JPH03218731A - Predictive electrocardiogram waveform recognizing device - Google Patents

Abstract

PURPOSE:To improve the recognition accuracy and the maximum likelihood of a partition point by using a pattern associative type neural network and a recurrent neural network, and executing the recognition of a partition by taking notice of an electrocardiogram waveform of that time and a timewise structure of the partition. CONSTITUTION:A leaning/recognition control part 16 fetches a value St of an electrocardiogram waveform of the present time point (t) of the electrocardiogram and its partition Kt from a standard waveform/partition pattern memory part 11, and uses them for the learning of internal parameters of a pattern associative type neural network 12 and a recurrent neural network 13. The network 12 sets a signal S't obtained by superposing the value St of the electrocardiogram waveform of the time point (t) and the output Nt of a noise generator 14, and a partition Kt-1 of the time point (t)-1 as input signals, sets the partition Kt of the time point (t) as a teacher signal, and updates the internal parameter of the network 12. The network 13 sets the partition Kt-1 of the time point (t)-1 as the input signal and sets the partition Kt of the time point (t) as the teacher signal, and updated the internal parameter of the network 13.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to an apparatus for recognizing classification of electrocardiogram waveforms.

[Conventional technology]

Conventionally, in order to recognize the division point, noise is removed from the input signal using a low-pass filter or band-pass filter, the result is differentiated, and the point where the value becomes 0 is determined as the peak of the R wave. Detect as a point. Further, in order to detect the peak point of the R wave, it is also possible to perform second-order differentiation on the output of the filter and detect the time point at which the output reaches a maximum as the peak point of the R wave. After finding the time point of the peak point of the R wave, the dividing point of the Q wave and S wave is found using the points where the input waveform crosses the 0 point four times. After that,
The dividing point of the T wave and P wave is found using the dividing point of the Q wave and S wave and the point where the input waveform intersects the 0 point. FIG. 9 is an explanatory diagram showing a general electrocardiogram waveform and its dividing points.

[Problem to be solved by the invention]

This type of conventional electrocardiogram waveform recognition apparatus recognizes other division points based on finding the peak point of the R wave, without paying attention to the temporal structure of the electrocardiogram waveform. Therefore, there is a drawback that if the peak point of the R wave is incorrectly recognized, all the division points will be incorrectly recognized.

In the present invention, pattern associative neural network learning is performed using a waveform in which noise is superimposed on a learning pattern, and a pattern associative neural network adapted to the temporal structure of an electrocardiogram waveform is used. As a result, the number of types of learning patterns can be increased by performing learning, and it is possible to prevent excessive learning from only a specific learning pattern. In addition, the predicted output of the recurrent neural network that was trained to predict the time series of the segmentation pattern at that time from the time series of the segmentation pattern up to one time ago is calculated based on the segmentation pattern from one time ago and the waveform pattern at that time. is input into a pattern associative neural network trained to recognize the time division.

By using two types of neural networks in this way, segment recognition is performed by focusing on the electrocardiogram waveform at that time and the temporal structure of the segment, which improves the recognition accuracy of segment points compared to conventional methods. Both accuracy and accuracy are improved.

[Means to solve the problem]

The predictive electrocardiogram waveform recognition device according to the first invention of the present application includes a standard waveform/section pattern memory section that holds a standard waveform pattern that is a standard electrocardiogram waveform pattern and divisions of the waveform, and an increased number of types of learning patterns. A noise generator that prevents excessive learning from only a specific learning pattern; a pattern associative neural network that recognizes the current time waveform classification from the predicted electrocardiogram waveform pattern one time ago; A recurrent neural network that predicts the segment at that time from the time series of previous segment patterns, a segment buffer that holds the predicted segment, and the standard pattern with the noise superimposed thereon is input to the pattern associative neural network unit. and a learning/recognition control unit that controls the learning of the recurrent neural network unit by giving a segmentation point as a teacher signal to provide a pattern to be output when the pattern associative neural network unit learns. .

A predictive electrocardiogram waveform recognition device according to a second invention of the present application includes, in addition to the first invention, an input waveform buffer that holds a waveform pattern in the vicinity of the current time.

[Effect]

Since the electrocardiogram waveform is determined by the structure of the heart,
(1) Generally has a chronological structure. In the present invention, this time-series structure is used in a waveform recognition device. Recognition can be performed more accurately by including predicted divisions than by relying only on the electrocardiogram waveform at the time point to be recognized. This recognition is achieved using a pattern associative neural network.

Prediction of the electrocardiogram classification at the time point you want to recognize is done using recurrent
Realized using neural networks. A recurrent neural network can learn the structure of a general time series, and has a function of predicting the classification of a point in time based on recognition results up to that point in time.

The pattern associative neural network used here includes, for example, "Nikkei Electronics" magazine issue 427 (19
The pattern is explained in detail in the article entitled "Using Neural Nets for Pattern Recognition, Signal Processing, and Knowledge Processing" (hereinafter referred to as cited document l) on pages 115 to 124 of August 1987). Associative neural networks can be used.

FIG. 7 shows the structure of this pattern associative neural network. As shown in FIG. 7, this pattern associative neural network has a hierarchical structure including an input layer 71, an intermediate layer 72, and an output layer 73.

Although the intermediate layer is one layer in this figure, it may be multilayered with two or more layers.

The output of the node in each layer of the pattern associative neural network is assigned a weight W to the node connected to that node.
The value of the sum of the products multiplied by is converted by a nonlinear function. In this way, the conversion characteristics of the pattern associative neural network are determined by the weights W. The value of weight W is determined by learning. Regarding learning methods,
For example, it can be performed using backward propagation, which is explained in detail in the cited literature.

The recurrent neural network used here includes, for example, Servan-Schreiber. D.
etc. al. ”EncOyamang Sequentia
Structure in Simple
RecurrentNetworks”Techni
cal Report CMU-CS-88-183,
Carne-gie Mellon University
Ty. (1988) (hereinafter referred to as cited document 2)
The recurrent neural network described in detail can be used.

FIG. 8 shows the structure of this Ricaresoto neural network. As shown in FIG. 8, this Ricaresoto neural network has a hierarchical structure including an input layer 81, a context layer 82, an intermediate layer 83, and an output layer 84. Although the intermediate layer is one layer in this figure, it may be multiple layers of two or more layers.

The output of a node in each layer of the Ricaresoto neural network is obtained by converting the sum of the nodes connected to that node by a weight W using a nonlinear function. In this way, the conversion characteristics of the recurrent neural network are determined by the weights W. Weight W
The value of is determined by learning. The learning method can be carried out using, for example, backward propagation, which is explained in detail in Reference I.

〔Example〕

A first embodiment corresponding to the first invention of the present application will be described with reference to the drawings. FIG. 1 is a diagram showing an embodiment of the invention.

The standard waveform/segmentation pattern memory section 11 holds (c) sets (N sets) of standard electrocardiogram waveforms and their segments. The learning/recognition control unit 16 acquires the value St of the electrocardiogram waveform at the current point in time and its division K from the standard waveform/segmentation pattern memory unit l1, and uses the pattern associative neural network and the recurrent neural network. Used for learning internal parameters with the network. This learning phase will be explained using FIG. 1 and FIGS. 2(a) and (b).

The pattern associative neural network 12 uses the electrocardiogram waveform value S at time point t, the output N of the noise generator 14,
The signal S', which is the superimposition of
, is an input signal, and the section K of time point t is a teacher signal. The internal parameters of the pattern associative neural network 12 are updated based on these input signals and teacher signals. This update can be performed using backward propagation, which is explained in detail in Reference 1.

The above-described learning operation is performed on the electrocardiogram waveform values S, and the classifications K, which are held in the standard waveform/segmentation pattern memory section 11.
Do about. If the error after learning is not small enough, the learning operation described above is repeated until the error becomes small enough.

Since N is the output of the noise generator 14, it becomes a different value each time learning is repeated, and the value St of the electrocardiogram and the output N of the noise generator 14 are different. The signal S' obtained by superimposing the signals S' and S' also has different values. This makes it possible to increase the types of input signals for learning. This completes the learning of the pattern associative neural network 12. Through this learning, the pattern associative neural network 12 can determine the value S of the electrocardiogram waveform at time point t, and the division Kt at time point t-1.
It has a function of recognizing the classification K of time point t based on -1.

In the Ricaresoto neural network 13, the section Kl-1 of the time point t-1 is used as an input signal, and the section K of the likely time point is used as a teacher signal. Based on these input signals and teacher signals, the internal parameters of the Ricaresoto neural network 13 are updated. This update can be performed using backward propagation, which is explained in detail in Reference 1. The learning operation described above is performed for the section K held in the standard waveform/section pattern memory section 1l. If the error after learning is not small enough, the learning operation described above is repeated until the error becomes small enough.

By this, Rikaresoto Neural Network 1
Learning in step 3 is complete. Through this learning, the recurrent neural network l3 is able to determine the division K of the time point t-1,
-1 time series to the division K of time point t. It has the ability to predict.

Next, the phase of recognizing electrocardiogram segments using the trained pattern associative neural network 12 and the trained recurrent neural network 13 will be described with reference to FIGS. 1 and 3.

To recognize the division at time point t, the value S of the electrocardiogram waveform to be recognized and the prediction K't of the division at time point t in the division buffer 15 (this prediction will be described later) are input to the pattern associative neural network 12. shall be. With these inputs, pattern associative neural network 12
As an output, a recognition result K of the classification of time point t is obtained. This recognition result K is input to the recurrent neural network 13. This input allows recurrent
Time point t+1 as the output of the neural network 13
A prediction K'l+1 of the partition is obtained.

This predicted value K'l+1 is stored in the classification buffer l5 and used for recognition of the classification at time point t+1. The partition buffer 15, which holds the partition 2 prediction of the next time point, is initialized prior to recognition.

Next, a 20th actual journey example corresponding to the second invention of the present application will be described with reference to the drawings. FIG. 4 is a configuration diagram showing an embodiment of the invention.

The standard waveform/segmentation pattern memory section holds sets (N sets) of standard electrocardiogram waveforms and their segments. Here, the time point [t el+
t + 82] (e+≧1, e2≧1). Generally, e1≠e2 may be satisfied, but here, for simplicity, it is assumed that 81=82 (=8). The input waveform buffer 47 stores the electrocardiogram value S' at a time point near the current time point t.
t (S I-* r S t-*+l z...,
S,, + S the-1 r S ++e).

The learning/recognition control unit 46 extracts the value S of the electrocardiogram waveform near the current time point t of the electrocardiogram waveform and its classification K,
It is used to learn the internal parameters of the pattern associative neural network 42 and the recurrent neural network 43. This learning phase will be explained using FIG. 4 and FIGS. 5(a) and (b).

In the four-turn associative neural network 42, the value S of the electrocardiogram waveform corresponding to the time point [t-e, t+e] near the current time point t stored in the input waveform buffer 47, and the noise generation Output N of the device 44, (N.,, Nt-e+1+..., N,,...
, Nt+*-1*N, +. ) and the superimposed signal S't
(S't-++r S't-*+1+...+S't+
...,S',. *-1 1 8' t+*) is used as an input signal, and the division Kt of time point t is used as a teacher signal.

Based on these input signals and teacher signals, internal parameters of the pattern associative neural network 42 are updated. This update can be performed using backward propagation, which is explained in detail in Reference 1. The above-described learning operation is performed on the electrocardiogram waveform value S' corresponding to the time point [t-e, t+e] near the time point t stored in the standard waveform/segmentation pattern memory section 41,
and category K. If the error after learning is not small enough, the learning operation described above is repeated until the pseudo difference becomes small enough. Through the above steps, the learning of the pattern associative neural network 42 is completed.

Through this learning, the pattern associative neural network 42 is able to control the time point [te, t10e] near the time point t.
The value S of the electrocardiogram waveform corresponding to , and the division K of the time point t-1
It has a function of recognizing the division K of time point t from t-1.

Learning of the recurrent neural network is performed at time point t-1 in the recurrent neural network 43.
Let the section K t-1 of the time point t be the input signal, and the section K of the time point t be the teacher signal. Based on these input signals and the teacher signal, the internal parameters of the Ricaresoto neural network 43 are updated.

This update can be performed using backward propagation, which is described in detail in reference 1. The above-described learning operation is performed for the section Kt held in the standard waveform/section pattern memory section 41. If the error after learning is not small enough, the learning operation described above is repeated until the error becomes small enough. This completes the learning of the Ricaresoto neural network 43. Through this learning, the recurrent neural network 43 has a function of predicting the division Kt of the time point t from the time series of the division Kt-1 of the time point t-1.

Next, the phase of electrocardiogram segment recognition using the trained pattern associative neural network 42 and the trained recurrent neural network 43 will be described with reference to FIGS. 4 and 6.

Upon recognition of the section at time point t, the input waveform is temporarily stored in the input waveform buffer 47.

In recognizing the division at time point t, the time point [t-e
.
2 input.

As a result, the pattern associative neural network 42 obtains as an output the recognition result K of the classification of the time point t. This recognition result K is input to the Ricaresoto neural network 43. As a result, in the Rikaresoto neural network 43, the time point t+1
A prediction K't+1 for the partition is obtained. This child measurement K't+
1 is stored in the segmentation buffer 45 and used to recognize the segment at time point t+1. The partition buffer 45, which holds predictions of the partition of the next time point, is initialized prior to recognition.

〔Effect of the invention〕

According to the present invention, segment points are recognized by paying attention to the temporal structure of the electrocardiogram waveform, so that there is an effect that both the precision and accuracy of segment point recognition are improved compared to conventional methods.

[BRIEF DESCRIPTION OF THE DRAWINGS] FIG. 1 is a block diagram showing the configuration of the first embodiment of the present invention, and FIGS. 2(a) and (b) are signals of the learning phase of the first embodiment of the present invention. 3 is an explanatory diagram showing the flow of signals in the recognition phase of the first embodiment of the present invention. FIG. 4 is a block diagram showing the configuration of the second embodiment of the present invention. FIGS. 5(a) and 5(b) are explanatory diagrams showing the signal flow in the learning phase of the second embodiment of the present invention, and FIG. 6 is the signal flow in the recognition phase of the second embodiment of the present invention. FIG. 8 is an explanatory diagram showing a configuration example of a recurrent neural network. FIG. 7 is a configuration example of a pattern M imaginary neural network. ...Pattern associative neural network unit, 13...Recall soto neural network unit, 14...Noise generator, 15...Division buffer, 16...
...Learning/recognition control unit, 4l...Standard waveform/segmentation pattern memory section, 42...Pattern associative neural network section, 43...Recoresoto neural network Part, 44...
Noise generator, 45... Division buffer, 46.
...Learning/recognition control unit, 47...Input waveform buffer.

Claims (1)

[Claims] 1. A standard waveform/segmentation pattern memory section that holds a standard waveform pattern that is a standard electrocardiogram waveform pattern and its waveform divisions, and a standard waveform/segmentation pattern memory section that stores the standard waveform pattern that is a standard electrocardiogram waveform pattern and the divisions of the waveform, and increases the types of learning patterns and allows only specific learning patterns to be used. A noise generator that prevents excessive learning, a pattern associative neural network that recognizes the waveform division at the current time based on the predicted electrocardiogram waveform pattern from the division one time ago, and a pattern associative neural network that recognizes the waveform division at the current time based on the division pattern from one time ago. a recurrent neural network that predicts the time segment from the series; a segment buffer that holds the predicted segment; and a pattern that should be output when the noise-superimposed standard pattern is input to the pattern associative neural network unit. A predictive electrocardiogram waveform comprising: a learning/recognition control section that causes a pattern associative neural network section to learn by giving a division point as a teacher signal to give a training signal, and controls the learning of the recurrent neural network section. recognition device. 2. The predictive electrocardiogram waveform according to claim 1, characterized in that during learning, the input of the pattern associative neural network unit is provided with an input waveform buffer that holds a standard pattern at a time near the current time. recognition device.