JP2010187993A - Blood pressure measuring device and program as well as recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a blood pressure measuring device which can estimate blood pressure with high accuracy by carrying out good learning to blood pressure for which truth or falsehood is ambiguous, a program, and a recording medium. <P>SOLUTION: In a step 110, a pulse wave signal and an electrocardiogram signal are analyzed and a characteristic amount required for blood pressure measurement is computed. In a step 130, two or more binary discrimination are performed. Namely, the characteristic amount is used for a discrimination function established for each blood pressure for every 1 mmHg, being under or equal to above a blood pressure value for the determination subject is determined, and a binary series, which has the number of digits corresponding to the number of blood pressure for the determination subject, is created. In a step 140, the binary series and a code table, which are created in the step 130, are compared, Hamming distances are calculated, and a blood pressure value corresponding to the minimum Hamming distance is chosen. In a step 150, the determination of reliability is performed by the number of learning data. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a blood pressure measurement device, a program, and a recording medium that are mounted on a vehicle and can measure blood pressure such as a driver without using a cuff or the like.

  Conventionally, when measuring blood pressure, both the auscultation method and the oscillometric method require tightening with a cuff, the device becomes larger, the subject is suffering, and continuous blood pressure measurement is required. There were problems such as being unable to do so.

  On the other hand, in recent years, measurement methods for calculating blood pressure using a pulse wave velocity, a pulse wave feature amount, and the like have been studied (see Patent Documents 1 to 10). For example, volume pulse waves can be measured using light, so that the apparatus can be miniaturized, the pain of the measurement subject can be eliminated, and continuous blood pressure measurement can be performed.

  Further, as a technique for estimating blood pressure, learning using multiple regression, neural network (NN), and the like have been proposed (see Patent Documents 11 and 12).

JP-A-10-295656 Japanese Patent Laid-Open No. 10-295657 Japanese Patent Laid-Open No. 11-318837 JP-T-2001-504362 Special table 2003-555 gazette Special table 2003-527149 gazette JP 2006-263354 A JP 2006-6897 A JP 2007-7075 A JP 2007-7077 A JP 2000-126142 A JP 2000-166885 A

  The techniques of Patent Documents 1 to 12 described above use, for example, the feature amount of a pulse wave or an electrocardiogram signal (for example, the second derivative of the pulse wave, the pulse wave velocity (PTT), the heart rate cycle (RR)), etc. Alternatively, blood pressure is estimated using multiple regression or NN techniques, but this is not always sufficient.

  In multiple regression, since estimation is performed linearly, estimation accuracy is limited, and learning by NN takes a long calculation time, and the learning data becomes a true value. There was a problem that it was difficult to apply to estimation.

That is, in blood pressure estimation, when learning a value measured with a commercially available cuff sphygmomanometer as a true value, the measurement accuracy (standard deviation of error) of the cuff sphygmomanometer varies from 4 to 8 mmHg. Based on this, we had to learn, and the estimation accuracy had been reduced. (Reference: JIS T 1115 (Non-invasive electronic blood pressure monitor), http://www.healthcare.omron.co.jp/medical/product/technique/pdf/technique_nO1.pdf)
In view of the above points, an object of the present invention is to provide a blood pressure measurement device, a program, and a recording medium that can estimate blood pressure with high accuracy with respect to blood pressure whose measurement accuracy is likely to vary.

  (1) The invention of claim 1 is a blood pressure measurement apparatus that estimates blood pressure using at least biological information obtained from a pulse wave signal among electrocardiogram signals and pulse wave signals detected from a living body. Learning means for obtaining a binarized determination of blood pressure from the feature value of the biological information for each predetermined blood pressure, which is obtained by previously learning the relationship between the feature value of the biological information and the blood pressure. And binarization for performing binarization determination on the predetermined blood pressure based on the feature amount of the biological information obtained by measurement using the learning identification unit when estimating the blood pressure. The determination unit and the binarization determination unit respectively perform binarization determination on the different predetermined blood pressures based on the feature amount of the biological information, and based on the results of the multiple binarization determinations Blood pressure to estimate the blood pressure Characterized by comprising a constant means.

  In the present invention, the blood pressure is calculated from the feature value of the biometric information for each predetermined blood pressure (for example, every 1 mmHg such as 100 mmHg or 101 mmHg) by machine learning the relationship between the feature value of the biometric information that is learning data and the blood pressure in advance. Learning discriminating means (for example, discriminant function: evaluation function: discriminant function) is generated. Next, when the blood pressure is actually estimated, the learning identification unit is used to determine a predetermined blood pressure (for example, every 1 mmHg such as 100 mmHg or 101 mmHg) based on the feature amount of the biological information obtained by the measurement. Then, binarization determination is performed respectively. Then, the blood pressure is estimated based on a plurality of determination data obtained as a result.

  In this way, in the present invention, binarization determination is performed for each predetermined blood pressure, and blood pressure is estimated based on the data of a plurality of times. Therefore, the true value is ambiguous compared to the conventional method (true By binarizing the blood pressure (which includes an error), the influence of ambiguous data is reduced, and the blood pressure can be estimated with high accuracy.

  Here, the feature amount is a characteristic value indicating a signal state obtained from an electrocardiogram signal or a characteristic value indicating a signal state obtained from a pulse wave signal (accordingly, an index indicating the state of a living body). Or a characteristic value that serves as an index indicating the state of a living body obtained from an electrocardiogram signal and a pulse wave signal. For example, one or a plurality of well-known feature values such as the following are used. it can.

  Time difference between electrocardiogram signal and pulse wave signal (PTT: see FIG. 3), heart rate (HR), heart rate interval (RRI), pulse wave AI (Augmentation Index: reflected wave (P2) in pulse wave and traveling wave ( P1) magnitude ratio (P2 / P1): see FIG. 4), velocity pulse wave (first-order derivative) feature quantities a1 to f1 (see FIG. 4), acceleration pulse wave (second-order derivative) feature quantity a ˜f (a wave indicating initial wave peak along time axis (contraction early positive wave), b wave (initial contraction negative wave), c wave (contraction middle re-rising wave), d wave (late systolic re-falling wave) E wave (extended initial positive wave), f wave (extended initial negative wave): see FIG.

  As the feature amount, in addition to the above-described feature amount, a body feature amount (a value indicating a feature of the body: for example, height, weight, sex, age), a biological impedance, etc. It can also be used to estimate the blood pressure.

  (2) The invention of claim 2 is characterized in that the learning identification means generates an identification function for binarizing the blood pressure by adaboost.

  In the present invention, the blood pressure binarization determination is performed using a known Adaboost technique which is a kind of boosting.

  This boosting is a method of constructing a high-performance learning machine by combining a number of learning machines (identifiers) with weak performance, and AdaBoost, the most common boosting algorithm, is not good at it. It is an algorithm that learns by applying weights that are updated exponentially to examples, and gradually increases the ability, and because it can obtain high general-purpose performance with a very simple procedure, it is currently widely used Yes.

  As for the method of Adaboost, JP 2008-9843 A, JP 2007-307358 A, JP 2008-140230 A, JP 2008-117391 A, Mathematical Analysis Laboratory, Vol. 1439, 2005. 111-127 and the like.

  (3) The invention of claim 3 creates a binary number sequence in which the determination results for each blood pressure are arranged from the result of the plurality of binarization determinations, and in the code table prepared in advance with this binary number sequence The blood pressure is estimated by comparing with a sequence for each blood pressure.

  The present invention relates to a blood pressure estimation method (for example, an ECCC (Error Correcting Output Coding) method: TG Dietterich and G. Bakiri., “Solving multiclass learning problems via error correcting output codes”, Journal of Artificial Intelligence Research, 2: 265-286, 1995))). In the present invention, a binary number sequence obtained from the result of multiple binarization determinations is compared with a number sequence for each blood pressure (for example, a binary number sequence) in the code table, and the degree of coincidence of the corresponding digits. From this, blood pressure can be estimated. In addition, in order to compare each digit of both number sequences, the number of digits of each number sequence is the same.

  For example, as shown in FIG. 7A, the code table is a numerical value of a predetermined digit of a numerical sequence set for each blood pressure gradually changes (for example, one digit) as the blood pressure increases (or decreases). Increments from 0 to 1). As the number sequence of the code table, the number sequences of “0” and “1” are conceivable. However, for weighting, for example, as shown in FIG.

  (4) According to the invention of claim 4, the minimum Hamming distance in both number sequences is obtained from the binary number sequence obtained from the result of the plurality of binarization determinations and the binary number sequence of the code table, and the minimum Hamming The blood pressure is estimated based on the distance.

  The present invention exemplifies a specific method using a code table. For example, if the numerical values of the corresponding digits in both number sequences are the same, it is 0, and if they are different, it is 1 and the determination result of each digit is added to obtain the Hamming distance, and the minimum value can be obtained as the minimum Hamming distance.

  Since it is considered that the blood pressure corresponding to the minimum hamming distance has high estimation accuracy, the present invention estimates the blood pressure based on the minimum hamming distance.

  (5) The invention of claim 5 is characterized in that the numerical sequences of the code table are set so that the numerical values are different with a plurality of widths (see Example 3).

  For example, as shown in FIG. 11, in the case where the numerical value of each digit in each numerical sequence of the code table is increased for each digit, the blood pressure estimation is more likely to occur when the digit number is increased by a plurality of digits (for example, two digits). The calculation burden is reduced (calculation time is shortened).

  (6) The invention of claim 6 is characterized in that a code table of a numerical sequence set in accordance with the condition of the subject is used as the code table (see Example 4).

  For example, when the subject is elderly, there is a low possibility that the blood pressure is low. In this case, as shown in FIG. 12, for example, a code table without a sequence for determining low blood pressure can be used. As a result, only high blood pressure is determined, so that the calculation time is shortened.

  (7) The invention of claim 7 is characterized in that the determination of binarization of the blood pressure is performed for each group of a predetermined blood pressure range (see Example 5).

  It is said that a commercially available cuff sphygmomanometer that is a true value in learning usually has an error of 4 to 8 mmHg. Therefore, for example, instead of performing binarization determination for each fine blood pressure such as 1 mmHg, binarization determination may be performed for each group (step) of 4 to 8 mmHg, for example. This also provides a certain degree of measurement accuracy and shortens the calculation time.

  (8) The invention of claim 8 is characterized in that the width of the group is narrowed within the range of the predetermined blood pressure, and the width of the group is widened outside the range of the predetermined blood pressure (see Example 6). ).

  The learning data is concentrated in an average blood pressure range (central portion of data: for example, 100 to 140 mmHg), and the number of data in this portion is large, and the number of data in other portions is small.

  Therefore, the range of blood pressure determination in a region with a large amount of learning data is narrowed (for example, 1 mmHg), and the range of blood pressure determination in a region with a small amount of learning data is widened (for example, 4 to 8 mmHg).

  Thereby, shortening of calculation time and ambiguity of reference blood pressure (learning data) can be reduced.

  In addition, it is preferable to set the predetermined blood pressure range to 100 to 140 mmHg because the calculation time can be shortened while ensuring the measurement accuracy. Further, when the width of the group is narrowed, it may be 2 mmHg or more, but may be 1 mmHg as a minimum unit.

  (9) The invention of claim 9 is characterized in that learning data in a predetermined blood pressure range is not used for learning in the prior learning in the learning identification means (see Example 7).

  Since a commercially available cuff sphygmomanometer usually has an error of 4 to 8 mmHg, in consideration of this, data within a predetermined range (for example, within ± 4 mmHg) is used from the blood pressure value of binarized learning data. It is not to be. This improves blood pressure estimation accuracy.

  (10) The invention of claim 10 is characterized in that the ratio of the number of binarized data in a range of different blood pressures is set within a predetermined range at the time of pre-learning in the learning discriminating means. (See Example 8).

  If the number of learning data in each blood pressure layer is significantly different, even if there is learning data with the same feature amount in the blood pressure to which the pulse wave originally belongs, the same applies to blood pressure to which the pulse wave does not originally belong. The probability that there exists learning data having a feature amount increases, and as a result, blood pressure estimation accuracy may decrease.

  Therefore, in the present invention, the ratio of the number of learning data in each blood pressure layer (for example, each blood pressure layer divided into less than 140 mmHg) divided into two by a predetermined blood pressure value for binarization determination is calculated. , Set within a predetermined range (for example, within 2 times). As a result, the probability that there is data having a similar feature amount in blood pressure that does not originally belong is reduced, and blood pressure measurement accuracy is improved.

  (11) The invention of claim 11 is characterized in that when the blood pressure is estimated, the reliability of the data is calculated, and the reliability of the data is displayed together with the blood pressure estimation result.

  The reliability (degree of reliability) of the estimated value of blood pressure varies depending on the number of learning data and the like. Therefore, accurate information can be conveyed to the measurer and the person to be measured by displaying the result together with the reliability.

  (12) The invention of claim 12 calculates the reliability of the data, and when the reliability is lower than a predetermined determination value, adopts a different method that is expected to have higher reliability, It is characterized by performing estimation.

  For example, the reliability is set according to the number of learning data. For example, when the number of learning data is greater than or equal to a predetermined value, the reliability is set high, and when the number is less than the predetermined value, the reliability is set low.

  Therefore, it is possible to determine whether or not to adopt a different method according to the reliability.

  (13) The invention of claim 13 is characterized in that the reliability of the data is determined based on the number of learning data or the minimum Hamming distance.

  The present invention exemplifies the basis of data reliability. For example, when the learning data is small, the reliability of the estimated blood pressure value is low because the learning data is insufficient, and the smaller the minimum Hamming distance, the higher the degree of coincidence with the code table and the higher the blood pressure reliability. it is conceivable that. Therefore, the reliability of blood pressure can be determined based on the number of learning data and the minimum Hamming distance.

  (14) The invention of claim 14 is characterized in that the different method is a method of estimation by multiple regression.

  The present invention illustrates different approaches. The estimated value of blood pressure when the number of learning data is small is considered to be less accurate than the estimated value of blood pressure by multiple regression. In other words, in multiple regression, a certain value is extrapolated and estimated using all data, so that a certain degree of accuracy can be ensured even in an area where the number of data is small. The accuracy decreases in a region where the number is small. Therefore, when the number of learning data is less than a predetermined value, the accuracy of blood pressure estimation is improved by adopting a technique based on multiple regression.

  (15) The invention of claim 15 is the blood pressure measurement apparatus for estimating blood pressure using at least biological information obtained from the pulse wave signal among the electrocardiogram signal and the pulse wave signal detected from the living body. Based on the feature quantity of the information, the first blood pressure estimation means for estimating the blood pressure by multiple regression and the discrimination means obtained by learning the relationship between the feature quantity of the biological information and the blood pressure, the blood pressure is calculated from the feature quantity. The second blood pressure estimating means to be estimated, the first estimated blood pressure estimated by the first blood pressure estimating means, and the second estimated blood pressure estimated by the second blood pressure estimating means are used to determine whether the predetermined reliability is high or low. Blood pressure determining means for determining an estimated blood pressure having a high degree of reliability by comparison based on the determination condition.

  In the present invention, a first blood pressure estimation unit that estimates blood pressure by multiple regression based on a feature amount of biological information, and a determination unit obtained by learning a relationship between the feature amount of biological information and blood pressure (for example, the learning identification) And the second blood pressure estimating means for estimating the blood pressure from the feature amount, the highly reliable blood pressure can be estimated by adopting the one with higher reliability by the determination.

  As the second blood pressure estimation means, in addition to the learning identification means by Adaboost, other machine learners such as a neural network (NN: Neural Network) described in Patent Documents 11 and 12, the SVM (Support Vector Machine), etc. A means for estimating blood pressure by the Ing technique can be employed.

  As the determination conditions for determining the level of the predetermined reliability, the number of learning data, the blood pressure range, the pulse wave feature amount range, the minimum Hamming distance (by the method of claim 1), and the like can be adopted.

  For example, when the learning data is small, it is considered that the reliability of the estimated blood pressure is low. Therefore, when the number of learning data is less than a predetermined value, the blood pressure obtained by multiple regression can be employed.

  As the feature amount, in addition to the feature amount obtained from the above-described electrocardiogram signal or pulse wave signal, a body feature amount that is a feature amount of biological information (a value indicating a feature of the body: for example, height, weight, sex, etc. , Age), biological impedance, and the like can also be used for the estimation of the blood pressure.

  (16) In the invention of claim 16, the second blood pressure estimation means is an identification means obtained by machine learning in advance of the relationship between the characteristic amount of the biological information, which is learning data, and blood pressure, Learning identification means for binarizing and discriminating blood pressure from the feature amount of the biological information for each blood pressure, and the feature amount of the biological information obtained by measurement using the learning identification means when estimating the blood pressure And binarization determination means for performing binarization determination on the predetermined blood pressure, and the binarization determination means for each of the predetermined blood pressures that differ based on the feature amount of the biological information. Blood pressure estimation means for performing binarization determination and estimating the blood pressure based on the result of the plurality of binarization determinations.

  The present invention exemplifies the second blood pressure estimating means. Therefore, the invention of claims 1 to 13 can be employed as the second blood pressure estimating means.

  (17) The invention of claim 17 is characterized in that, when the first estimated blood pressure obtained by the multiple regression is 100 to 140 mmHg, the second estimated blood pressure by the learning is adopted as an estimated value of blood pressure. To do.

  The blood pressure of 100 to 140 mmHg is a range of normal blood pressure, and it is considered that the number of learning data is large. In this case, an estimated value of blood pressure by learning is adopted.

  (18) The invention of claim 18 is characterized in that when the learning data is 5 or more per 1 mmHg, the second estimated blood pressure by the learning is adopted as an estimated value of blood pressure.

  The present invention exemplifies a preferable range of the number of learning data.

  (19) The invention of claim 19 is a program for realizing the function of the blood pressure measurement device according to any one of claims 1 to 18.

  Therefore, by executing this program with a microcomputer or the like, it is possible to estimate the blood pressure using the biological information described above.

  (20) A twentieth aspect of the invention is a computer-readable recording medium on which the program according to the nineteenth aspect is recorded.

  Therefore, using the program recorded on this recording medium, the blood pressure can be estimated as described above.

It is explanatory drawing which shows the system configuration | structure of the blood-pressure measurement apparatus in a vehicle. It is a block diagram which shows functions, such as a blood pressure measuring device It is a graph which shows an electrocardiogram signal and a pulse wave signal. It is a graph which shows a pulse wave signal and the signal which differentiated it. It is explanatory drawing which shows the procedure of the blood-pressure estimation in Example 1. FIG. It is explanatory drawing which shows the method of the binarization discrimination | determination of a blood pressure. (A) is explanatory drawing which shows a code table without weighting, (b) is explanatory drawing which shows a code table with weighting. 3 is a flowchart illustrating a blood pressure estimation process according to the first embodiment. It is explanatory drawing which shows the procedure of the blood-pressure estimation in Example 2. FIG. 10 is a flowchart illustrating a blood pressure estimation process in the second embodiment. It is explanatory drawing which shows the code | cord | chord table etc. which are used in Example 3. FIG. It is explanatory drawing which shows the code | cord | chord table etc. which are used in Example 4. FIG. It is explanatory drawing which shows the experimental result of Experimental example 1 (comparison of AdaBoost and multiple regression). It is explanatory drawing which shows the experimental result of Experimental example 2 (comparison of the estimation precision of each blood-pressure layer). It is explanatory drawing which shows the experimental result of Experimental example 3 (blood pressure estimation by a group). It is explanatory drawing which shows the experimental result of Experimental example 4 (minimum hamming distance and error). It is explanatory drawing which shows the experimental result of Experimental example 5 (effect of selection of learning data). It is explanatory drawing which shows the experimental result of Experimental example 6 (effect of learning data ratio).

  Next, an embodiment of the blood pressure measurement device of the present invention will be described.

  In the present embodiment, a blood pressure measurement device that is mounted on an automobile and measures the blood pressure of a driver will be described.

  a) First, the system configuration of the blood pressure measurement device according to the present embodiment will be described.

  As shown in FIG. 1, in this embodiment, a blood pressure measurement device 1 mounted on an automobile, a notification device 3 that notifies a driver or the like, a manual input unit 5 that manually inputs data, A pulse wave sensor 9 attached to the steering 7 and a pair of electrodes 11 and 13 attached to the steering 7 are provided.

  The blood pressure measurement device 1 is an electronic control device centered on a well-known microcomputer, and calculates blood pressure, controls the notification device 3, and the like based on signals from the pulse wave sensor 9, the electrodes 11, 13, and the like. .

  The notification device 3 includes a display 15 such as a liquid crystal that displays blood pressure and the like, and a speaker 17 that outputs the contents thereof by voice or the like.

  The manual input unit 5 is an input device such as a keyboard, a numeric keypad, or a remote controller capable of inputting individual body characteristic information such as weight, height, age, and sex manually. Note that data may be input using the display screen of the display 15 as a touch panel.

  The pulse wave sensor 9 is an optical sensor provided with a known light emitting element (LED) or light receiving element (PD). For example, the pulse wave sensor 9 irradiates light on a fingertip of a driver and uses the reflected wave to make a pulse wave. (Volume pulse wave) can be detected. Therefore, a pulse wave signal used for blood pressure calculation described later can be obtained from the pulse wave sensor 9.

  The pair of electrodes 11 and 13 are arranged on the left and right of the steering wheel 9 so as to come into contact with the left and right hands of the driver, respectively. These electrodes 11 and 13 are used as electrodes of an electrocardiograph that applies a voltage to obtain an electrocardiographic signal. Here, both the electrodes 11 and 13 and the blood pressure measuring device 1 serve as an electrocardiograph. Therefore, an electrocardiographic signal used for blood pressure calculation described later can be obtained using the electrodes 11 and 13.

  b) Next, the function of the blood pressure measurement device 1 and the like of this embodiment will be described in more detail with reference to a block diagram.

  As shown in FIG. 2, the blood pressure measurement device 1 of the present embodiment includes an electrocardiogram signal acquisition unit 25, a pulse wave signal acquisition unit 27, an electrocardiogram signal analysis unit 31, and an electrocardiogram and pulse wave signal analysis unit. 33, a pulse wave signal analysis unit 35, and a blood pressure calculation unit 37.

  Among these, the electrocardiogram signal acquisition unit 25 measures electrical excitement accompanying the activity of the heart as a potential difference (electrocardiogram signal) between the electrodes 11 and 13.

  The pulse wave signal acquisition unit 27 drives the pulse wave sensor 9 to acquire a pulse wave signal.

  The electrocardiogram signal analysis unit 31 analyzes the electrocardiogram signal and calculates, for example, a heartbeat interval (HR).

  As shown in FIG. 3, the electrocardiogram and pulse wave signal analysis unit 33 uses the electrocardiogram signal and the pulse wave signal to obtain a pulse wave propagation time (PTT) that is a delay time of the pulse wave signal with respect to the electrocardiogram signal. Is.

  As shown in FIG. 4, the pulse wave signal analysis unit 35 analyzes the pulse wave signal and performs first-order differentiation (velocity pulse wave), second-order differentiation (acceleration pulse wave), third-order differentiation, and fourth-order differentiation. At the same time, feature amounts (for example, velocity pulse waves a1 to k1, acceleration pulse waves a to f, etc.) and AI (volume AI) are calculated.

  The blood pressure calculation unit 37 obtains at least a feature amount obtained from the pulse wave signal among the pulse wave signal and the electrocardiogram signal as biological information, for example, a feature amount such as PTT, HR, velocity pulse wave, acceleration pulse wave, volume AI, and the like. The blood pressure is estimated (calculated) by learning, multiple regression, etc., which will be described later.

  As will be described later, an identification function for performing binarization determination of the blood pressure is generated and stored in advance in the blood pressure calculation unit 37 (for each predetermined blood pressure).

  b) Next, the procedure of blood pressure measurement in the present embodiment will be described.

  In the present embodiment, the blood pressure binarization is determined using a well-known Adaboost technique which is a kind of boosting.

  (1) First, the pre-learning mode in AdaBoost will be described.

  Here, a discriminant function for performing binarization determination of blood pressure is created by the Adaboost method. In other words, an identification function is created (to determine whether the feature value obtained during actual blood pressure measurement is less than or equal to the predetermined blood pressure) by learning in advance the feature value data when the blood pressure is known To do.

  As described below, the procedure for generating the discriminant function by Adaboost is disclosed in, for example, Research Institute of Mathematical Analysis Vol. 1439 2005 111-127.

  Hereinafter, a method for generating the discriminant function (F) will be described in detail.

  First, learning samples (x1, y1),..., (XN, yN) that are correctly identified as positive samples or negative samples from sample data with known blood pressure are prepared.

  For example, if a blood pressure (for example, 100 mmHg) is used as a boundary and less than 100 is divided into 100 or more, and a case where the blood pressure is less than 100 is mistaken and 100 or more is considered positive, a positive sample (correct answer) and a negative sample (incorrect Learning samples (x1, y2)... (XN, yN) with correct answers are prepared and learning is performed using the sample data.

  Note that xi represents learning sample data, that is, the plurality of feature amounts described above. Further, yi is a label of the learning sample, yi = 1 indicates a case where the learning sample is labeled as a positive sample (a sample indicating that the blood pressure is equal to or higher than a predetermined value: positive), and yi = −1 , Indicates that the learning sample is labeled as a sample indicating that it is a negative sample (for example, blood pressure is not greater than or equal to a predetermined value (ie, less than a predetermined value): false).

  Here, when the defining function is the following formula (1), the Adaboost algorithm can be written as follows:

1. The initial value of the weight is w ₁ (i) = 1 / N, and the discriminant function is F ₀ (x) = 0.

2. For t = 1 ・ ・ ・ T
(A) When the weighted error rate is expressed by the following equation (2), each weak learning machine (the following equation (3)) that minimizes the weighted error rate is selected.

(B) A coefficient (the following formula (4)) for the selected weak learning machine is calculated.

This learning machine is updated with _{F t = F t + 1 +} α t f t.

    (C) The weight is updated as shown in the following formula (5).

3. It discriminate | determines with the code | symbol of the discriminant function of following formula (6).

Therefore, by determining whether the discrimination function is positive or negative, it is possible to evaluate whether it is a positive sample.

  Here, the case where the discriminant function F is acquired by using the Adaboosting algorithm is illustrated, but instead, other machine learning methods such as a neural network and SVM (Support Vector Machine) are used. It can also be used.

  (2) Next, the procedure in the detection mode for actually measuring blood pressure will be described.

  As shown in FIG. 5 and FIG. 6, feature quantities of the electrocardiogram signal and the pulse wave signal are obtained from the signals from the pulse wave sensor 9 and the electrodes 11 and 13.

  Specifically, the above-described feature amounts (PTT, HR, velocity pulse wave feature amounts a1 to k1, acceleration pulse wave feature amounts a to f, AI (volume AI), etc.) are used to calculate blood pressure. calculate.

  Next, the binarization determination in the AdaBoost detection mode is performed a plurality of times.

  Specifically, the discriminant function generated in the pre-learning mode, that is, the discriminant function is used to binarize and discriminate whether the blood pressure is less than or equal to a predetermined blood pressure (for example, 80 mmHg), and from the feature quantity that is the parameter of the discriminant function, Whether the blood pressure is less than 80 mmHg or not is binarized. As a result, for example, if it is less than 80 mmHg, the determination result is set to 0, and if it is 80 mmHg or more, the determination result is set to 1.

  Similar determination is performed by gradually increasing the determination blood pressure for each 1 mmHg using each discriminant function generated in the pre-learning mode (performing binarization determination for each 1 mmHg). Here, the blood pressure is determined in the range of 80 to 160 mmHg.

  Then, the determination result by each discrimination function is arranged in order of increasing blood pressure to create a binary number sequence (for example, (0 ·· 011 ·· 1)).

  In other words, a binary number sequence is created so that the determined blood pressure increases as it goes to the right digit so that the leftmost digit of the binary number sequence indicates the determination result of 80 mmHg and the rightmost digit indicates the determination result of 160 mmHg. To do.

  Next, blood pressure is estimated by the ECOC method.

  Specifically, first, the numerical sequence is compared with a code table created in advance for each 1 mmHg. Usually, as shown in FIG.6 (b), it determines once (81 determinations for every 1 mmHg are set as 1 determination).

  In this code table, a binary number sequence having the same number of digits as the binary number sequence obtained by the determination is sequentially arranged for each determination blood pressure, and the binary number sequence is, for example, 80 according to the determination blood pressure. It is set for every 1 mmHg in the range of ~ 160.

  Specifically, in the code table, for example, as the blood pressure goes from 160 to 80, as shown in FIG. 7A (note that some digits are omitted), (0... 0001), (0. (0011), (0... 0111), and the number of digits indicating the number 1 in the binary number sequence is set to increase sequentially from the right side.

  Next, for each blood pressure, the binary number sequence obtained by the determination is compared with the numbers in each digit of the binary number sequence of the code table. Specifically, 0 is set when the numerical values of the respective digits match, and 1 is set when the numbers do not match, thereby obtaining a so-called Hamming distance.

  Then, a hamming distance is obtained for each blood pressure, a minimum value (minimum hamming distance) is obtained, and a blood pressure that becomes the minimum hamming distance is selected. Here, the reason for selecting the blood pressure that is the minimum Hamming distance is that it is considered that there is little error in blood pressure estimation.

  That is, here, the blood pressure that provides the minimum hamming distance is selected as the most probable blood pressure.

  -Next, the reliability is calculated.

  For example, the minimum hamming distance and the number of learning data are used as parameters. For example, when the learning data regarding the selected blood pressure is sufficiently larger than a predetermined value, the reliability is high. For example, the greater the number of learning data, the higher the reliability, and when the learning data is equal to or less than a predetermined determination value, the reliability may be set to 0 and not selected.

  -Next, the blood pressure determined in consideration of the reliability is displayed. At this time, data relating to reliability may also be displayed. For example, the reliability can be displayed divided into low, medium, and high according to the number of learning data. When the learning data is less than the predetermined determination value, it is preferable to display that fact (for example, “the blood pressure is not displayed because measurement reliability is low”) and the blood pressure is not displayed.

  c) Next, blood pressure measurement processing performed in the blood pressure calculation unit 37 according to the above-described procedure will be described.

  As shown in FIG. 8, in step 100 (S), a pulse wave signal and an electrocardiogram signal are acquired from the pulse wave sensor 9 and the electrodes 11 and 13.

  In the subsequent step 110, the pulse wave signal and the electrocardiogram signal are analyzed, and the characteristic quantities (for example, PTT, HR, AI, acceleration pulse waves a to f, etc.) necessary for the blood pressure measurement described above are calculated.

  In the following step 120, it is determined whether or not the feature quantity obtained in this way is an appropriate value (that is, an abnormal value that is not possible). If an affirmative determination is made here, the process proceeds to step 130. On the other hand, if a negative determination is made, the process returns to step 100.

  In step 130, binary discrimination is performed a plurality of times. That is, the feature amount is used for the discrimination function set for each blood pressure for each 1 mmHg, and it is determined whether the blood pressure value is less than or equal to the determination target (for example, less than 80 mmHg). Create a binary sequence with the corresponding number of digits.

  In the following step 140, the binary sequence created in step 130 is compared with the code table shown in FIG. 6B to obtain a Hamming distance, and a blood pressure value corresponding to the minimum Hamming distance is selected.

  In the subsequent step 150, the reliability is determined based on the number of learning data, for example. For example, when it is determined that the blood pressure value is 100 mmHg, it is determined whether there are 10 or more learning data numbers at 99 mmHg, 100 mmHg, and 101 mmHg. If it is determined that the number of learning data is equal to or greater than the predetermined value and the reliability is high, the process proceeds to step 170. On the other hand, if the reliability is determined to be low, the process proceeds to step 160.

  In step 170, since a highly reliable blood pressure is obtained, this blood pressure is determined as a measurement value.

  In the subsequent step 180, the blood pressure determined in step 170 is displayed on the display of the notification device 3, and the process is temporarily terminated.

  On the other hand, in step 160, it is determined that the reliability is low, and in step 160, the blood pressure is calculated by a multiple regression equation obtained in advance from the relationship between the feature amount of the biological information and the blood pressure based on the feature amount, In step 170, the blood pressure is determined as the current measurement. In step 180, the blood pressure is displayed on the display, and the process is temporarily terminated.

  The technique for calculating blood pressure by multiple regression will be described in Example 2 below, but if the reliability is low, it is displayed that blood pressure cannot be measured without performing blood pressure estimation by multiple regression. This processing may be temporarily terminated.

  d) As described above, in this embodiment, binarization determination using Adaboost is performed a plurality of times of binarization determination corresponding to each blood pressure value, and the binary sequence obtained thereby is represented in the code table. Compared with the binary sequence, the blood pressure corresponding to the minimum Hamming distance is determined as the current measurement value, so that even if the measurement value is ambiguous, the learning accuracy can be improved. As a result, the blood pressure measurement is performed with high accuracy. be able to.

  In addition, since the reliability of the blood pressure value obtained by the binarization determination is determined, there is an advantage that a highly accurate measurement value can be obtained from this point.

  Next, the second embodiment will be described, but the description of the same contents as the first embodiment will be omitted.

  Since the hardware configuration of this embodiment is the same as that of the first embodiment, the procedure for measuring blood pressure will be described.

  In this embodiment, the blood pressure is estimated with higher accuracy by performing both the process of estimating the blood pressure by learning and the process of estimating the blood pressure by multiple regression.

  a) Here, since the technique for estimating blood pressure by learning is the same as that in the first embodiment, an arithmetic expression for multiple regression used for calculating blood pressure will be described.

  For estimation of blood pressure by this multiple regression, for example, a method already proposed by the present inventor (see Japanese Patent Application No. 2007-154086) can be employed. This will be briefly described below.

  Here, using an arithmetic expression using the central vascular feature (PTT, pulse wave area) and the peripheral vascular feature (volume AI), for example, any one of the following expressions (7) to (11), The blood pressure (estimated blood pressure: EBP) is obtained.

・ When using PTT EBP = α ・ PTT + β ・ Volume AI + γ ・ ・ （7）
EBP = α · PTT + β · Volume AI + γ · (Pulse wave feature) + · · + · · · (8)
EBP = α · PTT + β · Volume AI + γ · d + δ · e + ε · f
+ Ζ · HR + · · · · · · (9)
・ When PTT is not used EBP = α ・ Pulse wave area + β ・ Volume AI + γ (10)
・ Estimated expression with higher accuracy EBP = α ・ PTT + β ・ Volume AI + γ ・ (Pulse wave feature) + δ ・ (Body feature)
+ Ε · (Impedance) + ζ (11)
In the above equations, α, β, γ, δ, ε, and ζ are coefficients. The pulse wave feature amounts are, for example, a1 and b1 (see FIG. 4). HR is a pulse, and a pulse interval (RRI) can be used instead of a pulse. The pulse wave area is the area of the portion surrounded by the waveform when the baseline of the pulse wave is zero. The body feature amount includes height, weight, sex, age, and the like.

  Hereinafter, each formula will be described.

  In the formula (7), PTT and volume AI are used for calculating the estimated blood pressure. Among these, PTT is a parameter indicating the vascular elasticity of the central blood vessel, and the volume AI is a parameter indicating the vascular elasticity of the peripheral blood vessel. Therefore, by using both parameters, information on the central and peripheral systems is taken in. The blood pressure can be estimated with high accuracy.

  In the equation (8), the pulse wave feature amount is used in addition to the PTT and the volume AI for calculation of the estimated blood pressure. Therefore, the blood pressure can be estimated with higher accuracy than the equation (7). The reason why the estimation accuracy is improved when the pulse wave feature amount is taken into account is that the effect of nonlinearity of the feature amount, the influence of the cardiac output, and the like are incorporated.

  In the formula (9), the estimated blood pressure is calculated using d, e, f, and HR of the acceleration pulse wave in addition to the PTT and the volume AI. Therefore, the blood pressure can be estimated with higher accuracy than the equation (7). Here, when the acceleration pulse wave d, e, f and HR are taken into account, the estimation accuracy is improved because of the effect of correcting the volume AI described later to the pressure AI and the influence of the cardiac output. is there.

  In the equation (10), the pulse wave area (S) and the volume AI are used to calculate the estimated blood pressure. Of these, S is considered to be a feature value indicating the vascular elasticity of the central blood vessel (estimated by experiment), and the volume AI is a feature value indicating the vascular elasticity of the peripheral blood vessel. The blood pressure can be accurately estimated without inputting the body weight.

  In the equation (11), in addition to the PTT and the volume AI, a pulse wave feature value, a body feature value, and impedance (for example, impedance measured between both hands) are used for calculating the estimated blood pressure. Therefore, the blood pressure can be estimated with higher accuracy than the equation (8). Body features such as height and weight have a correlation with blood flow, and impedance has a correlation with body weight and body fat (and thus also with blood flow). Thus, it is considered that blood pressure estimation accuracy is improved.

  b) Next, the procedure of blood pressure measurement in the present embodiment will be described.

  In this embodiment, blood pressure is estimated using the Adaboost and ECOC methods of the first embodiment, and blood pressure is estimated by multiple regression.

  As shown in FIG. 9, the feature quantity of the electrocardiogram signal and the pulse wave signal is obtained from the signals from the pulse wave sensor 9 and the electrodes 11 and 13.

  Specifically, the above-described feature amounts (PTT, HR, velocity pulse wave feature amounts a1 to k1, acceleration pulse wave feature amounts a to f, AI (volume AI), etc.) are used to calculate blood pressure. calculate.

  Next, blood pressure is calculated by performing binarization determination in the AdaBoost detection mode a plurality of times.

  That is, as in the first embodiment, the feature amount calculated this time (for example, PTT or HR) is applied to the discriminant function generated in the pre-learning mode, and binarization determination is performed for each blood pressure. A binary number sequence of discrimination function determination results is created. Then, the binary number sequence and the code table are compared to determine the minimum hamming distance, and the blood pressure that is the minimum hamming distance is selected.

  Further, the blood pressure is calculated from the characteristic amount (for example, PTT, HR, etc.) using the calculation formula for the multiple regression.

  Next, the blood pressure obtained by learning is compared with the blood pressure obtained by multiple regression to determine the blood pressure.

  For example, when the blood pressure obtained by multiple regression is 115 mmHg and the blood pressure obtained by learning is 112 mmHg, if the blood pressure obtained by multiple regression is in the range of 100 mmHg and 140 mmHg, the blood pressure obtained by learning is used as data with high reliability. To do. This is because the number of learning data in the range of 100 mmHg to 140 mmHg is large.

  Further, when the number of blood pressure data actually obtained by learning is 10, the blood pressure obtained by learning is adopted, and when the number is 3 or less, data obtained by multiple regression is adopted. This is because it is considered that the smaller the learning data, the lower the reliability of the calculated blood pressure.

  c) Next, blood pressure measurement processing performed according to the above-described procedure will be described.

  As shown in FIG. 10, in step 200 (S), a pulse wave signal and an electrocardiogram signal are acquired from the pulse wave sensor 9 and the electrodes 11 and 13.

  In the subsequent step 210, the pulse wave signal and the electrocardiogram signal are analyzed, and the characteristic quantities (for example, PTT, HR, AI, acceleration pulse waves a to f, etc.) necessary for the blood pressure measurement described above are calculated.

  In the following step 220, it is determined whether or not the feature quantity obtained in this way is an appropriate value (that is, an abnormal value that is impossible). If an affirmative determination is made here, the process proceeds to step 230. On the other hand, if a negative determination is made, the process returns to step 200.

  In step 230, blood pressure is estimated by multiple regression.

  In the following step 240, binary discrimination is performed a plurality of times using the AdaBoost, and blood pressure is estimated by learning.

  In subsequent step 250, the estimated value of blood pressure by multiple regression is compared with the estimated value of blood pressure by learning, for example, reliability is determined based on the number of learning data described above, and a highly reliable blood pressure value is adopted.

  In the subsequent step 260, the blood pressure determined in the step 250 is displayed on the display of the notification device 3, and the process is temporarily terminated.

  d) As described above, in this embodiment, blood pressure is estimated by multiple regression and blood pressure is estimated by learning, and a highly reliable blood pressure is adopted. Therefore, it is possible to always perform highly accurate blood pressure measurement. There is.

  In this embodiment, the discriminant function is generated by Adaboost as in the first embodiment, but other machine learning methods such as the above-described neural network and SVM can be used instead.

  Next, the third embodiment will be described, but the description of the same contents as the first embodiment will be omitted.

  In this embodiment, a code table having a width is used. That is, for example, a code table with two skips is used.

  In the first embodiment, as shown in FIG. 11A, each binary number sequence in the code table is set to increase by 1 from the right digit as the blood pressure decreases by 1 mmHg.

  However, in this embodiment, as shown in FIG. 11B, each binary number sequence in the code table is set so that 1 increases from the right by 2 digits as the blood pressure decreases by 2 mmHg. Therefore, for example, 100 mmHg and 101 mmHg are evaluated in the same number sequence.

  For example, when an ideal sequence of 100 mmHg is entered (0 ··· 00111 ··· 1), the Hamming distance (deviation from the code table) is 2 at 98, 1 at 100, and 2 at 102, and is simply 100 mmHg. Judgment is possible.

  For example, when an ideal sequence of 101 mmHg is entered (0..00011..1), the Hamming distance is 3, 98 at 1, 100 at 1, 102 at 1, and the Hamming distance is minimum at 100 mmHg and 102 mmHg. It can be determined as 101 mmHg.

  Thereby, since the reference to the code table is half as compared with the case of FIG. 11A (about 41 times, which is about half), there is an advantage that the calculation time can be shortened.

  Next, the fourth embodiment will be described, but the description of the same contents as the first embodiment will be omitted.

  In this embodiment, a code table suitable for the subject is used.

  For example, in the case of an elderly person, there is a low possibility that the blood pressure is 100 mmHg. Therefore, for example, as shown in FIG. 12, a code table without a binary sequence within the frame is used. In other words, for example, a code table with a small number of sequences can be used according to age (excluding the number sequence corresponding to the blood pressure when there is little possibility of binarized blood pressure).

  Thereby, calculation time can be shortened.

  Next, the fifth embodiment will be described, but the description of the same contents as the first embodiment will be omitted.

  In this embodiment, blood pressure estimation is performed as a group of predetermined blood pressure ranges.

  That is, as described below, for example, blood pressure is estimated for each step of 4 mmHg.

...
Less than 100mmHg / 104mmHg or more Less than 104mmHg / 108mmHg or more
...
Thereby, calculation time can be shortened.

  Next, the sixth embodiment will be described, but the description of the same contents as the first embodiment will be omitted.

  In this embodiment, blood pressure estimation is performed as a group of a predetermined blood pressure range, but the width of the group is narrow in a certain range, and the width of the group is widened in a certain range.

  That is, for example, in the case of 100 to 140 mmHg, the width of the group is as narrow as 1 mmHg, and in the case of less than 100 mmHg or in the range of 140 to 180 mmHg, the width of the group is as moderate as 4 mmHg. .

  The reason for setting in this way is that the reliability of the data is low when the learning data is small, and that the blood pressure data has a lot of learning data around the average value, and the learning data decreases as it goes away. This is because there is a phenomenon.

  Thereby, there is an advantage that calculation time can be shortened and ambiguity due to the number of learning data can be reduced.

  Next, although Example 7 is demonstrated, description of the content similar to the said Example 1 is abbreviate | omitted.

  Generally, a cuff sphygmomanometer that is a true value of learning usually has an error of 4 to 8 mmHg. Therefore, in this embodiment, in consideration of this point, data within a predetermined range (for example, within 4 mmHg width, within 8 mmHg width, etc.) is not used from the blood pressure value of the binarized learning data.

  Thereby, correct true value data can be used for learning, and blood pressure estimation accuracy is improved.

  Next, although Example 8 will be described, description of the same contents as in Example 1 will be omitted.

  In this embodiment, the number of learning data for binarization is within a certain range.

  When the number of learning data in each blood pressure layer is greatly different, blood pressure estimation accuracy may decrease. For example, when the ratio of the number of learning data is the same, such as the number of learning data of less than 100 mmHg is 1: 1, the accuracy of blood pressure estimation is high, but the number of data of less than 140 mmHg is 8: As shown in FIG. 1, when the ratio of the number of learning data is greatly different, the blood pressure value estimation accuracy tends to be low.

  The reason for this is that even if there is learning data with the same feature amount in the blood pressure to which the pulse wave originally belongs, the probability that there is learning data with the same feature amount in the blood pressure to which the pulse wave does not originally belong This is because the accuracy of blood pressure estimation decreases as a result.

  Therefore, in this embodiment, the ratio of the number of learning data in each blood pressure layer (for example, each blood pressure layer divided into two, less than 140 mmHg) divided by a predetermined blood pressure value for binarization determination is , Within a predetermined range (for example, within 2 times).

  Thereby, since the effect dragged by the number of data is reduced, blood pressure measurement accuracy is improved.

That is, when learning data is used, it is desirable to set in advance at what ratio data in a predetermined blood pressure range is used. In addition, as a preferable ratio of the number of data, for example, a degree within 5 times can be mentioned.
<< Experimental example >>
Next, experimental examples conducted for confirming the effects of the present invention will be described.

  This experiment was conducted using Matlab's Adaboost Toolbox (http://research.graphicon.ru/machine-learnining/gml-adaboost-matlab-toolbox.html) published on the Internet.

  In the following experiments, PTT, HR, AI, acceleration pulse waves a to f, body feature amounts (height, weight, sex, age), and impedance were adopted as feature amounts used for learning. Further, as feature quantities used for multiple regression, PTT, AI, pulse wave feature quantities (acceleration pulse waves a to f), body feature quantities (height, weight, sex, age), and impedance are adopted and used for multiple regression. As the formula, the formula (11) was adopted.

  Moreover, the weak learning machine used CART (Classification And Regression Trees (http://research.graphicon.ru/machine-learning/index.php), and the number was 20 pieces.

<Experimental Example 1> Comparison between AdaBoost and Multiple Regression In this experimental example, blood pressure is simply estimated by Adaboost and multiple regression for all data. Note that AdaBoost was implemented in 1 mmHg steps. The experimental results are shown in FIG.

  FIG. 13A shows a graph of estimated blood pressure and cuff blood pressure in learning data (number N = 345) for pre-learning, and FIG. 13B shows detection data in evaluation mode (number N = 86). A graph of estimated blood pressure and cuff blood pressure is shown.

  As is clear from FIG. 13B, it can be seen that AdaBoost has the same degree of accuracy as the multiple regression or better estimation accuracy than the multiple regression in the central part of blood pressure (standard range: 100 to 140 mmHg).

  Therefore, it is desirable to adopt the blood pressure value obtained by Adaboost rather than multiple regression at the center of blood pressure.

<Experimental Example 2> Estimated Accuracy of Each Blood Pressure Layer In this experimental example, similar to Experimental Example 1, the blood pressure estimation accuracy of multiple regression and Adaboost in each blood pressure layer was examined. The experimental results are shown in FIG.

  FIG. 14 (a) shows a graph of each blood pressure range and standard deviation in the learning data, and FIG. 14 (b) shows a graph of each blood pressure range and standard deviation in the evaluation data.

  From FIG. 14, it can be seen that in the blood pressure range of 110 to 130 mmHg and the learning data number N is 50 or more (5 or more per 1 mmHg), Adaboost has better accuracy than multiple regression.

  Therefore, when the above conditions are satisfied, it is desirable to adopt the blood pressure estimated by Adaboost.

<Experimental example 3> Blood pressure estimation by group (step) This experimental example uses data having a certain width (for example, a step width of 4 mmHg, etc.) instead of every 1 mmHg as the learning data when estimation is performed by Adaboost. Thus, blood pressure is estimated as a group. The experimental results are shown in FIG.

  FIG. 15A is a graph showing the relationship between the estimated blood pressure and the cuff blood pressure, and FIG. 15B is the relationship between the standard deviation of the error (converted with the average value of the error being 0) and the blood pressure at each step. It is a graph which shows.

  From FIG. 15, it can be seen that the estimation accuracy is better when the step width is smaller in the central portion of the blood pressure and the step width is larger in the peripheral portion.

  Therefore, it is desirable to decrease the blood pressure estimation step width in the central portion of the blood pressure, for example, in the range of 100 to 140 mmHg, and increase the step width in the outer region.

<Experimental example 4> Hamming distance and error In this experimental example, the relationship between the Hamming distance and the error of the estimated blood pressure (the difference between the cuff blood pressure and the estimated blood pressure) was examined. The experimental results are shown in FIG.

  FIG. 16A shows a graph in which the relationship between the minimum hamming distance and the error (mmHg) is examined, and FIG. 16B shows the relationship between the error when the minimum hamming distance is less than 10 and 10 or more. It is a graph.

  From FIG. 16, the error tends to increase as the minimum Hamming distance increases.

  Accordingly, it is preferable to employ a small hamming distance.

<Experimental example 5> Effect of selection of learning data This selection of learning data is to improve estimation accuracy by excluding data that may be wrong.

  In this experimental example, in consideration of the fact that the cuff sphygmomanometer normally has an error of 4 to 8 mmHg, data within a predetermined range is not used from the blood pressure value of binarized learning data. .

  In other words, since there is an error in the reference blood pressure itself that is the true value of learning, using incorrect data for learning may result in an incorrect result. For example, when binarization is performed at less than or equal to 100 mmHg, By removing 96-104 mmHg (± 4 mmHg) data that is highly likely to be incorrect from the learning data, data that is likely to be incorrect is removed, and learning is performed using data that is likely to be correct. It is. The experimental results are shown in FIG.

  FIG. 17A is a graph showing the relationship between estimated blood pressure and cuff blood pressure corresponding to data removal, and FIG. 17B is a graph showing the relationship between error and blood pressure corresponding to data removal. Yes, FIG. 17 (c) shows the standard deviation of error of estimated blood pressure corresponding to the number of data divided into each blood pressure layer and the data removal width (data removal width) (converted with the average value of the error as 0). It is a table.

  FIG. 17 shows that the accuracy is improved by 1 mmHg when binarized learning is performed without using the data inside the 4-8 mmHg width for learning.

<Experimental Example 6> Influence of Learning Data Ratio In this experimental example, the influence of the number of learning data on blood pressure estimation is examined in each blood pressure layer that classifies learning data by a predetermined blood pressure.

  In this experiment, when there was a difference in the number of data more than twice in each blood pressure layer, the experiment was performed with the larger number of data halved. The experimental results are shown in FIG.

  Specifically, for example, the number of learning data is 39 for 99 mmHg or less, 60 for 100 to 109 mmHg, 69 for 110 to 119 mmHg, 79 for 120 to 129 mmHg, 55 for 130 to 139 mmHg, 24 for 140 to 149 mmHg When there are 19 pieces of 150 mmHg or more, 302 pieces are below 140 mmHg and 43 pieces are 140 mmHg or more. Therefore, in this case, the data of the blood pressure layer of less than 140 mmHg is halved (for example, an odd or even number is selected from those having a low blood pressure).

  FIG. 18A is a graph showing the relationship between the estimated blood pressure and the cuff blood pressure when the number of data is changed, and FIG. 18B is a graph when the number of data divided into each blood pressure layer is changed. It is a table | surface which shows the standard deviation (The average value of an error is converted into 0) of the error of estimated blood pressure.

  As is clear from FIG. 18, when there is a difference in the number of data more than twice in each blood pressure layer, the ratio of the number of data (data ratio) is corrected compared to the case where the number of data is not adjusted (as it is) ( It was confirmed that the accuracy was improved.

  Therefore, it can be seen that it is desirable to make the number of data of each blood pressure layer closer.

  Needless to say, the present invention is not limited to the above-described embodiments, and can be implemented in various modes without departing from the scope of the present invention.

  (1) For example, as shown in FIG. 7B, a weighted number sequence can be used as the code table. In this case, if the corresponding number of digits is different, the numerical values of different code tables may be added. This has the advantage that the hamming distance of the greatly deviated data is increased, the deviated data is removed, and the accuracy can be improved.

  (2) Further, for example, the function of the blood pressure measurement device described above can be realized by processing executed by a computer program, and this program can be recorded on a recording medium.

  That is, the function for realizing the above-described program in the computer system can be provided as a program that is activated on the computer system side, for example. In the case of such a program, for example, the program is recorded on a computer-readable recording medium such as a flexible disk, a magneto-optical disk, a CD-ROM, a DVD, and a hard disk, and is used by being loaded into a computer system and started as necessary. be able to. In addition, the program may be recorded as a computer-readable recording medium such as a ROM or a backup RAM, and the ROM or the backup RAM may be incorporated into a computer system.

DESCRIPTION OF SYMBOLS 1 ... Blood pressure measuring device 3 ... Notification apparatus 5 ... Manual input part 7 ... Steering 9 ... Pulse wave sensor 11, 13 ... Electrode

Claims (20)

A blood pressure measurement device that estimates blood pressure using at least biological information obtained from a pulse wave signal among an electrocardiogram signal and a pulse wave signal detected from a living body,
Identification means for binarizing and discriminating blood pressure from the feature amount of the biological information for each different blood pressure based on the relationship between the feature amount of the biological information and blood pressure obtained by prior learning;
A binarization determination unit that performs binarization determination on a predetermined blood pressure based on a feature amount of the biological information obtained by measurement using the identification unit when estimating blood pressure;
Blood pressure estimation in which the binarization determination unit performs binarization determination on different blood pressures based on the feature amount of the biological information, and estimates the blood pressure based on the results of the multiple binarization determinations Means,
A blood pressure measurement device comprising:   The blood pressure measurement device according to claim 1, wherein the learning identification unit generates an identification function for binarizing the blood pressure by Adaboost.   A binary number sequence in which the determination results for each blood pressure are arranged from the result of the binarization determination of the plurality of times is created, and this binary number sequence is compared with the number sequence for each blood pressure in the code table prepared in advance. The blood pressure measurement apparatus according to claim 1, wherein the blood pressure is estimated.   A minimum hamming distance in both sequences is obtained from the binary sequence obtained from the result of the binary determination and the binary sequence in the code table, and the blood pressure is estimated based on the minimum hamming distance. The blood pressure measurement device according to claim 3, wherein the blood pressure measurement device is performed.   5. The blood pressure measurement device according to claim 3, wherein the numerical sequence of the code table is set so that a numerical value is different in a plurality of widths.   The blood pressure measurement device according to claim 3 or 4, wherein a code table having a numerical sequence set in conformity with the condition of the subject is used as the code table.   The blood pressure measurement device according to claim 1, wherein the determination of binarization of the blood pressure is performed for each group of a predetermined blood pressure range.   8. The blood pressure measurement according to claim 7, wherein the blood pressure is estimated by reducing the width of the group within a predetermined blood pressure range and increasing the width of the group outside the predetermined blood pressure range. apparatus.   The blood pressure measurement device according to any one of claims 1 to 8, wherein data in a predetermined blood pressure range is not used for learning in advance learning in the learning identification unit.   The blood pressure according to any one of claims 1 to 9, wherein a ratio of the number of binarized data in different blood pressure ranges is set within a predetermined range at the time of prior learning in the learning identification means. measuring device.   The blood pressure measurement device according to any one of claims 1 to 10, wherein when estimating the blood pressure, the reliability of data is calculated, and the reliability of the data is displayed together with a blood pressure estimation result.   The blood pressure is estimated by calculating a reliability of the data, and when the reliability is lower than a predetermined determination value, the blood pressure is estimated using a different method that is expected to have a higher reliability. 11. The blood pressure measurement device according to 11.   The blood pressure measurement device according to claim 11 or 12, wherein the reliability of the data is determined based on the number of learning data or a minimum Hamming distance.   The blood pressure measurement device according to claim 12 or 13, wherein the different method is an estimation method by multiple regression. In a blood pressure measurement device that estimates blood pressure using at least biological information obtained from the pulse wave signal among an electrocardiogram signal and a pulse wave signal detected from a living body,
First blood pressure estimating means for estimating blood pressure by multiple regression based on the feature amount of the biological information;
Second blood pressure estimating means for estimating blood pressure from the feature quantity using a discrimination means obtained by learning the relationship between the feature quantity of the biological information and blood pressure;
The first estimated blood pressure estimated by the first blood pressure estimating means and the second estimated blood pressure estimated by the second blood pressure estimating means are compared based on a determination condition for determining whether the predetermined reliability is high or low. Blood pressure determining means for determining the estimated blood pressure with high reliability,
A blood pressure measurement device comprising: The second blood pressure estimating means includes
An identification means obtained by previously learning the relationship between the feature value of the biological information, which is learning data, and the blood pressure, and binarizing and determining the blood pressure from the feature value of the biological information for each predetermined blood pressure Learning identification means;
Binarization discriminating means for performing binarization discriminating on the predetermined blood pressure based on the feature quantity of the biological information obtained by measurement using the learning discriminating means when estimating the blood pressure; ,
The binarization determination means performs binarization determination on each of the different predetermined blood pressures based on the feature amount of the biological information, and estimates the blood pressure based on the results of the multiple binarization determinations Blood pressure estimating means for
The blood pressure measurement device according to claim 15, comprising:   17. The blood pressure according to claim 15, wherein when the first estimated blood pressure obtained by the multiple regression is 100 to 140 mmHg, the second estimated blood pressure obtained by the learning is adopted as an estimated value of blood pressure. measuring device.   The blood pressure measuring device according to any one of claims 15 to 17, wherein when the learning data is 5 or more per 1 mmHg, the second estimated blood pressure obtained by the learning is adopted as an estimated value of blood pressure.   The program for implement | achieving the function of the blood pressure measurement apparatus in any one of the said Claims 1-18.   A computer-readable recording medium on which the program according to claim 19 is recorded.