Drowsiness detection device, detection method, and computer program

Abstract

To provide a dozing detection device which detects dozing accurately.SOLUTION: A dozing detection device 100 is a detection device for detecting dozing and comprises an arithmetic unit 1 which executes detection processing for detecting dozing. The detection processing includes obtaining an index value for detecting dozing from an RRI (R-R Interval) generated from an electrocardiographic signal of a subject, measuring a duration time in which the index value exceeds a first threshold, and detecting dozing of the subject when the duration time exceeds a second threshold.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a drowsiness detection device, a detection method, and a computer program. [Background technology]

[0002] As a technique for detecting whether a subject is dozing off, for example, Japanese Patent Laid-Open Publication No. 2015-226696 (hereinafter, Patent Document 1) proposes a method that uses the subject's electrocardiogram. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2015-226696 A Summary of the Invention

[0004] Such drowsiness detection is expected to be used when detecting drowsiness of a driver of a vehicle, etc. Therefore, further improvement of the detection accuracy is desired.

[0005] A drowsiness detection device according to one embodiment is a detection device for detecting drowsiness, and includes a calculation device that executes a detection process for detecting drowsiness, and the detection process includes obtaining an index value for drowsiness detection from an RRI (RR Interval) generated from a subject's electrocardiogram signal, timing the duration during which the index value exceeds a first threshold, and detecting the subject's drowsiness when the duration exceeds a second threshold.

[0006] A detection method according to one embodiment is a method for detecting dozing off, comprising: obtaining an index value for dozing off detection from an RRI generated from a subject's electrocardiogram signal; timing the duration during which the index value exceeds a first threshold; and detecting the subject's dozing off when the duration exceeds a second threshold.

[0007] A computer program according to one embodiment is a computer program for causing a computer to execute a detection process for detecting dozing off, the detection process including obtaining an index value for dozing off detection from an RRI generated from the subject's electrocardiogram signal, timing the duration during which the index value exceeds a first threshold, and detecting that the subject is dozing off when the duration exceeds a second threshold.

[0008] Further details will be described in the following embodiments. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram showing an outline of the configuration of a drowsiness detection device (hereinafter, detection device) according to an embodiment. [Diagram 2] FIG. 2(a) is a diagram showing an example of an electrocardiogram signal, and (b) is a diagram showing R wave data corresponding to the electrocardiogram signal of (a). [Diagram 3] FIG. 3 is a block diagram showing an example of the configuration of a calculation device of the detection device. [Figure 4] FIG. 4 is a diagram for explaining the Poincaré index, the T2 statistic, and the Q statistic. [Diagram 5] FIG. 5 is a diagram illustrating the canonical correlation analysis of the data set. [Figure 6] FIG. 6 is a diagram for explaining a specific example of a method for making a data set dynamic. [Figure 7] FIG. 7 is a flowchart showing an example of a method for detecting drowsiness by the detection device according to the first embodiment. [Figure 8] FIG. 8 is a flowchart showing an example of the determination process in step S300 of the flowchart in FIG. [Figure 9] FIG. 9 shows the results of a test conducted by the inventors to verify whether using the Q statistic or the T2 statistic as an index value provides better detection accuracy in the detection method according to the first embodiment. [Figure 10]FIG. 10 shows a Receiver Operating Characteristic (ROC) curve obtained when the HRV index value was used as it is as the index value used as the data set to be input to the classifier in the detection method according to the first embodiment of the present invention. [Figure 11] FIG. 11 shows an ROC curve obtained when only the Poincaré index value is used as the index value used as the data set to be input to the classifier in the detection method according to the first embodiment of the present invention. [Figure 12] FIG. 12 shows an ROC curve obtained when canonical variables obtained by performing a canonical correlation analysis on the values ​​of the HRV index and the Poincaré index are used as index values ​​used as a data set to be input to a classifier in the detection method according to the first embodiment developed by the inventors. [Figure 13] FIG. 13 shows an ROC curve obtained when time-series data having time-series changes obtained by dynamicizing the values ​​of the HRV index and the Poincaré index are used as index values ​​used as a data set to be input to a classifier in the detection method according to the first embodiment developed by the inventors. [Figure 14] FIG. 14 shows an ROC curve obtained when time-series data having time-series changes obtained by dynamicizing canonical variables obtained by performing canonical correlation analysis on the values ​​of HRV index and Poincaré index are used as index values ​​used as a data set to be input to a classifier in the detection method according to the first embodiment developed by the inventors. [Figure 15] FIG. 15 is a diagram showing the results of a test conducted by the inventors to verify how many beats of past data should be incorporated in the dynamic detection method according to the first embodiment. [Figure 16] FIG. 16 is a flowchart showing an example of a method for detecting drowsiness by the detection device according to the second embodiment. [Figure 17] FIG. 17 is a diagram showing the results of an evaluation experiment conducted by the inventors on the detection method according to the second embodiment for electrocardiogram signals including a state in which the subject is dozing. [Figure 18] FIG. 18 is a diagram showing the results of an evaluation experiment conducted by the inventors on the detection method according to the second embodiment for electrocardiogram signals from subjects who are always in an awake state. [Figure 19] FIG. 19 shows the results of an evaluation test conducted by the inventors when the autoencoder used in the detection method according to the second embodiment has an LSTM configuration and an autoencoder with self-attention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] <1. Overview of the Drowsiness Detection Device, Detection Method, and Computer Program>

[0011] (1) In one embodiment, a drowsiness detection device detects drowsiness and includes a calculation device that executes a detection process for detecting drowsiness. The detection process includes obtaining an index value for drowsiness detection from an RRI (RR Interval) generated from a subject's electrocardiogram signal, timing the duration during which the index value exceeds a first threshold, and detecting the subject's drowsiness when the duration exceeds a second threshold.

[0012] The index value for dozing detection is, for example, the following error or statistics. The first threshold is a value at which the index value can be determined as dozing. When it is determined that the index value exceeds the first threshold, the possibility that the subject is asleep is determined.

[0013] To obtain an index value for drowsiness detection, it is possible to use an autoencoder, a classifier using MSPC (Multivariate Statistical Process Control), a deep learning model using LSTM (long short-term memory), or a deep learning model using self-attention.

[0014] The second threshold is the duration for which a state determined to be possibly dozing can be determined as dozing, and is, for example, about 10 seconds. The fact that about 10 seconds can be adopted as the second threshold has been verified by an evaluation experiment performed by the inventors, which will be described later, that when the second threshold is set to 10 seconds, the subject's dozing and awakening from dozing can be detected with high accuracy (FIGS. 17 and 18). This makes it possible to detect dozing with higher accuracy than when dozing is determined based on the index value exceeding the first threshold. This was also verified by a verification test performed by the inventors, which will be described later.

[0015] (2) Preferably, obtaining an index value for drowsiness detection includes providing a value obtained from the RRI as an input value to a model that uses target data as an input value and outputs a value used to detect an anomaly in the target data, and obtaining an index value using an output value from the model. The model is a computation model, for example, a machine-learned model. Specifically, it is an autoencoder, a deep learning model using LSTM, a deep learning model using self-attention, etc. Another example of the model is a model that performs statistical processing on input values, for example, MSPC, etc. By using the model, a value required for calculating the index value can be obtained by output from the model.

[0016] (3) Preferably, the model is an autoencoder, and obtaining the index value includes providing the autoencoder with the RRI generated from the subject's electrocardiogram signal as input data, obtaining output data that is reconstructed data of the input data, and calculating the error of the output data from the RRI generated from the subject's electrocardiogram signal as the index value. This makes it possible to easily detect drowsiness with high accuracy.

[0017] (4) Preferably, the autoencoder is trained by providing the RRI generated from the electrocardiogram signal when the user is awake as training input data. In this case, the first threshold is set to a value different from the input data obtained from the RRI generated from the electrocardiogram signal when the user is awake, that is, a value somewhat larger than the input data obtained from the RRI generated from the electrocardiogram signal including the time when the user is drowsy, and the possibility of drowsyness can be determined by comparing the error with the first threshold.

[0018] (5) Preferably, the model is a classifier using MSPC, and obtaining the index value includes generating an index related to heart rate from an RRI generated from an electrocardiogram signal of the subject, inputting a data set including a plurality of variables HRV among the indexes to the classifier, and obtaining a statistic as the index value. The index related to heart rate is, for example, a heart rate variability (HRV) index.

[0019] The statistic is, for example, at least one of a Q statistic and a T2 statistic. The Q statistic is expressed by equation (5) in FIG. 4, and the T2 statistic is expressed by equation (4) in FIG. 4. In this case, the first threshold is a value at which the Q statistic and / or the T2 statistic can be determined as a dozing state, and is, for example, a control limit value. For example, the control limit value is a statistic obtained by inputting a data set generated only from RRI obtained from an awake person into a classifier. This makes it possible to determine the possibility of dozing by comparing the statistic with the first threshold.

[0020] (6) Preferably, the multiple HRV variables include at least one variable among variables included in the Poincaré index obtained from the RRI generated from the electrocardiogram signal of the subject. This has been verified by a verification test by the inventors to enable more accurate detection of drowsiness.

[0021] (7) Preferably, the data set includes a plurality of canonical variables obtained by performing a canonical correlation analysis on a plurality of HRV variables. This has been verified by a verification test by the inventors to enable more accurate detection of drowsiness.

[0022] (8) Preferably, the data set includes data having time series changes generated from an index related to heart rate obtained from each of a plurality of heart rates of the electrocardiogram signal of the subject. This has been verified by a verification test by the inventors to enable more accurate detection of drowsiness.

[0023] (9) Preferably, the data set includes data having time series changes obtained from multiple variables HRV among indices related to heart rate, including at least one of variables included in the Poincaré index obtained from the RRI generated from the electrocardiogram signal of the subject, and the data having time series changes is generated from indices related to heart rate obtained from each of multiple heart rates of the electrocardiogram signal of the subject for each of multiple canonical variables among canonical variables obtained by performing canonical correlation analysis on the multiple variables HRV. It has been verified by a verification test by the inventors that this enables more accurate detection of dozing.

[0024] (10) Preferably, obtaining the index value for drowsiness detection includes using at least one of an autoencoder, a classifier using MSPC, a deep learning model using LSTM, and a deep learning model using self-attention.

[0025] (11) A detection method according to an embodiment is a method for detecting dozing, comprising: obtaining an index value for dozing detection from an RRI generated from an electrocardiogram signal of a subject; measuring the duration during which the index value exceeds a first threshold; and detecting the subject's dozing when the duration exceeds a second threshold. This realizes the detection device described in (1) to (10).

[0026] (12) A computer program according to an embodiment is a computer program for causing a computer to execute a detection process for detecting dozing, the detection process including obtaining an index value for dozing detection from an RRI generated from an electrocardiogram signal of a subject, timing a duration during which the index value exceeds a first threshold, and detecting the subject's dozing when the duration exceeds a second threshold, thereby causing the computer to function as the detection device according to (1) to (10).

[0027] <2. Examples of Drowsiness Detection Device, Detection Method, and Computer Program>

[0028] [First embodiment]

[0029] 1 is a diagram showing an outline of the configuration of a drowsiness detection device (hereinafter, simply referred to as a detection device) according to the present embodiment. With reference to the figure, detection device 100 includes a calculation device 1 and a heart rate measuring device 2.

[0030] The detection device 100 is a detection device that detects whether a subject is dozing off based on the subject's heart rate. The subject is, for example, a driver of a vehicle. The vehicle is, for example, a bus, a taxi, a passenger car, a train, or the like. The arithmetic device 1 is, for example, at least one of a device that can be mounted on such a vehicle, a device that can be carried by the subject, or the like, and a server that can communicate with these devices via communication such as the Internet.

[0031] Detecting dozing includes detecting an awake state. That is, the state of the subject may be classified into two states, a dozing state and an awake state, and at least one of the states may be detected. Preferably, the detection device 100 detects both the dozing state and the awake state of the subject.

[0032] The arithmetic device 1 and the heart rate measuring device 2 are capable of communicating with each other. The communication may be wireless communication, for example. The wireless communication may be short-range wireless communication such as Bluetooth (registered trademark), communication via the Internet, or the like.

[0033] The heart rate measuring device 2 is a small, lightweight wearable device that is attached to the body of the subject P to measure the heart rate of the subject P. A plurality of electrodes 21 (three in FIG. 1) that are attached to the body surface of the subject P are connected to the heart rate measuring device 2. The three electrodes 21 are, for example, a positive electrode, a negative electrode, and a ground electrode.

[0034] An example of a wearable terminal that functions as the heart rate measuring device 2 is a smart watch that has a heart rate measuring function. Note that the wearable terminal itself may function as the computing device 1 and the heart rate measuring device 2.

[0035] Fig. 2(a) is a diagram showing an example of an electrocardiogram signal. In Fig. 2(a), the vertical axis indicates potential, and the horizontal axis indicates time. When the heartbeat is measured using electrodes 21, potential changes consisting of P to T waves as shown in Fig. 2(a) appear periodically. The peak of potential with the highest potential in the potential change in a unit period is called an R wave, and the heart beats at the timing of the R wave. Heartbeat measuring device 2 transmits R wave data indicating the R wave to calculation device 1.

[0036] Fig. 2(b) shows R-wave data corresponding to the electrocardiogram signal of Fig. 2(a). As shown in Fig. 2(b), the R-wave data is, for example, data representing a rectangular pulse train in which a period corresponding to an R-wave in the electrocardiogram signal (a period in which the signal intensity I exceeds a predetermined intensity threshold Ith) is set to "1" and other periods are set to "0."

[0037] The arithmetic device 1 functions as a detection device that detects that the subject is in a drowsy state based on the R wave data transmitted from the heart rate measuring device 2. The arithmetic device 1 is, for example, a communication terminal such as a smartphone or a personal computer, or is a device that is mounted on a device operated by the subject, such as an in-vehicle device, or that can be mounted on an in-vehicle device or the like.

[0038] Fig. 3 is a block diagram showing an example of a configuration of the arithmetic device 1. Referring to Fig. 3, the arithmetic device 1 includes a processing unit 10. The processing unit 10 is configured with a dedicated microcomputer or the like. The arithmetic device 1 also has a memory 12. The memory 12 stores a program 121 for performing calculations in the processing unit 10. The program 121 has program code for processing in the arithmetic device 1. The program 121 is read and executed by the processing unit 10 connected to the memory in which the program 121 is stored.

[0039] The program 121 may be provided as a program product recorded on a computer-readable non-transitory recording medium. The program 121 may be provided by downloading via a network. The program 121 may also be provided as an application program, or a part or all of the program may be included in the operating system (OS) of the computer.

[0040] In the following description, the processing in the arithmetic device 1 is assumed to be realized by the program 121, but as another example, it may be realized by circuit elements or other hardware.

[0041] The arithmetic device 1 has a communication unit 13 that communicates with the heart rate measuring device 2. The communication unit 13 receives, for example, a wireless signal indicating an R wave transmitted from the heart rate measuring device 2, and inputs the signal to the processing unit 10. Alternatively, the communication unit 13 may be a reading device that accesses a recording medium on which data indicating an R wave is recorded, and reads the data from the recording medium. In other words, obtaining data includes receiving data by communication and reading data from the recording medium.

[0042] The processing unit 10 executes a program 121 stored in the memory 12 to perform a detection process for detecting dozing. The detection process includes an RRI calculation process 101. The RRI calculation process 101 includes calculating an RRI (RR Interval), which is an R-wave interval, from R-wave data such as that shown in FIG. 2(b) input from the communication unit 13 to obtain RRI data. The RRI data is time-series data of RRI variables.

[0043] As another example, the heartbeat measuring device 2 may detect the RRI from the R wave data and output the heartbeat data including the RRI data. The communication unit 13 receives the heartbeat data output from the heartbeat measuring device 2. In this case, the processing unit 10 executes a process of extracting the RRI data included in the heartbeat data instead of the RRI calculation process 101. In other words, by the heartbeat measuring device 2 outputting the heartbeat data including the RRI data, the arithmetic device 1 does not need to perform the process of calculating the RRI. In addition, the data transmitted from the heartbeat measuring device 2 becomes RRI data, which is smaller in size than the wireless signal indicating the R wave, making data transmission easier.

[0044] The detection process includes an index value calculation process 102. The index value calculation process 102 includes calculating an index value for drowsiness detection using the RRI. In detail, the index value calculation process 102 includes using a calculation model 105, inputting a value obtained from the RRI to the model 105, and obtaining an output value from the model 105. The model 105 is a model that takes target data as an input value and outputs a value used to detect an abnormality in the target data.

[0045] Obtaining an index value for dozing detection includes, for example, generating an index related to a heart rate from the RRI. The index related to a heart rate is, for example, a heart rate variability (HRV) index (hereinafter, referred to as an HRV index).

[0046] HRV index includes the following 10 indices. 1) meanNN: Average value of RRI 2) SDNN: Standard deviation of RRI 3) Total Power: RRI variance 4) RMSSD: Root mean square of the difference between adjacent RRIs 5) NN50: The number of times the difference between adjacent RRIs exceeded 50 ms 6) LF: Low frequency (0.04-0.15Hz) power spectrum of PSD (Power Spectrum Density: RRI data power spectral density) 7) HF: High frequency (0.15-0.40Hz) power spectrum of PSD 8) LF / HF: Ratio of LF to HF 9) LFnu: Corrected LF defined as LF / (LF+HF) 10) HFnu: Corrected HF defined as HF / (LF+HF)

[0047] The RMSSD value is larger when the RRI fluctuates more. The RMSSD is used as an index of the activity of the autonomic nervous system. Normal RRI fluctuations are less than 50 milliseconds. Therefore, the NN50 is an index of severe RRI fluctuations. In other words, a large NN50 indicates a large RRI fluctuation.

[0048] LF is primarily an index of sympathetic nervous system activity, and HF is primarily an index of parasympathetic nervous system activity. Therefore, LF / HF represents the ratio of sympathetic nervous system activity to parasympathetic nervous system activity, with a larger value indicating a dominance of the sympathetic nervous system and a smaller value indicating a dominance of the parasympathetic nervous system. LFnu is an index value that emphasizes changes in sympathetic nervous activity, and HFnu is an index value that emphasizes changes in parasympathetic nervous activity.

[0049] Obtaining the index value for drowsiness detection may further include generating a Poincaré index. Fig. 4 is a diagram for explaining the Poincaré index, in which the horizontal axis is RRI and the vertical axis is RRI after one beat. The Poincaré index includes the following three indexes represented by ellipse E in Fig. 4. All three indexes indicate the magnitude of the fluctuation of the RRI. 1) SD1: Standard deviation of the minor axis A2 of ellipse E 2) SD2: Standard deviation of the major axis A1 of ellipse E 3) SD1 / SD2: Ratio of SD1 to SD2

[0050] Obtaining an index value for dozing off includes inputting a data set in which a plurality of indices among the top 10 indices of the HRV index are arranged in time series as variables to the model 105, and obtaining an output value from the model 105. In the detection device 100 according to the first embodiment, the model 105 is a computation model that performs statistical processing on the input value, and is, for example, a classifier that uses MSPC (Multivariate Statistical Process Control). That is, in the first embodiment, obtaining an index value for dozing off includes applying a classifier to a data set in which a plurality of indices among the top 10 indices of the HRV index are arranged in time series as variables.

[0051] MSPC is a statistical method based on principal component analysis, a multivariate analysis aimed at extracting features and reducing the dimensionality of data, by linearly combining variables to create new synthetic variables called principal components in order to capture correlations between variables.

[0052] In MSPC, principal components are set one after another in the following procedure: the first principal component is set in the direction that best expresses the data, and the second principal component is set in the direction that best expresses the data fluctuations that cannot be expressed by the first principal component in the space perpendicular to the first principal component. The direction that best expresses the data is the direction in which the variance of the principal component scores is the largest, and the principal component scores refer to the coordinates on the principal component axes, that is, the values ​​obtained by projecting the data into the space spanned by the principal components. In the example in Figure 4, the direction of the long axis A1 is the direction of the first principal component, and the direction of the short axis A2 is the direction of the second principal component.

[0053] Here, the data matrix X is defined as shown in formula (1) in Figure 4 using the number of variables P and the number of samples N. Each variable is assumed to be standardized. In addition, the singular value decomposition of the data matrix X is defined as shown in formula (2) in Figure 4 using the orthogonal matrices U and V, and the diagonal matrix S in which the singular values ​​s and r are arranged in descending order in the diagonal elements.

[0054] In this case, the number of principal components used is R, and the rth principal component is the loading matrix V R The rth column vr of R Using the rth column vr of

[0055] In the first embodiment, when a data set arranged in time series is input, the model 105 outputs at least one of Q statistics and T2 statistics according to MSPC. In the index value calculation process 102, the value output by the model 105 is set as an index value for drowsiness detection. The Q statistics and T2 statistics are indexes for evaluating the degree of deviation of a variable from a normal value.

[0056] The T2 statistic represents the distance of each sample from the mean in the principal component space obtained by compressing the original variables. In other words, the T2 statistic represents the deviation of the input data set from the mean. In detail, the T2 statistic is expressed by formula (4) in FIG. 4 using the r-th principal component tr expressed by formula (3) in FIG. 4 and the standard deviation σtr of the r-th principal component score tr. The T2 statistic corresponds to the Mahalanobis distance. In other words, the smaller the T2 statistic, the closer the sample is to the mean of the model construction data.

[0057] The Q statistic is an index that evaluates deviations from correlations between variables in an input data set. It can detect abnormalities in the correlation between variables. The Q statistic is calculated from the residuals, and is specifically expressed by equation (5) in Figure 4. The Q statistic is also called the squared prediction error, and is considered to represent the part of the data that cannot be expressed by a model. MSPC simultaneously monitors the T2 statistic and the Q statistic, and if either one exceeds the control limit, it is determined to be an abnormality.

[0058] The data set input to the model 105 may be one in which the values ​​of the 10 indices are arranged in a time series, with the values ​​of each indices as elements, or one in which some of the indices are arranged in a time series. Preferably, the data set input to the model 105 further includes at least one of the three indices above the Poincaré index as an element. For example, the elements of the input data set may be one in which the values ​​of a total of 13 indices, including the 10 indices above the HRV index and the three indices above the Poincaré index, are arranged in a time series. Also, the data set may be a normalized data set.

[0059] Preferably, in obtaining the index value for drowsiness detection, the processing unit 10 performs canonical correlation analysis on the data set, and inputs the data set consisting of the obtained canonical variables to the classifier. The canonical correlation analysis is a method for obtaining correlation between linearly transformed values ​​of a plurality of data, and the canonical correlation analysis includes multiplying the data set by a sample canonical coefficient.

[0060] The sample canonical coefficients are coefficients that express the individual differences inherent in the data set as weights. Specifically, the sample canonical coefficients are obtained by performing a canonical correlation analysis on the matrix (X) shown in Figure 5 and the corresponding one-hot matrix (Y). The matrix (X) is a matrix in which the data sets generated using the RRI obtained during wakefulness for each of N subjects, each numbered n (n = 1, 2, 3, ..., n), are vertically concatenated. A one-hot matrix refers to a vector in which only one element is set to 1 and the other elements are set to 0. Specifically, it is a concatenation of N vectors in which the time series data of the nth person is set to 1 and the rest are set to 0.

[0061] In the processing unit 10, at least some of the obtained canonical variables are extracted as a new data set, and input to the model 105. For example, the first, tenth, eleventh, and twelfth variables are extracted from the obtained canonical variables as a new data set, and input to the model 105. At this time, preferably, the new data set has variables obtained from at least one of the three indices of the Poincaré indices as elements.

[0062] Preferably, in obtaining the index value for drowsiness detection, the processing unit 10 dynamizes the data set, and inputs the obtained data set of time-series data having time-series changes to the model 105. The data set to be dynamized may be a data set consisting of canonical variables obtained by canonical correlation analysis, or may be a data set consisting of variables before being subjected to canonical correlation analysis.

[0063] The dynamicalization refers to generating data having a time series change for each element of a new data set by using values ​​of indices related to a plurality of heartbeats obtained from each of a plurality of heartbeats of an electrocardiogram signal of a subject. Specifically, the dynamicalization may be to generate data having a time series change by adding past data as an element.

[0064] A specific example of a method for dynamicizing the first data set D1 will be described with reference to Fig. 6. The first data set D1 is time-series data having elements of the values ​​of heart rates 1 to 5 for each of four variables, namely, a first variable, a tenth variable, an eleventh variable, and a twelfth variable.

[0065] In making the first data set D1 dynamic, for example, a second data set D2 corresponding to time series data one beat later is used for the first data set D1. The second data set D2 is time series data in which the elements are the values ​​of the first, tenth, eleventh, and twelfth variables from the first heartbeat to the heartbeats 2 to 5 one beat later in the first data set D1, which are the same as those in the first data set D1.

[0066] In the dynamicization, the third data set D3 is generated by concatenating the second data set D2 to the first data set D1. The third data set D3 is time series data consisting of 8 variables and 32 elements, including values ​​of heartbeats 1 to 4 of the first variable, the tenth variable, the eleventh variable, and the twelfth variable at the heartbeat of the first data set D1, and values ​​of heartbeats 2 to 5 one beat later of the first variable, the tenth variable, the eleventh variable, and the twelfth variable at the heartbeat.

[0067] In the dynamicization of the first data set D1, a data set of a heartbeat after the first heartbeat of the second data set D2 may be further combined with the first data set D1. The further beat may be, for example, two beats later. In the example of FIG. 6, when the data set two beats later is further combined, the generated data set becomes time series data consisting of 12 variables and 48 elements.

[0068] In the index value calculation process 102, a data set having time-series changes obtained by dynamicization, or a normalized version of that data set, is input to the model 105 to obtain Q statistics and T2 statistics as output values ​​of the model 105, and at least one of them is used as the index value.

[0069] The detection process includes a determination process 103. The determination process 103 includes comparing a plurality of index values ​​calculated from successive heartbeats by the index value calculation process 102 with a first threshold, and measuring a duration DT1 during which the index value exceeds the first threshold and a duration DT2 during which the index value falls below the first threshold. The index value is at least one of a Q statistic and a T2 statistic obtained by inputting a data set to a classifier.

[0070] The first threshold value is a value at which the Q statistic and / or the T2 statistic may be determined as a dozing state, and is, for example, a control limit value Q0 and / or a control limit value T0. The control limit value Q0 is, for example, a Q statistic obtained by inputting into the model 105 a data set generated from RRIs obtained only from an awake person and not including RRIs during dozing. The control limit value T0 is, for example, a T2 statistic obtained by inputting into the model 105 a data set generated from RRIs obtained only from an awake person and not including RRIs during dozing. As another example, the control limit value may be a value stored in advance as the Q statistic and / or the T2 statistic during wakefulness.

[0071] The determination process 103 includes determining whether the index value obtained from the electrocardiogram signal of the subject is in a state deviating from the index value obtained from the electrocardiogram signal in the wakeful state by comparing the index value with a first threshold. That is, in this case, when the index value exceeds the first threshold, it is determined that the subject is not in a wakeful state, that is, is in a dozing state. When the index value falls below the first threshold, it is determined that the subject is in a wakeful state, that is, is not in a dozing state.

[0072] Here, noise may be mixed into the electrocardiogram due to factors such as premature ventricular contraction (PVC), a common arrhythmia that can occur even in healthy individuals, and the subject's body movements. As a result, an accurate RRI may not be obtained. In such cases, even if the index value exceeds the first threshold, this may not necessarily mean that the subject is dozing. Therefore, a more accurate dozing detection is desired.

[0073] In addition, when the index value based on the obtained RRI falls below the first threshold, the driver is determined to be awake, but in reality, the driver may not have fully woken up from a doze. Therefore, more accurate detection of wakefulness is also desired. In other words, more accurate detection of doze is desired.

[0074] In this regard, in the judgment process 103, a duration DT1 exceeding the control limit value is measured for at least one of the Q statistics and the T2 statistics obtained as the index value. The judgment process 103 includes comparing the obtained duration DT1 with a second threshold value and judging that the duration DT1 exceeds the second threshold value as a state of dozing. Also, a duration DT2 below the control limit value is measured for at least one of the Q statistics and the T2 statistics. The judgment process 103 includes comparing the obtained duration DT2 with a second threshold value and judging that the duration DT2 exceeds the second threshold value as a state of wakefulness.

[0075] The second threshold is a threshold t0 for determining the duration of a state determined to have a possibility of dozing off as a dozing state and for determining the duration of a state determined to have a possibility of waking off as a waking state, and is, for example, about 10 seconds. Note that different thresholds may be used for determining dozing off and waking off.

[0076] That is, the detection device 100 detects the subject's dozing when the index values ​​of Q statistics and / or T2 statistics exceed the control limit value for a period of time that is equal to or longer than the threshold value t0, which is a duration that can be determined as dozing. This makes it possible to detect dozing with higher accuracy than when dozing is determined based on the index values ​​of Q statistics and / or T2 statistics exceeding the control limit value.

[0077] Furthermore, the detection device 100 detects the subject's wakefulness when the index values, Q statistics and / or T2 statistics, fall below the control limit value for a period of time that is equal to or longer than a threshold value t0, which is a duration that can be determined as a wakefulness state. This makes it possible to detect the wakefulness state with higher accuracy than when the index values, Q statistics and / or T2 statistics, fall below the control limit value, and in other words, to detect dozing with higher accuracy.

[0078] The detection process may include an output process 104. The output process 104 includes outputting the detection result. The output of the detection result may be one or more of audio output, display output, light output, vibration, and transmission to another device. The output process 104 includes transmitting a control signal to an output device (not shown) to instruct the output.

[0079] The output detection result may be a detection result of whether the subject is dozing or not dozing, i.e., is in an awake state. This makes it possible to notify in real time whether the subject is dozing or in an awake state.

[0080] An example of a method for detecting dozing off by the detection device 100 according to the first embodiment will be described with reference to the flowchart of Fig. 7. The flowchart of Fig. 7 is executed by the processing unit 10 of the arithmetic device 1. The detection method according to the first embodiment is started, for example, when R wave data is obtained.

[0081] 7, the processing unit 10 first initializes each parameter T1, T2, f, and st used in the process (step S101). The parameters T1 and T2 are parameters for the duration of at least one of the Q statistic and the T2 statistic obtained as the index value exceeding the control limit value and the duration of the Q statistic falling below the control limit value, respectively. The parameter f is a parameter representing whether the state is likely to be dozing or awake, obtained by the index value and the first threshold value. The parameter st is a parameter representing the determination result of whether the subject is dozing or awake.

[0082] The initial values ​​of parameters T1 and T2 are 0. The initial value of parameter f is 0, which indicates the possibility of awakening. The initial value of parameter st is 0, which indicates the judgment result of awakening.

[0083] The processing unit 10 acquires the RRI (step S103), and calculates the HRV index and the Poincaré index using the RRI (step S105). The processing unit 10 performs a canonical correlation analysis on a data set having the values ​​of at least some of the 13 indices as elements (step S107), and obtains a data set including canonical variables.

[0084] The processing unit 10 generates a data set having a time series change by dynamicizing the data set obtained in step S107 (step S109). The processing unit 10 normalizes the obtained data set (step S111) and inputs it to the model 105, which is a classifier using MSPC (step S113). As a result, Q statistics and T2 statistics are obtained, and at least one of them is indexed (step S115).

[0085] The processing unit 10 executes a judgment process using the statistics as the index value (step S300). The processing unit 10 repeats the above process to judge in real time whether the subject is asleep or awake as represented by the obtained R wave data.

[0086] 8, the processing unit 10 first compares the index value with a first threshold value (step S301). In the detection method according to the first embodiment, the statistical quantity used as the index value is compared with a control limit value.

[0087] If the index value exceeds the first threshold (YES in step S301) and the parameter f indicating the possibility of the state is 1, that is, indicating dozing (YES in step S303), the processing unit 10 adds the RRI at that time to the duration DT1 during which the index value exceeded the first threshold (step S305).

[0088] If the index value exceeds the first threshold (YES in step S301) and the parameter f indicating the possibility of the state is 0, that is, indicating awake (NO in step S303), the processing unit 10 initializes the duration DT1 and changes the parameter f to 1 indicating dozing (step S307), and adds the RRI at that time to the duration DT1 (step S305). This ends the measurement of the duration DT2 in the state where there is a possibility of awake, and starts the measurement of the duration DT1 in the state where there is a possibility of dozing.

[0089] The processing unit 10 compares the duration DT1 with a second threshold value, ie, a threshold value t0 of the duration. As a result, if the duration DT1 exceeds the threshold value t0 (YES in step S309), the processing unit 10 determines that the user is dozing, and sets the parameter st indicating the determination result to 1, which indicates dozing (step S311).

[0090] On the other hand, if the index value is below the first threshold (NO in step S301) and the parameter f indicating the possibility of the state is 0, that is, indicating awakening (YES in step S313), the processing unit 10 adds the RRI at that time to the duration DT2 during which the index value is below the first threshold (step S315).

[0091] If the index value is below the first threshold (NO in step S301) and the parameter f indicating the possibility of the state is 1, that is, indicating dozing (NO in step S313), the processing unit 10 initializes the duration DT2 and changes the parameter f to 0 indicating wakefulness (step S307), and then adds the RRI at that time to the duration DT2 (step S315). This ends the measurement of the duration DT1 in the state where there is a possibility of dozing, and starts the measurement of the duration DT2 in the state where there is a possibility of wakefulness.

[0092] The processing unit 10 compares the duration DT2 with a second threshold, ie, a threshold t0 of duration. As a result, if the duration DT2 exceeds the threshold t0 (YES in step S319), the processing unit 10 determines that the subject is awake, and sets the parameter st indicating the determination result to 0, which indicates awake (step S321).

[0093] If the parameter st indicating the determination result after the above processes is 1 indicating dozing (YES in step S323), the processing unit 10 executes output processing such as outputting a warning (step S325).

[0094] By performing the determination process shown in FIG. 8 in the detection method according to the first embodiment, even if the statistic, which is the index value, exceeds the first threshold, the subject is not determined to be dozing unless the duration DT1 exceeds the second threshold. Also, even if the index value falls below the first threshold after being determined to be dozing, the subject is not determined to be awake unless the duration DT2 after the fall does not exceed the second threshold. In other words, false detection of dozing due to noise in the RRI and the like is prevented, and detection of awakeness from a dozing state that is not completely awake is also prevented. Therefore, the subject's dozing is detected with high accuracy.

[0095] In order to verify which of the Q statistics and the T2 statistics is used as an index value with better detection accuracy in the detection method according to the first embodiment, the inventors detected dozing using the T2 statistics and the Q statistics of the same subject. FIG. 9 (A) shows the transition of the T2 statistics, and FIG. 9 (B) shows the transition of the Q statistics for the same electrocardiogram signal. In FIGS. 9 (A) and (B), the horizontal axis shows time, and the vertical axis shows the index value. The control limit value T0 of the T2 statistics is set to 10, and the control limit value Q0 of the Q statistics is set to 0.2. In FIGS. 9 (A) and (B), the subject is dozing during a period D from time T2 to T3.

[0096] In the results of Fig. 9(A), when the T2 statistics is used, the T2 statistics exceeds the control limit value T0 from time T0, and when it reaches time T1, the duration DT exceeds the threshold value t0, and dozing is detected. On the other hand, as shown in Fig. 9(B), in this measurement result, when the Q statistics is used, dozing is not detected even at time T1, where dozing is detected in (A).

[0097] Therefore, the results of FIG. 9 verify that when either the Q statistic or the T2 statistic is used as the index value, the T2 statistic is better.

[0098] Next, the inventor performed first to fifth tests to verify the detection accuracy of each of the following cases regarding index values ​​used as a data set to be input to model 105 in the detection method according to the first embodiment: when the value of the HRV index is used as is, when only the value of the Poincaré index is used, when a canonical variable obtained by performing a canonical correlation analysis on the values ​​of the HRV index and the Poincaré index is used, when time-series data having a time-series change obtained by dynamicizing the values ​​of the HRV index and the Poincaré index is used, and when time-series data having a time-series change obtained by dynamicizing the canonical variable obtained by performing a canonical correlation analysis on the values ​​of the HRV index and the Poincaré index is used.

[0099] 10 to 14 are diagrams showing the ROC (Receiver Operating Characteristic) curves obtained from Tests 1 to 5, respectively. The ROC curves are plots of FPR (False Positive Rate) on the horizontal axis and TPR (True Positive Rate) on the vertical axis, plotting TPR against FPR.

[0100] The inventors evaluated the data sets to be input to the model 105 using the AUC (area under the receiver operating characteristic curve), sensitivity, and specificity obtained from the first to fifth test results. The AUC is obtained by the area under the ROC curve. The AUC takes values ​​from 0 to 1, and the higher the AUC, the higher the detection performance. The sensitivity is the true positive rate for true positives and false negatives. The specificity is the true negative rate for false positives and true negatives.

[0101] In the first test shown in FIG. 10, AUC was 0.8349, sensitivity was 0.78, and specificity was 0.71. In the second test shown in FIG. 11, AUC was 0.8650, sensitivity was 0.78, and specificity was 0.88. In the third test shown in FIG. 12, AUC was 0.8434, sensitivity was 0.78, and specificity was 0.71. In the fourth test shown in FIG. 13, AUC was 0.8534, sensitivity was 0.89, and specificity was 0.71. In the fifth test shown in FIG. 14, AUC was 0.8904, sensitivity was 0.89, and specificity was 0.75.

[0102] Comparing the results of the first and second tests, the second test had higher AUC and specificity than the first test, and it was found that the detection accuracy was higher when the Poincaré index was used.Comparing the results of the first test with the third to fifth tests, it was found that the detection accuracy was higher than when the HRV index value was used as it was in all cases: when the canonical variables obtained by canonical correlation analysis of the values ​​of the HRV index and the Poincaré index were used, when time series data having time series changes obtained by dynamicizing the canonical variables obtained by canonical correlation analysis were used, and when time series data having time series changes obtained by dynamicizing the canonical variables obtained by canonical correlation analysis were used.

[0103] Furthermore, when the results of the third to fifth tests were compared, it was found that the detection accuracy was higher when time series data having time series changes obtained by dynamicizing the canonical variables obtained by canonical correlation analysis of the values ​​of HRV and Poincaré index was used than when canonical variables obtained by canonical correlation analysis were used, and also that the detection accuracy was higher when time series data having time series changes obtained by dynamicizing the canonical variables obtained by canonical correlation analysis was used than when time series data having time series changes obtained by dynamicizing them.

[0104] From the above results, it was verified that in the detection method according to the first embodiment, the detection accuracy is improved by using at least a part of the Poincaré index value as the index value rather than using the value of the HRV index as it is. Furthermore, it was verified that the detection accuracy is improved by using a canonical variable obtained by performing canonical correlation analysis on the index value rather than using the index value as it is. Furthermore, it was verified that the detection accuracy is improved by using time series data having a time series change obtained by dynamicizing the index value rather than using the index value as it is. Furthermore, it was verified that the detection accuracy is improved by using time series data having a time series change obtained by dynamicizing the canonical variable obtained by performing canonical correlation analysis on the index value rather than using the index value as it is.

[0105] In addition, in the dynamicization, in order to verify how many beats after the first data set the second data set should be linked to the first data set, the inventors evaluated the detection accuracy of the detection device 100 by changing the difference in the number of beats of the second data set from the first data set. The detection accuracy was evaluated using AUC. As a result, the results shown in FIG. 15 were obtained. In FIG. 15, the vertical axis represents AUC, and the horizontal axis represents the delay in the number of beats of the second data set from the first data set. From the results shown in FIG. 15, it can be seen that in this test, the detection accuracy was highest when the second data set that was eight beats behind the first data set was linked to the first data set. This verified that it is good to use a second data set that is delayed by about eight beats.

[0106] [Second embodiment]

[0107] In the detection device 100 according to the second embodiment, the model 105 is a machine-learned model, for example, an autoencoder. In the index value calculation process 102 according to the second embodiment, the RRI is given to the autoencoder as input data, and an index value for drowsiness detection is obtained using output data that is reconstructed data of the input data.

[0108] An autoencoder is an example of a deep learning model using LSTM (long short-term memory) and self-attention, and is a deep learning model that is machine-learned using training data to reconstruct input data and output values ​​close to the input data. Therefore, if the model 105 is an autoencoder, when data with a tendency different from the training data is input to the model 105, the error of the output data from the input data increases. In other words, when abnormal data is input to the model 105 that is machine-learned using normal data as training data, the error increases. The detection device 100 according to the second embodiment detects drowsiness by detecting abnormal data using this characteristic of the model 105. This makes it possible to easily detect drowsiness with high accuracy.

[0109] In the index value calculation process 102, the processing unit 10 uses the model 105, which is an autoencoder that is machine-learned using only the RRI data in the wakeful state as training data. That is, the training data includes the RRI data in the wakeful state, but does not include the RRI data in the dozing state. The model 105 may be stored in the calculation device 1, or may be stored in another device accessible by the communication unit 13.

[0110] The index value calculation process 102 according to the second embodiment includes inputting the RRI obtained from the R-wave data to the model 105 and obtaining output data. In the index value calculation process 102, the processing unit 10 stores the RRI every time it is obtained, and inputs a predetermined number of RRIs to the model 105. The predetermined number is, for example, 30. At this time, 30 pieces of output data are obtained from the model 105.

[0111] The index calculation process 102 includes calculating an error of the output data from the model 105 from the RRI input to the model 105. In the detection method according to the second embodiment, the obtained error is used as the index. The error is, for example, a mean square error.

[0112] The determination process 103 according to the second embodiment includes comparing the error calculated by the index value calculation process 102 with a first threshold and timing a duration DT1 during which the error exceeds the first threshold. The determination process 103 also includes timing a duration DT2 during which the error falls below the first threshold. The first threshold here is a value e0 that indicates the deviation from the input data to such an extent that the data is considered to be abnormal, and may be set in advance.

[0113] The processing unit 10 compares the error with the first threshold value to determine whether the input RRI is data with the same tendency as the RRI obtained from the R wave data in the wakeful state or data with a different tendency. The first threshold value, threshold value e0, is an error that can determine that the input data is different from the input data obtained only in the wakeful state, that is, a value that can determine that the subject is dozing. As a result, when the error is larger than threshold value e0, it is determined that the subject is not in an awake state, that is, there is a possibility that the subject is dozing, and when the error is below threshold value e0, it is determined that the subject is in an awake state.

[0114] The judgment process 103 according to the second embodiment also measures the duration DT1 during which the error exceeds the threshold e0 and the duration DT2 during which the error falls below the threshold e0. The judgment process 103 includes detecting the subject's dozing off when the obtained duration DT1 exceeds a second threshold, and detecting the subject's awakening when the duration DT2 exceeds the second threshold. The second threshold is a duration threshold t0 at which it can be determined that the subject is dozing off or awakening, and is, for example, about 10 seconds.

[0115] As a result, the detection device 100 according to the second embodiment also detects the subject's dozing when the state in which the error exceeds the threshold e0 continues for a duration equal to or longer than the threshold t0, which is a duration that can be determined as dozing. Also, the detection device 100 detects the subject's wakefulness when the state in which the error falls below the threshold e0 continues for a duration equal to or longer than the threshold t0, which is a duration that can be determined as wakefulness. As a result, dozing can be detected with higher accuracy than when dozing is determined when the error exceeds the first threshold or when wakefulness is determined when the error falls below the first threshold.

[0116] An example of a method for detecting dozing off by the detection device 100 according to the second embodiment will be described with reference to the flowchart of Fig. 16. The flowchart of Fig. 16 is executed by the processing unit 10 of the arithmetic device 1. The detection method according to the second embodiment also starts, for example, when R wave data is obtained.

[0117] 16, the processing unit 10 first initializes parameters T1, T2, f, and st used in the process (step S201).

[0118] The processing unit 10 acquires the RRI (step S203), standardizes it, and then stores it in memory (step S205). The processing unit 10 repeats the above calculation of the RRI until the RRI stored in the memory reaches a number suitable for input to the model 105, which is an autoencoder (NO in step S207).

[0119] When the number of RRIs stored in the memory reaches a suitable number for input to the model 105 (YES in step S207), the processing unit 10 inputs these RRIs to the model 105, which is an autoencoder (step S209). As a result, a value in which the RRI is reconstructed from the model 105 is obtained as output data. The processing unit 10 calculates the error of the input RRI from the output data of the model 105 (step S211), and sets it as an index value.

[0120] The processing unit 10 executes a judgment process using the error as an index value (step S300). The judgment process in step S300 is the process in FIG. 8. The processing unit 10 repeats the above process to judge in real time whether the subject is asleep or awake as represented by the obtained R-wave data. Here, by using the RRI as an input to the model 105 without calculating the HRV index or the like from the RRI, the index value can be easily obtained, and a decrease in the amount of information when various indexes are calculated from the RRI can be prevented, thereby improving the detection accuracy.

[0121] In addition, by performing the determination process shown in FIG. 8 in the detection method according to the second embodiment, even if the error, which is the index value, exceeds the first threshold, the subject is not determined to be dozing unless the duration DT1 exceeds the second threshold. Also, even if the index value falls below the first threshold after being determined to be dozing, the subject is not determined to be awake unless the duration DT2 after the fall does not exceed the second threshold. In other words, false detection of dozing due to noise in the RRI and the like are prevented, and detection of awakeness in a state that is not completely awake from a dozing state is also prevented. Therefore, the subject's dozing is detected with high accuracy.

[0122] [Evaluation test]

[0123] The inventors performed an evaluation test to evaluate the detection accuracy of the detection method according to the second embodiment. In the evaluation test, an error threshold e0, which is an index value, was set to 0.6, and a duration threshold t0 was set to 10 seconds.

[0124] 17 and 18 are diagrams showing the results of an evaluation experiment on the detection method according to the second embodiment, where the vertical axis shows the magnitude of error from the input data of the reconstructed data, and the horizontal axis shows time. Fig. 17 shows the detection results from an electrocardiogram signal including a state in which the subject is dozing, and Fig. 18 shows the detection results from an electrocardiogram signal in which the subject is always awake. In these diagrams, the gray areas indicate the periods in which the detection device 100 detected the subject's dozing.

[0125] In detail, in the error transition shown in Fig. 17, there is a range where the error exceeds the threshold e0 for 10 seconds or more, and dozing is detected in that range. In addition, when the error subsequently falls below the threshold e0 for more than 10 seconds, an awakening from dozing is detected. On the other hand, in the error transition shown in Fig. 18, although there is a time when the error exceeds the threshold e0, the state where it exceeds the threshold e0 does not continue for 10 seconds or more, so dozing is not detected.

[0126] 17 and 18, the detection method according to the second embodiment prevents a person from being detected as dozing when awake, even if the error exceeds the threshold e0 that indicates a dozing state due to noise, fluctuations in heart rate, etc. In other words, the accuracy of dozing detection can be improved.

[0127] The inventors performed an evaluation test of the detection method according to the second embodiment for the configuration of the autoencoder, LSTM and self-attention, and evaluated the configuration of the autoencoder. In this evaluation test, 26 subjects performed a driving experiment using a driving simulator, and the RRI calculated from the electrocardiogram signals obtained during the experiment was used as a data set. Each data set had a length of 15 to 20 minutes.

[0128] The autoencoder configuration was evaluated using the AUC (area under the receiver operating characteristic curve) of the dataset used. The AUC is obtained by the area under the locus of the true positive rate (TPS) and false positive rate (FPS) corresponding to each threshold e0 when the duration threshold t0 is fixed and the error threshold e0 is changed. The locus of the true positive rate and false positive rate is called the ROC curve. The AUC ranges from 0 to 1, and the higher the AUC, the higher the detection performance. In the evaluation test, the threshold t0 was changed to obtain the AUC for each value of the threshold t0.

[0129] Figure 19 is a graph showing the results of an evaluation test for an autoencoder configuration of LSTM and self-attention, with the vertical axis representing AUC and the horizontal axis representing the value of threshold t0. The results in Figure 19 verify that the detection accuracy of an autoencoder configuration of self-attention is higher than that of an autoencoder configuration of LSTM, regardless of the value of threshold t0.

[0130] <3. Notes> The present invention is not limited to the above-described embodiment, and various modifications are possible. [Explanation of symbols]

[0131] 1: Arithmetic device 2: Heart rate monitor 10: Processing section 12: Memory 13: Communications Department 21: Electrode 100:Detection device 101: RRI calculation process 102: Index value calculation process 103: Judgment process 104: Output processing 105: Model 121: Program A1:Long axis A2:Short axis D: Period D1: First data set D2: The second data set D3: The third data set E: Ellipse HF: Correction I : Signal strength Ith: Intensity threshold LF: Correction P: Subject Q0: Control limit T: Duration T0: Control limit e0: threshold t0: threshold

Claims

1. A detection device for detecting drowsiness, comprising: A computing device that executes a detection process for detecting drowsiness, The detection process includes: An index value for detecting drowsiness is obtained from the RRI (RR Interval) generated from the subject's electrocardiogram signal, measuring the duration during which the index value exceeds a first threshold; detecting the subject falling asleep when the duration exceeds a second threshold; Obtaining the index value for the drowsiness detection includes providing a value obtained from the RRI as an input value to a model that uses target data as an input value and outputs a value used to detect an anomaly in the target data, and obtaining the index value using an output value from the model; The model is an autoencoder, Obtaining the index value for the drowsiness detection includes: Providing an autoencoder with an RRI generated from the electrocardiogram signal of the subject as input data to obtain output data that is reconstruction data of the input data; Calculating an error in the output data from an RRI generated from the electrocardiogram signal of the subject as the index value. Drowsiness detection device.

2. The autoencoder is trained by providing the RRI generated from the electrocardiogram signal during wakefulness as training input data. The drowsiness detection device according to claim 1 .

3. A method for detecting drowsiness, comprising: Obtaining an index value for detecting drowsiness from an RRI generated from the subject's electrocardiogram signal; measuring the duration during which the index value exceeds a first threshold; detecting that the subject is dozing when the duration exceeds a second threshold; Obtaining the index value for the drowsiness detection includes providing a value obtained from the RRI as an input value to a model that uses target data as an input value and outputs a value used to detect an anomaly in the target data, and obtaining the index value using an output value from the model; The model is an autoencoder, Obtaining the index value for the drowsiness detection includes: Providing an autoencoder with an RRI generated from the electrocardiogram signal of the subject as input data to obtain output data that is reconstruction data of the input data; Calculating an error in the output data from an RRI generated from the electrocardiogram signal of the subject as the index value. Detection method.

4. A computer program for causing a computer to execute a detection process for detecting drowsiness, The detection process includes: Obtaining an index value for detecting drowsiness from an RRI generated from the subject's electrocardiogram signal; measuring the duration during which the index value exceeds a first threshold; detecting that the subject is dozing when the duration exceeds a second threshold; Obtaining the index value for the drowsiness detection includes providing a value obtained from the RRI as an input value to a model that uses target data as an input value and outputs a value used to detect an anomaly in the target data, and obtaining the index value using an output value from the model; The model is an autoencoder, Obtaining the index value for the drowsiness detection includes: Providing an autoencoder with an RRI generated from the electrocardiogram signal of the subject as input data to obtain output data that is reconstruction data of the input data; Calculating an error in the output data from an RRI generated from the electrocardiogram signal of the subject as the index value. Computer program.

Images