CN1539372A - Method and device for early diagnosis of heart disease based on high-frequency electrocardiographic waveform - Google Patents

Abstract

Method for early diagnosis of heart disease based on high-frequency ECG waveform: High-frequency ECG waveform obtained by high input impedance, low noise, high common-mode rejection ratio and broadband multi-channel ECG amplifier; 12-bit A/D conversion board; .Using wavelet transform to combine time and frequency domains for analysis to form HFECG (time-frequency-amplitude) three-dimensional spectrogram. Then extract the characteristic parameters reflecting the richness of the high-frequency components of HFECG from the QRS complex high-frequency three-dimensional spectrogram of the twelve-lead HFECG, including the high-end cut-off frequency, etc., and carry out the classification and diagnosis of the above-mentioned characteristic parameters by artificial neural network: they are in The comprehensive influence in an m-dimensional space composed of m characteristic parameters determines the high-frequency three-dimensional spectrum properties of the QRS complex. There is an m-dimensional surface that divides the space into two parts: the normal and abnormal. For early diagnosis of heart disease.

Description

Method and device based on the heart disease early diagnosis of high frequency electrocardiogram waveform

One. technical field

The present invention relates to a kind of method and device of the heart disease early diagnosis based on the high frequency electrocardiogram waveform, wherein relate to and adopt wavelet transformation from high frequency ECG (High Frequency Electrocardiogram, HFECG) extract fine feature in, and diagnose the method for differentiation with the neural network classification method, also comprise the circuit that obtains meticulous high frequency electrocardio signal.

Two. background technology

Trickle composition in the high frequency electrocardiogram waveform and many heart diseases comprise that ventricle left and right sides bundle branch block, myocardial ischemia, myocardial damage etc. have substantial connection.The applicant has applied for relevant patents such as high frequency ECG instrument.Normal person's cardiac muscle exists scrambling to a certain degree on form and function, this has caused the existence of these trickle compositions.And in many heart diseases, the scrambling of this transmitting medium of myocardial cell will significantly increase, and these compositions are also significantly increased, and presents characteristic separately.Chinese scholars has confirmed that by time-domain analysis and frequency-domain analysis high frequency ECG is very strong to the sensitivity of disease, has the significance of early diagnosis clinically for many years.

Wavelet transformation (Wavelet transform, WT) be a kind of linear operation, it carries out the decomposition of different scale to signal, the mixed signal that the different frequency of various weave ins is formed can be resolved into the block signal of frequency inequality, can be effectively applied to separate, improve the resolution in time-frequency two territories etc. as signal.It can combine analysis from time-frequency two territories with HFECG, forms (T/F-amplitude) three-dimensional spectrogram, therefrom extracts its characteristic parameter.

Neural network analysis system, with the imitation human brain function, finishing the work of similar human brain, is physiological true human brain neural network's 26S Proteasome Structure and Function, and certain theoretical abstraction, simplification and the simulation of some fundamental characteristics and a kind of information processing system of constituting.From systematic point of view, this artificial neural network is the self-adaptation nonlinear dynamical system that is made of by extremely abundant and perfect connection a large amount of neurons.Owing between the neuron different connecting modes is arranged, so can form the nerve network system of different structure form.Error backpropagation algorithm (BP algorithm) network and adaptive resonance theory algorithm (ART2) network all have good system identification and classification capacity, we come the characteristic parameter that extracts is carried out sort operation by means of these two kinds of artificial neural networks, thereby realize the early diagnosis of heart disease.

Three. summary of the invention

The purpose of this invention is to provide and a kind ofly obtain the early stage cardiac disease signal and differentiate the method and the device of diagnosis from the high frequency electrocardiogram waveform, carry out the method and the device of heart disease early diagnosis based on the high frequency electrocardiogram waveform, realize the heart disease early diagnosis, its realization is divided into three steps:

Obtain the method and the device of early stage cardiac disease signal from the high frequency electrocardiogram waveform:

1. the acquisition of high frequency ECG.The high frequency electrocardiogram waveform installs by high input impedance, low noise, high cmrr and broadband 12 passage ecg amplifiers by obtaining as lower device; 12 A/D change-over panels; Microsystem and ancillary equipment thereof constitute.

2. adopt wavelet transformation to extract the characteristic parameter that a series of reflection HFECG high frequency compositions enrich degree from the three-dimensional frequency spectrum of the QRS wave group high frequency of 12 lead HFECG (high frequency ECG), they are:

High-end cut-off frequency: the highest frequency of HFECG frequency band range;

The number at block signal peak and sloping number sum in the three-dimensional spectrogram: the waveform complexity of describing three-dimensional frequency spectrum;

Summit frequency: describe the frequency that ceiling capacity distributes.

3. artificial neural network is to the classification diagnosis of characteristic parameter

From any one independent characteristic parameter is to can not find the frequency spectrum that definite marginal value is distinguished patient and normal person, these parameters all exist certain relatedness with heart disease, and their combined influences in the space of a m dimension that is made of m characteristic parameter have determined the three-dimensional frequency spectrum character of high frequency of QRS wave group.That is to say that have the curved surface of m dimension, this curved surface is divided into two parts with the space: cardiac function normal and unusual.

Because artificial neural network has good learning capacity and to the capability of fitting of any nonlinear function, we have introduced a feed-forward type network, utilize improved error backpropagation algorithm (BP algorithm) training network, and the ART2 algorithm training network that teacher's refinement match is arranged, well the curved surface of this m dimension of match is distinguished normal and unusual high frequency electrocardiogram characteristic parameter, realizes the early diagnosis of heart disease.

Characteristics of the present invention are: the extraction of the detection of the high frequency electrocardio signal of uniqueness exploitation, advanced wavelet transformation characteristic parameter and the classifying and analyzing method of artificial neural network are organically combined, especially early diagnosis is significant to clinical cardiopathic diagnosis for a plurality of high frequency electrocardiogram characteristic parameter new methods that the match hypersurface is classified in hyperspace, this method is that a new way has been opened up in the further investigation of high frequency ECG, belongs to initiative both at home and abroad.

Four. description of drawings

Fig. 1: 12 passage ecg amplifier block diagrams, wherein Fig. 1 a is the The general frame of ecg amplifier, and Fig. 1 b is the structured flowchart of a channel amplifier, and Fig. 1 c is standard lead and orthogonal lead circuit diagram.

Fig. 2: be the 12 passage ecg amplifier electrical schematic diagrams of Fig. 1 b

Fig. 3: be three-dimensional spectrogram

Fig. 4: be improved error backpropagation algorithm (BP algorithm) network structure

Fig. 5: be the ART2 algorithm network structure of refinement match that the teacher is arranged

System hardware comprises three parts: high-performance 12 passage ecg amplifiers (comprising that 12 lead and 3 lead systems); 12 A/D change-over panels; Microsystem and ancillary equipment thereof.Its structured flowchart as shown in Figure 1.

The apparatus features of the acquisition of high frequency ECG is: comprise that low noise broadband, high input impedance, high cmrr ecg amplifier constitute, its design is one of biomedical electronics research topic, is the important step of ecg information being carried out data acquisition and processing.Among the present invention: shown in Figure 2, the amplifier of each passage is to be made of high performance integral measuring amplifier U3 (AD620), and its performance can meet design requirement.The circuit structure of 12 channel amplifiers is identical, the performance basically identical.Integral measuring amplifier AD620 provides main voltage gain (500 times) and common mode rejection ratio (more than the 90dB).(with U12, U15A, three operational amplifiers of U15B is core for baseline stability and automatic reset circuit, the peripheral components that adds them is formed) be an integral form feedback network, for the signal that is lower than its cut-off frequency, this circuit forces to make difference amplifier to be output as zero, reaches the purpose of baseline stability.This circuit has does not influence the difference amplifier common mode rejection ratio, need not install advantages such as capacitance before difference amplifier additional.Its effect that automatically resets is when the difference amplifier output voltage amplitude exceeds outside the window setting value, forces to make difference amplifier outfan baseline to be reset to zero rapidly.The linear light buffer circuit has the favorable linearity transmission characteristic by the offset-type linear signal isolation transmitter that two the approximately uniform photo-coupler U6 of characteristic, U7 (TIL117) and operational amplifier U5B, U8B (LF353) form.12 passage ecg amplifier lead systems constitute and to be different from three-path amplifier, more are different from the single channel amplifier, its require to connect simultaneously at every turn standard lead 12 road signals (be I, II, III, aVR, aVL, aVF, V ₁, V ₂, V ₃, V ₄, V ₅, V ₆), (X, Y Z), connect three road signals that frank leads (X ', Y ', Z ') to switch three road signals connect orthogonal lead simultaneously by electrical switch.

Electrocardiosignal is an analogue signal, and must convert digital signal to could be handled by computer acquisition.If the high fdrequency component of electrocardiosignal 10 μ v, the equivalent input noise of amplifier is 3 μ v, and when amplification 1000, then Shu Chu high fdrequency component is 10mv, and noise is 3mv, and signal to noise ratio is 3.3.Da Xiao signal to noise ratio as long as the figure place of A/D is enough big, is the high fdrequency component in the electrocardiosignal more accurately can be changed into digital quantity like this.It is 12 A/D change-over panels of 5v that the design adopts reference voltage, and its resolution is 5v/ (2 ¹²-1)=and 1.22mv, promptly can both be collected greater than the voltage of 1.22mv.This A/D change-over panel has 16 analog input channels,-5～+ the bipolar input of 5V, interruption of work mode, the highest sample frequency are 100KHz (being μ s conversion times 10), can satisfy 12 channel sample (the every passage 5KHz) designing requirement that sample frequency is up to 60K.

The used microcomputer of system is the III that runs quickly, the 128MB internal memory, and the 20GB hard disk, maximum clock frequency 733MHz is furnished with high-resolution (1024 * 768) colour video display unit that the SVGA card is supported, CD-ROM drive (50 speed) and laser printer.Be characterized in that the speed of service is fast, the analyzing and processing ability is strong.

Systems soft ware comprises two parts: wavelet transformation high frequency ECG characteristic parameter extraction algorithm and neural network classification algorithm.

As shown in Figure 3, utilize wavelet transformation that time-frequency two territories are combined analysis, form the three-dimensional spectrogram of HFECG (T/F-amplitude).From the three-dimensional spectrogram of the QRS wave group high frequency of 12 lead HFECG, extract the characteristic parameter that a series of HFECG of reflection high frequency compositions enrich degree again.

BP algorithm neural network structure as shown in Figure 4.Can prove, trilaminar BP network can match nonlinear function arbitrarily, for not increasing the complexity of network, we just determine that the network hidden layer is one deck.The neuron number N of the input layer of network is determined by the number of the characteristic parameter that we will study.We have selected seven characteristic parameters altogether, and input layer just needs seven like this.The selection of hidden neuron number M is according to empirical equation

$P<\underset{i=0}{\overset{N}{\mathrm{\&Sigma;}}}{C}_M^i-----(C_M^i=0,i>M)$

Wherein P is the number of samples in training storehouse.Formula (1) has illustrated that the hidden neuron number is relevant with the number of samples of training, its essence is that the complexity of non-linear relation of the equal calcaneus rete network of M and P match is relevant.System is complicated more, and the number of samples in required hidden neuron number and training storehouse is all many more.Through experimental debugging, find to get M=5, the diagnosis height during P ≈ 32 after network training converges faster and the training, fitting effect is better.The kind number decision that the output layer neuron number K of network need be classified by us.Because we only study healthy people and coronary disease patient's difference now, and specifically do not study the difference of various concrete diseases, so selected output neuron is one, 0 and 1 of its output valve is represented healthy and ill respectively.

ART2 neural network is actually the distance of meeting according to the analog quantity input pattern in model space classifies, and we utilize it that the high frequency ECG characteristic is classified.With the input unit of these characteristics as ART2 neural network, we find: the characteristic of many tested objects is after ART2 neural network is handled, and its output is juxtaposition.Be that the different pattern vector that their high frequency ECG characteristic of healthy people and cardiac constitutes intermeshes in the space, they are not separable two classes of simple linear just in model space.In order to solve this classification problem, we transform ART2 neural network, revise.Proposed a kind of new, fitting method after the first refinement.New ART2 neural network structure as shown in Figure 5, this structure has increased preprocessor, adapter and F3 mapper on original ART2 neural network basis.This neutral net also needs earlier through network structure and subclass are determined in the study of master sample, and when working then unknown input pattern being made comparisons with the subclass that has produced obtains classification results.

Example: following table is 4 cases of certain hospital, obtains 7 characteristic parameters from their HFECG, and

Use neural network classification, diagnostic result and practical situation meet fully.

Name	High-end cut-off frequency	Ratio W1	Ratio W2	Ratio W3	Absolute value P	Slope number+peak number	The summit frequency	Diagnostic result
									Du * *	??310	?0.806	?0.230	?1.032	?3.713	??35	??135	Unusually
Lee *	??165	?0.430	?0.104	?0.641	?7.723	??16	??120	Unusually
									Bears * *	??245	?0.900	?0.267	?0.892	?4.168	??27	??190	Normally
Zhu *	??215	?0.640	?0.170	?0.659	?4.458	??22	??130	Normally

Claims (4)

1. Method for early diagnosis of heart disease based on high frequency ECG waveform: high input impedance, low noise, high common mode rejection ratio and broadband multi-channel ECG amplifier; 12-bit A/D conversion board to obtain high frequency ECG Waveform; using wavelet transform to combine time and frequency domains to analyze to form HFECG (time-frequency-amplitude) three-dimensional spectrogram; and then extract a series of The characteristic parameters reflecting the richness of high-frequency components of HFECG are: High-end cut-off frequency: describes the highest frequency of the HFECG (time-frequency-amplitude) frequency band range;

The sum of the number of block signal peaks and the number of slopes in the three-dimensional spectrogram: describes the waveform complexity of the three-dimensional spectrum; Highest peak frequency: describes the frequency of the maximum energy distribution; Carry out the classification diagnosis of artificial neural network on the above characteristic parameters: Their comprehensive influence in an m-dimensional space composed of m characteristic parameters determines the high-frequency three-dimensional spectrum properties of the QRS complex. There is an m-dimensional surface that divides the space into two parts: the heart function normal and abnormal; In the BP network, it is determined that the hidden layer of the network is one layer; the number of neurons N of the input layer of the network is selected from the number of the above-mentioned characteristic parameters, and a total of seven characteristic parameters are selected, so that seven neurons in the input layer are required; The selection of the number of elements M is based on the empirical formula $\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; <</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>$ Among them, P is the number of samples in the training library, and the number of neurons in the hidden layer is related to the number of training samples. In the formula, both M and P are related to the complexity of the nonlinear relationship fitted by the network; the number of neurons in the output layer of the network The number K is determined by the number of species that need to be classified, and one output neuron is selected, and the output values of 0 and 1 represent healthy and diseased respectively. 2. The method for early diagnosis of heart disease based on high-frequency electrocardiographic waveform according to claim 1, characterized in that the hidden layer is one layer in the BP network, and M=5, P≈32. 3, by the method for the early diagnosis of heart disease based on high-frequency electrocardiogram wave form described in claim 1, it is characterized in that in ART2 neural network, for distinguishing healthy people and their high-frequency electrocardiogram feature data of heart disease patients constitute different The pattern vectors are interlaced in space, so it will be modified and corrected, and the method of refinement and then fitting will be used; this structure adds a preprocessor, a matcher and an F3 mapper to the original ART2 neural network; This kind of neural network first determines the network structure and subcategory through the study of standard samples, and then compares the unknown input mode with the generated subcategory to obtain the classification result. 4. The device for early diagnosis of heart disease based on high-frequency ECG waveform, characterized in that the amplifier of each channel is composed of high-performance integrated measurement amplifier AD620. The baseline stabilization circuit and the automatic reset circuit are an integral feedback network. For signals lower than its cut-off frequency, the circuit forces the output of the differential amplifier to be zero to achieve the purpose of baseline stabilization.

Images