JPH05207985A - Electrocardiogram waveform recognizing system - Google Patents

Abstract

PURPOSE:To learn a template waveform, to decide whether an error is within a threshold or not, and to recognize whether an input waveform is the same as the template waveform or not by varying a weight coefficient so as to decrease an error between output data and teach data. CONSTITUTION:Plural template waveforms of a nature waveform and an extrasystole waveform are selected, respectively, and stored in a data storage part 6. The template waveform is inputted to neural network, and as a result, an output waveform obtained in an output layer and a teacher signal are compared by an error deciding part 7, and each coupling coefficient is changed so that a square error decreases. By using the same waveform as the inputted template waveform for the teacher signal, an input waveform can be learned by the network. In the error deciding part 7, a square error of output data of the neural network 5 and template data stored in the data storage part 6 is calculated. The template waveform whose error becomes minimum is recognized as a waveform nearest to input data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an electrocardiogram waveform recognition system, and more particularly to an electrocardiogram waveform recognition system using a neural network.

[0002]

2. Description of the Related Art A conventional method for comparing electrocardiographic waveforms is, for example, when comparing QRS groups, the start point and the end point of the QRS group serving as a template are obtained by using a special division point recognition logic, and then the QRS group is compared. Parameters such as area, amplitude, time width, etc. are calculated, each parameter of the QRS group that is comparatively symmetric is calculated, both parameter values are compared, and if each parameter is within a certain error range, template QRS group There is a method to regard it as the same waveform as.

Further, the method of classification of electrocardiographic findings is P wave, Q
After the start point and the end point of the RS group, T wave, etc. are obtained by using the special division point recognition logic, parameters such as the amplitude value and time width of each spike wave are calculated, and each parameter is within the normal range. There is a way to find out. In order to perform normal finding classification, 12 waveforms are derived per examination (standard 1
The most common method is to obtain each parameter for each of the 2-lead electrocardiogram examination) and the 12-lead waveform and perform finding classification based on a certain classification standard (for example, Minnesota code classification standard).

[0004]

In the conventional method, the characteristics of the entire waveform are not sufficiently expressed only by limited numerical parameters such as amplitude, time width, and area. Therefore, even if the waveforms are different, there is not much difference in the numerical values of the parameters, and even if the waveforms can be regarded as the same waveform, if the noise is mixed, the parameter values will be affected and the waveforms will be judged as different. There was a point. In the classification of findings, parameters such as amplitude and time width are unstable in classification accuracy when they are very close to the classification reference value of findings, or when noise is mixed, an error occurs in parameter calculation. There were problems such as establishing findings classification.

[0005]

The electrocardiogram waveform recognition method of the present invention comprises a neural network which inputs and outputs digitized time-series electrocardiogram waveform data,
In the learning process of determining the weighting factor that connects the input and output of the neural network, the template waveform data of the electrocardiogram is used as the input data and the teacher data, and the weighting factor is changed so as to reduce the error between the output data and the teacher data. By learning the template waveform by this, by deciding whether the error between the output data and the template waveform data is within the threshold value in the recognition process, it is configured to recognize whether the input waveform is the same as the template waveform. ..

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

An electrocardiogram signal derived from a living body is amplified by an amplifier 1 and then converted into time series digital data by an A / D converter 2. For example, the sampling frequency in this embodiment is 125 Hz. In addition, the amplified electrocardiogram signal is also input to the R-wave detector 3 and Q of the electrocardiogram is detected.
The RS group time series generation position is detected. The data extraction unit 4 extracts data to be input to the neural network. In this embodiment, 80 m before and after the QRS group position
It was set to sec. That is, the number of data is 20 points before and after the QRS group detection position, and a total of 41 points including the detection points.

In the learning process, the extracted data is stored in the data storage unit 6. After inputting electrocardiogram data for a certain time, a plurality of heartbeats extracted by the data extraction unit 4 based on the QRS complex detection position detected by the R wave detection unit 3 within the input time and stored in the data storage unit 6 are stored. The electrocardiogram data of is displayed on the display unit 9. The operator selects a waveform serving as a template while observing the electrocardiogram data displayed on the display unit 9 and designates it from the operation unit 8. The template waveform is selected as a template even when the waveform is changing due to a change in body posture or the like even in the case of normal sinus rhythm, and each is also selected as a template even if the shape is different even during extrasystole. In this way, a plurality of template waveforms of the normal waveform and the extra systolic waveform are selected and stored in the data storage unit 6.

After selecting the template data, the neural network 5 is learned. FIG. 2 shows the neural network used in this embodiment. Input layer E
Is 41, the number of intermediate layers M is 10, and the number of output layers A is 41. In the learning process, the switch 11 is in a state in which the neural network 5 and the data storage unit 6 are connected, and the 41-point data of the template electrocardiogram waveform stored in the data storage unit is directly input to the input layer.

Back propagation is used as a learning algorithm. That is, in one learning, a certain template waveform is input to the neural network, and as a result, the output waveform obtained in the output layer is compared with the teacher signal by the error determination unit 7, and the square error is reduced. As described above, each coupling coefficient is changed. The coupling coefficient is a weighting coefficient that couples the elements of the input layer E and the intermediate layer M and the elements of the intermediate layer M and the output layer A, for example, w (E = i, M = j),
It is shown by w (M = i, A = j). By using the same waveform as the input template waveform for the teacher signal, the input waveform can be learned by the network.
The number of times of learning is performed 500 times for each template waveform stored in the data storage unit 6. The coupling coefficient at the end of learning is used in the recognition process.

In the recognition process, the 41 points of data before and after the QRS complex detection position are successively extracted by the data extraction unit 4 and input to the neural network 5 as in the learning process. The error determination unit 7 uses neural
The squared error between the output data of the network 5 and the template data stored in the data storage unit 6 is calculated,
The template waveform with the smallest error is recognized as the waveform closest to the input data, and among the template counters provided in the data storage unit 6, the counter corresponding to the template recognized as having the same waveform is incremented. .. Further, the occurrence time of the recognized waveform is stored in the data storage unit 6.

When the recognition process is completed, the template waveform data stored in the data storage unit, the count number of the counter corresponding to the template, and the histogram of the QRS group generated in a unit time are edited for each template and recorded in the recording unit 10. Recorded in. Recording examples are shown in FIGS. 3 and 4. This recording example is a result obtained by classifying the normal waveform into one pattern (FIG. 3) and the extra systolic waveform into four patterns (FIG. 4) using a total of five templates. The histogram and the number of occurrences are recorded. In FIG. 4, the number of occurrences of the template waveform corresponding to the designated pattern out of the four patterns of extra systolic waveforms is shown per unit time [in this case (A)] [every hour in this case]. It is aggregated and displayed. The number next to “TOTAL” under the histogram is the cumulative total number after the start of measurement.

[0013]

As described above, according to the present invention, it is possible to recognize whether the inputted electrocardiogram waveform is the same as the template waveform by using the neural network.
Compared with waveform recognition using numerical parameters such as amplitude value and time width, waveform feature extraction can be performed more faithfully, and waveform recognition can be performed accurately without being affected by noise contamination. Have.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing the structure of the neural network shown in FIG.

FIG. 3 is an explanatory diagram showing an example of recording according to the present embodiment.

FIG. 4 is an explanatory diagram showing an example of recording according to the present embodiment.

[Explanation of symbols]

 1 Amplifier 2 A / D Converter 3 R Wave Detecting Section 4 Data Extracting Section 5 Neural Network 6 Data Storage Section 7 Error Judgment Section 8 Operation Section 9 Display Section 10 Recording Section

Claims (5)

[Claims] 1. A template waveform of an electrocardiogram in a learning process, comprising a neural network which inputs and outputs digitized time-series electrocardiographic waveform data, and which determines a weighting factor connecting the input and output of the neural network. Data is used as input data and teacher data, and the template waveform is learned by changing the weighting coefficient so as to reduce the error between the output data and the teacher data.In the recognition process, the error between the output data and the template waveform data is within the threshold value. By determining if
An electrocardiogram waveform recognition method characterized by recognizing whether the input waveform is the same as the template waveform. 2. In the neural network, a plurality of template waveforms are learned in a learning process, and an input waveform is determined by determining whether an error between output data and each template data is within a threshold value in a recognition process. 2. The electrocardiogram waveform recognition method according to claim 1, wherein the template waveform is determined by determining which template waveform is close to. 3. The neural network, wherein, in the learning process, at least one normal waveform and at least one extrasystolic waveform QRS of the same subject.
By learning the data in the vicinity of the group and inputting the data in the vicinity of the QRS within the arbitrary time of the subject in the recognition process, and recognizing which QRS group the input waveform is close to The electrocardiogram waveform recognition method according to claim 2, wherein the number of extrasystoles that have occurred is counted. 4. The neural network learns one or a plurality of waveforms characteristic of an electrocardiographic finding in a learning process and determines whether an input waveform corresponds to the finding in a recognition process. The electrocardiogram waveform recognition method according to claim 1 or 2. 5. By using a plurality of neural networks according to claim 4, in the learning process, a neural network corresponding to at least one waveform characteristic of each of a plurality of different electrocardiographic findings. The network is trained, and in the recognition process, the input waveform is sequentially input to a plurality of neural networks,
An electrocardiographic waveform recognition method characterized by determining which electrocardiographic finding the input waveform corresponds to.