Non-Invasive Heart Murmur Detection via a Force Sensing-Based System and SVM Classification

Transcription

1 Int'l Conf. Health Informatics and Medical Systems HIMS'15 31 Non-Invasive Heart Murmur Detection via a Force Sensing-Based System and SVM Classification Samuel W. Rud, Nicholas St. Jacque, Aaron D. Vant and Jiann-Shiou Yang Department of Electrical Engineering, University of Minnesota, Duluth, MN 55812, USA Abstract Noninvasive heart murmur detection is the process of diagnosing a patient s heart condition without intrusion into the body. By providing a physician with effective and economical tools to help aid in the diagnosis could decrease misdiagnosis rates. This paper presents a study of detecting low frequency vibrations on the human chest via a FlexiForce force sensing technique and then correlate them to cardiac conditions via the Support Vector Machine (SVM) classification technique. The force sensor-based system is implemented and, with the consent of the University of Minnesota Institutional Review Board (IRB), the system was tested through clinical trials. A Support Vector Machine (SVM) learning algorithm is then used to train, tune, and classify signals. Our preliminary study indicates that a SVM is able to distinguish signals between normal and abnormal cardiac conditions. Keywords: Heart murmur; support vector machine, classification. I. INTRODUCTION EART murmurs are sounds caused by turbulent blood H flow through a heart s valve. Turbulence is present when the flow across the valve is excessive for the area of the open valve. It can be due to normal flow across a diseased valve, abnormally high flow across a normal valve, or a combination. Murmurs are classified as systolic, diastolic, or innocent to describe the location of the turbulent blood flow in the heart [1]. Noninvasive heart murmur detection is the process of diagnosing a patient s heart condition without intrusion into the body. The process has evolved in many different directions stemming from listening to the human chest with a stethoscope to using computers to detect heart conditions [2-4]. The simplest and most formal way of diagnosing heart murmurs is via a primary care physician s judgment on what he or she heard. However, primary care physicians accurately associate a murmur diagnosis with its pathology as little as 20% of the time. By providing a physician with effective and economical tools to help aid in the diagnosis could decrease misdiagnosis rates. The aim of this paper is to focus on detecting low frequency vibrations on the human chest using a force sensing technique and then correlate them to cardiac conditions via the Support Vector Machine (SVM) classification technique. We mainly target vibrations with frequencies ranging from 10 to 150 Hz. The SVM classification is used to devise a computationally efficient way of learning good separating hyperplanes in a high dimensional feature space. Since the invention of SVMs by Vapnik [5, 6], there have been intensive studies on the SVM for classification and regression (e.g., [7, 8]). Recently, the application of SVM to various research fields have also been reported in the literature (e.g., the intelligent transportation systems (ITS) [9, 10]). This paper will describe the methodology and results for detecting and classifying heart murmurs. II. METHODS A. Device, Circuits and Software The primary sensor device used is this study is the TekScan A201 FlexiForce sensor. It uses an aluminum dye that inversely changes resistance when force is applied on the dye. The sensor s range in force applied is approximately 0-1 pound, corresponding to a resistance range of 5 MΩ to 1 kω ( The FlexForce sensor device shown in Fig. 1 is used to measure force between two objects. Fig. 1 TekScan A201 FlexiForce Sensor The hardware part includes a drive circuit utilizing Texas Instrument s OPA228 operational amplifier. The FlexiForce electrical circuit is composed of an inverting operational amplifier with the FlexiForce sensor acting as a variable resistor as shown in Fig. 2. The circuit consists of (1) a Texas Instruments OPA4228N op-amp, (2) a 20 kω resistor, and (3) a FlexiForce 1 pound sensor. The circuit measures the load or force on the FlexiForce sensor as a voltage. An EKG circuitry, shown in Fig. 3, was designed to electrically isolate the patient ( The design keeps the leakage current to a minimum to prevent electrical shock. The heart would seize if enough current traversed it, therefore, electrical isolation is needed

2 32 Int'l Conf. Health Informatics and Medical Systems HIMS'15 to be ensured for this reason. The sensor s signal is amplified via a non-inverting amplifier with an OPA4228N op-amp. In addition, several analog high-pass filters are used in the system to eliminate any DC offset produced by the sensors. This allows for optimizing the range of frequencies existing in a system. The analog filters have a high enough corner frequency to eliminate any breathing cycles produced by a subject, but also low enough not to interrupt the target frequencies of the signal. The detailed information about our designed 1 st and 2 nd order high-pass filters, with corner frequencies at Hertz (Hz) and Hz, and their frequency responses are available upon request. The PCB, shown in Fig. 4, is composed of four first-order passive high-pass filters with some additional hardware and circuitry to deliver the driving signal to the appropriate sensor system. This board was designed and implemented using Pad2Pad software. It creates a low noise situation that a solderless bread board introduces. This allows for a smaller signal to noise ratio thus enabling a cleaner signal to be recorded. Since the signals that are being recorded are generally low amplitude, large amount of amplification is needed. In doing so, noise is introduced in the system and this PCB helps to minimize the noise effect. Fig. 2 FlexiForce drive circuit. Fig. 5 Bench and FlexiForce circuit. Fig. 3 EKG schematic. Fig. 4 Printed circuit board. The bench assembly in Fig. 5 provides a static surface for the FlexiForce system to keep the sensor under constant load. It also provides a firm surface to assist the sensor in creating a linear resistance characteristic. The bench mainly consists of a piece of poplar lumber measuring approximately two feet in length. This number is able to accommodate a human body or other equipment when the bench is placed over it. Sheet metal was bent to act as a guide for the vibration transmitter which is shown in Fig. 5 labeled pin. The legs, which act as risers, are constructed by the shaft of a bolt. This allows alterations to be made by spinning a nut on the riser for fine tuning of the system. The circuit that drives the FlexiForce sensor is able to rest on the top of the board as shown in Fig. 5. The signal generated by these circuits is processed by a Measurement Computing PCI-DAS6036 data acquisition card which interfaces with a custom Matlab software application to record the signals. The sensor device observes vibrations from the human chest with interest in the lower range of frequencies of 10 to 150 Hz. The signals are read into the computer through a data acquisition system and recorded. The software application dubbed MurmurPro, a graphical user interface (GUI) package that combines data acquisition, signal analysis, and signal detection into one package, allows us to view recorded signals, analyze and process the recorded signals, and later classify the signals using a Support Vector Machine (SVM) learning algorithm.

3 Int'l Conf. Health Informatics and Medical Systems HIMS'15 33 B. Testing Procedure The testing environment, with consent of the University of Minnesota Institutional Review Board s (IRB) Human Subjects Committee, was human patient testing (or clinical trials). The FlexiForce system acquired voltage signals indirectly through a vibration transmitter placed on the surface of the subwoofer. The vibration transmitter experiences vibrations from the speaker within and transfers them to the FlexiForce sensor which is positioned between the vibration transmitter and the adjustable bench. The circuitry for the FlexiForce sensor system existed on top of the height adjustable bench. C. Support Vector Machine Training and Testing The SVM is used to create an optimized boundary that separates between normal and abnormal cardiac conditions. It is a multi-step process as shown in Fig. 6 that includes signal processing, two-dimensional signal transformation, SVM tuning and optimization, and SVM Training. Signal processing entails filtering the signal to eliminate excessive noise and cropping the signal to a fixed number of heart beats. This is done to standardize each signal such that a two-dimensional representation can be achieved. The next process is then to tune the SVM. SVM tuning is a function provided by the Matlab s LS-SVM toolbox [11] that automatically tunes the SVM. It automatically tunes the hyper-parameters of the model with respect to the given performance measure. Hyper-parameters adjust the SVM model such that a desired plane can be achieved. The performance measure used in the software is called grid search, which finds the minimum cost/error of the model based on a two-dimensional hyper-parameter search. If the user of MurmurPro chooses to optimize the system, a Bayesian optimization algorithm is provided by LS-SVM. This will further tune the hyper-parameters based on Bayesian statistical methods [11]. Finally, the SVM is trained using the tuned hyper-parameters and the given data. This yields a two-dimensional visual representation of the SVM with each signal s data points and the optimized boundary plotted. frequency response plots was created using the Fast Fourier Transform (FFT) method. In the TS plot, S1 and S2 are distinguishable and separable while the FR plot shows a frequency range of 10 to 300 Hz. Note that S1 and S2 represent the timings of the sounds of a normal heart cycle, where S1 is the first sound and S2 is the second sound. For this study, it is assumed that the major and minor peaks in the signal occur in sync with S1 and S2 respectively. Fig. 8 shows the results obtained while testing an aortic stenosis recording played through the subwoofer. Fig. 8 illustrates the presence of extra vibration after S1. The FFT shows relatively the same frequency range as the normal heart recording in Fig. 7. However, several spikes are present that are not in the normal heart recording. Signal Processing SVM Optimization SVM Training SVM Training Flow Diagram Optimizer on Optimizer off 2-D Data Transformation SVM Tuning Fig. 6 Flow diagram of SVM training and tuning. III. RESULTS A. Clinical Trials The signals come from multiple patients in which the signals were categorized by the physician s diagnosis as normal or abnormal and pathological or non-pathological. Fig. 7 shows the results obtained while testing a normal heart recording played through the subwoofer. This figure contains two plots. The first plot is the filtered time-series (TS) plot (voltage (V) vs. time (s)) while the plot following is the frequency response (FR) (magnitude (db) vs. frequency (f)) plot of the filtered phonocardiogram. The Fig. 7 Normal heart recorded by the FlexiForce sensor.

4 34 Int'l Conf. Health Informatics and Medical Systems HIMS'15 B. SVM Classification The signals from the device include the following heart conditions: (a) aortic stenosis; (b) Austin Flint murmur; (c) diastolic murmur; (d) friction rub; (e) mitral valve prolapse; (f) systolic murmur; and (g) normal heart. Each SVM was first trained with a training set and then tested with a testing set. The testing set did not contain any of the same signals as the training set. Also, heart rates of the signals ranged from 60 bpm to 80 bpm. Fig 9 shows the trained SVM. Notice how normal heart signals tend to cluster together in a linear fashion, while the abnormal heart signals tend to be sporadic. This is due to the fact that normal heart signals should have a similar duration time and magnitude (depending on heart rate). The abnormal signals separate away from the normal signals, since the abnormal signals contain more vibrations giving a longer duration time and a larger magnitude. Fig. 8 Aortic Stenosos recorded by the FlexiForce sensor. Fig. 10 shows the tested SVM. With 14 test cases, varying from normal to abnormal, only 2 misclassifications occurred resulting in 14% misclassification. The two test cases that were misclassified were one normal and one abnormal. Fig. 10 Subwoofer FlexiForce tested SVM. IV. CONCLUSION This paper focuses on the study of heart murmur detection and classification. The experimental setup via a FlexiForce sensor together with hardware and software interfaces was developed to detect low frequency vibrations from a human chest (clinical trials) in the targeted frequency range of 10 to 150 Hz. The purpose of patient testing is to provide reasonable evidence that this system can perform on humans. The preliminary results show that the described technique has the capacity to capture distinguishable signals. During clinical trials, it was seen that the signal generated from a patient with a normal heart could be distinguished from a person with an abnormal heart via the FlexiForce sensor based system. Also, the SVM plots showed that it is possible to classify the signals. Unfortunately, the test data was limited to the amount of patients received. Also, there was no electro-cardiogram present at the clinical trials to establish the timings of the heart. This is due to the limits established by the IRB. Future developments of this study include improvements on the sensor system design and implementation. Also, improvements on the SVM twodimensional representation algorithm would further classify the signals more effectively. In addition, a real-time classification scheme could be devised to render an immediate diagnosis. Finally, extensive medical trials should be conducted to verify the sensor system and their respective classification accuracy rates. Fig. 9 Subwoofer FlexiForce trained SVM. REFERENCES [1] O. Epstein, et al, Clinical Examination, New York: Gower Medical Publishing, [2] Watrous, et. al, Computer-Assisted Detection of Systolic Murmurs Associated with Hypertrophy Cardiomyopathy, Texas Heart Institute Journal, vol. 31, no. 4, pp [3] K. Ejaz, et. al, A Heart Murmur Detection System Using Spectrograms and Artificial Neural Networks, Proc. IASTED Int l Conf. Circuits, Signals, and Systems, pp , [4] N. Andrisevic, et. al, Detection of Heart Murmurs Using Wavelet Analysis and Artificial Networks, Journal of Biomechanical Engineering, vol. 127, pp , November [5] V. N. Vapnik, The Nature of Statistical Learning Theory, New York, NY: Springer, 1995.

5 Int'l Conf. Health Informatics and Medical Systems HIMS'15 35 [6] V. N. Vapnik, An Overview of Statistical Learning Theory, IEEE Trans. Neural Networks, vol. 10, pp , September [7] K.R. Muller, A. J. Smola, G. Ratsch, B. Scholkopf, J. Kohlmorgen, and V. Vapnik, Using Support vector Support Machines for Time Series Prediction, in Advances in Kernel Methods, pp , Cambridge, MA: MIT Press, [8] S. R. Gunn, Support Vector Machine for Classification and Regression, Technical Report, University of Southampton, Southampton, K.K., May [9] D. Gao, J. Zhou, and L. Xin, SVM-based Detection of Moving Vehicles for Automatic Traffic Monitoring, Proc. IEEE 4 th Int l Conf. Intelligent Transportation Systems, pp , [10] R. Reyna, A. Giralt, and D. Esteve, Head Detection Inside Vehicles with a Modified SVM for Safer Airbags, Proc. IEEE 4 th Int l Conf. Intelligent Transportation Systems, pp , [11] K. Pelckmans, et. al., LS-SVM Toolbox User s Guide, Version 1.4, Dept. Of Electrical Engineering, Katholieke Universiteit Leuven, October 2002.