smartphone-Based Platform for Real-Time Cardiovascular Disease Detection Via Electrocardiogram Processing,” IEEE Trans. on Information Technology in Biomedicine, vol. 14, pp. 734-740, 2010.
[17] Akin Özçift, “Random forests ensemble classifier trained with data resampling strategy to improve cardiac arrhythmia diagnosis,” Computers in Biology and Medicine, vol. 41, pp. 265-271, 2011.
[18] C.H. Antink et al., “Reducing false arrhythmia alarms using robust interval estimation and machine learning,”2015 Computing in Cardiology Conference (CinC), IEEE, pp. 285-288, 2015.
[19] Rajpurkar Pranav, Awni Y. Hannun, Masoumeh Hagh-panahi, Codie Bourn, Andrew Y. Ng, “Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks,” 2017.
[20] U. R. Acharya, H. Fujita, O. S. Lih, Y. Hagiwara, J. H. Tan, M. Adam, "Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network", Inf. Sci., vol. 405, pp. 81-90, 2017.
[21] S. Kiranyaz et al., "Real-time patient-specific ECG classification by 1-D convolutional neural networks", IEEE Trans. Biomed. Eng., vol. 63, no. 3, pp. 664-675, Mar. 2016.
[22] U.R. Acharya, H. Fujita, S. L. Oh, A. Muhammad, J.H. Tan, K.C. Chua, “Automated detection of coronary artery disease using different durations of ECG segments with convolutional neural network,” Knowledge-Based Systems, vol. 132, pp. 62-71, Sep. 2017.
[23] M. M. Al Rahhal, Y. Bazi, H. AlHichri, N. Alajlan, F. Melgani, R. R. Yager, "Deep learning approach for active classification of electrocardiogram signals", Inf. Sci., vol. 345, pp. 340-354, Jun. 2016.
[24] M. K. Sarkaleh, A. Shahbahrami, "Classification of ECG arrhythmias using Discrete Wavelet Transform and neural networks", Int. J. of Comput. Sci. Eng. and Applicat. (IJCSEA), vol. 2, no. 1, pp. 1-13, 2012.
[25] Rine Nakanishi et al., “Machine Learning Adds to Standard Clinical & CAC Assessments in Predicting 10-Year Coronary Heart Disease & Cardiovascular Disease Deaths: Insight From the Coronary Artery Calcium Consortium of 66,636 Patients,” Circulation, vol. 136, A15137, Nov. 2017.
[26] Robi Goswami et al., “Coronary Artery Disease Learning and Algorithm Development Study: Early Analysis of Ejection Fraction Evaluation,” Journal of the American College of Cardiology, vol. 69, Issue 11 Supplement, March 2017.
[27] M. Rathi, B. Narasimhan, “Data Mining, Soft Computing, Machine Learning and BioInspired Computing for Heart Disease Classification/ Prediction – A Review,” International Journal of Advanced Research in Computer Science and Software Engineering 7(4), vol. 7, pp. 7-10, Apr. 2017
[28] Go AS, Hylek EM, Phillips KA, Chang Y, Henault LE, Selby JV, Singer DE, “Prevalence of diagnosed atrial fibrillation in adults: national implications for rhythm management and stroke prevention: the AnTicoagulation and Risk Factors in Atrial Fibrillation (ATRIA) Study,” JAMA. 2001.
[29] Pfeiffer, E. R., Tangney, J. R., Omens, J. H., McCulloch, “A. D. Biomechanics of cardiac electromechanical coupling and mechanoelectric feedback,” Journal of biomechanical engineering, 136, 021007, 2014.
[30] Flaker, G. C., Belew, K., Beckman, K., Vidaillet, H., Kron, J., Safford, R., et al. “Asymptomatic atrial fibrillation: demographic features and prognostic information from the Atrial Fibrillation Follow-up Investigation of Rhythm Management (AFFIRM) study,” American heart journal, 149, 657-663, 2005.
[31] Lian, J., Wang, L., Muessig, D., “A simple method to detect atrial fibrillation using RR intervals,” The American journal of cardiology, 107, 1494-1497, 2011.
[32] Lee, J., Nam, Y., Chon, K., “Time-varying coherence function for atrial fibrillation detection,” IEEE Trans. on Biomed. Eng., 60, 2783-2793, 2013.
[33] Jinseok, L., Reyes, B. A., McManus, D. D. Mathias, O., Chon K. H., “Atrial fibrillation detection using an iPhone 4S,” IEEE Trans. on Biomed. Eng, 60, 203-206, 2013.
[34] Das, R., Turkoglu, I., Sengur, A., “Effective diagnosis of heart disease through neural networks ensembles,” Expert Systems with Applications, 36, 7675–7680, 2009.
[35] Das, R., Sengur, A., “Evaluation of ensemble methods for diagnosing of valvular heart disease,” Expert Syst., 37, 5110-5115, 2010.
[36] Carlos A Morillo et al., “Atrial fibrillation: the current epidemic,” J Geriatr Cardiol, vol. 14(3), pp. 195-203, Mar. 2017.
[37] Daubechies I, Lu JF, Wu HT, “Synchrosqueezed wavelet transforms: An empirical mode decomposition-like tool,” Applied and Computational Harmonic Analysis, 2(30): 243-261, 2011.
[38] Breiman, L., J. H. Friedman, R. A. Olshen, and C. J. Stone, “Classification and Regression Trees,” Boca Raton, FL: Chapman & Hall, 1984.
[39]Vapnik. V., “Statistical learning theory,” Wiley Inc, 1998.
[40] Boser, B. E., Guyon, I. M. and V. Vapnik, “A training algorithm for optimal margin classifiers,” In Fifth Annual Workshop on Computational Learning Theory, Pittsburgh, ACM, 1992.
[41] E.L. Allwein, R.E. Schapire, and Y. Singer, “Reducing multiclass to binary: A unifying approach for margin classifiers,” Machine Learning: Proceedings of the Seventeenth International Conference, 2000.
[42] Hubel, D., and Wiesel, T. N., “Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex,” J. Physiol., 160, 106–154, 1962
[43] MATLAB and Signal Processing Toolbox Release 2017b, The MathWorks, Inc., Natick, Massachusetts, United States.
[44] MATLAB and Statistics and Machine Learning Toolbox Release 2017b, The MathWorks, Inc., Natick, Massachusetts, United States.
[45] MATLAB and Neural Network Toolbox Release 2017b, The MathWorks, Inc., Natick, Massachusetts, United States.