JP2014117537A - Waveform detecting device, waveform detecting method, and waveform detecting program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect a shape of a target signal waveform from an input signal with a high degree of precision.SOLUTION: A waveform detecting device 10 acquires a signal collected by an electrocardiographic sensor 20, and creates a waveform of a polygonal line approximate to a waveform of the signal. Furthermore, the waveform detecting device 10 calculates a feature quantity of extreme points where the inclination of straight lines adjacent to each other is inverted from positive to negative or vice versa, of intersections of straight lines forming the polygonal line. Still furthermore, the waveform detecting device 10 classifies the extreme points into a convex part with a convex shape and a flat base line part, using the feature quantity of the extreme points included in the polygonal line.

Description

  The present invention relates to a waveform detection device, a waveform detection method, and a waveform detection program.

  Noise may occur in a signal collected by a sensor or the like. For example, when detecting an electrocardiogram waveform from a signal collected by an electrocardiogram sensor worn by the subject, the electromyogram signal is superimposed on the input signal, or the electrocardiogram waveform is Noise is generated when the base line fluctuates.

  As an example of a method for reducing such noise, signal processing using a digital filter is applied. For example, after detecting an R wave in an ECG waveform having a sharp shape from an input signal using a high-pass filter, other characteristic waveforms such as a P wave, a Q wave, and a T wave are detected starting from the R wave. Is done.

Special table 2012-502670 gazette Special table 2007-524466 Special table 2010-516430 gazette Japanese Patent Laid-Open No. 3-201824

  However, the noise reduction method has a problem that the shape of the target signal waveform cannot be accurately detected from the input signal on which the noise is superimposed, as described below.

  That is, various parameters such as a cut-off frequency are set in the above digital filter in accordance with individual differences in heart rate variability and electrocardiogram waveforms. However, it is difficult to set parameters by predicting noise such as myoelectric signals and baseline fluctuations in advance. For this reason, when the myoelectric signal is superimposed on the input signal, the myoelectric signal passes through the high-pass filter, or when the baseline of the electrocardiogram waveform included in the input signal fluctuates, The corresponding signal component may pass through the high-pass filter. Thus, even if the above-described high-pass filter is used, noise may be erroneously detected as an R wave.

  An object of one aspect is to provide a waveform detection apparatus, a waveform detection method, and a waveform detection program that can accurately detect the shape of a target signal waveform from an input signal.

  The waveform detection apparatus according to one aspect includes an acquisition unit that acquires a signal collected by a sensor, and a generation unit that generates a waveform of a polygonal line that approximates the waveform of the signal. Furthermore, the waveform detection apparatus includes a calculation unit that calculates a feature amount of an extreme point at which the positive / negative of the slope of the straight lines adjacent to each other among the intersections of the straight lines forming the broken line is inverted. Furthermore, the waveform detection apparatus has a classification unit that classifies the extreme value points into convex portions having a convex shape or flat base line portions using feature values of the extreme value points included in the broken line.

  According to one embodiment, the shape of a target signal waveform can be accurately detected from an input signal.

FIG. 1 is a block diagram illustrating a functional configuration of the waveform detection apparatus according to the first embodiment. FIG. 2 is an example of a graph showing a low-pass waveform and a broken line waveform. FIG. 3 is a diagram illustrating an example of feature values of extreme points. FIG. 4 is a diagram illustrating an example of feature values of extreme points. FIG. 5 is a diagram illustrating a configuration example of data stored in the extreme value point storage unit. FIG. 6 is a diagram illustrating an example of a classification tree. FIG. 7 is a flowchart of the learning process according to the first embodiment. FIG. 8 is a flowchart illustrating the procedure of the classification process according to the first embodiment. FIG. 9 is an example of a graph showing a low-pass waveform and a broken line waveform. FIG. 10 is an example of a graph showing a low-pass waveform and a broken line waveform. FIG. 11 is a flowchart illustrating the procedure of the determination process according to Application Example 1. FIG. 12 is an example of a graph showing a low-pass waveform and a broken line waveform. FIG. 13 is a flowchart illustrating a procedure of RT pair candidate extraction processing according to Application Example 3. FIG. 14 is a flowchart illustrating a procedure of RT pair determination processing according to the application example 3. FIG. 15 is an example of a graph showing a low-pass waveform and a broken line waveform. FIG. 16 is an example of a graph showing a low-pass waveform and a broken line waveform. FIG. 17 is a diagram illustrating an example of the tendency of the appearance order of extreme values. FIG. 18 is a diagram illustrating an example of the tendency of the appearance order of extreme values. FIG. 19 is a diagram illustrating an example of the tendency of the appearance order of extreme values. FIG. 20 is a diagram illustrating an example of the tendency of the appearance order of extreme values. FIG. 21 is a flowchart illustrating a procedure of extreme value appearance order trend determination processing according to Application Example 4. FIG. 22 is an example of a graph showing the accuracy rate. FIG. 23 is an example of a graph showing signal waveforms of the acceleration sensor. FIG. 24 is a diagram illustrating an example of a classification tree. FIG. 25 is a schematic diagram illustrating an example of a computer that executes a waveform detection program according to the first and second embodiments.

  Hereinafter, embodiments of a waveform detection device, a waveform detection method, and a waveform detection program disclosed in the present application will be described in detail with reference to the drawings. Note that this embodiment does not limit the disclosed technology. Each embodiment can be appropriately combined within a range in which processing contents are not contradictory.

  FIG. 1 is a block diagram illustrating a functional configuration of the waveform detection apparatus according to the first embodiment. The waveform detection apparatus 10 shown in FIG. 1 detects the shape of the subject's electrocardiogram waveform, for example, an R wave or a T wave that forms a part of the electrocardiogram waveform, from a signal collected by the electrocardiographic sensor 20 worn by the subject. The waveform detection process to be executed is executed.

  As one aspect, the waveform detection apparatus 10 can be implemented by causing a desired computer to install a waveform detection program provided as package software or online software. For example, the above waveform detection program may be installed in a fixed terminal such as a personal computer. The waveform detection program may be installed in a mobile terminal device such as a smartphone, a mobile phone, a PHS (Personal Handyphone System), or a PDA (Personal Digital Assistants). As another aspect, the waveform detection apparatus 10 may be implemented as a dedicated device that is integrated with the electrocardiographic sensor 20 and provides the waveform detection function described above. As a further aspect, the waveform detection device 10 may be implemented as a Web server that provides a waveform detection service, or may be implemented as a cloud that provides the waveform detection service by outsourcing.

  As shown in FIG. 1, the waveform detection device 10 and the electrocardiographic sensor 20 are communicably connected from at least the electrocardiographic sensor 20 to the waveform detection device 10. The waveform detection device 10 and the electrocardiographic sensor 20 may be communicated with each other in a wired or wireless connection form. For example, when communication is performed via a network, any type of communication network such as the Internet, a local area network (LAN), or a virtual private network (VPN) can be employed. Although FIG. 1 illustrates the case where the number of electrocardiographic sensors 20 connected to the waveform detection apparatus 10 is one, a plurality of electrocardiographic sensors 20 can be connected to the waveform detection apparatus 10.

  The electrocardiographic sensor 20 is a device that collects a potential difference between a plurality of electrodes. By collecting such a potential difference, a bioelectric phenomenon having a voltage of about several mV, a frequency of about 0.1 to 200 Hz, and an impedance of about 1 to 20 kΩ, so-called electromotive force, is measured as an electrocardiographic signal. For example, the electrocardiographic sensor 20 is mounted to include an electronic circuit that detects and amplifies a potential difference between electrodes, a digital signal circuit that converts an analog signal of a potential difference into a digital signal at a predetermined sampling frequency, and the like.

  At least two or more disposable electrodes to which a conductive adhesive gel is attached are attached to the electrocardiographic sensor 20. The electrocardiogram signal is measured in a state where these disposable electrodes are mounted on the surface of the living body of the subject, for example, the chest. At this time, when the contact state changes between the disposable electrode and the body surface due to the body movement of the subject, the potential difference between the electrodes also changes as the contact state changes. In this case, noise is superimposed on the signal input from the electrocardiographic sensor 20 to the waveform detection device 10, and the baseline of the electrocardiographic waveform varies, for example. In addition, when the electrode is attached directly above the pectoral muscle, the myoelectric signal is also observed at the same time, so that the myoelectric signal is superimposed on the electrocardiographic signal. Since the myoelectric signal has a higher frequency than the electrocardiographic signal, it is difficult to detect the convex portion of the electrocardiographic waveform when the myoelectric signal is superimposed.

  The influence of these noises on the electrocardiographic signal varies among individuals depending on the subject, and also varies depending on the type and size of body movement. Furthermore, the effect of noise increases as the number of electrodes attached to the surface of the living body decreases or as a simple digital signal circuit is employed. For this reason, when noise is reduced using a digital filter such as a high-pass filter, it is difficult to appropriately set parameters such as a cutoff frequency.

  Therefore, the waveform detection apparatus 10 according to the present embodiment approximates the waveform of the input signal to a polygonal line waveform, and has an extreme value having an extreme value at which the positive and negative of the slopes of adjacent straight lines among the intersections of the straight lines forming the broken line are inverted. The extreme points are classified into a convex portion and a base line portion using feature values of the value points.

  Thus, in the waveform detection apparatus 10 according to the present embodiment, modeling is performed into a broken line waveform that approximates the waveform of the input signal. For this reason, it is possible to suppress a situation in which high-frequency noise such as a myoelectric signal is detected as an extreme point among noises superimposed on the electrocardiogram signal. Furthermore, in the waveform detection apparatus 10 according to the present embodiment, the extreme points having a convex shape are narrowed down to the extreme points having a convex shape among the intersections of the straight lines forming the polygonal waveform, and the extreme points are defined as the convex portions of the electrocardiographic waveform by the sharpness of the convex shape. Classify to baseline. Therefore, the “convex part” of the electrocardiogram waveform, that is, the part where the amplitude change is large in the one-beat ECG waveform and the “base line part”, ie, the flat part where the amplitude change is small in the one-beat ECG waveform It is possible to reduce the frequency of applying signal processing for extracting or blocking a specific frequency component to classify extreme points. As a result, as with a digital filter such as a high-pass filter, the detection accuracy is not affected by the filter design, for example, the setting of parameters such as the cut-off frequency, and extreme points even if noise occurs due to body movement or myoelectric signals Can be stably classified into convex portions or baseline portions.

  Therefore, according to the waveform detection apparatus 10 according to the present embodiment, the shape of the electrocardiographic waveform can be accurately detected from the input signal.

  As shown in FIG. 1, the waveform detection apparatus 10 includes an acquisition unit 11, a low-pass filter (LPF) 12, a generation unit 13, a calculation unit 14, an extreme point storage unit 15, and learning. Unit 16, learning result storage unit 17, and classification unit 18. Note that the waveform detection apparatus 10 may include various functional units included in a known computer in addition to the functional units illustrated in FIG. For example, when the waveform detection device 10 is implemented as a portable terminal, the function unit of the portable terminal such as an antenna, a carrier communication unit that performs communication via a carrier network, a GPS (Global Positioning System) receiver, and the like. May further be included.

  The acquisition unit 11 is a processing unit that acquires a signal collected by the electrocardiographic sensor 20. As one aspect, the acquisition unit 11 acquires a signal in which a potential difference between electrodes is collected in time series by the electrocardiographic sensor 20 at a constant period. Hereinafter, a signal input from the electrocardiographic sensor 20 to the waveform detection apparatus 10 may be referred to as an “input signal”. As another aspect, the acquisition unit 11 is an auxiliary storage device such as a hard disk or an optical disk or a removable memory card such as a USB (Universal Serial Bus) memory that accumulates signals collected from the subject using the electrocardiographic sensor 20. Signals can also be obtained from media. As a further aspect, the acquisition unit 11 can also acquire an input signal from the electrocardiographic sensor 20 or an external device on which the electrocardiographic sensor 20 is mounted via a network.

  The low-pass filter 12 is a digital filter that passes a signal component having a low frequency in the input signal acquired by the acquisition unit 11. The low-pass filter 12 can employ any filter circuit design method such as a biquadratic type or a Butterworth type. The cutoff frequency and order set in the transfer function of the low-pass filter 12 cut off the signal component of the myoelectric signal having a higher frequency than that of the electrocardiogram signal, while the R wave or T of the target electrocardiogram waveform. The value is preferably such that a signal component of a convex portion such as a wave is extracted.

  The generation unit 13 is a processing unit that generates a polygonal waveform that approximates the waveform of the low-pass signal output by the low-pass filter 12. As one aspect, the generation unit 13 time-divides the waveform of the low-pass signal output by the low-pass filter 12 at equal intervals. Hereinafter, the waveform of the low-pass signal is sometimes referred to as “low-pass waveform”, and the waveform of the broken line is sometimes referred to as “polygonal waveform”. Then, the generation unit 13 generates a regression line of the section from the digital value of the sampling point included in the low-pass waveform for each section of the low-pass waveform that is time-divided at equal intervals. Note that the division width in which the above sections are time-divided can be set to an arbitrary value by the designer of the waveform detection program, the user of the waveform detection apparatus 10, and the like, and different division widths are set between the sections. You can also

FIG. 2 is an example of a graph showing a low-pass waveform and a broken line waveform. The vertical axis of the graph shown in FIG. 2 indicates the amplitude representing the strength of the electrical stimulation, and the horizontal axis of the graph indicates time. In FIG. 2, the low-pass waveform is illustrated by a broken line, and the broken line waveform is illustrated by a solid line. As shown in FIG. 2, when approximation to a polygonal line is performed, the low-pass waveform is time-divided into equal intervals Ts, for example, a period in which the potential difference is sampled ten times by the electrocardiographic sensor 20. At this time, if a regression line is independently obtained for each time-divided section, the regression line is not continuous between the sections, and a polygonal waveform may not be obtained. Therefore, in order to make the regression line continuous between the sections, a regression line passing through the end point E ₀ of the regression line generated in the immediately preceding section is generated.

For example, the coordinates of the end point of the previous regression line are (t ₀ , S ₀ ), the coordinates of the sampling points of the low-pass waveform in the interval for obtaining the regression line are (t ₁ , S ₁ ) _,. , S _N ), the slope a and the intercept b of the regression line S = at + b can be calculated by the following equations (1) and (2). Here, “intercept b” represents the amplitude at the time when the approximation to the polygonal line waveform is started, that is, the intersection of the regression line and the horizontal axis. A series of regression lines in which the end point and the start point are shared between the sections, that is, a polygonal line waveform, can be obtained by the following formulas for calculating the slope a and the intercept b.

  The calculation unit 14 is a processing unit that calculates feature values of extreme points where the signs of the slopes of the straight lines adjacent to each other among the intersections of the straight lines forming the polygonal line are reversed. Here, the above-mentioned “extreme point” is a point where the slope of the broken line changes from positive to negative or from negative to positive among the intersections of the regression line and the regression line, and is therefore included in an electrocardiographic waveform of one beat. There is a high possibility of corresponding to a convex portion having a large amplitude such as an R wave or a T wave. In order to detect a convex portion having a large amplitude such as an R wave or a T wave contained in a single-beat electrocardiographic waveform from the polygonal line waveform by using the property of the extreme point, That is, an index representing the sharpness of the convex shape is calculated as the feature amount.

  Examples of the above-described feature amount include the following types (1) to (11). Specifically, (1) the difference in the inclination of the broken line, (2) the amount of change in the vertical axis with respect to the front extreme value, (3) the amount of change in the vertical axis with respect to the rear extreme value, (4) the inclination of the front broken line, It is possible to calculate, as feature quantities, a) an intercept of a front broken line, and (6) an inclination of a rear broken line. (7) Intersection of rear broken line, (8) Difference in vertical axis between intersection and extremal point of regression line, (9) Difference in slope of regression line, (10) Height of convex part, (11) A sharpness index or the like can be calculated as a feature amount.

  3 and 4 are diagrams illustrating an example of feature values of extreme points. The vertical axis of the graphs shown in FIGS. 3 and 4 indicates amplitude, and the horizontal axis of the graph indicates time. 3 and 4, the low-pass waveform is illustrated by a broken line, the broken line waveform is illustrated by a solid line, and the extreme points on the broken line are illustrated by black circles. In this case, before the extreme point for obtaining the feature value is sampled, it is described as “front”, and after the extreme point for obtaining the feature value is sampled, it may be described as “back”. is there.

For example, (1) “difference in polygonal line slope” means the difference between the slope of the regression line generated in the section ahead of the extreme point and the slope of the regression line generated in the section after the extreme point. Point to. For example, when attention is paid as a target for obtaining a feature quantity extreme points e _m shown in FIG. 3, the inclination of the extreme point e regression line generated in the front section of _m (i), the extreme point e _m The difference from the slope of the regression line (B) generated in the rear section is calculated as the feature quantity (1). Since the difference in inclination is smaller as the feature amount (1) is smaller, it can be evaluated that the convex shape of the local wave including the extreme point is sharper. Therefore, it can be estimated that the smaller the feature amount (1), the higher the possibility that the extreme point is a convex portion of the electrocardiogram waveform.

Further, (2) “the amount of change in the vertical axis with respect to the front extreme value” refers to the amount of change in the amplitude measured between the extreme value point at which the feature value is obtained and the extreme value point in front of the feature point. The “vertical change amount with respect to the rear extremum” refers to the amount of change in amplitude measured between the extremum point at which the feature value is obtained and the extremum point behind it. For example, when attention is paid to the extremum point e _m shown in FIG. 3, the feature amount of the amplitude value of the extreme point e _m of interest, the difference between the amplitude value of the front extreme point e _m-1 (2) Is calculated as On the other hand, is calculated as the feature amount of the difference between the amplitude value of the extreme point e _m of the target, the back of the amplitude value of the extreme point e _{m + 1} (3). Therefore, the same value as the feature amount of (2) of the extreme point e _m, extrema e _m-1 of the feature amount of (3). It can be evaluated that the convex shape of the local wave including the extreme point is sharper as the feature quantities of (2) and (3) are larger. Therefore, it can be estimated that there is a higher possibility that the extreme points are convex portions of the electrocardiogram waveform as the feature amounts of (2) and (3) are larger.

In addition, (4) “Inclination of front broken line”, (5) “Intersection of front broken line”, (6) “Inclination of rear broken line”, and (7) “Intersection of rear broken line” are sections in front of extreme points. This is a parameter relating to the regression line generated in step 1 or the regression line generated in the section after the extreme point. When focusing on extreme points e _m shown in FIG. 3, with the inclination of the extreme point e regression line generated in the front section of _m (i) is calculated as a feature amount of (4), the front of the regression The intercept of the straight line (A) is calculated as the feature quantity of (5). Further, the inclination of the extreme point e _m behind the generated regression line in section (b) is calculated as a feature amount of (6), characterized in sections of the rear of the regression line (b) is (7) Calculated as a quantity. In this way, the regression line intercept is used as a feature value for the slope of the regression line alone, whether the steep change in the polygonal line waveform is a baseline fluctuation due to body movement, or a convex part due to pulsation, that is, This is because it may not be possible to identify whether the wave is an R wave or a T wave.

Further, (8) “difference in the vertical axis direction between the intersection of the regression line and the extreme value point” means that the regression line and the feature value generated in the interval between the extreme value point for obtaining the feature value and the extreme value point in the front are obtained. It refers to the difference between the amplitude value of the intersection of the obtained extreme value point and the regression line generated in the interval between the extreme value point and the amplitude of the extreme value point of the low-pass waveform. For example, when attention is paid to the extremum point e _m shown in FIG. 4, with regression in front extreme point e _m-1 and interval of extreme point e _m of the target straight line (alpha) is calculated, the interest poles regression line (beta) is calculated by the value point e _m and rear extreme point e _{m + 1} sections. Among them, the time interval of the front extreme point e _m-1 and extreme point e _m of interest, since the matching the division width Ts of the low-pass waveform, same regression and the regression line (a) shown in FIG. 3 A straight line (α) is calculated. On the other hand, extreme point e _m and the time interval of the rear extreme point e _{m + 1} of interest is larger than the division width Ts of the low-pass waveform does not match the division width Ts of the low-pass waveform, shown in FIG. 3 Unlike the regression line (b), the regression line (β) is calculated in a section longer than the division width Ts of the low-pass waveform. On top of that, the amplitude value of the intersection point P _m of the regression line (alpha) and the regression line (beta), the difference between the amplitude value of the extreme point Q of the low-pass waveform is calculated as a feature amount of (8). It can be evaluated that the smaller the feature amount (8), the higher the degree of coincidence between the extreme point of the polygonal line waveform and the extreme point of the low-pass waveform. Therefore, it can be estimated that the feature value at the extreme point is more reliable as data as the feature value in (8) is smaller.

  In addition, (9) “difference in the slope of the regression line” means that the regression line generated in the interval between the extreme points and the extreme extreme points in front and the regression generated in the interval between the extreme points and the backward extreme points Refers to the difference in slope from a straight line. As shown in FIG. 4, the difference between the slope of the regression line (α) and the slope of the regression line (β) described in (8) feature amount is calculated as the feature amount (9). Since the difference in inclination is smaller as the feature amount (9) is smaller, it can be evaluated that the convex shape of the local wave including the extreme point is sharper. Therefore, it can be estimated that the smaller the feature amount of (9), the higher the possibility that the extreme point is a convex portion of the electrocardiogram waveform.

Also, (10) “convex height” refers to the height from the base line of vibration, and for example, the smaller one of the feature value (2) and the feature value (3) is adopted. When focusing on extreme points e _m shown in FIG. 3, since the (2) and smaller than the feature quantity of the values of characteristic quantity (2) of (3) of (3), characterized in (3) The value of the quantity is adopted as the feature quantity of (10). As described above, the smaller one of (2) and (3) is adopted as the convex portion height. When the larger value is adopted as the convex portion height, the regression line slope is negative. When extreme points that change to positive and extreme points where the slope of the regression line changes from positive to negative appear continuously, the feature value is the same between consecutive extreme points Because there is. In this way, if extreme points with the same feature amount are consecutive, whether each extreme point is an inflection point on the convex portion of the electrocardiogram waveform or an inflection point on the baseline. Cannot be identified. In order to identify these inflection points, the smaller one of (2) and (3) is adopted as the feature value of (10).

In addition, (11) “sharpness index” refers to a value obtained by dividing the feature value in (10) by the time width from the extreme extreme point to the backward extreme value point based on the extreme value point for obtaining the feature value. . In the example of FIG. 4 shows a front extreme point e _m-1 convex height of the extreme point e _m the time range of up to extreme point _{m + 1} backward, that is, FIG. 3 (10) A value obtained by dividing the feature amount is calculated as the feature amount of (11). It can be evaluated that the convex shape of the local wave including the extreme point is sharper as the feature amount of (11) is larger. Therefore, it can be estimated that the larger the feature quantity in (11), the higher the possibility that the extreme point is a convex portion of the electrocardiogram waveform.

  Thus, the calculation unit 14 can calculate the feature amounts (1) to (11) described above as feature points of extreme points. At this time, the calculation unit 14 does not necessarily calculate all the feature amounts (1) to (11), and calculates at least one feature amount among the feature amounts (1) to (11). do it. The calculation unit 14 can also calculate feature amounts for an arbitrary number and an arbitrary combination of the above-described feature amounts (1) to (11).

  Here, the feature value of the extreme value point can be used to detect the shape of the electrocardiographic waveform by performing a classification process of classifying the extreme value point of the polygonal line waveform into a convex part or a base line part. It can also be used as a learning data set for generating a determination model to be applied to the classification process. In the following, after describing the functional unit of the learning system that learns the determination model to be applied to the classification process using the learning data set, that is, the extreme point storage unit 15, the learning unit 16, and the learning result storage unit 17, A function unit of a classification system that classifies extreme points using the determination model obtained as a result, that is, the classification unit 18 will be described. FIG. 1 illustrates the case where the waveform detection apparatus 10 has both a learning system function unit and a classification system function unit. However, the waveform detection apparatus 10 does not have a learning system function unit and has only a classification unit 18. It doesn't matter.

  Returning to the description of FIG. 1, the extreme point storage unit 15 is a storage unit that stores various types of information related to extreme point. As an example of access to the extreme value point storage unit 15, when the feature value of the extreme value point is calculated by the calculation unit 14, the feature value for each extreme value point is registered in the extreme value point storage unit 15. . As another example, when a feature value of an extreme value point is newly registered, the extreme value point can be visually checked by a designer of a waveform detection program, a user of the waveform detection device 10, or the like. The correct answers of the classified classes are additionally registered in the extreme point storage unit 15. As a further example, when a new learning data set is completed by additionally registering a class in the feature value of the extreme point, all the learning data stored in the extreme point storage unit 15 is stored. The set is referred to by the learning unit 16.

  As an example of the data structure of the extreme point storage unit 15, data in which items such as the time when the extreme point is sampled, the feature amount of the extreme point, and the class to which the extreme point belongs can be adopted. FIG. 5 is a diagram illustrating a configuration example of data stored in the extreme point storage unit 15. As shown in FIG. 5, the feature quantities (1) to (11) are stored for each extreme point. In the example of FIG. 5, it means that extreme points appear at the 30th, 50th, 80th and 120th sampling points. Furthermore, in the example of FIG. 5, the extreme points sampled at the 30th and 120th points belong to the “base line portion”, and the extreme points sampled at the 50th and 80th points are “ It means that it belongs to "convex part".

  The learning unit 16 classifies whether the extreme point belongs to the convex portion class or the baseline portion class by using the feature value and class of the extreme point stored in the extreme point storage unit 15. It is a processing unit that learns a determination model to be applied to the classification process. Arbitrary algorithms such as boosting, neural network, support vector machine and the like can be employed for such machine learning of classification. Here, as an example, a case of generating a classification tree having feature points of extreme points as nodes is illustrated. In this case, the learning unit 16 classifies, for example, the nodes to be adopted as the nodes among the feature quantities (1) to (11) above, the hierarchy in which the nodes are arranged, and the threshold value set in each node. A classification tree is generated so that the correct answer rate of is the highest. Then, the learning unit 16 registers the generated classification tree, that is, the learning result of the determination model, in the learning result storage unit 17. Thus, when generating a classification tree, the accuracy of extreme point classification can be increased as the number of learning data sets increases.

  FIG. 6 is a diagram illustrating an example of a classification tree. When the classification tree shown in FIG. 6 is used as a determination model, it is first determined whether the (10) convex part height of the extreme point is less than 27.8 (step S11). Thus, it can be seen from the classification tree shown in FIG. 6 that the height of the convex portion, which is the feature amount of (10), is useful for broadly dividing the extreme points into the convex portion or the base line portion.

  At this time, if the height of the convex portion is less than 27.8, it is further determined whether or not the (10) convex portion height of the extreme point is less than 22.3 (step S12). Thus, the threshold value is lowered and the convex portion height is re-determined because there is still room for the extreme point to be a convex portion even if the convex portion height is less than 27.8. . When the height of the convex portion is less than 22.3, the extreme point is classified as a “base line portion”. On the other hand, if the height of the convex portion is 22.3 or more, it is further determined whether or not the slope of the extreme point (4) front broken line is less than -1.18 (step S13). At this time, when the inclination of the front broken line is less than −1.18, the extreme point is classified as “base line portion”, while when the inclination of the front broken line is −1.18 or more, Extreme points are classified as “convex parts”.

  When the height of the convex portion is 27.8 or more, it is further determined whether or not (4) the inclination of the front polygonal line of the extreme value point is −1.27 or more (step S14). The reason why the determination is continued by changing the type of feature amount is that there is still room for the extreme point not to be a convex portion even if the convex portion height is 27.8 or more. And when the inclination of a front broken line is -1.27 or more, an extreme point is classified into "convex part". On the other hand, when the inclination of the front broken line is less than −1.27, it is further determined whether or not the (10) convex portion height of the extreme point is less than 41.8 (step S15). At this time, when the height of the convex portion is less than 41.8, the extreme point is classified as the “base line portion”. Further, when the height of the convex portion is 41.8 or more, (8) it is further determined whether or not the difference in the vertical axis direction between the intersection of the regression line and the extreme value point is 9.0 or more. (Step S16). The purpose of this determination is to determine the reliability of data related to the feature value of the extreme point. For this reason, when the feature amount of (8) is excessively large, it is estimated that the reliability of the data is low and classification as a convex portion is avoided. In other words, when the difference in the vertical axis direction between the intersection of the regression lines and the extreme value point is 9.0 or more, the extreme value point is classified as the “base line portion”. On the other hand, when the vertical axis difference between the intersection of the regression lines and the extreme value point is less than 9.0, it is considered that the reliability of the data has been confirmed, and the extreme value point is classified as a “convex portion”. .

  In addition, although the case where the machine learning is used for generation of the determination model is illustrated here, it is not always necessary to use the machine learning for generation of the determination model, and the threshold determination is easily performed using one type of feature amount. It doesn't matter. For example, it is determined whether or not the height of the convex portion is equal to or greater than a predetermined threshold. If the height is equal to or greater than the threshold, the extreme point is classified as “convex”, while if the height is less than the threshold, It is not necessary to prepare a learning data set by using a simple determination model of classifying as “part”. At this time, if the height of the convex portion is equal to or greater than a predetermined threshold, it is further determined whether or not the difference in the vertical axis direction between the intersection of the regression line and the extreme value point is equal to or greater than the predetermined threshold. In this case, the extreme points may be classified as “convex portions” by narrowing down.

  Returning to the description of FIG. 1, the learning result storage unit 17 is a storage unit that stores the learning result of the determination model. As an example of access to the learning result storage unit 17, when the learning unit 16 learns a classification tree as a determination model, the classification tree is registered in the learning result storage unit 17. As another example, when the feature values of extreme points are classified by the classification unit 18 described later, the classification tree stored as the learning result of the determination model in the learning result storage unit 17 is referred to by the classification unit 18. . As an example of the data format stored in the learning result storage unit 17, the hierarchical structure of nodes included in the classification tree, the feature amount appearing in each node, and the threshold value of each node are described in the format of IF to THEN to sentence. Data can be adopted.

  The classification unit 18 is a processing unit that classifies the extreme point into a convex portion having a convex shape or a flat base line portion using the feature amount of the extreme point included in the polygonal line waveform. As one aspect, the classification unit 18 applies the extreme value point to the convex portion or the base line portion by applying the classification tree stored in the learning result storage unit 17 to the feature value of the extreme value point calculated by the calculation unit 14. Classify. Thereafter, the classification unit 18 stores the classification result of each extreme value point in the storage unit in the waveform detection device 10, displays it on a display device connected to the waveform detection device 10, or outputs it to an external device. .

  For example, as an example of a method for displaying the classification result, the appearance time of a convex portion such as an R wave or a T wave is obtained by synchronizing the time with at least one of an input signal waveform, a low-pass waveform, or a polygonal waveform. Can be displayed. Furthermore, the characteristics of the electrocardiogram signal, that is, the time interval of one beat is substantially the same, and a characteristic shape appears in the order of P wave, Q wave, R wave, S wave, and T wave in one beat electrocardiogram waveform. In addition, using the property of having periodicity, the interval between the R wave and the R wave, so-called RRI (RR Interval) and the heart rate are obtained and displayed from the appearance time of the extreme points classified as convex portions. You can also.

  Moreover, as an example of the output destination of the classification result, there is a terminal device used by a person related to a user who uses the waveform detection device 10, such as a caregiver or a doctor. Another example is a server on which a diagnostic program for diagnosing the presence or absence of heart disease or autonomic nerve malfunction from RRI or heart rate is implemented. In this case, the diagnosis result obtained by diagnosing the presence or absence of the subject's heart disease or autonomic nerve malfunction by the diagnostic program operated on the server can be displayed on the display device of the waveform detection apparatus 10. If there is an abnormality in the diagnosis result, the diagnosis content can be output to a terminal device used by a person who uses the waveform detection apparatus 10. In this way, the terminal device used by the parties related to the user who uses the waveform detection device 10 using the RRI and heart rate obtained from the classification result, as well as the classification result of each extreme point, and further the above-described diagnosis result. By outputting to, monitoring service for out-of-hospital, for example, staying at home or sitting, becomes possible.

  Various integrated circuits and electronic circuits can be employed for various functional units such as the acquisition unit 11, the low-pass filter 12, the generation unit 13, the calculation unit 14, the learning unit 16, and the classification unit 18 illustrated in FIG. For example, examples of the integrated circuit include ASIC (Application Specific Integrated Circuit) and FPGA (Field Programmable Gate Array). Examples of the electronic circuit include a central processing unit (CPU) and a micro processing unit (MPU).

  Moreover, the following devices can be employed for various storage units such as the extreme point storage unit 15 and the learning result storage unit 17 shown in FIG. For example, a semiconductor memory element such as a random access memory (RAM), a read only memory (ROM), or a flash memory can be employed. A storage device such as a hard disk or an optical disk can also be employed.

[Process flow]
Next, a processing flow of the waveform detection apparatus 10 according to the present embodiment will be described. Here, after (1) learning processing executed by the waveform detection apparatus 10 is described, (2) classification processing is described.

(1) Learning Process FIG. 7 is a flowchart illustrating the procedure of the learning process according to the first embodiment. This learning process is a process executed when “learning mode” is selected from “learning mode” for learning a determination model or “classification mode” for classifying extreme points, for example.

  As illustrated in FIG. 7, the acquisition unit 11 acquires a signal in which the potential difference between the electrodes is collected in time series by the electrocardiographic sensor 20 (step S101). Subsequently, the low-pass filter 12 removes high-frequency noise by passing a low-frequency signal component of the signal acquired in step S101 (step S102).

  And the production | generation part 13 produces | generates the polygonal line waveform approximated to the low-pass waveform from which the high frequency noise was removed by step S102 (step S103). Subsequently, the calculation unit 14 has an extreme point where the slope of the broken line changes from positive to negative or from negative to positive at the intersection of the regression line and the regression line included in the broken line waveform generated in step S103. It is determined whether or not (step S104).

  At this time, when there is an extreme point (step S104 Yes), the calculation unit 14 calculates the feature amount of each extreme point, for example, the feature amount of (1) to (11) above (step S105). . If there is no extreme point (No in step S104), the process from step S105 to step S109 is skipped, and the process is terminated as it is. Examples of cases where such extreme points do not exist include situations where the user does not wear the electrocardiographic sensor 20, or the wearing method is inappropriate and the input signal does not include an electrocardiographic signal.

  Subsequently, the calculation unit 14 obtains a correct answer of the classification result in which the shape of the extreme value point is classified by visual observation of the low-pass waveform or the broken line waveform by the designer of the waveform detection program, the user of the waveform detection device 10, and the like ( Step S106).

  Thereafter, the calculation unit 14 registers the feature value of the extreme point and the correct answer of the classification result in the extreme point storage unit 15 in association with each extreme point (step S107).

  After that, the learning unit 16 uses the feature value and class of the extreme point stored in the extreme point storage unit 15, and the extreme point belongs to the convex portion class or the baseline portion class. A determination model to be applied to the classification process for classifying the monster is learned (step S108).

  Then, the learning unit 16 registers the learning result of the determination model generated in step S108 in the learning result storage unit 17 (step S109), and ends the process.

  As described above, in the learning process illustrated in FIG. 7, as long as the waveform detection apparatus 10 is set to the “learning mode”, the processes in steps S101 to S109 are repeatedly executed.

  In the flowchart shown in FIG. 7, the case where the process is executed every time a signal is acquired from the electrocardiographic sensor 20 at a constant period is illustrated, but the feature of each extreme point is stored in the extreme point storage unit 15. The amount may be registered and the process may be executed from step S106. The feature value of each extreme point and the correct answer of the classification result may be registered in advance in the extreme point storage unit 15, and the process may be started from step S108.

(2) Classification Processing FIG. 8 is a flowchart illustrating a classification processing procedure according to the first embodiment. This classification process is, for example, a process executed when “classification mode” is selected from “learning mode” or “classification mode”.

  As illustrated in FIG. 8, the acquisition unit 11 acquires a signal in which the potential difference between the electrodes is collected in time series by the electrocardiographic sensor 20 (step S201). Subsequently, the low-pass filter 12 removes high-frequency noise by passing a low-frequency signal component of the signal acquired in step S201 (step S202).

  And the production | generation part 13 produces | generates the polygonal line waveform approximated to the low-pass waveform from which the high frequency noise was removed by step S202 (step S203). Subsequently, the calculation unit 14 has an extreme point where the slope of the broken line changes from positive to negative or from negative to positive at the intersection of the regression line and the regression line included in the broken line waveform generated in step S203. Is determined (step S204).

  At this time, when there is an extreme point (step S204 Yes), the calculation unit 14 calculates the feature amount of each extreme point, for example, the feature amount of (1) to (11) above (step S205). . If there is no extreme point (No in step S204), the process from step S205 to step S207 is skipped, and the process is terminated as it is. Examples of cases where such extreme points do not exist include situations where the user does not wear the electrocardiographic sensor 20, or the wearing method is inappropriate and the input signal does not include an electrocardiographic signal.

  Subsequently, the classifying unit 18 classifies the extreme points as convex portions or baseline portions by applying the determination model stored in the learning result storage unit 17 to the feature values of the extreme points calculated in step S205 ( Step S206).

  Thereafter, the classification unit 18 displays the classification result of each extreme point classified in step S206 on a display device connected to the waveform detection device 10 or outputs it to an external device (step S207). finish.

  As described above, in the classification process illustrated in FIG. 8, as long as the waveform detection device 10 is set to the “classification mode”, the processes in steps S201 to S207 are repeatedly executed.

[Effect of Example 1]
As described above, the waveform detection apparatus 10 according to the present embodiment approximates the waveform of the input signal to a polygonal line, and poles that invert the positive and negative of the slopes of adjacent straight lines among the intersections of the straight lines forming the broken line. The extreme points are classified into a convex part and a base line part using feature values of extreme points having values.

  Thus, in the waveform detection apparatus 10 according to the present embodiment, modeling is performed into a broken line waveform that approximates the waveform of the input signal. For this reason, it is possible to suppress a situation in which high-frequency noise such as a myoelectric signal is detected as an extreme point among noises superimposed on the electrocardiogram signal. Furthermore, in the waveform detection apparatus 10 according to the present embodiment, the extreme points having a convex shape are narrowed down to the extreme points having a convex shape among the intersections of the straight lines forming the polygonal waveform, and the extreme points are defined as the convex portions of the electrocardiographic waveform by the sharpness of the convex shape. Classify to baseline. For this reason, it is possible to reduce the frequency of applying signal processing for extracting or blocking a specific frequency component to classify extreme points into “convex portions” and “baseline portions” of the electrocardiogram waveform. As a result, as with a digital filter such as a high-pass filter, the detection accuracy is not affected by the filter design, for example, the setting of parameters such as the cut-off frequency, and extreme points even if noise occurs due to body movement or myoelectric signals Can be stably classified into convex portions or baseline portions.

  Therefore, according to the waveform detection apparatus 10 according to the present embodiment, the shape of the electrocardiographic waveform can be accurately detected from the input signal.

  Although the embodiments related to the disclosed apparatus have been described above, the present invention may be implemented in various different forms other than the above-described embodiments. Therefore, another embodiment included in the present invention will be described below.

[Application Example 1]
For example, in the above-described first embodiment, the case where the low-pass waveform is time-divided at equal intervals has been illustrated, but can be changed to an arbitrary division width for each time-division section. As an example, when the degree of deviation between the low-pass waveform and the polygonal line waveform is equal to or greater than a predetermined threshold, the waveform detection apparatus 10 can retry the approximation to the polygonal line by reducing the division width of the time division section.

  9 and 10 are examples of graphs showing the low-pass waveform and the broken line waveform. The vertical axis of the graphs shown in FIGS. 9 and 10 indicates the amplitude representing the intensity of the electrical stimulation, and the horizontal axis of the graph indicates time. In FIG. 9 and FIG. 10, the low-pass waveform is illustrated with a broken line, and the broken line waveform is illustrated with a solid line, as in the example described so far. In addition to this, in FIG. 9 and FIG. 10, the degree of divergence between the low-pass waveform and the broken line waveform is shown by a one-dot chain line. FIG. 9 shows a polygonal line waveform when the waveform of the low-pass signal is time-divided at equal intervals. FIG. 10 shows time-division until the divergence between the low-pass waveform and the polygonal line waveform is less than the threshold. A broken line waveform in which the division width of the section to be readjusted is illustrated.

  When the low-pass waveform is time-divided at equal intervals, there are cases where the approximation to the broken line fails at a location where the change in the low-pass waveform is large. For example, in the circled portion shown in FIG. 9, the degree of divergence between the low-pass waveform and the broken line waveform shown by the alternate long and short dash line is greater than in other locations. In this case, an inflection point appears as an R wave in the low-pass waveform, while no extreme point corresponding to the R wave appears in the broken line waveform. For this reason, the extreme point corresponding to the R wave is not detected from the polygonal line waveform, and there is a possibility that the R wave cannot be detected.

In order to suppress such a situation, first, the waveform detection apparatus 10 calculates the degree of divergence between the low-pass waveform and the broken line waveform every time a broken line waveform that approximates the low-pass waveform is generated. As an example of such a divergence degree, the divergence degree M from the low-pass waveform is given by the sum of squares of the differences using the following equation (3). In the following equation (3), “U _n ” indicates the amplitude value of the low-pass waveform, and “S _n ” indicates the amplitude value of the broken line waveform.

  Then, the waveform detection apparatus 10 determines whether or not the divergence degree M calculated by substituting the amplitude value of the low-pass waveform and the amplitude value of the broken line waveform into the above equation (3) is equal to or greater than a predetermined threshold value. judge. The threshold value to be compared with the divergence degree M can be set by the designer of the waveform detection program, or can be calculated from the accuracy of the classification requested by the designer or the user.

  At this time, if the divergence degree M is equal to or greater than the threshold value, the waveform detection apparatus 10 executes a change that reduces the division width of the time-division section to a predetermined ratio, for example, half before the change. For example, in the example of FIG. 9, since the division width is 10 samplings, half of the 5 samplings are changed to the division width. In addition, the waveform detection apparatus 10 time-divides the low-pass waveform with the changed division width. After that, the waveform detection apparatus 10 calculates the divergence degree between the low-pass waveform and the polygonal line waveform, and repeatedly executes the change of the division width and the approximation to the polygonal line until the divergence degree becomes less than the threshold value.

  As described above, when the division width of the section to be time-divided is adjusted according to the degree of deviation between the low-pass waveform and the polygonal line waveform, as shown in FIG. And we can see that the gap between the two is getting smaller.

  FIG. 11 is a flowchart illustrating the procedure of the determination process according to Application Example 1. Here, the case where the process of adjusting the division width of the section to be time-divided is incorporated into step S203 of the classification process shown in FIG. 8, but can be incorporated into step S103 of the learning process shown in FIG. .

  As illustrated in FIG. 11, the acquisition unit 11 acquires a signal in which the potential difference between the electrodes is collected in time series by the electrocardiographic sensor 20 (step S201). Subsequently, the low-pass filter 12 removes high-frequency noise by passing a low-frequency signal component of the signal acquired in step S201 (step S202).

  Then, the generation unit 13 time-divides the low-pass waveform from which the high-frequency noise has been removed in step S202 with a predetermined division width (step S301). Subsequently, the generation unit 13 generates a regression line that approximates the low-pass waveform of the section for each time-divided section (step S302). As a result, a polygonal line waveform is generated from the low-pass waveform. Subsequently, the generation unit 13 determines whether or not the degree of divergence M between the low-pass waveform and the polygonal line waveform is equal to or greater than a predetermined threshold (step S303).

  Here, when the divergence degree M is equal to or greater than the threshold (Yes in step S303), the generation unit 13 executes a change to reduce the division width of the time-division section to the predetermined ratio of the division width used in step S301. (Step S304). If the divergence degree M is less than the threshold (No in step S303), the process of step S304 is skipped and the process proceeds to step S204.

  After that, the generation unit 13 repeatedly executes the processes of steps S301 to S304 described above until the deviation degree M is less than the threshold value (step S303 Yes), and when the deviation degree M is less than the threshold value (step S303 No). The process proceeds to step S204.

  Thereafter, the calculation unit 14 determines whether there is an extreme point at which the slope of the broken line changes from positive to negative or from negative to positive at the intersection of the regression line and the regression line included in the broken line waveform generated in step S302. It is determined whether or not (step S204).

  At this time, when there is an extreme point (step S204 Yes), the calculation unit 14 calculates the feature amount of each extreme point, for example, the feature amount of (1) to (11) above (step S205). . If there is no extreme point (No in step S204), the process from step S205 to step S207 is skipped, and the process is terminated as it is. Examples of cases where such extreme points do not exist include situations where the user does not wear the electrocardiographic sensor 20, or the wearing method is inappropriate and the input signal does not include an electrocardiographic signal.

  Subsequently, the classifying unit 18 classifies the extreme points as convex portions or baseline portions by applying the determination model stored in the learning result storage unit 17 to the feature values of the extreme points calculated in step S205 ( Step S206).

  Thereafter, the classification unit 18 displays the classification result of each extreme point classified in step S206 on a display device connected to the waveform detection device 10 or outputs it to an external device (step S207). finish.

  In this way, by adjusting the division width of the time-division section according to the degree of divergence between the low-pass waveform and the polygonal line waveform, it is possible to suppress a situation where the low-pass waveform and the polygonal line waveform diverge and the extreme points are classified incorrectly.

[Application 2]
In addition, in order to reduce the influence of individual differences, the waveform detection device 10 can be configured to further include a normalization unit that normalizes the amplitude of the waveform, thereby suppressing a reduction in classification performance.

  That is, there are individual differences in the amplitude and shape of the waveform of the electrocardiogram signal. For this reason, among the feature values of extreme points shown in FIG. 5, there is a feature amount whose extreme point classification performance deteriorates when the subject changes. For example, (10) the height of the convex portion is a feature amount representing the amplitude of the waveform, and if only samples of persons with strong cardiopulmonary function are registered in the data set, the subject with normal or weak cardiopulmonary function is shown in FIG. Even if the classification tree shown is applied as it is, classification of extreme points may fail. Therefore, in order to reduce the influence of individual differences as much as possible, the degradation of the classification performance is suppressed by scaling the waveform amplitude.

  First, at the time of learning, the waveform detection apparatus 10 calculates the median convex height included in the learning data set stored in the extreme point storage unit 15 for each individual, and then calculates the individual convex height. The low-pass waveform is normalized by dividing the amplitude of the low-pass waveform of each individual using the median of. Then, the waveform detection apparatus 10 applies approximation to a polygonal line and calculation of feature values of extreme points to the normalized low-pass waveform to generate a classification tree.

  On the other hand, at the time of determination, the waveform detection apparatus 10 detects a convex portion by classifying by an extreme point feature amount for a polygonal waveform before normalization, targeting an appropriate section at the beginning of the determination target section, Calculate the median value. Thereafter, the waveform detection apparatus 10 sequentially normalizes the polygonal line waveform obtained by approximating the waveform of the signal input from the electrocardiographic sensor 20 to the polygonal line with the calculated median height of the convex portion, and the poles of the normalized polygonal line waveform Using the value points, the feature values of the extreme points are calculated and the extreme points are classified. In addition, although the case where normalization was performed using the median value of the convex part height was illustrated here, normalization may be performed using the average value of the convex part height, and other types of feature quantities Normalization may be performed using the median or average value of.

[Application Example 3]
Further, the waveform detection apparatus 10 can further classify extreme points whose classification result is a convex portion into R-wave and T-wave pairs by using the waveform characteristics peculiar to the electrocardiogram signal. That is, the waveform detection apparatus 10 utilizes the feature that the time interval of one beat of the electrocardiographic waveform is substantially the same and has periodicity. Hereinafter, the R wave and the T wave that appear in an electrocardiogram waveform of one beat may be referred to as “RT pair”.

  FIG. 12 is an example of a graph showing a low-pass waveform and a broken line waveform. The vertical axis of the graph shown in FIG. 12 indicates the amplitude representing the strength of the electrical stimulation, and the horizontal axis of the graph indicates time. In FIG. 12, similarly to the example described so far, the low-pass waveform is illustrated by a broken line, and the broken line waveform is illustrated by a solid line. In addition to this, in FIG. 12, extreme points classified into convex portions using the classification tree shown in FIG. 6 are plotted above the graph, and a whisker-like spike waveform indicates the extreme values of the convex portions. Corresponds to a point. Note that the extreme points plotted in the graph of FIG. 12 are distinguished by their line types, the section indicated by a thick line indicates that the RT pair has been determined, and the section indicated by a thin line indicates that the RT pair has not been determined. It shows that there is. In the determined interval of the RT pair, the RRI can be obtained by using the already obtained extreme point of the convex portion as a clue.

  Here, as shown in FIG. 12, when the periodicity of the electrocardiogram waveform is used, a time interval corresponding to the obtained RRI is set starting from the rightmost R wave of the determined interval of the RT pair, and the RT pair The time at which the R wave appears in the undetermined section can be estimated. Based on the estimated time, the appearance probability of the RT pair in the prediction section where the extreme point of the RT pair is predicted to appear is set to “1”. For example, the width of the prediction interval is set to a range of 10% from the estimated time of RRI for the front of the estimated time, while the range of 50% is set to the rear of the estimated time, and is linked to the RRI. Thereby, erroneous detection of RT pairs can be reduced by selecting extreme points included in the prediction interval as RT pairs.

  Next, with reference to FIG. 13, a procedure of processing for extracting RT pair candidates based on the periodicity of the electrocardiogram waveform from the extreme points classified as convex portions will be described in detail. FIG. 13 is a flowchart illustrating a procedure of RT pair candidate extraction processing according to Application Example 3. This candidate extraction process is a process executed after step S206 of the classification process shown in FIG.

  As illustrated in FIG. 13, the classification unit 18 determines whether or not the extreme point is classified as “convex” after the extreme points are classified in step S206 (step S401). At this time, if the extreme point is a “convex part” (Yes in step S401), the classification unit 18 determines that the current time, that is, the appearance time of the extreme point is the RT pair this time after the R wave appeared last time. It is further determined whether or not it has reached the “shortest appearance time” set as the shortest time that may appear as (step S402).

  If the shortest appearance time has been reached (Yes in step S402), the classification unit 18 sets the current time as the longest time that may appear as an RT pair this time after the R wave appeared last time. It is further determined whether or not the set “maximum appearance time” has passed (step S403).

  At this time, when the longest appearance time has elapsed (Yes in step S403), the classification unit 18 resets the shortest appearance time and the longest appearance time of the next RT pair from the time when the R wave appeared last time (step S404). ). Thereafter, the classification unit 18 initializes the RT pair candidate by deleting the RT pair candidates that have been held in an internal memory (not shown) (step S405).

  For example, the longest appearance time is set to 40 bpm, which is an example of the lowest value in the range measured as the human heart rate, that is, 1.5 seconds based on 40 times per minute. The shortest appearance time is set to 240 bpm, which is an example of the highest value in the range measured as the human heart rate, that is, 0.25 seconds based on 240 times per minute.

  On the other hand, when the longest appearance time has not passed (No in step S403), the classification unit 18 determines whether or not the extreme point is a convex portion first detected in the appearance range of the shortest appearance time and the longest appearance time. That is, it is determined whether or not any RT pair candidate is stored in the internal memory (step S406).

  At this time, if the extreme point is the first detected convex portion in the appearance range (step S406 Yes), the classification unit 18 extracts the extreme point as a RT pair candidate (step S409).

  On the other hand, when the extreme point is not the first convex portion detected in the appearance range (No in step S406), the classification unit 18 determines that the elapsed time elapsed from the appearance time of the extreme point first extracted as the RT pair. It is further determined whether or not the RRI half value is, for example, 750 ms (= 1.5 / 2) or less (step S407).

  If the elapsed time from the extraction of the first RT pair candidate is equal to or less than the half value of RRI (Yes in step S407), the classification unit 18 displays the appearance time of the extreme point extracted as the RT pair immediately before It is further determined whether or not the elapsed time has elapsed from a predetermined threshold value, for example, 150 ms or less (step S408).

  Here, if the elapsed time since the immediately preceding RT pair candidate is extracted is equal to or less than the threshold (Yes in step S408), the classification unit 18 extracts the extreme point as an RT pair candidate (step S409). ), The process is terminated.

  On the other hand, if the elapsed time since the first RT pair candidate is extracted is not less than the half value of RRI, or if the elapsed time since the immediately preceding RT pair candidate is extracted is not less than the threshold (step S407 No or step S408 No). ), The extreme point is not extracted as an RT pair candidate, and the process is terminated as it is.

  Next, a procedure of processing for determining an RT pair from extreme points extracted as RT pair candidates will be described with reference to FIG. FIG. 14 is a flowchart illustrating a procedure of RT pair determination processing according to the application example 3. This RT pair determination process is a process executed when the time measured by the waveform detection device 10 has passed the longest appearance time.

  As shown in FIG. 14, when there is only one RT pair candidate (step S501 Yes), the classification unit 18 classifies the extreme points of the convex portion extracted as RT pair candidates into “T waves”. In step S502, the extreme point having the highest convex height among the extreme points classified as the base line portion before the convex portion extracted as the RT pair candidate is classified as “R wave”. (Step S503).

  If there are two RT pair candidates (Yes in step S504), the classifying unit 18 classifies the extreme points of the convex portion previously extracted as RT pair candidates into “R waves” (step S504). In step S505, the extreme points of the convex portion extracted later as RT pair candidates are classified as “T waves” (step S506).

  If there are three or more RT pair candidates (No in step S504), the classification unit 18 sets the convex portion height to the top two of the extreme points of the convex portion extracted as RT pair candidates. The extreme points to be entered are extracted (step S507). Then, the classifying unit 18 classifies the extreme points of the convex portion previously extracted as RT pair candidates among the two extreme points extracted in step S507 into “R waves” (step S508), and later The extracted extreme points of the convex portions are classified as “T waves” (step S509).

  As described above, the extreme points classified into the convex portions can be further classified into the R wave and the T wave of the electrocardiogram waveform by the above RT pair candidate extraction processing and RT pair determination processing. Furthermore, since the RT pair candidates are extracted by utilizing the periodicity of the electrocardiogram signal, it is possible to filter the noise components of the convex portions that do not correspond to the R wave and the T wave. For example, in the example of FIG. 12, the extreme point detected first in the section where the RT pair is uncertain is the range where the appearance probability is not set to “1”, that is, before the shortest appearance time. Since it appears, the convex portion of the extreme point can be excluded from the RT pair candidates.

  Note that, here, a case has been illustrated in which the RRI is fixed and the appearance range of the RT pair is set, but the RRI may be variable. For example, each time an RT pair is determined, the waveform detection apparatus 10 updates RRI statistical values such as the average and median RRI of the section in which the RT pair has been determined, and uses the updated RRI statistical values. The RT pair appearance range can also be set.

[Application Example 4]
Moreover, the waveform detection apparatus 10 can improve the RT pair classification performance by using the tendency of the order of appearance of extreme types of extreme points, that is, the order of occurrence of local maxima or minima.

  15 and 16 are examples of graphs showing a low-pass waveform and a broken line waveform. The vertical axis of the graphs shown in FIGS. 15 and 16 indicates the amplitude representing the strength of the electrical stimulation, and the horizontal axis of the graph indicates time. 15 and 16 show low-pass waveforms. In addition to this, in FIG. 15 and FIG. 16, extreme points classified into convex portions using the classification tree shown in FIG. 6 are plotted above the graph, and a whisker-like spike waveform is shown in the convex portion. Corresponds to the extreme points of.

  As shown in FIG. 15 and FIG. 16, the RT pair appears in the order in which the subject detects a local extreme value of “convex downward” and then detects a local extreme value of “convex upward”. In some cases, the maximum extreme value of “convex upward” is detected, and then the maximum extreme value of “convex upward” is detected. When these “convex upward” and “convex downward” are expressed by broken line parameters, “convex upward” corresponds to the maximum where the inclination of the broken line changes from positive to negative, and “convex downward” indicates the inclination of the broken line. Corresponds to a minimum that changes from negative to positive.

  Therefore, the waveform detection device 10 automatically detects the tendency of the order in which the extreme values appear in the first time interval of the time series data of the input signal, and detects it in the subsequent time series data classification process. The classification performance can be improved by selecting extreme points that match the trend of the extreme value appearance order.

  For example, the extreme points are classified into RT pairs by the procedure using the periodicity described above. Specifically, the polarity of the broken line is checked, and 1 is added to iMaximal1 when the first convex portion is the maximum, and 1 is added to iMinimal1 when the first convex portion is the minimum. The second convex portion is processed in the same manner, and 1 is added to iMaximal2 or iMinimal2. The above process is repeated a predetermined number of times, for example, for 5 beats, and the order of appearance of the extreme values of the first and second convex parts is compared by comparing the magnitudes of iMaximal1 and iMinimal1 and the magnitudes of iMaximal2 and iMinimal2. Determine the trend.

  For example, when the above processing is applied to the two low-pass waveforms shown in FIGS. 15 and 16, the results shown in FIGS. 17 to 18 and FIGS. 19 to 20 are obtained. 17 to 20 are diagrams illustrating an example of the tendency of the appearance order of extreme values. For example, in the case of the low-pass waveform shown in FIG. 15, as shown in FIG. 17, iMinimal1 is measured 5 times in 5 beats as the extreme value of the first convex portion corresponding to the R wave, As shown in FIG. 18, iMaximal2 is measured 5 times in 5 beats as the extreme value of the second convex portion corresponding to the T wave. That is, the tendency of the appearance order of the two convex portions is not “iMaximal1> iMinimal1” and “iMaximal2> iMinimal2”, and therefore, the order of the minimum and maximum is detected. On the other hand, in the case of the low-pass waveform shown in FIG. 16, as shown in FIG. 19, iMaximal1 is measured five times in five beats as the extreme value of the first convex portion corresponding to the R wave. As shown at 20, iMaximal2 is measured five times during five beats as the extreme value of the second convex portion corresponding to the T wave. That is, the tendency of the order of appearance of the two convex portions is “iMaximal1> iMinimal1” and “iMaximal2> iMinimal2”, and therefore, the order of maximum and maximum is detected.

  Next, the procedure of the trend determination process for determining the trend of the order in which extreme values appear will be described with reference to FIG. FIG. 21 is a flowchart illustrating the procedure of the trend determination process according to Application Example 4. As illustrated in FIG. 21, when the R wave of the RT pair is “maximum” (Yes in step S <b> 601), the waveform detection device 10 determines the appearance order of RTs and the types of extreme values illustrated in FIGS. 17 and 18. In the correspondence table, iMaximal1 is incremented by one (step S602). On the other hand, when the R wave of the RT pair is “minimum” (No in step S601), the waveform detection apparatus 10 increments the number of iMinimal1 measurements in the correspondence table by one (step S603).

  On the other hand, when the T wave of the RT pair is “maximum” (step S604 Yes), the waveform detection apparatus 10 increments the number of iMaximal2 measurements by one in the correspondence table (step S605). On the other hand, when the T wave of the RT pair is “minimum” (No in step S604), the waveform detection apparatus 10 increments the number of iMinimal2 measurements in the correspondence table by one (step S606).

  Then, until the appearance order of the extreme value of the RT pair is measured a predetermined number of times, for example, five times (No in step S607), the waveform detection apparatus 10 repeatedly executes the processing from step S601 to step S606 described above.

  Thereafter, when the appearance order of the extreme values of the RT pair is measured a predetermined number of times (Yes in step S607), the waveform detection device 10 executes the following process. That is, when the measurement result is “iMaximal1> iMinimal1” (step S608 Yes), the waveform detection apparatus 10 sets the extreme value pattern of the R wave to “maximum” (step S609). On the other hand, when the measurement result is not “iMaximal1> iMinimal1” (No in step S608), the extreme value pattern of the R wave is set to “minimum” (step S610).

  Further, when the measurement result is “iMaximal2> iMinimal2” (step S611 Yes), the waveform detection apparatus 10 sets the extreme value pattern of the T wave to “maximum” (step S612). On the other hand, if the measurement result is not “iMaximal2> iMinimal2” (No in step S611), the T-wave extreme value pattern is set to “minimal” (step S613).

  In this way, by determining the trend of the extreme value appearance order of the RT pair according to the flowchart shown in FIG. 21, extreme points that do not match the trend of the extreme value appearance order may be excluded from the RT pair candidates. it can. For example, in the example of FIG. 12, since the extreme value patterns of the R wave and the T wave are both set to “maximum”, the extreme value point detected third in the interval in which the RT pair is not determined. Since the extreme value is “minimum”, the third extreme point can be excluded from the RT pair candidates.

[Accuracy rate]
FIG. 22 is an example of a graph showing the accuracy rate. The histogram of (a) shown in FIG. 22 shows the accuracy rate when the first embodiment is used, that is, when a classification tree is used. The histogram of (b) shown in FIG. 22 shows the accuracy rate when Application Example 2 is used, that is, when the height of the convex portion is normalized. The histogram of (c) shown in FIG. 22 illustrates the accuracy rate when the application example 3 is used, that is, when extreme points are classified into RT pairs using the periodicity of the electrocardiogram waveform. The histogram of (d) shown in FIG. 22 shows the accuracy rate when the application example 3 and the application example 4 are used in combination, that is, when the tendency of the extreme value appearance order is used for the RT pair classification. Yes. Further, the histogram of (e) shown in FIG. 22 illustrates the accuracy rate when the RRI interval is variable when combining the application example 3 and the application example 4. As shown in FIG. 22, even when only the classification tree is used, a high accuracy rate of “93.9%” can be exhibited. However, the application example 2, the application example 3, The accuracy rate can be improved each time Application Example 4 is combined, and finally the accuracy rate can be improved to “99.1%”. In the first embodiment and the first to fourth application examples, the waveform detection device 10 can be mounted in any combination.

[Scope of application]
In the first embodiment and the first to fourth application examples described above, the case where the input signal collected by the sensor is an electrocardiogram signal is exemplified. However, the waveform detection apparatus 10 also when the input signal is other than the electrocardiogram signal. Within the scope of application.

  For example, “stationary state” and “walking state” can be discriminated from signals collected by an acceleration sensor or a gyro sensor. For example, an acceleration sensor is attached to the user, and the electrocardiographic sensor 20 shown in FIG. 1 is replaced with an acceleration sensor to output an acceleration signal. For example, a three-axis (X, Y, Z) acceleration sensor may be used as the acceleration signal, and the magnitude of a combined vector of three-direction vectors or the magnitude of acceleration on the vertical axis may be used. The acceleration sensor converts an acceleration signal from an analog value to a digital value, approximates with a broken line in the generation unit according to the block diagram shown in FIG. 1, and calculates a feature amount in the calculation unit.

  Here, as for the acceleration during walking, convex waves appear upward and convex downward in a sine wave shape. Therefore, in the classification of the stationary state and the walking state, it is assumed that the stationary state is classified as a base line portion and the walking state is classified as a convex portion. In other words, since there is almost no change in acceleration in a stationary state, the height of the convex portion of the feature value of the calculated extreme value point is small, and the inclination of the broken line tends to be small. On the other hand, in the walking state, the height of the convex portion of the feature value of the extreme value point and the inclination of the broken line are large. Therefore, it is easy to discriminate the extreme value included in the stationary time interval from the extreme value included in the walking time interval.

  FIG. 23 is an example of a graph showing signal waveforms of the acceleration sensor. FIG. 24 is a diagram illustrating an example of a classification tree. The vertical axis of the graph shown in FIG. 23 indicates the amplitude representing the strength of the electrical stimulation, and the horizontal axis of the graph indicates time. FIG. 23 shows a broken line waveform of acceleration when changing from a stationary state, a walking state, and a stationary state. FIG. 24 illustrates a classification tree obtained by calculating the feature amount of the extreme point with respect to the waveform illustrated in FIG. 23 and obtained by the learning unit. When acceleration signals are input to the classification tree shown in FIG. 24, the extreme points included in the base line portion and the extreme points included in the convex portion can be classified by the determinations in step S21 and step S22, and the time between the stationary state and the walking state can be classified. It can be identified as a section.

  As described above, the waveform detection apparatus 10 can be generally applied to a process of classifying a signal waveform into a “base line portion” that is a relatively flat portion and a characteristic wave portion “convex portion”.

[Distribution and integration]
In addition, each component of each illustrated apparatus does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured. For example, the acquisition unit 11, the low-pass filter 12, the generation unit 13, the calculation unit 14, the learning unit 16, or the classification unit 18 may be connected as an external device of the waveform detection device 10 via a network. Further, the waveform detection device 10 described above is obtained by having the acquisition unit 11, the low-pass filter 12, the generation unit 13, the calculation unit 14, the learning unit 16, or the classification unit 18 connected to each other through a network connection. You may make it implement | achieve the function of.

[Waveform detection program]
The various processes described in the above embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. In the following, an example of a computer that executes a waveform detection program having the same function as that of the above embodiment will be described with reference to FIG.

  FIG. 25 is a schematic diagram illustrating an example of a computer that executes a waveform detection program according to the first and second embodiments. As illustrated in FIG. 25, the computer 100 includes an operation unit 110a, a speaker 110b, a camera 110c, a display 120, and a communication unit 130. Further, the computer 100 includes a CPU 150, a ROM 160, an HDD 170, and a RAM 180. These units 110 to 180 are connected via a bus 140.

  As shown in FIG. 25, the HDD 170 stores in advance a waveform detection program 170a that exhibits the same functions as those of the acquisition unit 11, the generation unit 13, the calculation unit 14, and the classification unit 18 described in the first embodiment. . The waveform detection program 170a may be integrated or separated as appropriate, like the components of the acquisition unit 11, the generation unit 13, the calculation unit 14, and the classification unit 18 illustrated in FIG. In other words, all data stored in the HDD 170 need not always be stored in the HDD 170, and only data necessary for processing may be stored in the HDD 170.

  Then, the CPU 150 reads the waveform detection program 170 a from the HDD 170 and develops it in the RAM 180. Thus, as shown in FIG. 25, the waveform detection program 170a functions as a waveform detection process 180a. The waveform detection process 180a appropriately expands various data read from the HDD 170 in an area allocated to itself on the RAM 180, and executes various processes based on the expanded data. The waveform detection process 180a is performed by the acquisition unit 11, the generation unit 13, the calculation unit 14, and the classification unit 18 illustrated in FIG. 1, for example, FIG. 7 to FIG. 8, FIG. And the processing shown in FIG. In addition, each processing unit virtually realized on the CPU 150 does not always require that all processing units operate on the CPU 150, and only a processing unit necessary for the processing needs to be virtually realized.

  Note that the waveform detection program 170a is not necessarily stored in the HDD 170 or the ROM 160 from the beginning. For example, each program is stored in a “portable physical medium” such as a flexible disk inserted into the computer 100, so-called FD, CD-ROM, DVD disk, magneto-optical disk, or IC card. Then, the computer 100 may acquire and execute each program from these portable physical media. In addition, each program is stored in another computer or server device connected to the computer 100 via a public line, the Internet, a LAN, a WAN, etc., and the computer 100 acquires and executes each program from these. It may be.

DESCRIPTION OF SYMBOLS 10 Waveform detection apparatus 11 Acquisition part 12 Low pass filter 13 Generation part 14 Calculation part 15 Extreme point storage part 16 Learning part 17 Learning result storage part 18 Classification part

Claims (7)

An acquisition unit for acquiring a signal collected by the sensor;
A generator that generates a waveform of a polygonal line that approximates the waveform of the signal;
A calculation unit that calculates a feature amount of an extreme point at which the positive / negative of the slope of a straight line adjacent to each other among the intersections of the straight lines forming the broken line is inverted;
A waveform detection apparatus comprising: a classification unit that classifies the extreme value points into convex portions having a convex shape or flat base line portions using feature amounts of the extreme value points included in the broken line. The generator is
Time-sharing the waveform of the signal,
Generate a polygonal waveform by generating a regression line that approximates the waveform of the section for each time-divided section,
Calculate the degree of divergence between the waveform of the signal and the waveform of the broken line,
2. The waveform detection apparatus according to claim 1, wherein a process of generating a polygonal line waveform by reducing a division width of a time-division section is repeatedly performed until the degree of deviation is less than a predetermined threshold. A normalization unit that normalizes the waveform of the polygonal line generated by the generation unit using the feature amount of the extreme point calculated by the calculation unit;
3. The waveform detection apparatus according to claim 1, wherein the calculation unit calculates a feature amount of the extreme value point from a broken line waveform normalized by the normalization unit. The classification unit includes:
Using the time interval for one ECG waveform estimated from the appearance time of the extreme points classified as the convex portion, the appearance range in which the R wave and T wave included in the ECG waveform can appear is predicted. ,
Extracting extreme points whose appearance time is within the appearance range among the extreme points classified into the convex portions as candidates for the R wave and the T wave,
The extreme points extracted as candidates are further classified into the R wave and the T wave based on the order of appearance of extreme points extracted as candidates for the R wave and the T wave. The waveform detection apparatus according to claim 1. A determining unit that determines a tendency of the order in which the extreme values of the R wave and the T wave appear in order from the classification result of the R wave and the T wave by the classification unit;
The classification unit includes:
The R wave and T wave candidates are extracted by excluding extreme points having extreme values that do not match the tendency determined by the determining unit from the extreme points classified into the convex portions. The waveform detection apparatus according to claim 4. Computer
Get the signal collected by the sensor,
Generate a polygonal waveform that approximates the waveform of the signal,
Calculate the feature amount of the extreme point where the positive / negative of the slope of the straight lines adjacent to each other among the intersections of the straight lines forming the broken line is inverted,
A waveform detection method comprising: performing each process of classifying an extreme point into a convex portion having a convex shape or a flat base line portion using a feature amount of the extreme point included in the broken line. On the computer,
Get the signal collected by the sensor,
Generate a polygonal waveform that approximates the waveform of the signal,
Calculate the feature amount of the extreme point where the positive / negative of the slope of the straight lines adjacent to each other among the intersections of the straight lines forming the broken line is inverted,
A waveform detection program for executing each processing for classifying an extreme point into a convex portion having a convex shape or a flat base line portion using a feature amount of the extreme point included in the broken line.

Images