SEGMENTATION ON PHONOCARDIOGRAM Professor Kuh EE645 FINAL PROJECTS By Rebecca Longstreth

DEFINITIONS Phonocardiogram Cardio cycle S1 S2 S3 and S4 Graph representing sounds made by the human heart Cardio cycle Represented by S1, S2, S3 and S4 S1 Orignates at the closure of the atrioventricular valves recordable between 91hz and 179 hz S2 Originates at the aortic an pulmonic valves Can reach 200 hz S3 and S4 Represent cardiac wall vibrations Not all audible, low density

DEFINITIONS S1 and S2 Sound goes left to right S3 and S4 Right to left

Reason for Segmentation Tool to assist doctors in diagnosing heart malfunctions accurately Prevents mistakes in diagnoses Improved quality care for patients Doctors experience benefits everyone

Choices of Segmentation Fourier transform Laplace transform Wavelet decomposition

Fourier transform Reasons not used: Phonocardiograms are not smooth wave signals Not linear time invariance Non-stationary Non-harmonic

Laplace transform Reasons not used: Initial condition does not exist

Wavelet decomposition Preferred method of analyzing phonocardiograms Time limited and frequency limited Utilize digital filters and down sampling

Wavelet decomposition conditions Normalize the amplitude of all artifacts before transformation to avoid amplification errors Short segmentation window for high frequency Long segmentation window for low frequency Short lengths of data are considered due to file size considerations

Phonocardiogram anomalies Fetal acoustic signals inconsistent from one sound to the next Wave shape can change Frequency could shift Amplitude, duration and position in the cardiac cycle can change Background noise of the mother and the shielding affect of the womb

GOAL Isolate one cycle for analysis Identify S1 and S2 Boundaries of S1 and S2 Systolic and diastolic periods Systolic: time interval from beginning of S1 to the beginning of S2 Diastolic: time interval from beginning of S2 to the beginning of S1 Associate frequency spectrum

Research According to Cardiologists interviewed There is no way to recognize S1 and S2 using only a phonocardiogram Location of the probe can radically affect the signal strengths of S1 and S2 on the same individual

Finding S1 and S2: Method 1 Compare EKG and the phonocardiogram simultaneously to ensure accurate designation of S1 and S2 S1 occurs shortly after the EKG peak The first PCG signal following the EKG peak is S1 S2 occurs shortly after the 2nd EKG peak

Finding S1 and S2: Method 1 Disadvantage Need EKG requires expensive bulky and fragile equipment to perform Need Robust, portable, easy to use low cost equipment

System for heart sound classification

Wavelet Decomposition Heart sounds (sampled at an 8kHz sample rate, 16 bits/sample) are first hand segmented into 4096 sample segments, each consisting of a single heartbeat cycle.

Feature Reduction & Denoising Figure 1. A Simple Heart Sound Classification System .Each segment is transformed using a 7 level wavelet decomposition, based on a Coifman 4th order wavelet kernel (relative symmetry and fast execution). The resulting transform vectors, 4096 values in length, are reduced to 256 element feature vectors by discarding the 4 levels with shortest scale. Neural network in the classifier reduces noise. The magnitudes of the remaining coefficients in each vector are calculated, then normalized by the vector’s energy.

Classification Each feature vector is classified using a three layer neural network (256 input nodes, 50 hidden nodes, and 5 output nodes).

Results And Discussion The system was evaluated using heart sounds corresponding to normal mitral valve prolapse(MVP) coarctation of the aorta (CA), ventricular septal defect (VSD), and pul-monary stenosis (PS). The classifier was trained using 10 shifted versions (over a range of 100 samples) of a single heartbeat cycle from each type.

Figure 2. Representative Heart Sounds (left to right) Without Added Noise, with Noise Variance 1000, and with Noise Variance 3000, x-axis is the number of sample segments

Figure 2. Representative Heart Sounds (left to right) Without Added Noise, with Noise Variance 1000, and with Noise Variance 3000,x-axis is the number of sample segments

Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2, x-axis is the number of feature vectors

Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2, x-axis is the number of feature vectors

Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2,x-axis is the number of feature vectors

Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2,x-axis is the number of feature vectors

Feature vectors with additive noise The feature vectors produced for these examples in Figure 3. key features remain relatively stable even with addiditive noise.

Figure 4. Classification Accuracy (in Percent) as a Function of the Variance of the Added Noise

Figure 5. Classification Accuracy as a Function of Signal-to-Noise Ratio (in dB)

the sounds differ widely (e. g the sounds differ widely (e.g., by a factor of approximately 16:1 comparing a typical normal heartbeat with one exhibiting VSD). Accounting for this variation, classification accuracy as a function of signal-to- noise ratio (SNR) is shown in Figure 5. For an SNR above 31dB (which is easily obtainable under most practical circumstances) classification accuracy is 100%.

REFERENCES Barschdorff, D., U. Femmer, and E. Trowitzsch (1995, Sept. 10-13). Automatic phonocardiogram signal analysis in infants based on wavelet transforms and artificial neural networks. In Computers in Cardiology 1995, pp. 753–756. IEEE, Vienna, Austria. http://www.ida.liu.se/~rtslab/publications/2001/reed01a-eurosim.pdf Barschdorff, D., U. Femmer, and E. Trowitzsch (1995, Sept. 10-13). Automatic phonocardiogram signal analysis in infants based on wavelet transforms and artificial neural networks. In Computersin Cardiology 1995, pp. 753–756. IEEE, Vienna, Austria. Donnerstein, R. L. and V. S. Thomsen (1994, September). Hemodynamic and anatomic factors affecting the frequency content of Still’s innocent murmur. The American Journal of Cardiology 74, 508–510. Durand, L.-G. and P. Pibarot (1995). Digital signal processing of the phonocardiogram: review of the most recent advancements. Critical Reviews in Biomedical Engineering 23(3/4), 163–219. El-Asir, B., L. Khadra, A.H. Al-Abbasi, and M.M.J. Mohammed (1996, Oct. 13-16). Multireso-lution analysis of heart sounds. In Proc. of the Third IEEE Int’l Conf. on Elec., Circ., and Sys., Volume 2, pp. 1084–1087. Rodos, Greece. Rajan, S., R. Doraiswami, R. Stevenson, and R. Watrous (1998, Oct. 6-9). Wavelet based bank of correlators approach for phonocardiogram signal classification. In Proc. of the IEEE-SP Int’l Symp. on Time-Frequency and Time-Scale Analysis, pp. 77–80. Pittsburgh, PA.

REFERENCES Shino, H., H. Yoshida, K. Yana, K. Harada, J. Sudoh, and E. Harasawa (1996, Oct. 31 - Nov. 3). Detection and classification of systolic murmur for phonocardiogram screening. In Proc. of the 18th Int’l Conf. of the IEEE Eng. in Med. and Biol. Soc., Volume 1, pp. 123–124. Amsterdam, The Netherlands. http://www.cinc.org/Program/p7b-1.htm THE ANALYSIS OF HEART SOUNDS FOR SYMPTOM DETECTION AND MACHINE-AIDED DIAGNOSIS, Todd R. Reed, Nancy E. Reed and Peter Fritzson, The Netherlands H. Liang, S. Lukkarinen, and I Hartimo, Heart Sound Segmentation Algorithm Based on Heart Sound Envelogram, Helsinki, Finland H. Liang, S. Lukkarinen, and I Hartimo, A Boundary Modification Mehtod for Sound Segmentation Algorithm, Helsinki, Finland Abdelhani Djebbari, and Fethi Bereski Reguig, Short-time Fourier Transform Analysis of the Phonocardiogram Signal, Algiers