Ventilation difficulty prediction apparatus, ventilation difficulty prediction method, and program

Abstract

To provide a ventilation difficulty prediction apparatus, a ventilation difficulty prediction method, and a program capable of accurately predicting the occurrence of ventilation difficulty.SOLUTION: A ventilation difficulty prediction apparatus comprises: an acquisition unit that acquires biometric data based on biometric information; and a prediction unit that predicts an occurrence of a ventilation difficulty event of a subject based on output data obtained by inputting the biometric information detected from the subject into a trained model that has been trained using training data in which the biometric data is used as input data and statistical data related to the biometric data is used as output data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a ventilation difficulty prediction device, a ventilation difficulty prediction method, and a program. [Background technology]

[0002] Cardiac arrest during general anesthesia can occur, for example, through hypoxemia caused by difficult ventilation. Difficult ventilation during general anesthesia is a serious complication that can lead to hypoxemia and cardiac arrest. A major cause of difficult ventilation is pharyngeal spasm, which can be caused by insufficient anesthesia for pain or stimuli. Risk factors for pharyngeal spasm include the patient's young age, airway hyperresponsiveness, and the anesthesiologist's inexperience.

[0003] Conventionally, artificial intelligence (hereinafter referred to as AI) that predicts hypoxemia has been developed as a technology for preventing cardiac arrest during general anesthesia. For example, AI that predicts hypoxemia uses measurement data such as the patient's transcutaneous oxygen saturation.

[0004] However, during general anesthesia, patients are generally administered oxygen at a high concentration of around 40%, so once hypoxemia occurs, oxygen consumption in the body progresses and the time until cardiac arrest becomes shorter. For this reason, it is more important to ensure a time limit until cardiac arrest occurs than to predict hypoxemia. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Lundberg SM, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng. 2018;2:749-760. [Non-patent document 2] Park JB, et al. Machine learning-based prediction of intraoperative hypoxemia for pediatric patients. PLoS One. 2023;18:e0282303 Summary of the Invention [Problem to be solved by the invention]

[0006] If ventilation difficulties could be predicted before hypoxemia occurs, it would provide a time window for cardiac arrest. However, when predicting ventilation difficulties, for example, even if a decrease in ventilation volume is detected, the decrease in ventilation volume may simply be due to a change in the ventilator settings. Furthermore, ventilation conditions change depending on factors such as the progress of surgery and the presence or absence of medication, making it difficult to define ventilation difficulties. For this reason, it may not be possible to predict the occurrence of ventilation difficulties until just before they occur.

[0007] The present invention has been made in consideration of the above-mentioned circumstances, and has an object to provide a ventilation difficulty prediction device, a ventilation difficulty prediction method, and a program that can quickly predict the occurrence of ventilation difficulty. [Means for solving the problem]

[0008] [1] In order to solve the above problem, one aspect of the present invention is a ventilation difficulty prediction device comprising: an acquisition unit that acquires biometric data based on biometric information; and a prediction unit that predicts the occurrence of a ventilation difficulty event in a subject based on output data obtained from a trained model trained using training data in which data based on the biometric data is used as input data and statistical data related to the biometric data is used as output data, and that measures the subject and uses the input data to obtain biometric data based on the biometric information.

[0009] [2] In another aspect of the present invention, the biological information includes at least one of heart rate, systolic blood pressure, transcutaneous oxygen saturation, expired carbon dioxide partial pressure, electroencephalogram index, ST change in electrocardiogram lead II, maximum inspiratory pressure, tidal volume, or minute ventilation.

[0010] [3] In another aspect of the present invention, the trained model is trained using, as input data, biometric data based on important biometric information that is clinically important among the biometric information.

[0011] [4] In another aspect of the present invention, the important biological information includes biological information of heart rate, systolic blood pressure, partial pressure of expired carbon dioxide, maximum inspiratory pressure, tidal volume, and minute ventilation.

[0012] [5] In another aspect of the present invention, the trained model is trained using parameters based on the sensitivity and false positive rate of the occurrence of a ventilation difficulty event.

[0013] [6] In order to solve the above problem, one aspect of the present invention is a ventilation difficulty prediction method in which a computer acquires biometric data based on biometric information, and predicts the occurrence of a ventilation difficulty event in a subject based on output data obtained from a trained model trained using training data in which data based on the biometric data is used as input data and statistical data related to the biometric data is used as output data, using biometric data based on the biometric information obtained by measuring the subject as input data.

[0014] [7] In order to solve the above problem, one aspect of the present invention is a program that causes a computer to acquire biometric data based on biometric information, and predict the occurrence of a ventilation difficulty event in a subject based on output data obtained using a trained model that has acquired biometric data based on biometric information obtained by measuring the subject as input data and statistical data related to the biometric data as output data. [Effects of the Invention]

[0015] According to the ventilation difficulty prediction device, ventilation difficulty prediction method, and program of the present invention, it is possible to quickly predict the occurrence of ventilation difficulty. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a diagram illustrating an example of the configuration of a ventilation difficulty prediction device 1 according to an embodiment. [Figure 2] FIG. 2 is a diagram conceptually illustrating the function of a prediction model 11. [Figure 3] 10 is a flowchart showing an example of the processing of the ventilation difficulty prediction device 1. [Figure 4] 10 is a graph showing an example of time series data of a first index value (T2 statistics) and time series data of a second index value (Q statistics) output by a prediction model 11. DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, a ventilation difficulty prediction device, a ventilation difficulty prediction method, and a program according to an embodiment will be described with reference to the drawings.

[0018] FIG. 1 is a diagram showing an example of the configuration of a ventilation difficulty prediction device 1 of an embodiment. The ventilation difficulty prediction device 1 of the embodiment is a device that predicts ventilation difficulty that will occur in a patient P under general anesthesia. The ventilation difficulty prediction device 1 is capable of communicating with, for example, a biosensor 2 that monitors the condition of the patient P and measures biometric information. The patient P is an example of a subject. The ventilation difficulty prediction device 1 may be provided in, for example, a personal computer, a tablet, or a smartphone.

[0019] The biosensor 2 measures, for example, the heart rate (hereinafter also referred to as HR), systolic blood pressure (hereinafter also referred to as SBP), partial pressure of expired carbon dioxide (hereinafter also referred to as EtCO2), maximum inspiratory pressure (hereinafter also referred to as PIP), tidal volume (hereinafter also referred to as TV), and minute ventilation (hereinafter also referred to as MV) of the patient P as bioinformation. The minute ventilation (MV) generated by the biosensor 2 is the total value of ventilation volume for one minute. The bioinformation is, for example, vital signs and information related to the artificial respirator.

[0020] The biosensor 2 generates time-series biodata from the measured bioinformation. The bioinformation may include percutaneous oxygen saturation (SpO2), bioencephalogram index (BIS), and ST change (ST2) in electrocardiogram lead II (the portion from the end of the S wave to the beginning of the T wave). The biosensor 2 transmits the generated time-series biodata to the ventilation difficulty prediction device 1.

[0021] The ventilation difficulty prediction device 1 includes, for example, a storage unit 10 and a control unit 20. The storage unit 10 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory), a flash memory, a hard disk, an optical disk, or the like. The storage unit 10 stores a prediction model 11. The prediction model 11 is a trained model trained using training data in which time-series biological data is used as input data and statistical data related to the time-series biological data is used as output data.

[0022] 2 is a diagram conceptually illustrating the function of the prediction model 11. The prediction model 11 has, for example, an input layer, an intermediate layer, and an output layer. For example, time-series biological data (HR, SBP, EtCO2, PIP, TV, MV) is input to the input layer of the prediction model 11 as input data. A first index value and a second index value are output from the output layer as output data.

[0023] The intermediate layer may be, for example, a multi-layer neural network connecting the input layer and the output layer, or a network based on multivariate statistical process control (MSPC), which is a linear model. 2 The second index value is, for example, a Q statistic, which is the square of the distance of the subspace spanned by the principal component.

[0024] When the prediction model 11 is trained, for example, by hyperparameter search in the intermediate layer, control limits for the first index value and the second index value at which a difficult ventilation event is predicted to occur are determined. When the first index value or the second index value exceeds the control limit, a difficult ventilation event is predicted to occur. The memory unit 10 stores the control limits together with the prediction model 11.

[0025] The prediction model 11 is generated based on, for example, recorded vital signs and biological information related to the ventilator. The prediction model 11 is generated based on, for example, multiple time-series biological data and whether or not a difficult ventilation event has occurred for the patient whose biological information was acquired to generate the time-series biological data. The prediction model 11 is trained using parameters based on, for example, the sensitivity of the difficult ventilation event occurrence and the false positive rate. A specific example of the prediction model 11 will be described later in the section on verifying the effectiveness of difficult ventilation prediction.

[0026] The control unit 20 includes, for example, an acquisition unit 21 and a prediction unit 22. The control unit 20 is realized by, for example, a hardware processor such as a CPU (Central Processing Unit) executing a program (software). Part or all of the control unit 20 may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit), or may be realized by a combination of software and hardware.

[0027] The program may be stored in advance in a storage device such as an HDD (Hard Disk Drive) or flash memory (a storage device with a non-transitory storage medium, which may be memory unit 10), or may be stored in a removable storage medium (non-transitory storage medium) such as a DVD or CD-ROM, and installed by inserting the storage medium into a drive device.

[0028] The acquiring unit 21 acquires time-series biological data based on biological information obtained by measuring a patient P who is a subject, and transmitted by the biological sensor 2. The time-series biological data is data in which biological data based on biological information measured by the biological sensor 2 is arranged in time series. The biological sensor 2 transmits the time-series biological data at predetermined time intervals, for example, at one-minute intervals. The acquiring unit 21 acquires the time-series biological data every time it is transmitted by the biological sensor 2.

[0029] The prediction unit 22 generates input data based on the time-series biometric data acquired by the acquisition unit 21. For example, the prediction unit 22 performs mathematical processing on the time-series biometric data to set the mean to 0 and the variance to 1, thereby generating input data that serves as feature quantities. The prediction unit 22 inputs the generated input data to the prediction model 11 stored in the storage unit 10.

[0030] The prediction unit 22 outputs output data to the prediction model 11 to which input data has been input. The prediction unit 22 outputs output data to the prediction model 11, for example, of a time period (hereinafter, predicted time period) in which it is predicted that patient P will develop difficult ventilation at the predicted time of occurrence of a difficult ventilation event (hereinafter, predicted occurrence time). The prediction unit 22 may determine the predicted occurrence time and predicted time period as appropriate. The predicted occurrence time may be, for example, 3 minutes or 5 minutes from the current time, and may be determined depending on the content of the surgery, the condition of patient P, etc.

[0031] The prediction unit 22 may set the predicted time period as, for example, a time period from 5 minutes to 1 minute before the predicted occurrence time. The predicted time period may be any other time period, but it is preferable to set it as a time period as early as possible before the predicted occurrence time, and it is preferable that the final time of the predicted time period be 1 minute or more before the predicted occurrence time. The prediction unit 22 predicts whether or not the patient P will develop dysventilation at the predicted occurrence time based on the output data output by the prediction model 11.

[0032] When the prediction unit 22 determines that the patient P will develop difficult ventilation during the predicted time period, it outputs an alert to a user such as a doctor or technician. The prediction unit 22 may output the alert in any manner. For example, the prediction unit 22 may issue a warning sound from a speaker provided near the user as the alert, or may display a warning message or a warning mark on a display visible to the user.

[0033] Next, we will explain the processing of the ventilation difficulty prediction device 1. Fig. 3 is a flowchart showing an example of the processing of the ventilation difficulty prediction device 1. The ventilation difficulty prediction device 1 executes the processing shown in Fig. 3, for example, at regular intervals, for example, every minute. The ventilation difficulty prediction device 1 first acquires time-series biological data transmitted by the biological sensor 2 using the acquisition unit 21 (step S101) and starts measurement.

[0034] Next, the prediction unit 22 reads out the prediction model 11 stored in the storage unit 10, and generates input data based on the time-series biological data acquired by the acquisition unit 21. The prediction unit 22 inputs the generated input data to the prediction model 11 (step S103). Next, the prediction unit 22 causes the prediction model 11 to output output data (step S105), and executes monitoring. The output data includes a first index value and a second index value.

[0035] Next, the prediction unit 22 determines whether the first index value or the second index value output by the prediction model 11 for the prediction time period exceeds the control limit (step S107). If it is determined that at least one of the first index value or the second index value exceeds the control limit, the prediction unit 22 predicts that ventilation difficulty will occur in the patient P at the predicted occurrence time, and outputs an alert (step S109). After issuing the alert, the prediction unit 22 continues monitoring (step S111).

[0036] On the other hand, if the prediction unit 22 determines that neither the first index value nor the second index value exceeds the control limit, the ventilation difficulty prediction device 1 continues monitoring (step S111). After the prediction unit 22 issues an alert in step S109, the ventilation difficulty prediction device 1 may simply end the processing shown in FIG.

[0037] In step S111, the prediction unit 22 that continues monitoring determines whether or not to continue measurement (step S113). If the prediction unit 22 determines to continue measurement, the acquisition unit 21 returns the process to step S101 and acquires unacquired time-series biological data (step S101). If the prediction unit 22 determines not to continue measurement, the ventilation difficulty prediction device 1 ends the process shown in FIG. 3.

[0038] The ventilation difficulty prediction device 1 of the embodiment predicts the occurrence of a ventilation difficulty event at a predicted occurrence time within a predicted time period. The predicted time period is set, for example, one minute or more before the predicted occurrence time. This allows the occurrence of a ventilation difficulty event to be predicted at an early time period. Therefore, the occurrence of ventilation difficulty can be quickly predicted.

[0039] (Verification of the effectiveness of ventilation difficulty prediction) Next, the results of verifying the effectiveness of difficult ventilation prediction will be described. To verify the effectiveness of difficult ventilation prediction, a prediction model 11 was generated. The prediction model 11 is generated by determining whether or not a difficult ventilation event has occurred and the time of occurrence using a data set in which, for example, time-series biological data of an actual patient under general anesthesia is used as input data and a first index value and a second index value are used as output data. The occurrence of a difficult ventilation event was defined by two anesthesiologists independently reviewing the anesthesia records.

[0040] Specifically, the anesthesiologists checked vital signs, ventilator data, medication records, surgical progress, and free-form comments at the time the ventilation rate decreased to determine whether it was due to difficult ventilation or an artifact. Artifacts are causes other than difficult ventilation, such as changes in ventilator settings or measurement errors. If the two anesthesiologists disagreed on their judgments, they discussed the matter and reached a consensus.

[0041] Next, we will explain an example of an experiment to predict the occurrence of difficult ventilation events. In this experiment, we used a prediction model11 generated using a dataset with time-series biological data measured on pediatric patients aged 1 year or older and under 15 years old who underwent general anesthesia using a supraglottic device between 2018 and 2023 as input data.

[0042] Time-series vital signs were extracted every minute from the patient's electronic anesthesia record. Two anesthesiologists independently reviewed vital signs, ventilator data, medication records, surgical progress, and free text for cases in which tidal volumes fell below 80% of the average for the previous 3 minutes, and labeled these as difficult ventilation events or artifacts.

[0043] Clinically important vital information was used to generate time-series vital data, including heart rate (HR), systolic blood pressure (SBP), partial pressure of exhaled carbon dioxide (EtCO2), maximum inspiratory pressure (PIP), tidal volume (TV), and minute ventilation (MV).

[0044] First, a training dataset was created using time-series biodata from patients without difficult ventilation events or artifacts among 457 patients who underwent general anesthesia between 2018 and 2022. Because ventilator settings are often changed immediately after the start and end of surgery, a dataset with time-series biodata acquired 10 minutes after the start of surgery and up to 10 minutes before the end of surgery as input data was used for learning.

[0045] For the prediction model 11, we adopted MSPC, an anomaly detection model compatible with time series data. 2 The Q statistic (first index value) and the Q statistic (second index value) were used. When the first index value or the second index value exceeded the pre-set control limit, it was determined that an abnormality had occurred.

[0046] In the training dataset, we used 5-fold cross-validation and grid search to find the number of principal components (≦number of input features), T 2 The control limits for the Q statistic and the Q statistic were determined. Considering the time it takes for the anesthetic agent to take effect, a delay of about one minute is considered necessary to prevent progression to hypoxemia. Therefore, if an abnormality is detected between five minutes and one minute before ventilation becomes difficult, it is considered a TP (true positive), and if an abnormality is detected at any other time, it is considered an FP (false positive).

[0047] Next, prediction model 11 was evaluated using sensitivity and false positive rate as effectiveness indicators. Prediction model 11 was evaluated using a test dataset that used a time series dataset of patients who underwent general anesthesia in 2023 as input data. As a result, the sensitivity of the test dataset was 57%, and the false positive rate was 0.65 times / hour.

[0048] An example in which the onset of a difficult ventilation event can be predicted one minute before the event actually occurs is shown below. FIG. 4 shows the first index value (T 2 4 is a graph showing an example of time series data of the first index value (T statistic) and time series data of the second index value (Q statistic). 2 The time series data for the second index value (Q statistic) is shown in the upper row, and the time series data for the second index value (Q statistic) is shown in the lower row.

[0049] In the example shown in FIG. 4, a difficult ventilation event occurs in the patient at time Te. In this example, if time Te is the predicted time of occurrence, the predicted time period is the time period from the predicted time period start time Tz1 to the predicted time period end time Tz2. The second index value shown in the lower part of FIG. 4 exceeds the second control limit Th2 in the predicted time period from the predicted time period start time Tz1 to the predicted time period end time Tz2. Therefore, it was possible to predict the occurrence of a difficult ventilation event in the patient at the predicted time Te, in the predicted time period more than one minute before the predicted time Te.

[0050] The above describes an embodiment of the present invention with reference to the drawings. However, the ventilation difficulty prediction device, ventilation difficulty prediction method, and program are not limited to the above-described embodiment, and various modifications, substitutions, combinations, and / or design changes can be made without departing from the spirit of the present invention.

[0051] Furthermore, the effects of the above-described embodiments of the present invention are described as examples. Therefore, the embodiments of the present invention may also achieve other effects that a person skilled in the art can recognize from the description of the above-described embodiments in addition to the above-described effects. [Explanation of symbols]

[0052] 1. Ventilation difficulty prediction device 2. Biometric sensors 10 Storage section 11 Predictive Models 20 Control Unit 21 Acquisition Department 22 Prediction Department P patient

Claims

1. an acquisition unit that acquires biometric data based on biometric information; a prediction unit that predicts an occurrence of a difficult ventilation event of the subject based on output data obtained using a trained model trained with training data in which data based on the biological data is used as input data and statistical data related to the biological data is used as output data and that uses biological data based on the biological information obtained by measuring the subject as input data; Ventilation difficulty prediction device.

2. The biological information includes at least one of a heart rate, a systolic blood pressure, a transcutaneous oxygen saturation, a partial pressure of expired carbon dioxide, an electroencephalogram index, an ST change in electrocardiogram lead II, a maximum inspiratory pressure, a tidal volume, and a minute ventilation volume. The ventilation difficulty prediction device according to claim 1.

3. The trained model learns biometric data based on clinically important biometric information from the biometric information as input data. The ventilation difficulty prediction device according to claim 1.

4. The important biological information includes biological information of heart rate, systolic blood pressure, partial pressure of exhaled carbon dioxide, maximum inspiratory pressure, tidal volume, and minute ventilation. The ventilation difficulty prediction device according to claim 3.

5. The trained model is trained using parameters based on the sensitivity and false positive rate of the occurrence of a ventilatory difficulty event. The ventilation difficulty prediction device according to claim 1.

6. The computer Obtaining biometric data based on biometric information; a trained model trained using training data in which data based on the biological data is used as input data and statistical data related to the biological data is used as output data, and the trained model predicts the occurrence of a difficult ventilation event in the subject based on output data obtained using biological data based on the biological information obtained by measuring the subject as input data; Methods for predicting ventilation difficulties.

7. On the computer, Obtaining biometric data based on biometric information; A trained model trained using training data in which data based on the biological data is used as input data and statistical data related to the biological data is used as output data predicts the occurrence of a difficult ventilation event in the subject based on output data obtained using biological data based on the biological information obtained by measuring the subject as input data. program.

Images