Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, 2005
Development of an Intelligent PDA-based Wearable Digital Phonocardiograph Matias Brusco and Homer Nazeran Department of Electrical and Computer Engineering, The University of Texas, El Paso Texas 79968, USA Email:
[email protected] Abstract — Last year at the EMBC 2004 in San Francisco, we presented a paper entitled: “Digital Phonocardiography: a PDA-based Approach”, which introduced the development of a PDA-based biomedical instrument capable of acquiring, processing, and analysis of heart sounds. In this paper we present a system, which is not only able to record and display the heart sounds in a Pocket PC but also apply several signal processing and statistical techniques to segment the these signals into four parts (the first heart sound, the systole, the second heart sound and the diastole) and implement feature extraction methods for classification purposes. Classification has been achieved using a Multilayer Perceptron (MLP) Artificial Neural Network (ANN). The system was used to classify a number of normal and abnormal heart sounds (normal, aortic regurgitation, aortic stenosis, mitral regurgitation and mitral stenosis) and validate the effectiveness of the statistical segmentation and the feature extraction methods in this environment. Keywords – Digital phonocardiography; heart sounds signal analysis; personal digital assistants (PDAs); wearable biomedical instruments; cardiac valvular diseases.
I. INTRODUCTION The heart sounds result mainly from the interaction of three events: the contraction and relaxation of the atria and ventricles, blood flow, and valve movements. Cardiac auscultation (interpretation by a physician of the heart sounds), is a fundamental part of cardiac diagnosis. Very often, heart auscultation allows the detection of abnormal function of the heart well before the appearance of any symptoms and before a definitive diagnosis can be made. However, detecting relevant symptoms and forming a diagnosis based on sounds heard through a stethoscope is a skill that can take years to acquire and refine. When the heart sounds are displayed graphically, the technique is known as phonocardiography. As phonocardiography is noninvasive and provides valuable information on the integrity and function of the heart valves as well as its hemodynamics performance, it has a high potential for detecting heart diseases [1-5]. Intracardiac phonocardiography combined with digital signal processing techniques strongly renewed the interest in studying the acoustic transmission of the heart sounds and murmurs [1-5]. Main characteristics of heart sounds: such as their timing relationships and components, frequency content, location, time of occurrence in the cardiac cycle and envelope shape of murmurs, can be quantified using digital signal processing techniques [6, 7].
0-7803-8740-6/05/$20.00 ©2005 IEEE.
The application of conventional and advanced digital signal processing techniques to the analysis of heart sounds requires computer-based capabilities, necessitating the use of relatively powerful processors, which are found in desktop or laptop computers. However, when portability and clinical versatility are important considerations, the search for a new portable device capable of robust and reliable analysis of heart sounds is not only justified but highly desirable. This paper, for the first time, addresses some of the central issues involved in modern digital phonocardiography to demonstrate that the realization of a PDA-based intelligent phonocardiograph is now possible. II. MATERIALS Figure 1 shows the system and its components, the Escope electronic stethoscope and the PDA Pocket PC iPAQ hp5550. The electronic stethoscope has three main controls: on/off, bell/diaphragm and volume. The connection between the electronic stethoscope and the Pocket PC is established using a regular audio cable, mono output (from the stethoscope) and stereo input (to the Pocket PC). The Pocket PC screen can be activated by using either the stylus or fingers. The latter option is best suitable for the case of recording since the clinician will have one hand busy (holding the stethoscope) leaving the other one free for holding the Pocket PC and to press the button record. The light weight of the PDA Pocket PC (7.6 oz) allows the physician to hold it for long periods of time without getting tired. In case the Pocket PC no longer needs to be used, the physician can wear it on the Pocket PC belt case and perform routine tasks without carrying the PDA all the time. Also, the long period of battery charge (average of 8 hours of permanent use/day) leaves no room for concerns when autonomy is the issue. The system is capable of recording and displaying heart sounds data. Acquisition is done using the electronic stethoscope with its analog output connected to the microphone input of the Pocket PC. The sound format is fixed at 8 KHz, mono, 8 bit resolution. However, the system is able to read any other file with a different sound format as long as it is in the mono format. The display limit is fixed to Automatic mode, which means the display will show as much data as have been loaded. The end-user can modify some of the settings for more convenience, otherwise default settings are applied.
3506
Figure 1: Components of the PDA-based digital phonocardio -graph. Zooming enables and defines the level of focus on graphs in PDA applications. Zoom style defines the type of zooming to perform. Heart sounds data can not only be acquired and saved, but also can be viewed, processed and classified. By just pressing either the open or save button a dialog window displays where the user chooses the file to be read from or saved into, respectively. The only requirement is that the file be in .wav format and the sound quality be in mono. Data manipulation and analysis built into the system are subdivided into: A) B) C) D) E)
The segmentation algorithm separates the heart sound signal into four parts: the first heart sound, the systole, the second heart sound and the diastole. This algorithm is based on the normalized average Shannon energy of the PCG signal. The segmentation algorithm implemented in LabVIEW was designed as follows: First, the heart sound signal was passed through a pass-band filter in order to work with the frequency components corresponding to those of S1 and S2. The signal was down sampled to reduce the number of computations and consequently the computation time. Second, the Shannon energy over the filtered and downsampled signal was calculated and passed through a peak detector, which found the location and amplitude of the peaks by comparing the Shannon energy signal with a threshold. The threshold was automatically chosen from the RMS (Root Mean Square) of the Shannon energy signal. Third, extra peaks were to be removed. This removal process was achieved by considering the distance between two adjacent peaks. The remaining peaks after a two-stage-extrapeak removal were either S1 or S2. Discrimination between S1 and S2 was done based on the mean of the final distances. After S1 and S2 have been detected, systole duration (interval between S1 and S2 in the same cycle), diastole duration (interval between S2 and S1 of the next cycle) and heart rate (number of cycles per minute) were calculated. Fig. 2 shows an abnormal heart sound (top display, green/yellowish color), S1 components (blue traces of the bottom display) and S2 components (red tracings in the bottom display).
Filtering Amplitude and Power Spectrum Heart Sound Segmentation Vector of Powers Formation Classification III. METHODS
A. Segmentation of Heart Sound Signals The segmentation of the phonocardiogram signal is the first step and the most important step in the analysis and automatic diagnosis of heart sounds. The majority of the attempts in segmenting the heart sounds use the ECG (electrocardiogram) signal and/or carotid pulse as reference signal. In our system we implemented the segmentation based on the heart sound signal itself without a reference to the ECG signal by using a set of normal recordings, hence eliminating the need for application of ECG electrodes and amplifiers.
Figure 2: Heart sound signal segmentation.
3507
B. Vector of Powers Formation and Classification In a study Özgür et al [10] performed signal segmentation by the wavelet detail coefficients at the second and sixth levels into 64 windows, each window containing 128 discrete data points. However, since the duration of a single period is not constant (it varies with heart rate), the size of the window should decrease or increase with lower or higher heart rates, respectively. Also the sub-bands should be chosen by considering the concentration of energy for normal and abnormal heart sounds. By selecting S1 as the starting point, a window which contained a single period of the heart sound was formed. Two sub-bands were considered to be important for feature extraction, one band (the low frequency contents band) which mainly represents normal components (62.5 Hz – 125 Hz) and another band (the high frequency contents band) that identified itself more with abnormal components (250 Hz – 500 Hz). The feature vector corresponding to a single period comprised of the power of the sub-windows of the decomposition levels generating a vector with 64 (32+32) elements. Not only was the frequency content considered in this algorithm but also the time information on each one of these bands was included.
Figure 3: Discriminating sub-bands (lower trace) and the feature vector (upper trace) created from them. TABLE I LIST OF RECORDINGS UTILIZED FOR TRAINING AND TESTING
Classification was performed using two MLP ANNs trained by the back propagation algorithm, two layers, 64 inputs, 3 and 5 outputs (for classification into 3 and 5 categories, respectively). Classification into three categories refers to: 1) Normal heart sound 2) Diastolic murmur 3) Systolic murmur Classification into five categories refers to: 1) Normal heart sound 2) Aortic Regurgitation 3) Aortic Stenosis 4) Mitral Regurgitation 5) Mitral Stenosis Figure 3 shows an example of a feature vector formation for a case of aortic regurgitation. In the lower part of the display two sub bands, light blue for the normal components and fuchsia for the abnormal components are clearly observed. The upper part of the display represents the feature vector extracted from these sub-bands, which are fed into the ANN. IV. RESULTS AND DISCUSSION A collection of 42 heart sounds with a total of 263 cycles, having an average of 62.6 cycles per recording, was selected. Table 1 is a summary of the recordings utilized for the training and testing of the ANN. Table 2 is a summary of the results obtained using the segmentation algorithm and classification using the Vector of Powers.
Sample Record
Training
Testing
Total
Normal Heart Sound
3
6
9
Aortic Regurgitation
2
22
4
Aortic Stenosis
7
5
12
Mitral Regurgitation
6
5
11
Mitral Stenosis
3
3
6
Total
21
21
42
TABLE II SUMMARY OF RESULTS Right cycles detected
Heart Rate error
Correct Specific
Correct General
Diagnosis
Diagnosis
79.3 %
16.0 %
47.6 %
71.4 %
Wrong
No
Diagnosis
Diagnosis
9.5 %
19.0 %
The automatic segmentation algorithm was found to be effective in segmenting the phonocardiogram signal into four parts. The algorithm showed 79.32 % success rate in 42 recordings. This result included 259 cycles of heart sounds. This is a good starting point for further analysis and classification of heart sound signals. With this segmentation, several features were extracted from each segment, including peak intensity, peak location, systole duration, diastole duration and heart rate. The heart rate error was 16.02%.
3508
Heart sound features were determined for classification into 3 and 5 categories. Classification into 3 categories included normal heart sound, diastolic murmur and systolic murmur. Classification into 5 categories included normal heart sound, mitral regurgitation, mitral stenosis, aortic regurgitation and aortic stenosis. There were 42 records for heart sound analysis, 21 for training the artificial neural network and 21 for testing.
which allows physicians to extract important information from heart sounds. The band-decomposition method of heart sounds for feature extraction and classification seems to be a reliable approach for detection of cardiac abnormalities. The accuracy may be further enhanced by improvement of these methods or even by combining them with other signalcharacterization methods. ACKNOWLEDGMENT
The implementation of the artificial neural network provided with these feature vectors showed promising results. In the case of classification into 5 categories, the achieved accuracy was 47.62 %. For the case of classification into 3 categories, the achieved accuracy was 71.43%. The improved performance in classification into 3 categories was due to the broader range of values for each class, facilitating the identification (the higher the number of sub-classifications, the lower the accuracy of the network). To evaluate the overall performance of the system as a screening device, the sensitivity, specificity and accuracy parameters were calculated. Once again better results were seen in the case of classification into 3 categories. There is still much work to be done in this area. The next step is to polish and refine the system to improve the final results and outcomes in order to obtain high values of accuracy, ideally above 95% for all cases. In order to achieve this desirable level of accuracy, work needs to be done in the following areas: x
Enlarging the heart sounds database with recordings from real patients.
x
Refining the segmentation technique.
x
Experimenting with feature vectors of different sizes.
x
Retraining the artificial neural changing its internal variables.
network
The authors would like to acknowledge the support of Texas Instruments Fund and would like to thank Dr Sergio Cabrera, Texas Instruments Foundation Professor of Digital Signal Processing for his continued support of this research. REFERENCES [1] Rangayyan, R. M., and Lehner, R.J., “Phonocardiogram Signal Analysis: A review”, Critical Reviews in Biomedical Engineering, 15, 211, 1988. [2] Durand L. G., and Pibarot P., “Digital Signal Processing of the Phonocardiogram: Review of the Most Recent Advancements”, Critical Reviews in Biomedical Engineering, 1995. [3] Tilkian C., “Understanding Heart sounds and Murmurs, with an introduction to lung sounds”, W.B. Saunders Company, fourth edition, 2001. [4]
“Auscultation Skills: Breath and Heart Sounds”, Springhouse Corporation, 1999.
[5]
Tavel, M.E., “Clinical Phonocardiography and External Pulse Recording”, Chicago: Year Book Medical Publishing, 3rd Ed, 1978.
[6] Durand L. G., “Comparison of Pattern Recognition Methods for Computer – Assisted Classification of Spectra of Heart Sounds in Patients with a Porcine Bioprosthetic Valve Implanted in the Mitral Position”, IEEE Transactions in Biomedical Engineering, vol. 37, No.12, December 1990. [7] Obaidat M. S., and M. M. Matalgah, “Performance of the Short Time Fourier Transform and Wavelet Transform to Phonocardiogram Signal Analysis”, Proceedings of the ACM/SIGAPP Symposium on Applied Computing: technological challenges of the 1990's. pp: 856 – 862, 1990. [8] Özgür, S., Zümary, D., and Tamer Ö., “Classification of heart Sounds By Using Wavelet Transforms”, Proceedings of the Second Joint EMBS/BMES Conference, Houston, TX, USA, pp. 128-129, 2002.
V. CONCLUSIONS The phonocardiogram (PCG) signal is a complex signal. PCG signal contains numerous nonstationary or transient characteristics such as timing of the heart sounds, their components, locations in the cardiac cycle and a fast change of frequency as time progresses. To find these characteristics, time-frequency analysis techniques could be deployed to provide better diagnostic results than the Fourier transform. In developing our system, powerful signal processing methods such as short-time Fourier transform, wavelet transform and wavelet decomposition techniques were deployed as a foundation to develop a user-friendly interface
3509
