Hypotension risk prediction device, hypotension risk prediction system, hypotension risk prediction method, and hypotension risk prediction program

Abstract

To provide a low-cost hypotension prediction device which allows medical staff with little experience to recognize that a patient is at a high risk of hypotension.SOLUTION: The hypotension risk prediction device predicts a hypotension risk of a patient from biometric information obtained from the patient, and comprises: an image accepting section configured to accept input of a biometric information monitor image displayed on a monitoring screen 30 of a biometric information monitor that acquires the biometric information of the patient; a computing section configured to compute information about the hypotension risk of the patient upon input of the accepted biometric information monitor image; and an output section configured to output the computed information about the hypotension risk, the computing section being provided with a trained computation model obtained through machine learning using teacher data including, as input, biometric information monitor images of a subject and, as output, information relating to whether hypotension occurs in the subject, so that information about the hypotension risk is computed when the biometric information monitor image is input.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a hypotension risk prediction device, a hypotension risk prediction system, a hypotension risk prediction method, and a hypotension risk prediction program. [Background technology]

[0002] When general anesthesia is administered, sudden changes in breathing and hemodynamics occur, which may result in postinduction hypotension (PIH), which carries the risk of serious complications such as cerebral infarction.

[0003] Hypotension is also a problem in many situations, not just when inducing general anesthesia, but also in the operating room, intensive care unit, emergency situations, and sudden changes in the ward, etc. It is recognized as a major problem that it is difficult for non-experts to determine which cases are at high risk of hypotension.

[0004] Traditionally, PIH has been predicted by doctors such as anesthesiologists by integrating patient information such as age and gender, medication information such as the type and amount of anesthetic used, physical information of the patient obtained during the examination, and data from vital signs monitors.

[0005] Furthermore, Patent Document 1 proposes a method for monitoring a patient's arterial pressure and alerting medical personnel to predicted future hypotension events in the patient, which involves multiple steps to predict hypotension for a certain threshold setting. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Special Publication No. 2023-515151 [Non-patent literature]

[0007] [Non-Patent Document 1] Smilowitz, NR, Gupta N, Ramakrishna H, Guo Y, Berger JS, Bangalore S, et al., Trends in Perioperative Major Adverse Cardiovascular and Cerebrovascular Events associated with Non-Cardiac Surgery, JAMA Cardiol, 2017;2(2):181-187. [Non-patent document 2] Salmasi V, Maheshwari K, Dongsheng Y, Mascha EJ, Singh A, Sessler DI, et al., Relationship between Intraoperative Hypotension, Defined by Either Reduction from Baseline or Absolute Thresholds, and Acute Kidney and Myocardial Injury after Noncardiac Surgery: A Retrospective Cohort Analysis, Anesthesiology, 2017;126(1):47-65. [Non-patent document 3] Selvaraju et.al., Grad-cam: Visual explanations from deep network via gradient-based localization, ICCV, 2017, 618-626. Summary of the Invention [Problem to be solved by the invention]

[0008] However, the above PIH predictions do not necessarily reflect the patient's actual condition. Properly integrating the above information, including data from patient monitors, to determine the risk of PIH requires advanced expertise and experience. However, the diagnostic experience and skills of anesthesiologists and other physicians vary, making it difficult for inexperienced doctors and nurses to predict the onset of PIH. As a result, inexperienced medical professionals may miss the dangerous signs displayed on patient monitors.

[0009] Furthermore, the hypotension prediction method proposed in Patent Document 1 and elsewhere requires multiple steps and requires dedicated equipment and software, which are expensive and therefore costly to introduce into medical settings.

[0010] Therefore, there is a need for a low-cost hypotension prediction device that enables even inexperienced medical personnel, rather than experienced physicians, to recognize that a patient is at high risk of hypotension. [Means for solving the problem]

[0011] The gist of the present invention is as follows. (1) A hypotension risk prediction device for predicting a hypotension risk of a patient based on biological information acquired from the patient, an image receiving unit that receives an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; a calculation unit that calculates information regarding a risk of hypotension of the patient when the received biological information monitor image is input; and an output unit that outputs information related to the calculated hypotension risk Equipped with the calculation unit has a trained calculation model that has been subjected to machine learning using training data including, as an input, a biological information monitor image of the subject and, as an output, information regarding whether or not the subject has developed hypotension, so that information regarding the hypotension risk is calculated when the biological information monitor image is input. Hypotension risk prediction device. (2) The hypotension risk prediction device according to (1) above, wherein the information relating to the hypotension risk includes information displayed as an image, a number, a letter, a symbol, a sound, or a combination thereof. (3) The information regarding the hypotension risk includes the level of the patient's risk of developing hypotension, the patient's risk score for developing hypotension, a heat map showing the areas among the partial areas included in the vital sign monitor image that affect the calculation of the information regarding the hypotension risk, or a combination thereof, in the hypotension risk prediction device described in (1) or (2) above. (4) A hypotension risk prediction device described in any of (1) to (3) above, wherein the vital sign monitor image is a captured image of the monitoring screen of the vital sign monitor, a captured image of the monitoring screen displayed on the display of the hypotension risk prediction device connected to the vital sign monitor, or a captured image of the monitoring screen displayed on the display of a computer connected to the vital sign monitor. (5) A hypotension risk prediction device described in any of (1) to (4) above, wherein the vital sign monitor image includes an electrocardiogram waveform, an arterial pressure waveform, a transcutaneous oxygen saturation waveform, an exhaled carbon dioxide concentration waveform, or a combination thereof. (6) The hypotension risk prediction device according to any one of (1) to (5) above; The vital sign monitor; A hypotension risk prediction system comprising: (7) A method for predicting a hypotension risk of a patient based on biological information acquired from the patient, comprising: receiving an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; inputting the received biological information monitor image and calculating information regarding the patient's risk of hypotension; and outputting information about the calculated hypotension risk; Including, the calculation is performed using a trained calculation model that has been subjected to machine learning using training data including, as an input, a biological information monitor image of the subject and, as an output, information regarding the presence or absence of hypotension in the subject, so that information regarding the hypotension risk is calculated when the biological information monitor image is input. Methods for predicting hypotension risk. (8) A hypotension risk prediction program for predicting a hypotension risk of a patient based on biological information acquired from the patient, an image receiving function for receiving an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; and a calculation function for calculating information regarding the patient's risk of hypotension when the received biological information monitor image is input; Including, the calculation function has a trained calculation model that has been subjected to machine learning using training data that includes, as an input, a biological information monitor image of the subject and, as an output, information on whether the subject has developed hypotension, so that information on the hypotension risk is calculated when the biological information monitor image is input. Hypotension risk prediction program. [Effects of the Invention]

[0012] According to the present invention, it is possible to provide a low-cost hypotension prediction device that enables even inexperienced medical personnel to recognize that patients who have previously been judged to be at high risk of hypotension by experienced doctors looking at vital signs monitors are at high risk of hypotension. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a schematic diagram showing an example of the configuration of a hypotension prediction system including the present hypotension risk prediction device. [Figure 2] FIG. 2 is a diagram showing an example of the configuration of the hypotension risk prediction device. [Figure 3] FIG. 3 shows an example of a biological information monitor image measured on a patient. [Figure 4]FIG. 4 is a biological information monitor image in which the colors other than the color displaying the biological information waveform of the biological information monitor image of FIG. 3 have been converted to black. [Figure 5] FIG. 5 is a configuration diagram of an example of a training data collection system when collecting training data TD used in machine learning. [Figure 6] FIG. 6 is a configuration diagram of an example of a machine learning system when machine learning is performed. [Figure 7] FIG. 7 is a schematic diagram showing an example of the calculation model M1. [Figure 8] FIG. 8 is a schematic diagram showing another example of the calculation model M1. [Figure 9] FIG. 9 is a configuration diagram of an example of a hypotension prediction system when predicting a patient's hypotension risk using the hypotension risk prediction device. [Figure 10] FIG. 10 is a heat map calculated based on the vital sign monitor image of FIG. 4 using the hypotension risk prediction device. [Figure 11] FIG. 11 shows a heat map calculated using a biological information monitor image including only one type of arterial pressure waveform in generating a trained calculation model and predicting the risk of hypotension, superimposed on the biological information monitor image. [Figure 12] Figure 12 shows a heat map calculated using a vital sign monitor image containing two types of waveforms, an electrocardiogram waveform and an arterial blood pressure waveform, from top to bottom, in generating a trained calculation model and predicting the risk of hypotension, as well as the vital sign monitor image. [Figure 13] Figure 13 shows a heat map calculated using a vital sign monitor image containing three types of waveforms, namely, an electrocardiogram waveform, an arterial pressure waveform, and a transcutaneous oxygen saturation waveform, from top to bottom, for generating a trained calculation model and predicting the risk of hypotension, as well as the vital sign monitor image. DETAILED DESCRIPTION OF THE INVENTION

[0014] The present disclosure relates to a hypotension risk prediction device that predicts a patient's risk of hypotension based on biological information acquired from the patient, the device comprising: an image receiving unit that receives input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires the patient's biological information; a calculation unit that calculates information related to the patient's hypotension risk when the received biological information monitor image is input; and an output unit that outputs the calculated information related to the hypotension risk, wherein the calculation unit has a trained calculation model that has been subjected to machine learning using training data that includes, as input, the biological information monitor image of the subject and, as output, information related to whether the subject has developed hypotension, so that, when the biological information monitor image is input, the information related to the hypotension risk is calculated.

[0015] Conventional deep learning models use algorithms that refer to numerical values obtained from waveform electrical signals.This device uses images obtained from vital signs monitoring screens that medical professionals actually refer to, and creates a trained calculation model (prediction model) using the same data that medical professionals see in their daily medical practice as training data.

[0016] The hypotension risk prediction device (hereinafter referred to as the present device) disclosed herein can predict the risk of hypotension using vital signs displayed on the vital signs monitoring screen of a vital signs monitor, which has not traditionally been used by physicians to predict PIH. This device allows even inexperienced medical professionals to detect dangerous signs that appear on the vital signs monitoring screen and to appropriately manage hypotension. The present device is also inexpensive because it can use conventional vital signs monitors without the need for expensive dedicated equipment and software.

[0017] This device allows even inexperienced doctors, nurses, and other unskilled medical professionals to recognize that a patient is at high risk of hypotension, so for patients at high risk of developing hypotension, they can reconsider the method of administering general anesthetics, wait for an experienced doctor to arrive, or make sufficient preparations such as vasopressor drugs, thereby preventing the onset of hypotension.

[0018] This device can be used in cases where general anesthetics can be administered, or in cases where procedures pose a risk of hypotension. According to the 2022 Statistics on Social Medical Treatment Procedures compiled by the Ministry of Health, Labor and Welfare, 140,000 surgical operations using general anesthesia are performed annually in Japan. The incidence of perioperative neurovascular complications is reported to be 3.0% (Non-Patent Document 1), which means that in Japan alone, 4,200 people suffer from complications annually. This device is expected to reduce the incidence of these complications.

[0019] Furthermore, since this device uses images from a vital sign monitor as input, it can be used easily and at low cost across manufacturers of vital sign monitors that have different formats for the vital sign data they measure. The images from the vital sign monitor are still images.

[0020] The hypotension prediction accuracy (AUC: Area Under the ROC Curve) of this device is preferably 0.52 or higher, more preferably 0.58 or higher, even more preferably 0.62 or higher, even more preferably 0.65 or higher, even more preferably 0.67 or higher, and even more preferably 0.69 or higher. AUC is one of the evaluation indices for binary classification tasks. By having the above-mentioned preferable prediction accuracy, this device can be used for screening cases that are clinically high risk.

[0021] In this application, hypotension refers to a mean arterial pressure (MAP) of less than 65 mmHg. That is, in this device, a MAP of 65 mmHg is set as the threshold for hypotension.

[0022] Table 1 shows the experimental results of examining AUC for each MAP threshold using this device. For a dataset containing vital signs monitor images, data on the occurrence of hypotension, and MAP values, the AUC was calculated by changing the output labels for the occurrence and absence of hypotension events depending on the threshold.

[0023] [Table 1]

[0024] The highest prediction accuracy (AUC) was observed when the MAP threshold was set at 65 mmHg. Non-Patent Document 2 reports that maintaining MAP at 65 mmHg or higher reduces the incidence of perioperative complications, and the MAP threshold of 65 mmHg is the same as the value that medical professionals are clinically aware of.

[0025] The method for measuring prediction accuracy (AUC) when considering the MAP threshold is explained below. The true positive rate (TPR) = (predicted occurrence of hypotension) / (correct answer: hypotension occurred), and the false positive rate (FPR) = (predicted occurrence of hypotension) / (correct answer: no hypotension occurred) are used to calculate the true positive rate (TPR) and false positive rate (FPR), and a receiver operating characteristics curve (ROC) is obtained by plotting the true positive rate (TPR) on the vertical axis and the false positive rate (FPR) on the horizontal axis, which shows the relationship between TPR and FPR.

[0026] A learning model was created by machine learning using patient monitor images containing one cycle of the arterial pressure waveform from subjects with hypotension and one cycle of the arterial pressure waveform from subjects without hypotension as training data. In the preliminary experiment shown in Table 1, patient monitor images from 53 subjects with hypotension and 29 subjects without hypotension were used as training data. A single cycle of the arterial pressure waveform refers to one cycle of the 7-9 cycles contained in the patient monitor image. Heart rates are approximately 60-100 beats per minute, and approximately 1-1.4 cycles of the arterial pressure waveform are displayed per second. However, the horizontal axis of the patient monitor monitor screen displays approximately 7 seconds, so approximately 7-9 cycles of the waveform are displayed on the monitoring screen. When measuring the AUC for determining the MAP threshold, one cycle of the waveform was extracted from the approximately 7-9 cycles.

[0027] The created learning model is input into a validation dataset, and the output is whether hypotension occurred (P) or not (N). The learning model for classifying hypotension as occurring (P) or not (N) is used to classify patient monitor image data containing one cycle of arterial pressure waveforms from multiple patients, and the predicted probability of hypotension occurring (P) for each patient monitor image is output. The predicted probability of hypotension occurring (P) is used as a threshold to obtain an ROC curve showing the true positive rate (TPR) and false positive rate (FPR). The AUC can be calculated from the obtained ROC curve. In the preliminary experiment shown in Table 1, patient monitor image data containing one cycle of arterial pressure waveforms from 23 patients with hypotension and 13 patients without hypotension were used.

[0028] FIG. 1 is a schematic diagram showing an example of the configuration of a hypotension prediction system 100 (hereinafter also referred to as the present system) including the present device 10. The present device 10 can be connected to a vital sign monitor that measures vital signs of a patient or subject. A patient is someone whose hypotension risk is predicted using the present device 10, and a subject is someone from whom training data is obtained in the present device 10.

[0029] Biological information monitors have been commonly used in medical settings and can display measured biological information waveforms. The biological information waveforms are preferably electrocardiogram waveforms, arterial pressure waveforms, transcutaneous oxygen saturation (hereinafter also referred to as SPO2) waveforms, exhaled carbon dioxide concentration waveforms, or a combination thereof, and more preferably electrocardiogram waveforms, arterial pressure waveforms, SPO2 waveforms, or a combination thereof. Any biological information monitor can be used as long as it can display the above-mentioned preferred biological information waveforms on the monitoring screen 30, such as the intelliVue MX800 manufactured by Philips. The horizontal axis of the biological information waveform graph displayed on the monitoring screen 30 represents time, and the graph is a moving image that moves from right to left on the screen in real time. A still image of the biological information monitor image obtained from the moving image displayed on the monitoring screen 30 can be used as input.

[0030] The biological information waveform contained in the biological information monitoring image to be learned is one or more types, preferably two or more types, more preferably three or more types, and even more preferably four types. The biological information waveform contained in the biological information monitoring image to be learned is preferably an electrocardiogram waveform, an arterial pressure waveform, an SPO2 waveform, an exhaled carbon dioxide concentration waveform, or a combination thereof, more preferably an electrocardiogram waveform, an arterial pressure waveform, an SPO2 waveform, or a combination thereof. The biological information waveform contained in the biological information monitoring image may be a waveform for multiple cycles contained in the entire biological information monitoring image, or a waveform for one cycle out of a plurality of waveform cycles.

[0031] Many models of patient monitors display four vital signs waveforms: electrocardiogram, arterial pressure, SPO2, and respiratory status, in addition to numerical values indicating blood pressure, etc. In daily medical practice, the three signs, electrocardiogram, SPO2, and respiratory status, are displayed on the patient monitor in almost all cases, while the four signs, electrocardiogram, arterial pressure, SPO2, and respiratory status, may be displayed in cases requiring attention at the discretion of medical professionals.

[0032] Experienced medical professionals determine treatment plans by taking into account the patient's condition based on information from these four types of waveforms. However, inexperienced medical professionals may pay attention only to the values displayed on the vital signs monitor and miss important information obtained from the waveforms, which may increase the risk of complications such as hypotension. Depending on the patient's underlying disease, the importance of certain vital signs waveform information may vary. By using deep learning to analyze one or more, preferably two or more, more preferably three or more, and even more preferably four of the above waveforms, it becomes possible to determine treatment plans similar to those of experienced medical professionals.

[0033] The device 10 and the vital sign monitor may be connected by wire or wirelessly, for example, via Bluetooth (registered trademark), which allows for the transmission and reception of vital sign monitor image data displayed on the monitoring screen 30 of the vital sign monitor, or may be connected via a network 20 which allows for the transmission and reception of vital sign monitor image data. The network 20 is a communications network such as the Internet. When the device 10 and the vital sign monitor are connected as described above, the vital sign monitor image can be displayed on the display of the device 10 or a display connected to the device 10, and the vital sign monitor image displayed on the display can be captured (screenshot).

[0034] Preferably, the vital sign monitor image is a captured image of the monitoring screen of the vital sign monitor, a captured image of the monitoring screen displayed on the display of this device connected to the vital sign monitor, or a captured image of the monitoring screen displayed on the display of a computer other than this device connected to the vital sign monitor.

[0035] The device 10 and the vital sign monitor do not need to be connected. If the device 10 and the vital sign monitor are not connected, the vital sign monitor image displayed on the monitoring screen 30 may be captured by the camera of the device 10, or may be captured by a camera such as a video camera, digital camera, tablet, or smartphone, and the captured vital sign monitor image may be loaded into the device 10.

[0036] The device 10 is an information processing device such as a computer or a server. The device 10 can be a personal computer, tablet, smartphone, or the like. The device 10 may also be an electronic medical record terminal equipped with a display. The device 10 has a calculation unit that calculates information regarding the patient's risk of hypotension based on a vital sign monitor image received by an image receiving unit.

[0037] The device 10 may be configured as a single information processing device, or may be a collection of multiple physically separate information processing devices. In this case, each of the multiple information processing devices may have the same functions, or may have the functions of the single device 10 in a distributed manner.

[0038] FIG. 2 is a diagram showing an example of the configuration of the device 10. The device 10 has an image receiving unit with a receiving (receiving) function for acquiring a biological information monitor image of a patient or subject measured by a biological information monitor. The device 10 also has a calculation unit with a learning function for training a calculation model and a calculation function for calculating information related to the patient's risk of hypotension using the trained calculation model, and an output unit with an output function for outputting the calculated information related to the hypotension risk. The device 10 may include an image receiving unit 11, a memory unit 12, a display unit 13, an operation unit 14, and a processing unit 15.

[0039] The device 10 can acquire a patient monitor image from a patient monitor via the image receiving unit 11. The image receiving unit 11 can be a communication device implemented as hardware, firmware, communication software such as a TCP / IP driver or a PPP driver, or a combination of these. Communication via the communication device can be wireless or wired. The device 10 can acquire patient monitor image data from the patient monitor via the communication device. The communication device may receive patient monitor image data from the patient monitor via serial communication via a USB cable. The communication device may also have an interface circuit for short-range wireless communication according to a communication method such as Bluetooth (registered trademark) and may receive radio waves from the patient monitor. The communication device may also have a receiving circuit for receiving various signals corresponding to patient monitor image data via infrared communication or the like. The communication device may also have a communication interface circuit for a wired LAN.

[0040] The image receiving unit 11 may include an input / output device that detachably holds a portable storage medium instead of or in addition to the communication device. In this case, the input / output device acquires the biological information monitor image data stored in the portable storage medium and supplies the acquired biological information monitor image data to the processing unit 15.

[0041] The storage unit 12 is, for example, a semiconductor memory device such as a ROM or RAM. The storage unit 12 may be, for example, a magnetic disk, an optical disk, a magneto-optical disk, a nonvolatile semiconductor memory, or any other storage device capable of storing data. The storage unit 12 stores an operating system program, a driver program, an application program, data, and the like used for processing in the processing unit 15. The computer programs stored in the storage unit 12 may be downloaded online and installed in the storage unit 12, or may be installed in the storage unit 12 from a computer-readable portable recording medium such as a CD-ROM or DVD-ROM using a known setup program. The device 10 may be connected to the storage unit 12 and processing unit 15 implemented in the cloud to perform the functions of the storage unit 12 and processing unit 15.

[0042] The data stored in the storage unit 12 includes a calculation model M1, training data TD, etc., which will be described later. The storage unit 12 can also store data of a biological information monitor image acquired from a biological information monitor, or data of a processed biological information monitor image. The storage unit 12 may also temporarily store data related to a predetermined process.

[0043] The display unit 13 is a display. The display unit 13 may be a liquid crystal display, an organic EL display, etc. The display unit 13 displays the biological information monitor image data supplied from the processing unit 15, the results of calculation and processing of information regarding the patient's risk of hypotension, etc.

[0044] The information about the hypotension risk calculated by the present device is preferably an image, a numerical value, a character, a symbol, a sound, or a combination thereof. More preferably, the information about the hypotension risk is the level of the hypotension risk, the hypotension risk score, a heat map (saliency map) showing the areas of each partial area included in the vital sign monitor image that affect the calculation of the information about the hypotension risk, or a combination thereof.

[0045] The heat map can visualize areas that affect the calculation of information related to hypotension risk. For example, red indicates a high impact, yellow indicates a medium impact, green indicates a low impact, and blue indicates no impact. When red or yellow is displayed on the heat map, medical professionals can recognize that the area has a high impact on the calculation of information related to hypotension risk. The heat map may be displayed separately from the patient information monitor image, or may be displayed overlaid or semi-transparently on the patient information monitor image. The heat map may be calculated by any method, but for example, GradCAM (Gradient-weighted Class Activation Mapping, Non-Patent Document 3) can be used.

[0046] The operation unit 14 can be a keyboard, a mouse, and / or a pointing device such as a touch panel. A user of the device 10 can operate the device 10 using the operation unit 14. When operated by the user of the device 10, the operation unit 14 generates a signal corresponding to the operation. The generated signal is then supplied to the processing unit 15 as an instruction from the user. The biological information monitor image displayed on the display unit 13 can be captured automatically at a predetermined timing, preferably every 7 seconds, or at any timing by the operation unit 14.

[0047] The processing unit 15 is a processing device that loads the operating system program, driver program, application program, control program, etc. stored in the storage unit 12 into memory and executes instructions included in the loaded programs. The processing unit 15 is, for example, an electronic circuit such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), DSP (Digital Signal Processor), or GPU (Graphics Processing Unit), or a combination of various electronic circuits.

[0048] Processing unit 15 may be realized by integrated circuits such as ASICs (Application Specific Integrated Circuits), PLDs (Programmable Logic Devices), FPGAs (Field Programmable Gate Arrays), MCUs (Micro Controller Units), etc. Although processing unit 15 is illustrated as a single component in Figure 2, processing unit 15 may also be a collection of multiple physically separate processors. For example, multiple processors operating cooperatively in parallel to execute instructions may be implemented.

[0049] The processing unit 15 can function as a reception processing unit 151, a conversion processing unit 152, a learning processing unit 153, a calculation processing unit 154, and an output processing unit 155 by executing various commands included in an application program (control program) stored in the storage unit 12. Examples of the functions of the reception processing unit 151, the conversion processing unit 152, the learning processing unit 153, the calculation processing unit 154, and the output processing unit 155 will be described with reference to FIGS. 3 to 9.

[0050] The receiving processor 151 receives biological information monitor image data transmitted from the biological information monitor via the image receiver 11 and stores the received biological information monitor image data in the storage unit 12 .

[0051] The biological information monitor image, which is input data for the teacher data TD, is received and processed by the receiving processor 151. The receiving processor 151 uses the data of the biological information monitor image of the subject as input data, generates teacher data TD with a binary value of whether or not the subject has experienced hypotension as the correct answer label, and stores the data in the memory unit 12. The binary value of whether or not the subject has experienced hypotension can be a binary value where 1 indicates the presence of hypotension and 0 indicates the absence of hypotension, or vice versa.

[0052] Figure 5 shows a configuration diagram of an example of a training data collection system for collecting training data TD to be used in machine learning. A biological information monitor equipped with a monitoring screen 30 measures the subject's biological information waveform and acquires a biological information monitor image. The subject's presence or absence of hypotension is also acquired. The subjects are divided into two groups, one that developed hypotension and one that did not, and deep learning is performed to create a learning model.

[0053] Data 121 of the biological information monitor image of the measured subject and data 125 of whether the subject has developed hypotension can be stored in the storage unit 12. The stored data can be displayed on the display unit 13.

[0054] As described above, data on the biological information monitor image and data on the presence or absence of hypotension onset 125 are collected as teacher data TD. The data on the presence or absence of hypotension onset 125 may be a binary value, with 0 representing the absence of hypotension onset in the subject and 1 representing the presence of hypotension onset, or vice versa.

[0055] In the present device 10, the input data can be only the data of the vital signs monitor image, but other data may also be included as input data, such as the subject's and patient's age, sex, height, weight, systolic blood pressure, mean blood pressure, diastolic blood pressure, RCRI (Revised Cardiac Risk Index), HEART score (History, ECG, Age, Risk factor, Troponin), etc.

[0056] The conversion processing unit 152 can process the biological information monitor image data stored in the storage unit 12 so as to display only the biological information waveform portion to be measured, and store the processed image in the storage unit 12. The process of displaying only the biological information waveform portion preferably includes converting the color other than the color displaying the biological information waveform to black.

[0057] Figure 3 shows an example of a biological information monitor image measured on a patient. The biological information monitor image in Figure 3 displays a graph with the horizontal axis representing time and the vertical axis representing the intensity of the biological information waveform. The biological information waveforms included in the biological information monitor image shown in Figure 3 are, from top to bottom, an electrocardiogram waveform, an arterial pressure waveform, an exhaled carbon dioxide concentration waveform, and an SPO2 waveform.

[0058] FIG. 4 shows a biological information monitor image in FIG. 3 in which all colors except those displaying the biological information waveform have been converted to black. FIG. 4 shows the biological information monitor image in FIG. 3 after extraction processing to display only the portion of the biological information waveform to be measured in order to extract the waveform and its characteristics from the biological information monitor image. This extraction processing prevents erroneous detection of components other than the waveform, further improving prediction accuracy. The image in FIG. 4 in which only the biological information waveform has been extracted can be obtained, for example, by a program that converts the color space of the image from RGB to HSV and performs processing to specify red, green, and blue wavelengths and leave only those portions, and the device 10 can be equipped with such a program.

[0059] Respiratory status is an important factor affecting the vital signs waveforms displayed on a patient monitor. Inhalation (breathing in) and exhalation (breathing out) have significantly different effects on blood pressure. It is known that the amount of blood returning from the venous system to the heart increases during inhalation, while it decreases during exhalation. Therefore, to capture changes on a patient monitor that occur depending on the respiratory status, both the inhalation and exhalation phases must be displayed on the same patient monitor. In this specification, "phase" refers to the time step or moment at which a specific physiological event occurs in the context of a patient monitor. The inhalation and exhalation phases refer to the respective stages in the respiratory cycle. In other words, displaying both the inhalation and exhalation phases on the same patient monitor means that a single patient monitor image contains both the inhalation and exhalation timings.

[0060] Spontaneous breathing is reported to occur at approximately 12 to 20 breaths per minute (Charilaos Chourpiliadis et al., Physiology, Respiratory Rate, StatPearls, Treasure Island, 2023), with each cycle of inspiration and expiration taking 3 to 5 seconds. In other words, a time axis of at least 6 seconds is required to capture one breathing cycle. The time span (span) of a vital sign monitor image is preferably 6 to 10 seconds, more preferably 7 to 9 seconds, and even more preferably approximately 7 seconds. The waveforms typically viewed by physicians and other medical professionals on vital sign monitors are displayed over a span of approximately 7 seconds. By setting the vital sign monitor image to a span of approximately 7 seconds, processing such as cropping a predetermined span of the screen is unnecessary, and the same amount of information can be obtained as input data as when a physician makes a judgment using only a screen capture.

[0061] The following describes the learning process performed by learning processing unit 153. Learning processing unit 153 generates a trained calculation model M1 trained using teacher data TD stored in storage unit 12, and stores the generated trained calculation model M1 in storage unit 12.

[0062] A configuration diagram of an example of a machine learning system when performing machine learning is shown in Figure 6. During machine learning, data 121 of measured biological information monitor images and data 125 on the presence or absence of hypotension, which are stored in the storage unit 12, are used as training data TD.

[0063] A learning device composed of a conversion processing unit 152 and a learning processing unit 153 preprocesses the biological information monitor image data 121 as desired, and trains a calculation model using training data TD, which uses the preprocessed biological information monitor image data as input data and has hypotension occurrence / non-occurrence data 125 as output labels. Preprocessing is an extraction process that extracts only the waveform components of the biological information monitor image described above. A calculation model M1 can be generated by updating the weighting variables of the learning model so that the information regarding the hypotension occurrence risk calculated (estimated) based on the input biological information monitor image data has less error or is more consistent with the hypotension occurrence / non-occurrence data 125, which is the output label. Deep learning is used to generate the calculation model M1 by machine learning, and learning methods such as backpropagation, feedback alignment, direct feedback alignment, synthetic gradient, target prop, difference target prop, and bootstrap can be used.

[0064] When the data 125 on the occurrence or non-occurrence of hypotension is a binary value of high or low for the hypotension risk corresponding to the data of the vital sign monitor image, the weighting variables of the learning model can be updated to generate a calculation model M1 so that the information on the hypotension risk calculated (estimated) based on the input vital sign monitor image data matches the output label, where 1 indicates high and 0 indicates low.

[0065] The biological information monitor images of the subjects and patients are acquired as time series data, preferably once every 7 seconds. The biological information monitor images of the subjects are measured for the number of subjects.

[0066] The unit of electrical signal for an electrocardiogram waveform is mV, and when displaying an electrocardiogram waveform on the monitoring screen of a patient monitor, 1 mV generally corresponds to 10 mm. Because normal electrocardiogram values are generally 1 to 3.5 mV, the vertical axis of the electrocardiogram waveform on the patient monitor image is preferably 10 to 35 mm. The intelliVue MX800 has a 19-inch screen, and when displaying four patient waveforms, the vertical axis of each waveform can be displayed at 20 to 40 mm. When the electrical signal is weak and the waveform size is small, the vertical axis can be enlarged up to approximately 40 mm to view the waveform in detail.

[0067] The unit of blood pressure for arterial pressure waveforms is mmHg. When displaying an arterial pressure waveform on a patient monitor screen, 1 mmHg may correspond to, for example, approximately 0.1 to 0.2 mm. Normal blood pressure is 120 / 80 mmHg or less, while hypertension is 140 / 90 mmHg or more. The difference between systolic and diastolic blood pressure corresponds to the vertical axis of the arterial pressure waveform, which is approximately 40 to 50 mmHg. Therefore, if 1 mmHg corresponds to 0.1 to 0.2 mm, the vertical axis of the arterial pressure waveform on the patient monitor image will be approximately 4 to 10 mm. The intelliVue MX800 has a 19-inch screen, and when displaying four patient waveforms, the vertical axis of each waveform can be displayed in 20 to 40 mm. For low blood pressure or to view the waveform in detail, the vertical axis can be enlarged up to approximately 40 mm.

[0068] SpO2 is measured in percent (%), with a normal value generally between 99% and 100%. There is no specific standard for the size of the SpO2 waveform displayed on the patient monitor's monitoring screen. This is because SpO2 is calculated from absorbance based on the difference in transmittance between two wavelengths, 660 nm and 880 nm, and there is no direct correlation between the SpO2 value and waveform size. The waveform displayed on the patient monitor's monitoring screen represents the pulse wave as a result of fluctuations in the amount of transmitted light. The intelliVue MX800 has a 19-inch screen, and when displaying four vital signs, the vertical axis of each waveform can be displayed in a range of 20 to 40 mm. If the pulse wave is weak, the vertical axis can be automatically enlarged up to approximately 40 mm.

[0069] The unit of expiratory carbon dioxide concentration is mmHg, and its normal value is generally 35 to 45 mmHg. The upper display limit on the monitoring screen of a patient monitor is generally approximately 50 mmHg. The intelliVue MX800 has a 19-inch screen, and when displaying four vital signs, each waveform can be displayed in a range of 20 to 40 mm. When displaying an expiratory carbon dioxide concentration waveform on the monitoring screen of a patient monitor, the vertical axis of each waveform can generally be equivalent to 30 mm. The vertical axis of the expiratory carbon dioxide concentration waveform on the monitoring screen of a patient monitor is preferably 20 to 30 mm. When the expiratory carbon dioxide concentration is low, it can be automatically enlarged to a maximum of approximately 40 mm.

[0070] Table 2 shows an example of the resolution of the vital sign monitor image used for input in this device. [Table 2]

[0071] Fig. 7 is a schematic diagram showing an example of a calculation model M1. The calculation model M1 shown in Fig. 7 is a neural network model having an input layer M1-1, a hidden layer M1-2, and an output layer M1-3. A biological information monitor image of the training data TD is input to each of the so-called "neurons" of the input layer M1-1. The number of neurons in the hidden layer M1-2 may be more or less than the number of neurons in the input layer M1-1.

[0072] Fig. 8 is a schematic diagram showing another example of the calculation model M1. The calculation model M1 shown in Fig. 8 is a deep neural network model including a convolutional neural network having an input layer M1-1, a convolutional layer M1-2, a pooling layer M1-3, and an output layer M1-4. The convolutional layer M1-2 and the pooling layer M1-3 may each have two or more layers. A biological information monitor image of the training data TD is input to each neuron of the input layer M1-1.

[0073] The learning processing unit 153 generates or updates a calculation model in which the weights of each neuron in the neural network are learned by performing known machine learning using the training data TD. The learning processing unit 153 may also generate or update a calculation model in which the weights of each neuron in the multi-layered neural network are learned by performing known deep learning using the training data TD.

[0074] 9 shows an example configuration diagram of the present system 100 when predicting information related to a patient's risk of hypotension using the present device 10. The predictor, which is made up of the conversion processing unit 152, learning processing unit 153, and output processing unit 155, preprocesses data of a biological information monitor image measured by a biological information monitor as desired, and calculates information related to the patient's risk of hypotension using the preprocessed biological information monitor image data as input data using the trained calculation model M1.

[0075] The calculation processing unit 154 executes calculation processing. The calculation processing unit 154 uses the received biological information monitor image data as input data and the trained calculation model M1 to calculate information related to the patient's risk of hypotension.

[0076] The calculation processing unit 154 transmits the calculation result to the output processing unit 155. The output processing unit 155 outputs information on the calculation result. The output of the information is to display it on the display unit 13 and / or to transmit it to another device, etc.

[0077] The calculation process can be performed as desired by the conversion processing unit 152, the calculation processing unit 154, and the output processing unit 155. In the calculation process, when data of a patient's biological information monitor image is acquired from a biological information monitor, information on the patient's risk of hypotension corresponding to the acquired biological information monitor image is calculated.

[0078] The present disclosure also provides a method for predicting a hypotension risk of a patient based on biological information acquired from the patient, the method comprising: receiving an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; inputting the received biological information monitor image and calculating information regarding the patient's risk of hypotension; and outputting information about the calculated hypotension risk; Including, the calculation is performed using a trained calculation model that has been subjected to machine learning using training data including, as an input, a biological information monitor image of the subject and, as an output, information regarding the presence or absence of hypotension in the subject, so that information regarding the hypotension risk is calculated when the biological information monitor image is input. This invention relates to a method for predicting the risk of hypotension.

[0079] The present disclosure also provides a hypotension risk prediction program for predicting a hypotension risk of a patient based on biological information acquired from the patient, the program comprising: an image receiving function for receiving an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; and a calculation function for calculating information regarding the patient's risk of hypotension when the received biological information monitor image is input; Including, the calculation function has a trained calculation model that has been subjected to machine learning using training data that includes, as an input, a biological information monitor image of the subject and, as an output, information on whether the subject has developed hypotension, so that information on the hypotension risk is calculated when the biological information monitor image is input. The present invention is directed to a hypotension risk prediction program. The description of each component included in the above-mentioned hypotension risk prediction device can be applied to each component included in the above-mentioned hypotension risk prediction method and hypotension risk prediction program.

[0080] (Embodiment) The device 10 used a neural network as a trained calculation model to estimate a patient's risk of hypotension. The vital sign monitor connected to the device 10 was the intelliVue MX800 manufactured by Philips. Machine learning to generate the trained calculation model used training data, such as vital sign monitor images (7-second span) of 53 subjects with hypotension episodes and vital sign monitor images (7-second span) of 29 subjects without hypotension episodes, as input values. Machine learning was performed using the presence or absence of hypotension episodes corresponding to the vital sign monitor images as output labels. In the following embodiment, vital sign monitor images including all waveforms for multiple cycles included in the 7-second span monitoring screen of the vital sign monitor were used.

[0081] The calculation model in the present device 10 may be a neural network, a random forest, a support vector machine (SVM), or the like.

[0082] The output label of the training data TD was set to a binary value of 1 if the subject had experienced hypotension and 0 if the subject had not experienced hypotension, and the neural network calculation model was machine-learned.

[0083] FIG. 10 shows a heat map of hypotension risk calculated using the device 10 based on the patient information monitor image of FIG. 4. The heat map indicates, in red, yellow, green, and blue, the areas of the patient information monitor image of FIG. 4 that affect the calculation of information related to hypotension risk. The heat map of FIG. 10 has been scaled differently from the patient information monitor image of FIG. 4 due to resizing to a predetermined size, e.g., 128x128 pixels, for a model such as a convolutional neural network. Resizing can be performed using, for example, the cv2.resize function. The aspect ratio, which is the ratio of the vertical and horizontal dimensions of the image, may change during resizing. This is also true for the heat maps shown below.

[0084] (Prediction accuracy depending on the number of biological waveform types) Figure 11 shows a diagram in which a heat map calculated using a patient monitor image containing only one type of arterial pressure waveform is superimposed on a patient monitor image in generating a trained calculation model and predicting the risk of hypotension. In the machine learning to generate a trained calculation model, patient monitor images (7-second span) containing one type of arterial pressure waveform measured with a patient monitor from 53 subjects who had experienced hypotension and patient monitor images (7-second span) containing one type of arterial pressure waveform from 29 subjects who had not experienced hypotension were used as training data, and machine learning was performed using the presence or absence of hypotension corresponding to the patient monitor image as the output label.

[0085] Figure 12 shows, from top to bottom, a heat map calculated using a patient monitor image containing two waveforms, electrocardiogram waveform and arterial pressure waveform, and the patient monitor image when generating a trained calculation model and predicting the risk of hypotension. In the machine learning to generate the trained calculation model, the input values were patient monitor images (7-second span) containing two waveforms, electrocardiogram waveforms and arterial pressure waveforms, of 53 subjects who had experienced hypotension, measured with a patient monitor, and patient monitor images (7-second span) containing two waveforms, electrocardiogram waveforms and arterial pressure waveforms, of 29 subjects who had not experienced hypotension, measured with a patient monitor, and machine learning was performed using the presence or absence of hypotension corresponding to the patient monitor image as the output label.

[0086] Figure 13 shows a heat map calculated using a patient monitor image containing, from top to bottom, three types of waveforms: electrocardiogram waveform, arterial pressure waveform, and SPO2, in generating a trained calculation model and predicting the risk of hypotension, as well as the patient monitor image. In the machine learning to generate the trained calculation model, the input values were training data: patient monitor images (7-second span) containing three types of waveforms: electrocardiogram waveform, arterial pressure waveform, and SPO2 waveform, measured with a patient monitor, from 53 subjects who had experienced hypotension, and patient monitor images (7-second span) containing three types of waveforms: electrocardiogram waveform, arterial pressure waveform, and SPO2 waveform, from 29 subjects who had not experienced hypotension, and machine learning was performed using the presence or absence of hypotension corresponding to the patient monitor image as the output label.

[0087] Figures 10 and 4 show heat maps calculated using patient monitor images containing, from top to bottom, four waveforms: electrocardiogram waveform, arterial pressure waveform, expired carbon dioxide concentration waveform, and SPO2 waveform, in generating a trained calculation model and predicting the risk of hypotension, and the patient monitor images. In the machine learning to generate the trained calculation model, the input values were patient monitor images (7-second span) containing four waveforms: electrocardiogram waveform, arterial pressure waveform, expired carbon dioxide concentration waveform, and SPO2 waveform, measured with a patient monitor, of 53 subjects who had experienced hypotension, and patient monitor images (7-second span) containing four waveforms: electrocardiogram waveform, arterial pressure waveform, expired carbon dioxide concentration waveform, and SPO2 waveform, of 29 subjects who had not experienced hypotension, and machine learning was performed using the presence or absence of hypotension corresponding to the patient monitor image as the output label.

[0088] Table 3 shows the AUC and prediction accuracy rate of this device depending on the number of waveform types. The prediction accuracy rate obtained from three or four types of biological information waveforms was better than the prediction accuracy rate obtained from one or two types of biological information waveforms. By using biological information monitor images containing multiple types of waveforms, a higher accuracy rate was obtained for predicting the occurrence of hypotension.

[0089] The prediction accuracy rate was calculated by using the device 10 to estimate a binary value of high or low risk of hypotension for a dataset of vital signs monitor images of patients known to have experienced hypotension. For each of the following waveform types: one type (arterial pressure waveform), two types (arterial pressure waveform, electrocardiogram waveform), three types (arterial pressure waveform, electrocardiogram waveform, SPO2 waveform), and four types (arterial pressure waveform, electrocardiogram waveform, SPO2 waveform, expired carbon dioxide concentration waveform), data from vital signs monitor images measured on 23 patients who experienced hypotension and 13 patients who did not experience hypotension were used.

[0090] [Table 3]

[0091] (Reference example: Accuracy of predicting hypotension using age, gender, etc.) Table 4 shows the results of hypotension prediction using systolic blood pressure, diastolic blood pressure, RCRI, and HEART score. The accuracy of hypotension prediction using systolic blood pressure, diastolic blood pressure, RCRI, and HEART score was low.

[0092] [Table 4] [Explanation of symbols]

[0093] 10. Hypotension risk prediction device 11 Image Reception Section 12 Storage section 121 Patient Monitor Image Data 125 Data on the occurrence of hypotension 13 Display section 14 Control section 15 Processing section 20 Network 30 Monitoring screen 100 Hypotension Risk Prediction System 151 Receiving processing unit 152 Conversion processing section 153 Learning processing unit 154 Discrimination processing unit 155 Output Processing Unit M1 calculation model TD teacher data

Claims

1. A hypotension risk prediction device that predicts a hypotension risk of a patient based on biological information acquired from the patient, an image receiving unit that receives an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; a calculation unit that calculates information regarding a risk of hypotension of the patient when the received biological information monitor image is input; and an output unit that outputs information related to the calculated hypotension risk Equipped with the calculation unit has a trained calculation model that has been subjected to machine learning using training data including, as an input, a biological information monitor image of the subject and, as an output, information regarding whether or not the subject has developed hypotension, so that information regarding the hypotension risk is calculated when the biological information monitor image is input. Hypotension risk prediction device.

2. The hypotension risk prediction device according to claim 1 , wherein the information relating to the hypotension risk includes information displayed as an image, a number, a letter, a symbol, a sound, or a combination thereof.

3. The hypotension risk prediction device of claim 1, wherein the information regarding the hypotension risk includes a level of the patient's risk of developing hypotension, a score of the patient's risk of developing hypotension, a heat map showing areas among each partial area included in the vital sign monitor image that affect the calculation of the information regarding the hypotension risk, or a combination thereof.

4. The hypotension risk prediction device of claim 1, wherein the biological information monitor image is a captured image of a monitoring screen of the biological information monitor, a captured image of a monitoring screen displayed on a display of the hypotension risk prediction device connected to the biological information monitor, or a captured image of a monitoring screen displayed on a display of a computer connected to the biological information monitor.

5. The hypotension risk prediction device according to claim 1 , wherein the biological information monitor image includes an electrocardiogram waveform, an arterial pressure waveform, a transcutaneous oxygen saturation waveform, an exhaled carbon dioxide concentration waveform, or a combination thereof.

6. The hypotension risk prediction device according to any one of claims 1 to 5, The vital sign monitor; A hypotension risk prediction system comprising:

7. A method for predicting a hypotension risk of a patient based on biological information acquired from the patient, comprising: receiving an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; inputting the received biological information monitor image and calculating information regarding the patient's risk of hypotension; and outputting information about the calculated hypotension risk; Including, the calculation is performed using a trained calculation model that has been subjected to machine learning using training data including, as an input, a biological information monitor image of the subject and, as an output, information regarding the presence or absence of hypotension in the subject, so that information regarding the hypotension risk is calculated when the biological information monitor image is input. Methods for predicting hypotension risk.

8. A hypotension risk prediction program that predicts a hypotension risk of a patient based on biological information acquired from the patient, an image receiving function for receiving an input of a biological information monitor image displayed on a monitoring screen of a biological information monitor that acquires biological information of the patient; and a calculation function for calculating information regarding the patient's risk of hypotension when the received biological information monitor image is input; Including, the calculation function has a trained calculation model that has been subjected to machine learning using training data that includes, as an input, a biological information monitor image of the subject and, as an output, information on whether the subject has developed hypotension, so that information on the hypotension risk is calculated when the biological information monitor image is input. Hypotension risk prediction program.

Images