Pulse wave signal analysis device, pulse wave signal analysis method, and computer program

Abstract

Provided is a pulse wave signal analysis device capable of accurately predicting, by a non-invasive method, a result of diagnosis based on a blood test. The analysis device includes: an acquisition unit that acquires a pulse wave signal corresponding to a pulse of a living body; a pulse wave waveform generation unit that generates a pulse wave on the basis of the acquired pulse wave signal; a feature amount extraction unit that extracts a feature amount of the generated pulse wave; and a model generation unit that performs machine learning with respect to the extracted feature amount by using a result of determination based on a blood test on the living body, and generates a prediction model for predicting, from a feature amount of a newly input pulse wave, a result of diagnosis based on a blood test. A pulse wave is generated on the basis of a newly acquired pulse wave signal, a feature amount of the generated pulse wave is extracted, and a result of diagnosis based on a blood test on the living body is predicted with respect to the extracted feature amount on the basis of the generated prediction model.

Description

Pulse wave signal analysis device, pulse wave signal analysis method, and computer program

[0001] The present disclosure relates to a pulse wave signal analysis device, a pulse wave signal analysis method, and a computer program, and particularly relates to a pulse wave signal analysis device, a pulse wave signal analysis method, and a computer program for predicting the results of a diagnosis based on a blood test.

[0002] In recent years, the aging of society has progressed, and the increase in heart failure and decline in renal function have been increasing. Among these, heart failure has a high mortality rate, and as of 2019, it has been reported that 50% of heart failure patients have died within 5 years. Also, along with the increase in hypertension due to aging, the number of heart failure patients has been increasing, and as of 2017, the number of heart failure patients worldwide has been reported to exceed 60 million.

[0003] In addition, the number of patients with chronic kidney disease, which is one of the declines in renal function, is estimated to be about 13 million in Japan, and it has become a common disease.

[0004] Also, it is known that heart failure and decline in renal function can be caused by lifestyle-related diseases. For example, acute heart failure is often caused by lifestyle-related diseases. Furthermore, it is also known that many lifestyle-related diseases deteriorate the kidneys. Therefore, it can be said that discovering lifestyle-related diseases not only leads to the early detection of heart failure, decline in renal function, etc., but also leads to preventing these symptoms. In particular, obesity in childhood and early adolescence has a very high possibility of leading to adult obesity, so it can be said that implementing preventive medical check-ups for lifestyle-related diseases during these periods leads to the early detection and prevention of heart failure, decline in renal function, etc.

[0005] For the above reasons, it is desirable to diagnose symptoms such as heart failure, decline in renal function, and lifestyle-related diseases at an early stage.

[0006] Japanese Unexamined Patent Application Publication No. 2016 - 40693

[0007] In the diagnosis and treatment of heart failure, blood tests typically use serum BNP and NT-proBNP levels. Similarly, impaired renal function is assessed using eGFR levels in blood tests. However, obtaining these values ​​requires blood tests, which are invasive and pose a significant risk of injury and burden to the patient. Furthermore, blood tests require specialized equipment for blood analysis.

[0008] It should be noted that techniques for analyzing the signs of arrhythmia using heart rate and respiratory rate have been known for some time (see, for example, Patent Document 1). However, conventional techniques cannot predict various types of information obtained through invasive methods such as blood tests using non-invasive methods.

[0009] Furthermore, measuring LDL cholesterol levels through blood tests is necessary when diagnosing lifestyle-related diseases. Therefore, there are problems with blood tests, such as them being invasive and requiring specialized equipment.

[0010] Furthermore, as mentioned above, while it is desirable to conduct lifestyle-related disease prevention health checkups during childhood and early adolescence, the need for invasive diagnosis makes it difficult to conduct mass examinations of children, for example, at schools.

[0011] This disclosure has been made in view of the above issues, and its purpose is to provide a pulse wave signal analyzer, a pulse wave signal analyzer, and a computer program that can accurately predict the results of a diagnosis based on blood tests in a non-invasive manner.

[0012] To solve the above problems, according to one aspect of this disclosure, a pulse wave signal analysis device is provided, comprising: an acquisition unit that acquires a pulse wave signal corresponding to the pulse rate of a living organism; a pulse wave waveform generation unit that generates a pulse wave based on the acquired pulse wave signal; a feature extraction unit that extracts feature quantities of the generated pulse wave; and a model generation unit that performs machine learning on the extracted feature quantities using the results of a judgment based on a blood test on the living organism, and generates a predictive model that predicts the result of a diagnosis based on a blood test from newly input pulse wave feature quantities.

[0013] Here, the diagnosis may be defined as a diagnosis of lifestyle-related disease, cardiovascular disease, or chronic kidney disease.

[0014] Furthermore, the feature quantity may include at least one of the following: the rise time of the maximum peak of the pulse wave, the pulse rate, the peak interval, and the increase index.

[0015] Furthermore, the aforementioned feature quantities may include statistics on the rise time of the maximum peak of the pulse wave, pulse rate, peak interval, and increase index.

[0016] Here, the statistic may include at least one of the following: standard deviation, mean, maximum, minimum, median, and coefficient of variation.

[0017] Furthermore, the model generation unit may include a calculation unit that calculates a cutoff value for predicting the results of the diagnosis.

[0018] According to another aspect of this disclosure, a pulse wave analysis device is provided, comprising: an acquisition unit for acquiring a pulse wave signal corresponding to the pulse rate of a living organism; a pulse wave waveform generation unit for generating a pulse wave based on the acquired pulse wave signal; a feature extraction unit for extracting feature quantities of the generated pulse wave; a prediction unit for predicting the result of a diagnosis based on a blood test on the living organism based on a prediction model generated by the pulse wave analysis device described above, using the extracted feature quantities; and an output unit for outputting the result predicted by the prediction unit.

[0019] Another aspect of this disclosure provides a method for analyzing pulse wave signals, comprising the steps of: acquiring a pulse wave signal corresponding to the pulse rate of a living organism; generating a pulse wave based on the acquired pulse wave signal; extracting feature quantities of the generated pulse wave; and generating a predictive model that predicts the results of a diagnosis based on a blood test, by performing machine learning on the extracted feature quantities using the results of a judgment based on a blood test of the living organism.

[0020] According to another aspect of the present invention, a computer program is provided that causes a computer to function as an analysis device for the pulse wave signal described above.

[0021] According to another aspect of the present invention, a computer-readable storage medium is provided which stores a computer program that causes a computer to function as an analysis device for the pulse wave signal described above.

[0022] According to this disclosure, a pulse wave signal corresponding to the pulse rate of a living organism is acquired, a pulse wave is generated based on the pulse wave signal, and features of the pulse wave are extracted. Then, using the results of a blood test performed on the living organism, machine learning is performed on the features to generate a predictive model that predicts the results of a blood test diagnosis from the newly input pulse wave features. Therefore, the results of a blood test diagnosis can be predicted accurately using a non-invasive method.

[0023] Furthermore, by implementing the embodiments of this disclosure in biosensors such as blood pressure monitors, it becomes possible to easily predict lifestyle-related diseases. Therefore, it is possible to predict lifestyle-related diseases in various situations, such as when conducting examinations of students at schools.

[0024] This is a block diagram showing the functional configuration of a pulse wave signal analysis device according to one embodiment. This is a block diagram showing the functional configuration of a control unit according to one embodiment. (a) is a flowchart showing the procedure of the method in the learning phase according to one embodiment, and (b) is a flowchart showing the procedure of the method in the prediction phase according to one embodiment. This is a figure showing an example of measurement results when blood pressure is measured on the upper arm, where (a) shows cuff pressure and (b) shows blood pressure pulse wave amplitude. (a) shows an example of measurement results when blood pressure is measured on the upper arm, and (b) shows an example of measurement results when blood pressure is measured at the wrist. This is a figure showing an example of pulse wave features. This is a figure showing an example of pulse wave features. This is a conceptual diagram for explaining a decision tree. This is a figure showing an example of an ROC curve. This is a graph showing the number of patients for each LDL-C value. This is a graph showing the number of patients for each LDL-C value in the learning data. This is a graph showing the number of patients for each LDL-C value in the validation data. This is a graph showing the number of patients for each LDL-C value in the test data.

[0025] The embodiments of this disclosure will be described in detail below with reference to the drawings. In the following description, "pulse wave" refers to waveform data corresponding to biological information generated by the beating of the heart. For example, the concept of a pulse wave in this embodiment includes, but is not limited to, blood pressure pulse waves measured by a blood pressure monitor that measures on the upper arm or wrist, waveforms based on oxygen saturation obtained by analyzing facial pulses using video pulse wave measurement (VPG) with a camera, pulse waves due to changes in blood pressure measured by non-contact devices such as millimeter-wave radar, and pulse waves measured based on changes in volume from photoelectric sensors such as fingertip volume pulse waves. The embodiments of this disclosure may include any of these, but the following embodiments will be described using blood pressure pulse waves as an example.

[0026] Figure 1 is a block diagram showing the functional configuration of a pulse wave signal analysis device according to an embodiment of the present disclosure. The analysis device 100 functions as a device for analyzing the heart condition of a living organism such as a subject, and is configured such that a pulse wave detection unit 104 and an analysis unit 101 are connected to each other and can communicate signals. In this case, the communication means may be wired or conform to a wireless communication standard such as Bluetooth®. The pulse wave detection unit 104 and the analysis unit 101 may be integrated into a single device or configured separately. If the pulse wave detection unit 104 and the analysis unit 101 are configured separately, they may be configured to be detachable from each other.

[0027] The analysis unit 101 comprises a storage unit 102, an analog-to-digital conversion circuit (A / D conversion circuit) 111, a control unit 112, an input unit 116, a notification unit 118, and a power supply 120. These components are connected via a bus 128. The analysis unit 101 further includes an amplification circuit 110 connected to the A / D conversion circuit 111.

[0028] The pulse wave detection unit 104 non-invasively detects pulse wave signals in accordance with the heartbeat of a living organism and can be configured as a contact-type or non-contact-type biosensor. In one embodiment, the analysis device 100 can be configured as a blood pressure monitor. In this case, the pulse wave detection unit 104 is configured to include a cuff that is wrapped around a predetermined part of the arm. The cuff is provided with an inflatable bag-shaped member. The blood pressure monitor has a pre-set compression pressure to block blood flow in the arteries located inside the cuff when the cuff is wrapped around the arm, etc. When detecting a pulse wave, the pressure inside the bag-shaped member is increased to the compression pressure, and then gradually decreased at a predetermined rate to measure blood pressure. The blood vessels compressed by the cuff vibrate in accordance with the beating of the heart. Therefore, during the process of decreasing blood pressure in blood pressure measurement, the pressure inside the bag-shaped member can be detected sequentially, and the pulse wave can be detected from the pressure change.

[0029] In one embodiment, the pulse wave detection unit 104 may be configured as a combination of a blood pressure monitor and a wireless electrocardiogram transmitter, and the analysis unit 101 may be configured as an information processing device such as a personal computer. In this case, the pulse wave detection unit 104 may include an input unit such as a button for instructing the execution of an operation to measure blood pressure, and a display unit for displaying the measurement results.

[0030] Alternatively, the pulse wave detection unit 104 may be configured as a fingertip pulse wave detection device. In this case, the pulse wave detection unit 104 can detect pulse waves by irradiating infrared light onto the fingertips of the hands or feet of a living body and measuring the change in the volume of blood flowing through the fingertips.

[0031] Alternatively, the pulse wave detection unit 104 may be configured as a non-contact device such as a camera or a mobile terminal. In this case, the pulse wave detection unit 104 can detect pulse waves by measuring changes in brightness caused by changes in blood flow based on the captured image.

[0032] Alternatively, the pulse wave detection unit 104 may be configured as another optical heart rate measurement device. In this case, the pulse wave detection unit 104 may be configured to have an optical heart rate sensor. The pulse wave detection unit 104 can detect a pulse wave by irradiating light onto a predetermined position such as the arm using an LED and measuring the amount of light reflected from hemoglobin in the blood vessels.

[0033] Next, the various functions of the analysis unit 101 will be described.

[0034] The amplification circuit 110 is configured to amplify the analog electrical signal supplied from the pulse wave detection unit 104. The A / D conversion circuit 111 is configured to convert the analog electrical signal output from the amplification circuit 110 into a digital signal.

[0035] The storage unit 102 is configured to store programs for executing the processing according to this embodiment, data used in arithmetic processing, or data of arithmetic results. Specifically, the storage unit 102 can be configured as a non-volatile storage device such as a USB flash drive, removable hard disk, magnetic disk, or optical disk. The storage unit 102 stores programs that are read and executed by the computer's processor. The program includes a machine learning program 106 and a prediction program 108. The machine learning program 106 performs machine learning using pulse wave waveform data as input and generates a prediction model. The prediction program 108 uses the prediction model generated as described above to predict lifestyle-related diseases, etc., based on the subject's pulse wave waveform data.

[0036] The memory unit 102 also stores data on the results of blood tests performed on living organisms. In the following description, this data will be referred to as the result data. The result data will be used in machine learning to generate a predictive model, as will be described later.

[0037] The control unit 112 performs calculation processing based on the electrical signals of the pulse wave output from the pulse wave detection unit 104. The control unit 112 includes a CPU (Central Processing Unit) 122, a RAM (Random Access Memory) 124, and a ROM (Read Only Memory) 126. The CPU 122 controls each part and transfers data according to the program stored in the ROM 126. The RAM 124 temporarily stores biological information such as pulse wave waveform data, and various data generated during the execution of the program in the CPU 122. The CPU 122 is configured to realize various functions by executing a program. By executing the machine learning program 106, the CPU 122 realizes the function of generating a prediction model, and by executing the prediction program 108, it realizes the function of predicting lifestyle-related diseases, cardiovascular diseases, or chronic kidney diseases.

[0038] Here, lifestyle-related diseases may include, but are not limited to, those in which the link between lifestyle habits and disease has been clearly established, such as the examples below: (1) Diseases caused by dietary habits: non-insulin-dependent diabetes mellitus, obesity, hyperlipidemia (excluding familial cases), hyperuricemia, cardiovascular diseases (excluding congenital cases), colorectal cancer (excluding familial cases), periodontal disease, etc. (2) Diseases caused by exercise habits: non-insulin-dependent diabetes mellitus, obesity, hyperlipidemia (excluding familial cases), hypertension, etc. (3) Diseases caused by smoking: squamous cell carcinoma of the lung, cardiovascular diseases (excluding congenital cases), chronic bronchitis, emphysema, periodontal disease, etc. (4) Diseases caused by alcohol consumption: alcoholic liver disease, etc.

[0039] Furthermore, cardiovascular diseases may include heart failure, stroke, and angina pectoris.

[0040] The input unit 116 is configured as a keyboard, buttons, dials or switches for inputting information such as commands and set values for information processing from the user, or an input interface for inputting data. The notification unit 118 is configured as a liquid crystal display, a speaker, or the like for outputting notification information based on the result of information processing by the control unit 112. In addition, the analysis device 100 includes a power supply 120 for supplying power to each component within the device.

[0041] Note that one or more of the functions included in the analysis unit 101 may be implemented in the pulse wave detection unit 104. Also, each of the CPU 122, RAM 124, and ROM 126 included in the control unit 112 may be one component or a combination of multiple components.

[0042] Next, referring to FIG. 2, the functional configuration of the analysis device 100 will be described. These functional configurations are implemented by the CPU 122 of the control unit 112 reading and executing the machine learning program 106 and the prediction program 108 stored in the storage unit 102. The control unit 112 includes a pulse wave signal acquisition unit 202, a pulse wave waveform generation unit 204, a feature amount extraction unit 206, a model generation unit 208, a prediction unit 210, and an output unit 211.

[0043] The pulse wave signal acquisition unit 202 is configured to acquire a pulse wave signal corresponding to the pulse of a living body. Specifically, the pulse wave signal acquisition unit 202 acquires a digital pulse wave signal corresponding to the heartbeat of the living body measured by the pulse wave detection unit 104. The pulse wave waveform generation unit 204 is configured to generate image data of a pulse wave waveform from the digital pulse wave signal.

[0044] The feature amount extraction unit 206 is configured to extract the feature amounts of the image data of the pulse wave waveform. The model generation unit 208 is configured to perform machine learning on the extracted feature amounts by using the result of determination based on a blood test for the living body, and generate a prediction model used for predicting the result of diagnosis based on the blood test of the living body.

[0045] The prediction unit 210 is configured to predict the result of a diagnosis based on a blood test of a living body based on the prediction model generated as described above. The output unit 211 is configured to output the result predicted by the prediction unit 210. Here, the output of the result includes various forms such as display on a screen and printing on a recording medium such as paper.

[0046] Next, the procedure of the process executed by the analysis device 100 will be described. The process executed in this embodiment is divided into a learning phase of performing machine learning to generate a prediction model and a prediction phase of predicting the diagnosis result of a blood test from a newly input pulse wave signal using the prediction model. Hereinafter, referring to FIGS. 3(a) and 3(b), the processes performed in these phases will be described.

[0047] (Learning Phase) FIG. 3(a) is a flowchart showing the flow of operations executed by the analysis device 100 in the learning phase. In the learning phase, the analysis device 100 performs machine learning to generate a prediction model.

[0048] First, a blood test is performed on a living body, and the result of determination of lifestyle-related diseases or the like is input to the analysis device 100 as determination result data. The determination result data is stored in the storage unit 102 in association with an identifier of the living body (subject identifier) (step S0). Table 1 shows an example of the determination result data of one subject.

[0049]

[0050] In Table 1, as indices, LDL cholesterol (LDL-C; Low Density Lipoprotein cholesterol) (mg / dL), human brain natriuretic peptide precursor N-terminal fragment (NT-proBNP; N-terminal pro-brain natriuretic peptide) (pg / mL), estimated glomerular filtration rate (eGFR; Estimated Glomerular Filtration Rate) (ml / min / 1.73m 2), and hemoglobin A1c (HbA1c) (%) are used. LDL-C is an indicator used to confirm the progression of arteriosclerosis, NT-proBNP is an indicator used to diagnose cardiovascular diseases such as heart failure, eGFR is an indicator used to confirm renal function such as chronic kidney disease, and HbA1c is an indicator used to diagnose diabetes. LDL-C and HbA1c are used as indicators to assess the risk of lifestyle-related diseases.

[0051] The following indicators may be used, but are not limited to, those used in diagnosing lifestyle-related diseases: (1) Indicators obtained by lipid abnormality tests: triglycerides, HDL cholesterol, LDL cholesterol (2) Indicators obtained by blood tests: fasting blood glucose or HbA1c (3) Indicators obtained by liver function tests: GOT, GPT, γ-GTP

[0052] Furthermore, indicators of cerebral vascular stenosis obtained from carotid artery ultrasound, MRA (magnetic resonance angiography), and cerebral angiography can be used as indicators for diagnosing cerebrovascular disorders such as stroke. In addition, indicators of myocardial ischemia obtained from electrocardiogram (ECG), cardiac catheterization, coronary CT, and myocardial scintigraphy can be used as indicators for diagnosing ischemic heart diseases such as angina pectoris.

[0053] A cutoff value is set as a numerical criterion for determining whether a value is normal or abnormal. In this example, two patterns of cutoff values ​​are set for LDL-C: 100 and 140. The judgment result is set with a flag value indicating whether it is abnormal or normal, based on the comparison of the numerical values ​​of the indicators obtained from the blood test.

[0054] Next, in step S1, the pulse wave signal acquisition unit 202 acquires a digital pulse wave signal corresponding to the pulse rate of the living body measured by the pulse wave detection unit 104.

[0055] Next, in step S2, the pulse wave waveform generation unit 204 generates pulse wave waveform image data from the digital pulse wave signal.

[0056] Next, in step S3, the feature extraction unit 206 extracts features from the pulse wave waveform image.

[0057] Next, in step S4, the model generation unit 208 uses the results of the judgment based on the blood test on the living organism to perform machine learning with the feature values ​​as input data and generate a predictive model. This predictive model is a model for determining a diagnostic result (for example, normal or abnormal) based on the indicators of the blood test from one or more features. The generated predictive model is stored in the storage unit 102.

[0058] (Prediction Phase) Figure 3(b) is a flowchart showing the flow of operations performed by the analysis device 100 in the prediction phase. In the prediction phase, lifestyle-related diseases and the like are predicted based on the prediction model for the newly measured pulse wave signal.

[0059] In step S11, the pulse wave signal acquisition unit 202 acquires a digital pulse wave signal corresponding to the pulse rate of the living body measured by the pulse wave detection unit 104.

[0060] Next, in step S12, the pulse wave waveform generation unit 204 generates pulse wave waveform image data from the digital pulse wave signal.

[0061] Next, in step S13, the feature extraction unit 206 extracts features from the pulse wave waveform.

[0062] Next, in step S14, the prediction unit 210 predicts the diagnosis of lifestyle-related diseases, etc., based on the extracted features and the prediction model stored in the memory unit 102. The prediction result is data showing the diagnosis result (for example, normal or abnormal) based on the indicators of the blood test. Then, in step S15, the output unit 211 outputs the prediction result. The output result may be, for example, a screen display of a message indicating the possibility of lifestyle-related diseases, information indicating normal or abnormality for a specific indicator of the blood test, or a screen display showing the probability that the subject has a lifestyle-related disease as a percentage.

[0063] (Machine Learning) Next, the machine learning procedure when an automatic blood pressure monitor is used as the pulse wave detection unit 104 will be explained. As an example, the case in which the pulse wave detection unit 104 is configured as a combination of an electrocardiogram transmitter and an automatic blood pressure monitor, and simultaneous measurement of electrocardiogram and blood pressure is performed will be explained. Figure 4 shows an example of measurement results when blood pressure is measured in the upper arm by the analysis device 100. Figure 4(a) shows the change in cuff pressure (mmHg) with respect to time (s) when measuring the blood pressure pulse wave using the oscillometric method with an automatic blood pressure monitor as curve 402. Figure 4(b) shows the blood pressure pulse wave amplitude (mmHg) output from the electrocardiogram transmitter. During the period when the intravascular pressure exceeds the cuff pressure, the blood vessels dilate, and the change in the volume of the blood vessels increases the internal pressure of the cuff. In the oscillometric method, blood pressure is determined by observing the fluctuations in cuff pressure synchronized with the heartbeat during the process of depressurizing the cuff. The fluctuations in cuff pressure observed here are measured as the blood pressure pulse wave.

[0064] In Figure 4(b), the pulse wave waveform 404 begins to increase in amplitude from the point where the cuff pressure curve 402 is at its maximum. Blood pressure measurement is performed within the range of frame 406, which includes the period from the increase to the decrease in amplitude. The analysis device 100 extracts data intervals from this range that have a predetermined number of local maximum amplitude values ​​(peaks), including the period before and after the maximum amplitude value. The data interval with 10 peaks enclosed by frame 408 is used as the target of analysis. The number of peaks in the data interval to be analyzed is not limited to 10; it may include 5 (total peaks 11) to 8 (total peaks 17) before and after the maximum peak.

[0065] Next, with reference to Figure 5, the method for measuring pulse wave characteristics will be explained. Figure 5(a) is a graph showing an example of the results of measuring blood pressure in the upper arm using the analysis device 100. As indicated by the arrow on the left, the cuff pressure at the point when the pulse wave rapidly increases and the intravascular pressure exceeds the cuff pressure is measured as the systolic blood pressure. Also, as indicated by the arrow on the right, the cuff pressure at the point when the change in the pulse wave disappears is measured as the diastolic blood pressure.

[0066] Figure 5(b) shows an example of measurement results when blood pressure is measured at the wrist using a wrist-worn blood pressure monitor. In this figure, the horizontal axis represents time (s). Curve 502 shows the cuff pressure (mmHg) of the automatic blood pressure monitor, and curve 504 shows the blood pressure pulse wave amplitude (mmHg) output from the electrocardiogram transmitter. When measuring blood pressure at the wrist, unlike when measuring it on the upper arm, the cuff pressure gradually increases, while the amplitude of the pulse wave waveform gradually increases and decreases. A pressure sensor is provided between the cuff pressure and the wrist surface of the wrist-worn blood pressure monitor. By measuring the pressure near the artery during the process of increasing the cuff pressure using the pressure sensor, the fluctuation in cuff pressure is measured as a blood pressure pulse wave.

[0067] Next, referring to Figure 6, we will explain how to extract features from the pulse wave waveform. Figure 6 is a graph of the pulse wave waveform. Curve 602 is the pulse wave waveform, with the horizontal axis representing time (s) and the vertical axis representing blood pressure (mmHg). DBP (diastolic blood pressure) is the pressure exerted on the blood vessel wall by the blood flow towards the heart when blood returns from the body to the heart (diastole). a shows the peak value of the driving pressure wave, which increases with the increase in blood flow velocity. The reflected wave arrives at the point when peak value a is observed, and the blood pressure is further informed. b shows the peak value observed by the reflected wave from the periphery when the blood flow is decelerating. This value is the pressure exerted on the aortic wall by the blood flow ejected from the heart when blood is sent from the heart to the aorta and the rest of the body (systole), and is also called SBP (systolic blood pressure).

[0068] Feature T5 represents the time from the point at which blood pressure begins to rise (DBP) until it reaches 5% of the peak value b. Feature T20 represents the time from the point at which blood pressure begins to rise (DBP) until it reaches 20% of the peak value b. Feature T-time represents the time (s) from the point at which blood pressure begins to rise until it reaches 50% of the peak value b. Feature U-time represents the time (s) from the point at which blood pressure begins to rise until it reaches the peak value b.

[0069] Figure 7 shows another example of features extracted from a pulse wave. AP (Augmented Pressure) is the reflected wave component and is determined by b-a. The feature AI (Augmentation Index) is defined as the ratio of AP to the peak value b.

[0070] Furthermore, features such as pulse rate (PR), RR interval (the interval between R waves generated during ventricular excitation), advancement index (AI), cardio-ankle vascular index (CAVI), pulse wave velocity (PWV), and ankle-brachial index (ABI) can be used.

[0071] As statistical measures for these features, the mean, standard deviation (SD), maximum value, minimum value (Min), median, and coefficient of variation (CV) of pulse wave intervals can be used. The coefficient of variation (CV) is the standard deviation divided by the mean, and this value is used to relatively evaluate the variability between data points.

[0072] Figure 8 conceptually illustrates the process of performing machine learning using decision tree analysis as an example of machine learning. Note that Figure 8 is a conceptual diagram of a decision tree and has been simplified for ease of explanation. In decision tree analysis, the target variable is set as "normal" and "abnormal". "Abnormal" indicates a judgment result such as lifestyle-related disease. The leftmost node in the diagram is the top-level node that initially receives the entire dataset and is called the root node. The branches represent the answers (yes / no) to the questions. The node at the very end where the branching stops shows the final predicted value and is called a leaf node.

[0073] The explanatory variables in the decision tree were features extracted from the subjects' pulse wave waveforms. The parenthetical notation in the feature notation indicates the statistics of that feature, including the standard deviation (SD), coefficient of variation (CV), or minimum value (Min). In the example in Figure 8, the explanatory variables are set as the standard deviation of pulse rate (PR(SD)), the coefficient of variation of AI (AI(CV)), the minimum value of the RR interval (*1 / 32(s)) (RRI(Min)), the standard deviation of T-Time (TT(SD)), and the standard deviation of U-Time (UT(SD)).

[0074] Feature data is associated with judgment result data by subject identifiers. In machine learning, the pulse wave waveform feature data and corresponding subject judgment result data included in the training data are taken as input, and a decision tree is output that reproduces the prediction as well as possible based on this data. First, the entire training data is used to create the root node. Next, the best split is found. In the split, the values ​​of the features that determine whether the judgment result is yes or no are specified. Here, features that split the data into the purest subsets may be selected based on criteria such as entropy. Next, the training data is split based on the selected features. Then, the best split is found again in each subset, and the splitting continues. This is repeated recursively until the leaf nodes are reached. In this way, a decision tree is created.

[0075] The analysis of subject data using test data is performed from left to right in the diagram. First, it is determined whether the PR (SD) is 0.425 or less. If the result is true, it is determined whether the AI ​​(CV) is 0.111 or less. If the result is true, it is determined to be "normal," and if it is false, it is determined to be "abnormal."

[0076] On the other hand, if PR(SD) is greater than 0.425, it is determined whether RRI(Min) is 26.629 or less. If the result is true, it is determined to be "normal". If the result is false, it is determined whether TT(SD) is 3.636 or less. If this result is true, it is determined to be "normal". Also, if this result is false, it is determined whether UT(SD) is 0.726 or less. If this result is true, it is determined to be "normal". On the other hand, if this result is false, it is determined to be "abnormal".

[0077] (Example 1) Next, we will describe the first example in which machine learning was performed using a decision tree. A decision tree is a supervised learning analysis method that creates a tree structure model of the explanatory variables (variables that explain the objective) that influence the target variable (variable to be predicted), and analyzes the data using this tree structure. In this example, LightGBM, published by Microsoft Corporation, was used as the algorithm for creating the decision tree.

[0078] First, blood tests were performed on the subjects, and for each indicator, the cutoff value was compared with the measured value. Based on this comparison, a judgment result (normal or abnormal) was determined. The number of subjects varied for each indicator: 138 for LDL-C, 133 for NT-proBNP, 146 for eGFR, and 109 for HbA1c. Judgment result data showing the judgment results was created in this way and entered into the analysis device, associated with the subject identifier.

[0079] Next, the blood pressure and pulse wave of each subject were measured using an Omron HEM-907 blood pressure monitor. The blood pressure and pulse wave were measured three times during each examination, and the features were quantified. For each subject's measurement, a table of feature data was generated, as shown in the example in Table 2 below. The feature data was also stored in memory, associated with the subject identifier. In Table 2, the first column shows the name of the feature, and the first row shows the name of the statistic.

[0080]

[0081] For subjects judged normal and those judged abnormal based on blood tests, 70% of the judgment result data was used as training data and 30% as test data. Machine learning was then performed using the training data, and a predictive model was generated using a decision tree.

[0082] Using the prediction model generated by the decision tree in this way, we evaluated the predictions using test data.

[0083] When performing machine learning, there are random elements in setting the initial values ​​of parameters. Therefore, the entire process from the learning phase to the prediction phase was considered one run, and the performance was evaluated by averaging five runs while changing the initial values. The mean ± standard deviation of the evaluation results is shown in Table 3.

[0084]

[0085] In Table 3, the AUC (Area Under the Curve) is obtained based on the ROC curve (Receiver Operating Characteristic curve). The ROC curve is represented as a curve as shown in the example in Figure 9. ROC curve 902 is a graph plotting the true positive rate on the vertical axis and the false positive rate on the horizontal axis. The true positive rate is the percentage of data predicted to be abnormal for a certain indicator that actually had an abnormal diagnostic result (correct value) in the blood test. The false positive rate is the percentage of data predicted to be abnormal for a certain indicator that actually had a normal diagnostic result (correct value) in the blood test. The area of ​​the gray region below the ROC curve represents the AUC value. Both the horizontal and vertical axes take values ​​between 0.0 and 1.0. Therefore, the minimum value of AUC is 0.0 and the maximum value is 1.0, and the closer the AUC is to 1.0, the better the prediction accuracy.

[0086] The blood test indicators used were LDL-C (cutoff values ​​of 100 and 140), NT-proBNP (cutoff values ​​of 55, 125, and 300), eGFR (cutoff values ​​of 45 and 60), and HbA1c. For each indicator, a binary classification of normal or abnormal was performed.

[0087] Furthermore, sensitivity is the proportion of data where the machine learning model predicted a normal result out of all data where the blood test result was normal. Specificity is the proportion of data where the machine learning model predicted an abnormal result out of data where the blood test result was abnormal. From the results in Table 3, it can be seen that the prediction accuracy for lifestyle-related diseases, etc., using this embodiment is high.

[0088] (Example 2) Next, we will describe a second example in which machine learning was performed using One-Class-SVM. One-Class-SVM is an analytical method that does not classify data into two classes, but rather determines whether data is similar or not based on whether it belongs to one class, and learns from this. One-Class-SVM is an unsupervised learning method that learns only the features of normal data and uses the boundary between the normal class and the abnormal class as a criterion to detect data that deviates from the criterion as abnormal. Therefore, it is possible to detect outliers even without correct labels such as "normal" or "abnormal".

[0089] In this example, validation was performed using the holdout method. In the holdout method, the feature data was divided into training data, validation data, and test data. A model was then generated using the training data, and validation was performed using the validation data each time a model was generated. During the validation phase, hyperparameters to control the model were adjusted. The hyperparameters include nu, which controls the proportion of above-average data, and gamma, which specifies the degree to which support vectors influence the boundary. For hyperparameter adjustment, Optuna, an open-source software framework for automating hyperparameter optimization provided by Preferred Networks, was used. The test data was used to finally test the predictive model.

[0090] First, blood tests were performed on 138 subjects. A cutoff value of 140 was set for the indicator LDL-C, and the measured values ​​were compared with the set value. Based on the results of this comparison, a diagnosis (normal or abnormal) was determined. In this way, judgment result data showing the judgment results was created and input into the analysis device. Figure 10 is a graph showing the number of patients for each LDL-C value (mg / dL). In the figure, the horizontal axis shows the LDL-C value category, and the vertical axis shows the number of patients. There were 126 patients with LDL less than 140, and 11 patients with LDL-C of 140 or more.

[0091] This data was divided into training data, validation data, and test data. Figure 11 shows the number of patients for each LDL-C value in the training data. In the figure, there are 126 patients with an LDL value less than 140 and 11 patients with an LDL-C value of 140 or more. First, all the training data is set as a cluster of normal data, and the data is mapped to a feature space in a high-dimensional space so that the origin belongs to the cluster of abnormal data. Through this operation, the training data is mapped so that it is placed far from the origin. On the other hand, data that is not similar to the original training data is grouped near the origin. This property is used to distinguish between normal and abnormal data.

[0092] Figure 12 shows the number of patients for each LDL-C value in the training data compared to the validation data. For the validation data, data from two patients was used for each LDL-C value range, which was divided into intervals of 10 mg / dL.

[0093] Figure 13 shows the number of patients for each LDL-C value in the test data. For patients with an LDL-C value of less than 140, data from two patients was used from each category, and for patients with an LDL-C value of 140 or more, data from nine patients was used.

[0094] Next, the blood pressure pulse wave of each subject was measured using an Omron HEM-907 blood pressure monitor. The blood pressure pulse wave was measured three times during each examination, and its characteristic features were quantified. The following values ​​were used as characteristic features obtained from the blood pressure pulse wave.

[0095] (1) Maximum / minimum pulse rate (maximum value divided by minimum value) (2) Maximum / minimum standard deviation of inter-peak interval (IPI; Inter-Peak-Interval) (IPI(SD)) (3) Maximum / minimum standard deviation of inter-systolic peak interval (ISI; Inter-Systolic-Peak-Interval) (ISI(SD)) Next, based on the above features (1) to (3), 15 features shown in Table 4 were defined.

[0096]

[0097] Next, five features were selected from the above fifteen features to generate a predictive model for predicting the outcome of a blood test diagnosis (normal / abnormal). The predictive model generated by this operation 15 C5 equals 3003 possibilities. For these prediction models, the results of diagnoses based on blood tests were predicted using validation data and test data. Prediction models with high AUC values ​​for both the prediction results using test data and the prediction results using validation data were evaluated as models that can better adapt to new data, i.e., models with high generalization performance.

[0098] Table 5 shows the evaluation results for the predictive models that were evaluated as having high generalization performance.

[0099]

[0100] In the table, accuracy indicates the percentage of prediction results that match the true value. Sensitivity is calculated by dividing the number of correct predictions where both the actual value and the predicted value were "abnormal" by the total number of data points where the actual value was "abnormal". Specificity refers to the number of correct predictions where both the actual value and the predicted value were "normal" by the total number of data points where the actual value was "normal". The second column from the right in the table, AUC, shows the AUC value when predictions were made using the test data, and the far right column, "Validation-AUC," shows the AUC value when predictions were made using the validation data. The values ​​for accuracy, sensitivity, specificity, AUC, and validation-AUC were obtained by a prediction model trained and generated using the multiple features shown in the leftmost column.

[0101] The results in Table 5 show that by combining existing features to generate new ones, it is possible to create a predictive model with higher generalization performance.

[0102] (Example 3) Next, a third example in which machine learning was performed using a decision tree will be described. In this example, LightGBM was used as the algorithm for creating the decision tree.

[0103] First, blood tests were performed on the subjects, and for each indicator, the cutoff value was compared with the measured value. Based on this comparison, a diagnosis (normal or abnormal) was determined. The number of subjects varied for each indicator: 115 for LDL-C, 110 for NT-proBNP, 126 for eGFR, and 87 for HbA1c. In this way, diagnoses were created, and the result data was entered into the analysis device, associated with the subject identifier.

[0104] Next, the blood pressure pulse wave of each subject was measured using an A&D UM-212 blood pressure monitor. The blood pressure pulse wave was measured three times during each examination, and its features were quantified. For each subject's measurement, a table of feature data, as shown in Table 2 above, was generated.

[0105] For subjects judged normal and those judged abnormal based on blood tests, 85% of the judgment result data was used as training data and 15% as test data. Machine learning was then performed using the training data, and a predictive model was generated using a decision tree.

[0106] Using the prediction model generated by the decision tree in this way, we evaluated the predictions using test data.

[0107] The entire process from the learning phase to the prediction phase was considered one run, and performance was evaluated by averaging five runs while changing the initial values. The mean ± standard deviation of the evaluation results is shown in Table 6.

[0108]

[0109] The blood test indicators used were LDL-C (cutoff values ​​of 100 and 130), NT-proBNP (cutoff values ​​of 55, 125, and 300), eGFR (cutoff values ​​of 45 and 60), and HbA1c (cutoff values ​​of 6.5 and 5.5). For each indicator, a binary classification of normal or abnormal was performed.

[0110] The results in Table 6 show that this embodiment has a high accuracy in predicting lifestyle-related diseases.

[0111] While embodiments of the present disclosure have been described above, the apparatus, its components, and the steps of the method described above according to these embodiments can be implemented by hardware or by a combination of software and hardware. Whether the functions of the apparatus or its components are performed by hardware or by software depends on the design constraints of this embodiment. Those skilled in the art can implement the functions of the components described above for specific applications using various methods, and such modifications are also within the scope of the present disclosure.

[0112] The apparatus described in this embodiment is merely illustrative and can be implemented in other ways. For example, the components described above are logical divisions of functions, and the components may be divided in other ways during implementation. Furthermore, two or more of the components described above may be integrated into one, and each component may exist physically independently.

[0113] When the functions described in this embodiment are implemented in software form, the computer program for realizing those functions can be stored on a computer-readable storage medium. The computer program may include several instructions for instructing the computer to function as all or some of the components described in this embodiment.

[0114] 100 Analysis device 101 Analysis unit 102 Memory unit 104 Pulse wave detection unit 106 Machine learning program 108 Prediction program 110 Amplification circuit 111 A / D conversion circuit 112 Control unit 113 Atrial fibrillation detection unit 114 Memory unit 116 Input unit 118 Notification unit 120 Power supply 128 Bus 202 Pulse wave signal acquisition unit 204 Pulse wave waveform generation unit 206 Feature extraction unit 208 Model generation unit 210 Prediction unit 211 Output unit

Claims

1. A pulse wave signal analysis device comprising: an acquisition unit for acquiring a pulse wave signal corresponding to the pulse rate of a living organism; a pulse wave waveform generation unit for generating a pulse wave based on the acquired pulse wave signal; a feature extraction unit for extracting feature quantities of the generated pulse wave; and a model generation unit for generating a predictive model that predicts the result of a diagnosis based on a blood test, by performing machine learning on the extracted feature quantities using the results of a judgment based on a blood test of the living organism, and predicting the result of a diagnosis based on a blood test from newly input pulse wave feature quantities.

2. The pulse wave signal analyzer according to claim 1, wherein the diagnosis is a diagnosis of a lifestyle-related disease, a cardiovascular disease, or a chronic kidney disease.

3. The pulse wave signal analyzer according to claim 1, wherein the feature quantity includes at least one of the rise time of the maximum peak of the pulse wave, pulse rate, peak interval, and increase index.

4. The pulse wave signal analyzer according to claim 1, wherein the feature quantities include statistics on the rise time of the maximum peak of the pulse wave, pulse rate, peak interval, and increase index.

5. The pulse wave signal analyzer according to claim 4, wherein the statistical measure includes at least one of the standard deviation, mean, maximum, minimum, median, and coefficient of variation.

6. A pulse wave signal analysis device comprising: an acquisition unit for acquiring a pulse wave signal corresponding to the pulse rate of a living organism; a pulse wave waveform generation unit for generating a pulse wave based on the acquired pulse wave signal; a feature extraction unit for extracting feature quantities of the generated pulse wave; a prediction unit for predicting the result of a diagnosis based on a blood test on the living organism based on the extracted feature quantities and a prediction model generated by the pulse wave signal analysis device described in claim 1; and an output unit for outputting the result predicted by the prediction unit.

7. A method for analyzing a pulse wave signal, comprising the steps of: acquiring a pulse wave signal corresponding to the pulse rate of a living organism; generating a pulse wave based on the acquired pulse wave signal; extracting feature quantities of the generated pulse wave; and generating a predictive model that predicts the result of a diagnosis based on a blood test from newly input pulse wave feature quantities by performing machine learning on the extracted feature quantities using the result of a judgment based on a blood test of the living organism.

8. The method for analyzing pulse wave signals according to claim 7, wherein the diagnosis is a diagnosis of a lifestyle-related disease, a cardiovascular disease, or a chronic kidney disease.

9. The method for analyzing a pulse wave signal according to claim 7, wherein the feature quantity includes at least one of the rise time of the maximum peak of the pulse wave, pulse rate, peak interval, and increase index.

10. The method for analyzing a pulse wave signal according to claim 7, wherein the feature quantities include statistics on the rise time of the maximum peak of the pulse wave, pulse rate, peak interval, and increase index.

11. The method for analyzing a pulse wave signal according to claim 10, wherein the statistic includes at least one of the standard deviation, mean, maximum, minimum, median, and coefficient of variation.

12. A computer program that causes a computer to function as a pulse wave signal analysis device according to any one of claims 1 to 6.

13. A computer-readable storage medium storing a computer program that causes a computer to function as a pulse wave signal analysis device according to any one of claims 1 to 6.

Images