Feature Analysis Of PCG Signals And The Value Of Feature In The Auxiliary Diagnosis

Cardiac auscultation is one of the most basic methods for doctors to estimate cardiacfunction, but also difficult to master. Phonocardiogram examination could get moreinformation about cardiac vascular and needn’t so much accumulate experience as cardiacauscultation. So phonocardiogram study has become one of the most popular research topicsin recent years, especially in the fields of PCG signals feature analysis and pattern recognition.However, the complex physiological mechanism leads to the slow progress in the study onPCG signals. In this paper, the envelope curve of PCG signals is extracted and used tocompute characteristic parameters as well as classifying PCG signals by pattern recognitiontechnology, and at last a kind of heart disease diagnosis system based on PCG signals isestablished.Firstly, a new method based on wavelet transform to extract PCG signals envelope curveis presented and used to help compute characteristic parameters. Decompose low frequencysignals into three layers by DB4at first and then reconstruct the approximation coefficients inthe third layer as the envelope curve. In contrast to other common means of extracting PCGsignals envelope curve, the method is proved to have many advantages as simple algorithm,smooth curve and outstanding feature point. Next the curve is used to locate the position of S1and S2, by the help of which the initial positions and peak points could be fixed effectively.At last, the parameters as heart rate, S1/S2, and D/S are computed.Secondly, pattern recognition technology such as Support Vector Machine and BP neuralnetwork are adapted to classify PCG signals. SVM is selected to discriminate normal andabnormal heart sounds by training the envelop area and wavelet energy as two characteristicparameters. In order to test the accuracy of discriminating normal and abnormal heart sounds,70heart sounds are collected and analyzed. The experiment demonstrates that the accuracyrate is more than90%, which is very useful in many aspects. After that the high frequencymitral murmur is decomposed into eight sub bands by wavelet packet in three levels. Powerspectrum information entropy of every sub band is computed and forms an8-D characteristicvector. The vector as the modal input of SVM and BP neural network classifies the sample signals into two categories as mitral stenosis and mitral incompetence. The classificationresult shows that power spectrum information entropy places a significant role in thecharacterization of high frequency murmur especially mitral murmur.Finally, a design thought about one kind of PCG auxiliary diagnosis system is presented.The system contains three steps: collect PCG signals from human body and then transmitthem wirelessly to PC for signal analysis and processing, from which get a result whether thesignal is abnormal or not and at the same time obtain several characteristic parameters. Theobtained information is feedback to the physician to confirm the diagnosis results with otheroutpatient information for reference, from what we can see that PCG signals occupy animportant position in auxiliary diagnosis and taking precautions against heart disease.