Depression risk evaluation device, depression risk evaluation model generation device, depression risk evaluation meter, depression risk evaluation system, depression risk evaluation method, depression risk evaluation model generation method, and program

Abstract

According to the present invention, a storage unit stores a depression risk evaluation model for calculating an evaluation value indicating the risk of depression by using a feature amount calculated from a prescribed frequency component obtained by decomposing brain waves by means of time-frequency analysis and having a value that changes in accordance with a change in the frequency of occurrence of a phase reset phenomenon. A feature amount calculation unit calculates the same type of feature amount as the feature amount used in the depression risk evaluation model from a prescribed frequency component obtained by decomposing the measured brain waves of a person-to-be-evaluated by means of time-frequency analysis. An evaluation unit calculates an evaluation value indicating the risk of depression of the person-to-be-evaluated by using the feature amount, which has been calculated by the feature amount calculation unit, as the input to the depression risk evaluation model.

Description

Depression risk assessment device, depression risk assessment model generation device, depression risk assessment meter, depression risk assessment system, depression risk assessment method, depression risk assessment model generation method and program

[0001] The present invention relates to a depression risk assessment device, a depression risk assessment model generation device, a depression risk assessment meter, a depression risk assessment system, a depression risk assessment method, a depression risk assessment model generation method, and a program. This application claims priority based on Japanese Patent Application No. 2023-119047, filed in Japan on July 21, 2023, the contents of which are incorporated herein by reference.

[0002] Depression is a very common mental illness, and anyone can suffer from it. Diagnosis of depression is primarily based on interviews and observations by specialists. However, early detection and treatment are crucial for depression, and biomarkers that can detect possible illness at an early stage are highly desirable. As a candidate for an easy-to-use biomarker, biomarkers using electroencephalography (EEG) have been studied. Examples include research on EEG at the population level for both depressed patients and healthy individuals (see, for example, Non-Patent Document 1), technology for assessing the severity of depression based on EEG (see, for example, Patent Document 1), and technology for identifying signs of depression using information including EEG (see, for example, Patent Document 2).

[0003] JP 2020-081006 A JP 2022-106068 A

[0004] Antora Dev et. al," Exploration of EEG-Based Depression Biomarkers Identification Techniques and Their Applications: A Systematic Review," IEEE Access, Vol.10, pp.16756-16781, Jan. 2022

[0005] In recent years, machine learning models have made it possible to accurately classify depressed patients and healthy individuals using electroencephalograms and other data. However, it is difficult to easily detect pre-depressed patients or assess depression risk using biomarkers for the early detection of depression. This is likely because the discrimination is based on EEG brain activity, which reflects the cognitive decline already caused by depression. For example, the technologies disclosed in Patent Document 1 and Non-Patent Document 1 do not detect pre-depressed patients or assess depression risk for the early detection of depression. Furthermore, when using stimulation-evoked electroencephalograms (EEGs), which are brain waves generated when a stimulus is applied to the subject, measurement takes time and requires a large amount of equipment and manpower, making measurement difficult and not highly accurate. For example, the technology disclosed in Patent Document 2 uses EEGs generated when the subject is speaking to identify signs of depression, which requires time for measurement and makes it difficult to perform measurement alone.

[0006] In view of the above circumstances, the present invention aims to provide a depression risk assessment device, a depression risk assessment model generation device, a depression risk assessment meter, a depression risk assessment system, a depression risk assessment method, a depression risk assessment model generation method, and a program that can easily assess the risk of depression.

[0007] One aspect of the present invention is a depression risk assessment device comprising: a memory unit that stores a depression risk assessment model for calculating an assessment value representing the risk of depression using features that are calculated from predetermined frequency components obtained by decomposing electroencephalograms using time-frequency analysis and whose values ​​change in accordance with changes in the frequency of occurrence of the phase resetting phenomenon; a feature calculation unit that calculates feature values ​​of the same type as the feature values ​​used in the depression risk assessment model from the predetermined frequency components obtained by decomposing the measured electroencephalograms of the subject using time-frequency analysis; and an assessment unit that calculates an assessment value representing the risk of depression of the subject using the feature values ​​calculated by the feature calculation unit as input to the depression risk assessment model.

[0008] One aspect of the present invention is the above-mentioned depression risk assessment device, wherein the assessment value represents whether the risk of depression is low or high, or the severity of depression, and the risk of depression includes the possibility that the person is already depressed, the possibility that the person is in a pre-depressive state, the possibility that the person is not currently depressed but will become depressed in the future, or the possibility that the person is not currently depressed but will likely become depressed if they do become depressed.

[0009] One aspect of the present invention is the depression risk assessment device described above, wherein the feature quantity is any one of the occurrence frequency of the phase resetting phenomenon, the normalized occurrence frequency, the average phase velocity, the normalized amplitude, the rate of increase of the occurrence frequency, the average phase velocity, or the normalized amplitude rate of increase, and the temporal fluctuation of the occurrence frequency, the average phase velocity, or the normalized amplitude rate of increase.

[0010] One aspect of the present invention is the above-mentioned depression risk assessment device, wherein the depression risk assessment model is generated based on a pair of the features obtained from each frequency component of the electroencephalogram of the subject and the presence or absence of depression in the subject or the severity of depression in the subject.

[0011] One aspect of the present invention is a depression risk assessment model generation device that includes: a feature calculation unit that calculates a feature, the value of which changes in accordance with a change in the frequency of occurrence of the phase resetting phenomenon, from the frequency components of each frequency obtained by decomposing the electroencephalograms of each of a plurality of subjects by time-frequency analysis; a selection unit that selects frequency components that have a high correlation with the risk of depression, based on the feature obtained from each frequency component of the electroencephalograms of each of the plurality of subjects and the presence or absence of depression or the severity of depression of the subjects; and a model generation unit that generates a depression risk assessment model for calculating an assessment value representing the risk of depression, using the feature obtained from the frequency components selected by the selection unit and the presence or absence of depression or the severity of depression of the subjects.

[0012] One aspect of the present invention is a depression risk assessment meter comprising: a memory unit that stores a depression risk assessment model for calculating an assessment value representing the risk of depression using features that are calculated from predetermined frequency components obtained by decomposing electroencephalograms by time-frequency analysis and whose values ​​change with changes in the frequency of occurrence of the phase resetting phenomenon; an electroencephalogram measurement unit that measures the electroencephalograms of a subject; a feature calculation unit that calculates feature values ​​of the same type as the feature values ​​used in the depression risk assessment model from the predetermined frequency components obtained by decomposing the measured electroencephalograms by time-frequency analysis; an evaluation unit that calculates an assessment value representing the risk of depression of the subject by using the feature values ​​calculated by the feature calculation unit as input to the depression risk assessment model; and an output unit that outputs the assessment result by the evaluation unit.

[0013] One aspect of the present invention is a depression risk assessment system comprising: a first feature calculation unit that calculates, from the frequency components of each frequency obtained by decomposing the electroencephalograms of each of a plurality of subjects by time-frequency analysis, a feature whose value changes in accordance with a change in the frequency of occurrence of the phase resetting phenomenon; a selection unit that selects frequency components that have a high correlation with the risk of depression based on the feature obtained from each frequency component of the electroencephalograms of each of the plurality of subjects and the presence or absence of depression or the severity of depression of each subject; a model generation unit that generates a depression risk assessment model for calculating an assessment value representing the risk of depression, using the feature obtained from the selected frequency components that are the frequency components selected by the selection unit and the presence or absence of depression or the severity of depression of each subject; a second feature calculation unit that calculates, from the selected frequency components obtained by decomposing the measured electroencephalograms of the subjects by time-frequency analysis, feature values ​​of the same type as the feature values ​​used in the depression risk assessment model; and an evaluation unit that calculates an assessment value representing the risk of depression of the subject by using the feature calculated by the second feature calculation unit as input to the depression risk assessment model.

[0014] One aspect of the present invention is a depression risk assessment method comprising: a reading step of reading from a storage unit a depression risk assessment model for calculating an assessment value representing the risk of depression using features calculated from predetermined frequency components obtained by decomposing electroencephalograms by time-frequency analysis and whose values ​​change with changes in the frequency of occurrence of the phase resetting phenomenon; a feature calculation step of calculating, from the predetermined frequency components obtained by decomposing the measured electroencephalograms of the subject by time-frequency analysis, features of the same type as the features used in the depression risk assessment model; and an assessment step of calculating an assessment value representing the risk of depression of the subject by using the features calculated in the feature calculation step as input to the depression risk assessment model.

[0015] One aspect of the present invention is a depression risk assessment model generation method comprising: a feature calculation step of calculating, from frequency components of each frequency obtained by decomposing the electroencephalograms of each of a plurality of subjects by time-frequency analysis, a feature whose value changes in accordance with a change in the frequency of occurrence of the phase resetting phenomenon; a selection step of selecting frequency components that have a high correlation with the risk of depression based on the feature obtained from each frequency component of the electroencephalograms of each of the plurality of subjects and the presence or absence of depression or the severity of depression in each of the subjects; and a model generation unit that generates a depression risk assessment model for calculating an assessment value representing the risk of depression, using the feature obtained from the frequency components selected in the selection step and the presence or absence of depression or the severity of depression in each of the subjects.

[0016] One aspect of the present invention is a program for causing a computer to function as the depression risk assessment device described above.

[0017] One aspect of the present invention is a program for causing a computer to function as the depression risk assessment model generation device described above.

[0018] The present invention makes it possible to easily determine the risk of depression.

[0019] 1 is a diagram showing an example of time changes in phase, normalized amplitude, and relative phase velocity in the frequency components of an electroencephalogram. FIG. 1 is a diagram showing frequency characteristics of PRR (phase resetting rate). FIG. 2 is a diagram showing t-values ​​of differences in PRR means for each frequency. FIG. 3 is a diagram showing the correspondence between depression scales and PRRs at frequencies with large t-values. FIG. 4 is a diagram plotting PRRs at two frequencies with large t-values. FIG. 5 is a diagram showing the correlation between PRRs and depression scales for each frequency. FIG. 6 is a diagram showing the correspondence between the average or linear sum of PRRs and depression scales. FIG. 7 is a diagram showing the correlation between PRR increase rates and depression scales for each frequency. FIG. 8 is a diagram showing the correspondence between PRR increase rates and depression scales. FIG. 9 is a diagram showing the correlation between PRR standard deviations and depression scales for each frequency. FIG. 10 is a diagram showing the correspondence between the linear sum of PRR standard deviations and depression scales. FIG. 11 is a diagram showing the correlation between PRR coefficients of variation and depression scales for each frequency. FIG. 12 is a diagram showing the correspondence between PRR linear sums of variation coefficients and depression scales. FIG. 13 is a diagram showing an example configuration of a depression risk assessment system according to an embodiment. FIG. 14 is a flow chart showing the processing of a depression risk assessment model generation device according to an embodiment. Fig. 1 is a flow chart showing the processing of the depression risk assessment device of an embodiment. Fig. 2 is a diagram showing the hardware configuration of the depression risk assessment device of an embodiment. Fig. 3 is a diagram showing the relationship between depression scales. Fig. 4 is a diagram showing the correlation coefficient between the normalized PRR average and the depression scale for each frequency. Fig. 5 is a diagram showing the correspondence between the estimated value and the actually measured value of the depression scale obtained by a linear regression model. Fig. 6 is a flow chart showing the detection processing of the phase resetting phenomenon in the depression risk assessment system of an embodiment.

[0020] An embodiment of the present invention will be described in detail below with reference to the drawings. In this embodiment, the risk of depression is assessed based on feature quantities obtained from electroencephalograms measured for a short period of time while the patient is in a resting state with their eyes closed. The risk of depression refers to the possibility that a person is already currently depressed, such as the possibility that they are actually depressed but have not yet been diagnosed, the possibility that they are in a state of being on the verge of depression (pre-depression) but not yet definitively diagnosed with depression, the possibility that a healthy person is not currently depressed but is at risk of developing depression in the future, or the possibility that depression is likely to become severe if triggered by some trigger, and the estimated severity of depression in the case of a depressed patient.

[0021] First, we describe the analysis of electroencephalograms (EEGs) of healthy controls (HCs) and patients with major depressive disorder (MDD). For this analysis, we used Dataset A and Dataset B, which contain electroencephalogram (EEG) data from multiple HCs and multiple MDDs. Dataset A is a publicly available dataset from researchers at the University of New Mexico (https: / / nemar.org / dataexplorer / detail?dataset_id=ds003478). Dataset A includes EEG data from 71 HCs, 43 MDDs, and 2 borderline cases. However, data with abnormalities or deficiencies were excluded from the analysis. Furthermore, Dataset A includes Beck Depression Inventory (BDI) scores as an index of depressive symptom severity (depression severity). Dataset B is a publicly available dataset from researchers at Lanzhou University in China (https: / / modma.lzu.edu.cn / data / application / , EEG_128channels_resting_lanzhou_2015). Dataset B includes EGG data from 29 HC patients and 24 MDD patients. Dataset B also includes values ​​from the Patient Health Questionnaire-9 (PHQ-9) as index data indicating the severity of depressive symptoms (severity of depression).

[0022] First, each EEG data was passed through a band-pass filter to reduce power supply noise and cut off DC components. The EEG data passed through the band-pass filter was then wavelet transformed and decomposed into frequency components with 20 frequency intervals per octave. The frequency components obtained by wavelet transformation indicate the time-dependent changes in the EEG of the corresponding frequency. The jth frequency (j is an integer equal to or greater than 1) among these frequencies was designated as f ｊ and each frequency f obtained from the EEG data ｊ The frequency components of E ｊ Each frequency component E ｊ The time changes of the amplitude and phase of each frequency component E were calculated. The amplitude was normalized so that the average over time and frequency was 1. Furthermore, ｊThe phase velocity of the time series is calculated from the amount of phase change at ｊ The relative phase velocity was calculated by subtracting 2πf ｊ is the frequency f ｊ is the phase velocity of the sinusoidal wave.

[0023] FIG. 1 shows examples of time-varying phase, normalized amplitude, and relative phase velocity. Note that the example shown in FIG. 1 uses EEG data that is not included in Data Set A or Data Set B. FIG. 1(a) shows an example of time-varying phase, FIG. 1(b) shows an example of time-varying normalized amplitude, and FIG. 1(c) shows an example of time-varying relative phase velocity. In FIG. 1, for each of channels AF3 and O2, the frequency f ｊ The frequency component E is 12.6 Hz ｊ As can be seen from these figures, the relative phase velocity (Fig. 1(c)) occasionally increases or decreases significantly, and at its peak, the amplitude (Fig. 1(b)) changes from decreasing to increasing. Such changes often occur simultaneously across multiple channels, and the phases of these channels then become aligned, which is thought to be due to the phase resetting phenomenon. The phase resetting phenomenon occurs throughout the brain, but as shown in Fig. 1, its occurrence can be seen as a sudden change in the phase velocity of the frequency component, resulting in the appearance of a peak, and then, almost simultaneously with this sudden change, the amplitude decreases and then changes back to increasing.

[0024] As described above, since phase resetting occurs anywhere in the brain, it is impossible to predict the combination of channels (electrodes) that will result in phase alignment. Therefore, EEGs are generally measured using many electrodes, and phase resetting is detected when phase alignment occurs between any of the electrodes. However, in this embodiment, by detecting a peak in the phase velocity of the frequency component, and then a decrease in amplitude almost simultaneously with this change, followed by an increase, it is possible to detect the occurrence of phase resetting even in EEGs from a single channel (electrode).

[0025] Each frequency component E ｊThe number of occurrences of the phase resetting phenomenon is counted for one period (1 / f ｊ The frequency (occurrence rate) of these events per second was calculated, which is called the Phase Resetting Rate (PRR).

[0026] When the PRR for each frequency was calculated for each subject included in Data Set A and Data Set B, there were differences between the datasets that were thought to be due to differences in the EEG, measurement environment, and even measurement data processing (noise reduction, etc.). Therefore, the absolute values ​​of the obtained features (mainly PRR) are not very meaningful. However, there were many commonalities in the qualitative results. These are explained below.

[0027] FIG. 2 is a diagram showing the frequency characteristics of PRR. FIG. 2 shows the average PRR (hereinafter referred to as "PRR average") for each frequency for the categories of HC and MDD. FIG. 2(a) shows the average PRR for each frequency for HC and MDD obtained from Dataset A. For Dataset A, EEG data from channel CPz was used. FIG. 2(b) shows the average PRR for each frequency for HC and MDD obtained from Dataset B. For Dataset B, EEG data from channel CPz was used. As shown in FIG. 2, both Dataset A and Dataset B show similar shapes. For each of Dataset A and Dataset B, the PRR average difference, which is the difference between the average PRR for HC and the average PRR for MDD, was calculated for each frequency.

[0028] FIG. 3 shows the t-values ​​of the PRR mean difference for each frequency. FIG. 3(a) shows the t-values ​​of the PRR mean difference for each frequency obtained from the EEG data of channel CPz of Dataset A. FIG. 3(b) shows the t-values ​​of the PRR mean difference for each frequency obtained from the EEG data of channel Fz of Dataset B. The t-values ​​are approximately 2.6 or greater, and p<0.01. As shown in FIG. 3, there were differences in the frequencies at which the t-values ​​were large between Dataset A and Dataset B. Dataset A had large t-values ​​at 9.3 Hz and 14.7 Hz, while Dataset B had large t-values ​​at 3.5 Hz and 6.6 Hz. The PRRs at frequencies with large t-values ​​were plotted against the depression scale, as shown in FIG. 4.

[0029] FIG. 4 is a diagram showing the correspondence between the depression scale and PRR at frequencies with large t-values. FIG. 4(a) shows the relationship between the BDI obtained from Dataset A and PRR at a frequency of 14.7 Hz. FIG. 4(b) shows the relationship between the PHQ-9 obtained from Dataset B and PRR at a frequency of 6.6 Hz. As shown in FIG. 4, the distribution of PRR differs to some extent between HC and MDD. However, since there are some mixed areas, it is difficult to accurately classify whether or not someone is depressed based on this alone. However, by combining this with other information, it may be possible to use it as useful information for classifying the risk of depression, such as whether or not someone is currently depressed. In this case, PRR can be used as an evaluation value.

[0030] Figure 5 is a plot of the PRRs at two frequencies with large t values ​​in Figure 3. Figure 5(a) shows the correspondence between the PRR at 9.3 Hz and the PRR at 14.7 Hz obtained from Data Set A. Figure 5(b) shows the correspondence between the PRR at 3.5 Hz and the PRR at 6.6 Hz obtained from Data Set B. According to Figure 5, the distributions of HC and MDD are somewhat different, but there are also some mixed areas, making accurate classification difficult. However, this may be useful information for classifying the risk of depression. For example, the difference in PRRs at two frequencies with large t values ​​can be used as an evaluation value for determining the risk of depression, such as whether a person is currently depressed, whether a healthy person is likely to develop depression in the future, or whether depression is likely to become more severe if it occurs.

[0031] FIG. 6 shows the correlation between PRR and depression scale for each frequency. FIG. 6(a) shows the correlation between PRR and BDI for each of HC and MDD obtained from Dataset A for each frequency. From FIG. 6(a), it can be seen that in MDD, the correlation between PRR and BDI is strong around 3.3 to 3.5 Hz. FIG. 6(b) shows the correlation between PRR and PHQ-9 for each of HC and MDD obtained from Dataset B for each frequency. From FIG. 6(b), it can be seen that in MDD, the correlation between PRR and PHQ-9 is strong around 2.6 Hz and 4.8 Hz. For each subject, the linear sum of PRR at these frequencies with strong correlation was calculated.

[0032] Figure 7 shows the correspondence between the mean or linear sum of PRRs and the depression scale at frequencies where there is a strong correlation between PRR and the MDD depression scale. Figure 7(a) shows the correspondence between the BDI of each subject obtained for Dataset A and the mean PRR at 3.3 to 3.5 Hz. Figure 7(b) shows the correspondence between the PHQ-9 of each subject obtained for Dataset B and the linear sum of PRRs at 4.8 Hz and 2.6 Hz. When calculating the linear sum, the coefficient was set to 1 or -1 depending on the positive or negative correlation coefficient, and the absolute value of the PRR was added. Figure 7 shows that the severity of MDD can be estimated by using these PRR mean and linear sum as evaluation values. On the other hand, since there is almost no correlation with HC, they cannot be used to diagnose depression, but healthy individuals with large PRR mean or linear sum values ​​are thought to be more likely to develop severe depression. In this way, the mean and linear sum of PRRs can be used as evaluation values ​​for determining depression risk.

[0033] In addition to PRR, the following features (1) to (3) have similar properties.

[0034] (1) Difference or sum of PRRs at multiple frequencies. The difference between PRRs at multiple frequencies can also be used as a feature. For example, the rate of increase in PRR with respect to frequency (PRR difference / frequency difference) fluctuates more dramatically than PRR, but at certain frequencies it shows a high correlation with the MDD depression scale.

[0035] FIG. 8 shows the correlation between the PRR increase rate and the depression scale for each frequency for HC and MDD categories. Here, the PRR increase rate is the increase rate between PRRs at adjacent frequencies. FIG. 8(a) shows the correlation between the PRR increase rate obtained from Dataset A and the BDI for each frequency. FIG. 8(b) shows the correlation between the PRR increase rate obtained from Dataset B and the PHQ-9 for each frequency. As shown in FIG. 8, there are frequencies where the correlation between the PRR increase rate and the MDD depression scale is high.

[0036] FIG. 9 is a diagram showing the correspondence between the PRR increase rate and the depression scale for each subject at frequencies with high correlation in FIG. 8. FIG. 9(a) shows the correspondence between the PRR increase rate and BDI for each subject obtained from Dataset A. FIG. 9(a) shows the correspondence between the average PRR increase rate at 11.1 Hz and 20.7 Hz, which are highly correlated with the MDD depression scale in FIG. 8(a), and the BDI. FIG. 9(b) shows the correspondence between the linear sum of the PRR increase rate for each subject obtained from Dataset B and the PHQ-9. FIG. 9(b) shows the correspondence between the linear sum of the PRR increase rate at 2.7 Hz, 4.1 Hz, and 14.7 Hz, which are highly correlated with the MDD depression scale in FIG. 8(b), and the PHQ-9. As shown in FIG. 9, for MDD, a linear relationship is observed between the depression scale and the average and linear sum of the PRR increase rate. Therefore, the average or linear sum of the PRR increase rate can be used as an evaluation value for determining the risk of depression.

[0037] (2) The magnitude of temporal fluctuation of PRR. PRR changes over time, and the magnitude of this change is also correlated with the depression scale. Figure 10 is a diagram showing the correlation between the standard deviation of PRR and the depression scale for eight 30-second intervals, for each frequency. Figure 10(a) shows the correlation between the standard deviation of PRR for each HD and MDD category obtained from Dataset A and the BDI, for each frequency. Figure 10(b) shows the correlation between the standard deviation of PRR for each HD and MDD category obtained from Dataset B and the PHQ-9, for each frequency. As shown in Figure 10, there are frequencies where the correlation between the standard deviation of PRR and the MDD depression scale is high.

[0038] FIG. 11 shows the correspondence between the linear sum of the standard deviations of PRRs of each subject at frequencies with high correlation in FIG. 10 and the depression scale. When calculating the linear sum, the coefficient was set to 1 or -1 depending on whether the standard deviation of PRR was positive or negative, and the absolute value of the standard deviation was added. FIG. 11(a) shows the correspondence between the linear sum of the standard deviations of PRRs of each subject obtained from Dataset A and the BDI. In FIG. 11(a), the linear sum was calculated by (standard deviation of PRR at 7.1 Hz) - (standard deviation of PRR at 3 Hz) + (standard deviation of PRR at 25.5 Hz). FIG. 11(b) shows the correspondence between the linear sum of the standard deviations of PRRs of each subject obtained from Dataset B and the PHQ-9. In FIG. 11(b), the linear sum was calculated by (standard deviation of PRR at 8.1 Hz) - (standard deviation of PRR at 4.4 Hz) + (standard deviation of PRR at 11.9 Hz). 11, for MDD patients, a linear relationship is observed between the depression scale and the linear sum of the standard deviations of PRR. Therefore, the linear sum of the standard deviations of PRR can be used as an evaluation value for determining the risk of depression.

[0039] Figure 12 shows the correlation between the coefficient of variation of PRR and the depression scale for eight 30-second intervals for each frequency. Figure 12(a) shows the correlation between the coefficient of variation of PRR for each HD and MDD category obtained from Dataset A and the BDI for each frequency. Figure 12(b) shows the correlation between the coefficient of variation of PRR for each HD and MDD category obtained from Dataset B and the PHQ-9 for each frequency. As shown in Figure 12, there are frequencies where the correlation between the coefficient of variation of PRR and the depression scale for MDD is high.

[0040] FIG. 13 shows the correspondence between the linear sum of the PRR coefficients of variation of each subject at frequencies with high correlation in FIG. 12 and the depression scale. When calculating the linear sum, the coefficient was set to 1 or -1 depending on whether the PRR coefficient of variation was positive or negative, and the absolute values ​​of the coefficients of variation were added. FIG. 13(a) shows the correspondence between the linear sum of the PRR coefficients of variation of each subject obtained from Data Set A and the BDI. In FIG. 13(a), the linear sum was calculated using (coefficient of variation of PRR at 9 Hz) - (coefficient of variation of PRR at 3 Hz) + (coefficient of variation of PRR at 7.1 Hz). FIG. 13(b) shows the correspondence between the linear sum of the PRR coefficients of variation of each subject obtained from Data Set B and the PHQ-9. In FIG. 13(b), the linear sum was calculated using (coefficient of variation of PRR at 8.1 Hz) - (coefficient of variation of PRR at 4.4 Hz) + (coefficient of variation of PRR at 11.9 Hz). 13, for MDD patients, a linear relationship is observed between the depression scale and the linear sum of the coefficients of variation of PRR. Therefore, the linear sum of the coefficients of variation of PRR can be used as an evaluation value for determining the risk of depression.

[0041] The variation of phase resetting intervals (VPRI) can also be used as a feature representing the magnitude of temporal fluctuation of PRR. VPRI represents the variation in the occurrence of the phase resetting phenomenon. VPRI is calculated as follows:

[0042] Let x be the interval between occurrences of phase resetting in EEGs of a certain frequency. x is always positive and can take on a very large value. Therefore, the interval between occurrences x is logarithmically transformed to y = log x. Then, the average value ŷ and standard deviation ŷ of y are calculated, and their respective exponents are used to calculate the average x̂ and standard deviation x̂ of the interval between occurrences of phase resetting, as shown in the following formula.

[0043]

[0044] The reciprocal of the average occurrence interval x corresponds to the frequency of the phase resetting phenomenon. In this case, VPRI is given by the following formula:

[0045]

[0046] VPRI is normalized by the average occurrence interval, so VPRI represents the degree of variation in the occurrence interval of the phase resetting phenomenon, without depending on the occurrence frequency of the phase resetting phenomenon.

[0047] (3) Features that change in conjunction with PRR. For example, the mean phase velocity and normalized amplitude change according to PRR, and therefore are correlated with depression scales. However, the correlation is generally weaker than that with PRR.

[0048] The usefulness of this embodiment at this point can be summarized as follows.

[0049] (1) It is useful for assisting the diagnosis of depression. While it is not possible to accurately diagnose whether a patient is depressed or healthy based solely on the PRR of a specific frequency, significant differences are observed between MDD and HC at many frequencies. Therefore, by using these as features and classifying them with a classifier, it may be possible to distinguish with sufficient accuracy. However, because it is highly dependent on the dataset, it is necessary to measure EEG in as similar a measurement environment as possible.

[0050] Previous studies have shown that existing methods can achieve nearly 100% accuracy using deep learning, so they are not superior in terms of accuracy. However, there have been no reports of similar accuracy being achieved using a different dataset from the one used for learning, and it is believed that accuracy would decrease significantly if a different dataset were used. Therefore, it cannot be said that this embodiment is significantly inferior. Furthermore, many existing methods are likely to capture cognitive decline caused by depression, and there have been no examples of successful early detection of depression. This embodiment uses PRR, which changes to reflect daily changes in depressed mood, and therefore may be useful for early diagnosis. For example, it is believed that the accuracy of early detection can be improved by combining it with existing technology that estimates changes in depressed mood through continuous electroencephalography (EEG) measurements.

[0051] (2) Useful for measuring the severity of depression. Since it showed a high correlation with the depression scale for MDD patients, it is believed that it can be used as an objective biomarker for the severity of depression. Because the BDI and PHQ-9 are assessment methods using questionnaires, erroneous results will be obtained if the patient does not answer honestly, but such errors will not occur in the present embodiment. Although it is not suitable for assessing whether a person is healthy, healthy people who show evaluation values ​​close to those of severely depressed people may be prone to depression, or if they do develop depression, their depression may become severe, and therefore it may be possible to use it for risk assessment.

[0052] While existing technologies include research into estimating values ​​such as BDI, the accuracy is not very high, and there are no reports that they can be used to estimate the severity of depression in patients. Furthermore, most of these techniques use machine learning, and no single feature is known to show a high correlation with the severity of depression. While machine learning techniques are highly dependent on the data set, this embodiment shows the same results across two different data sets, suggesting its versatility. The absolute value of PRR depends on the measurement environment, so adjustments are necessary accordingly.

[0053] Next, we will explain the analysis of EEG data using the above public dataset, as well as the analysis of EEG data based on experiments. The experiments were conducted as follows.

[0054] (1) Eighteen participants in Participant Group A underwent 14-channel EEG measurements and completed questionnaires on the SDS (Self-Rating Depression Scale) and PHQ-9 depression scales. Nine of these participants underwent the same measurements again approximately one month later.

[0055] (2) The same measurements as in (1) were conducted on 24 participants in Participant Group B. However, the BDI was also included in the depression scale questionnaire. These 24 participants were selected based on a preliminary questionnaire to ensure that the depression scale scores were distributed as evenly as possible.

[0056] Figure 18 shows the relationship between depression scales. Figure 18(a) shows the relationship between the SDS score and the PHQ-9 score obtained for each participant in Participant Group A and Participant Group B. Figure 18(b) shows the relationship between the SDS score and the BDI score obtained for each participant in Participant Group B. The open circle plots indicate participants who exceed both the criteria for severe depression, i.e., an SDS of 55 or more and a PHQ-9 of 15 or more. The cross plots indicate participants who meet either the criteria of an SDS of 55 or more and a PHQ-9 of 15 or more. The filled circle plots indicate HC participants.

[0057] The SDS is the most commonly used depression scale in Japan. Compared to other scales, the SDS has a wider range of scores from normal to borderline, and is therefore considered suitable for screening people at high risk of depression. On the other hand, the BDI is suitable for assessing the severity of depression. The PHQ-9 is somewhere in between. Although the PHQ-9 has only nine questions, as shown in Figure 18(a), the PHQ-9 and SDS correspond linearly, and the accuracy is not poor. Furthermore, as shown in Figure 18(b), the SDS also corresponds linearly to the BDI.

[0058] The analysis results of the experiment will now be explained. The following analysis was conducted on the PRR and VPRI of each participant group A and B. The results of the analysis for PRR and VPRI were almost the same, but PRR was superior. Therefore, the analysis results using PRR will be described here.

[0059] First, the electroencephalogram data EEG of each channel was passed through a bandpass filter and then wavelet-converted, and decomposed into 80 frequency components E (20 points per octave) from 2 to 32 Hz. ｉ The jth frequency (j in this experiment is an integer between 1 and 80) of the EEG ｊ The frequency components of E ｉｊ Each frequency component E ｉｊ After calculating the PRR for each of the above-mentioned analyses, several normalization procedures were performed and the correlation between the results and SDS was examined. The following normalization procedures are highly effective:

[0060] Channel C ｉ Frequency f ｊ Frequency component E ｉｊ The PRR obtained from x ｉ (f ｊ ) Also, channel c ｉ All frequencies f ｊ About x ｉ (f ｊ ) is the average of a ｉ , standard deviation is s ｉ As shown in the following equation (3), the frequency component E ｉｊ x ｉ (f ｊ ) to the average a ｉ and then subtract the standard deviation s ｉ The y obtained by dividing ｉ (f ｊ ) to channel c ｉ Frequency f ｊ Normalized PRR for ｉｊ It was decided.

[0061]

[0062] The above operation is believed to reduce the influence of differences in the contact state of each electrode and other measurement environments, as well as individual differences in the PRR value. ｉ (f ｊ ) instead of channel c ｉ Frequency f ｊ Frequency component E ｉｊ Use the VPRI obtained from

[0063] Next, the frequency f ｊ All 14 channels (electrodes) c ｉ Normalized PRR ｉｊ The average value of the normalized PRR was obtained as the normalized PRR average. In addition, the 14 channels were divided into three groups according to the region: the frontal region (AF3, AF4, F3, F4, F7, F8), the middle region (FC5, FC6, T7, T8), and the occipital region (P7, P8, O1, O2). ｉ Normalized PRR ｉｊ The average value of frequency f ｊThe normalized PRR average for each region was calculated. It is believed that using these features can reduce the effects of electrode position deviation and individual differences.

[0064] Furthermore, the frequency f ｊ The difference between the normalized PRR average for each region of the frontal region and the normalized PRR average for each region of the occipital region was calculated as the normalized PRR average difference for each region, and used as a feature. ｊ Normalized PRR of all channels per ｉｊ and the normalized PRR average maximum difference, which is the difference between the maximum value of ｊ Normalized PRR of all channels per ｉｊ The normalized PRR average minimum value difference, which is the difference between the minimum value of and the normalized PRR average, was also used as a feature.

[0065] 19 is a diagram showing the correlation coefficient between the normalized PRR average and the SDS score for each frequency. As shown in FIG. 19, there is a strong negative correlation at 2.8 Hz and 17.9 Hz, and a strong positive correlation at 3.6 Hz and 5.1 Hz. Thus, the normalized PRR at the frequency with a strong correlation with the SDS score was selected as an input feature for the depression risk assessment model.

[0066] Next, a depression risk assessment model will be described. Machine learning can be used to construct an assessment model that estimates an SDS score from features selected as having a strong correlation with the depression scale. However, the more complex the machine learning, such as deep learning, the less transparent the estimation process becomes and the less explainable the reason for the validity of the estimated value. Furthermore, there is a concern that the model may be adapted only to a specific data set, reducing its generality. Therefore, an example will be shown in which the simplest linear regression model is used as an assessment model that estimates an SDS score from features.

[0067] The features (explanatory variables) input to the model were eight features selected from those based on the average values ​​for all electrodes (all channels) shown in Figure 19 and those based on the average values ​​for each region, such as the frontal and occipital regions. These were combined with different frequencies and channel types (electrode positions) to ensure independence as much as possible. Linear multiple regression analysis was then performed using 51 data sets obtained for participant group A and participant group B to determine the coefficients of each feature to be input to the linear regression model and the intercept of the linear regression model. Note that the coefficients of features with a negative correlation with the SDS score were set to negative values, and absolute values ​​were added.

[0068] Figure 20 shows the correspondence between estimated values ​​and actual measured values ​​of the depression scale obtained by the linear regression model. Figure 20 is a graph plotting the estimated SDS values, which are the output when feature quantities obtained from participants in participant group A and participant group B are input into the linear regression model, against the actual measured values ​​obtained from the questionnaire. The correlation coefficient between the two was 0.793. Open circles indicate participants who exceeded both the criteria for severe depression, i.e., an SDS of 55 or higher and a PHQ-9 of 15 or higher, while crosses indicate participants who exceeded either criterion. Furthermore, black circles indicate HC participants. At least participants who exceeded both criteria (open circles) could be clearly distinguished from HCs by the model's output values. Furthermore, participants with moderate depression (SDS of 48-54) could also be relatively well distinguished. For these, there was no significant difference in error compared to when the SDS was estimated from the PHQ-9 score.

[0069] The public dataset used in the analysis presented above is mostly comprised of HC and MDD groups, and contains very few borderline individuals (those who have not been diagnosed with depression but are at high risk of depression). In the above experiment, a correlation was obtained between SDS scores with a wide borderline and a linear sum of features based on normalized PRR and normalized VPRI (hereinafter referred to as normalized VPRI). This indicates that normalized PRR, normalized VPRI, and features based on these can be used to assess depression risk.

[0070] It has already been discovered that there is a strong correlation between PRR values ​​and the intensity of depressive mood (depression level). However, PRR values ​​vary considerably from person to person. Therefore, it is conceivable to use the results of continuous measurement of each individual's EEG over a certain period of time to individually optimize a model that estimates depression level based on PRR-related features. However, "depression level" is a numerical representation of the depressive mood, which fluctuates from day to day, and is somewhat different from the "depression scale" commonly used to assess the risk and severity of depression. While it is thought that measuring depression level over a certain period of time can estimate a depression scale, it is not suitable for applications such as identifying individuals at high risk of depression using a single EEG measurement. In contrast, in this embodiment, a model is generated based on the relationship between EEG and depression level in a group, rather than a model applied to an individual. Therefore, using PRR-related features obtained from multi-channel EEG as input to the model is an effective method for reducing the influence of individual differences and differences in measurement environment. This makes it possible to accurately estimate the risk of depression using a small number of EEG measurements, essentially a single EEG measurement.

[0071] Next, an embodiment of the present invention will be described. Fig. 14 is a diagram showing the configuration of a depression risk assessment system 1 according to one embodiment of the present invention. The depression risk assessment system 1 includes a depression risk assessment model generation device 2, a depression risk assessment device 3, and an electroencephalogram (EEG) measurement device 4.

[0072] The depression risk assessment model generation device 2 includes a storage unit 21, an acquisition unit 22, a feature calculation unit 23, a selection unit 24, a model generation unit 25, and an output unit 26. The depression risk assessment model generation device 2 is realized by a computer device such as a server computer or a personal computer, for example.

[0073] The storage unit 21 stores various data including training data and a depression risk assessment model. The training data is data used to train the depression risk assessment model. The training data includes data on the electroencephalogram of the subject, the time of measurement of the electroencephalogram, and attributes of the subject. The attributes of the subject indicate the category of HC or MDD, and the value of a depression scale such as BDI, PHQ-9, or SDS.

[0074] The depression risk assessment model includes types of assessment features, frequencies of frequency components from which the assessment features are obtained, an operation for calculating an assessment value from the assessment features, and a range of assessment values ​​for each depression risk. The depression risk assessment model is configured with the types of assessment features and the operation for calculating an assessment value set in advance, but before learning the depression risk assessment model, the frequencies of the frequency components from which the assessment features are obtained and the range of assessment values ​​for each depression risk have not been determined.

[0075] The evaluation feature is a feature highly correlated with depression risk. The type of evaluation feature is PRR or a feature that changes with changes in PRR. Features that change with changes in PRR include, for example, the PRR increase rate, such as the difference between PRRs at different frequencies or the frequency difference between PRRs, the standard deviation or coefficient of variation of PRR, the temporal fluctuation of PRR (the range of change in PRR per predetermined unit time), and normalized PRR. The temporal fluctuation of PRR includes the standard deviation or coefficient of variation of PRR, VPRI, normalized VPRI, etc. Furthermore, the average value of normalized PRR or normalized VPRI across all channels, the average value of a site group, the difference between the average values ​​between different site groups, and the difference between the maximum or minimum value of all channels and the average value of all channels can also be used as a feature. Examples of site groups include the frontal lobe, the middle region, and the occipital region. Furthermore, instead of PRR, average phase velocity, normalized amplitude, etc. can also be used.

[0076] The calculation for calculating the evaluation value may involve using the evaluation feature as the evaluation value itself, calculating the average of the evaluation feature at n frequencies (n is an integer of 2 or greater), or calculating a linear sum of the evaluation feature at n frequencies (n is an integer of 2 or greater). The linear sum may be the sum of absolute values ​​of the evaluation feature. Alternatively, any machine learning method, such as a neural network, may be used as the calculation for calculating the evaluation value. Here, a case will be described in which a first evaluation model and a second evaluation model are used as depression risk assessment models, and the calculation for calculating the evaluation value is a linear sum. The first evaluation model outputs an evaluation result indicating whether the depression risk is high or low. A low depression risk corresponds to a high probability of HC, a low likelihood of developing depression in the future, or a low likelihood of the depression becoming severe if it does occur. A high depression risk corresponds to a high probability of MDD, a possibility of being in a state of pre-depression (not yet definitively diagnosed with depression), or a possibility of not currently suffering from depression but of developing depression or of being prone to severe depression if it does occur. The second evaluation model outputs an evaluation result indicating the estimated severity of depression.

[0077] The acquisition unit 22 acquires multiple pieces of training data. For example, the acquisition unit 22 may receive the training data from another computer device connected via a network, or may read the training data from a computer-readable recording medium. The acquisition unit 22 stores the acquired multiple pieces of training data in the storage unit 21.

[0078] The feature calculation unit 23 reads the training data from the storage unit 21. The feature calculation unit 23 performs time-frequency analysis on the EEGs represented by the read training data, thereby decomposing the EEGs into frequency components at predetermined frequency intervals within the target frequency range. For example, the target frequency range is from 2 Hz to 32 Hz, and the feature calculation unit 23 decomposes the EEGs into frequency components at 20 frequency intervals per octave. Because frequencies below 2 Hz require longer measurement times and frequencies above 32 Hz require a high-performance electroencephalograph, the target frequency range for analysis is set to 2 Hz to 32 Hz. However, lower or higher frequencies may also be targeted. Examples of time-frequency analysis include wavelet transform, Hilbert analysis, and short-time Fourier transform. The feature calculation unit 23 calculates evaluation features from the temporal changes in the EEGs represented by each frequency component.

[0079] The selection unit 24 identifies frequencies that are highly associated with depression risk among the evaluation feature quantities of frequencies in the frequency range to be analyzed obtained for the plurality of learning data. That is, the selection unit 24 identifies n frequencies that are different between HC and MDD among the evaluation feature quantities of each frequency in the frequency range to be analyzed obtained for the plurality of learning data. ｐ pieces (n ｐ is an integer of 1 or more), ｐ Furthermore, the selection unit 24 selects n evaluation features for each frequency in the frequency range to be analyzed obtained from the plurality of training data sets that are highly related to the depression scale for MDD. ｑ pieces (n ｑ is an integer of 1 or more), ｑ Identify the ｐ is the number of evaluation features used in the calculation of the first evaluation model, and n ｑ is the number of evaluation features used in the calculation of the second evaluation model.

[0080] The model generating unit 25 determines whether the category is HC or MDD obtained from the learning data and the first evaluation frequency f ｐThe model generating unit 25 calculates the range of the evaluation value for each depression risk based on the relationship with the evaluation value calculated by substituting the evaluation feature quantity f for the calculation of the first evaluation model. For example, the model generating unit 25 may calculate a threshold value that is the boundary between the range in which the depression risk is determined to be low and the range in which the depression risk is determined to be MDD. The model generating unit 25 calculates the range of the evaluation value for each depression risk based on the relationship with the evaluation feature quantity f for the calculation of the first evaluation model. For example, the model generating unit 25 may calculate a threshold value that is the boundary between the range in which the depression risk is determined to be low and the range in which the depression risk is determined to be MDD. ｐ and a range of evaluation values ​​for each depression risk, and stores them in the storage unit 21. The model generation unit 25 also uses the depression scale obtained from the MDD learning data and the second evaluation frequency f ｑ The model generating unit 25 calculates the range of the evaluation value for each severity of depression based on the relationship with the evaluation value calculated by substituting the evaluation feature quantity f for the calculation of the second evaluation model. The depression scale can be used as the severity of depression. The model generating unit 25 calculates the range of the evaluation value for each severity of depression based on the relationship with the evaluation feature quantity f for the calculation of the second evaluation model. ｑ The range of the evaluation value for each severity of depression is set and stored in the storage unit 21.

[0081] The output unit 26 reads out the depression risk assessment model generated by the model generation unit 25 from the storage unit 21 and outputs it. The output depression risk assessment model may be transmitted to the depression risk assessment device 3 or another computer device connected to the depression risk assessment device 3, or may be written to a recording medium. The output destination of the first evaluation model and the output destination of the second evaluation model may be different.

[0082] The depression risk assessment device 3 includes a storage unit 31, an acquisition unit 32, a feature calculation unit 33, an evaluation unit 34, an output unit 35, and a display unit 36. The depression risk assessment device 3 can be realized by a computer device such as a server computer, a personal computer, a smartphone, a tablet, or a dedicated device.

[0083] The memory unit 31 stores various information including the depression risk assessment model, electroencephalogram data of the subject measured by the electroencephalogram measuring device 4, and information on estimated depression risk. The acquisition unit 32 acquires the depression risk assessment model generated by the depression risk assessment model generation device 2 and stores it in the memory unit 31. The acquisition unit 32 may receive the depression risk assessment model from the depression risk assessment model generation device 2 or a computer device connected via a network, or may read it from a recording medium. Furthermore, the acquisition unit 32 receives electroencephalogram data of the subject from the electroencephalogram measuring device 4 and writes it to the memory unit 31. Data on the measurement time is added to the electroencephalogram data. It should be noted that the depression risk assessment model does not necessarily have to be generated by the depression risk assessment model generation device 2.

[0084] The feature calculation unit 33 receives from the storage unit 31 the type of evaluation feature added to the first evaluation model and the first evaluation frequency f ｐ The one or more first evaluation frequencies f ｐ The first evaluation feature w which is the evaluation feature of the type read out in each ｐ from the electroencephalogram data written in the storage unit 31. Furthermore, the feature calculation unit 33 receives from the storage unit 31 the type of evaluation feature added to the second evaluation model and the second evaluation frequency f ｑ The one or more second evaluation frequencies f ｑ The second evaluation feature w which is the evaluation feature of the type read out in each ｑ is calculated from the electroencephalogram data written in the storage unit 31.

[0085] The evaluation unit 34 adds the first evaluation feature value w calculated by the feature value calculation unit 33 to the first evaluation model read from the storage unit 31. ｐ Substituting the first evaluation value V ｐ The evaluation unit 34 calculates the first evaluation value V ｐ The evaluation unit 34 compares the second evaluation model read from the storage unit 31 with the second evaluation feature value w calculated by the feature value calculation unit 33, and determines the depression risk of the evaluation subject. ｑSubstituting the second evaluation value V ｑ The evaluation unit 34 calculates the calculated second evaluation value V ｑ and determine the subject's risk of depression, as represented by the severity of depression.

[0086] The output unit 35 outputs the evaluation result by the evaluation unit 34. The evaluation result is one or both of the depression risk obtained using the first evaluation model and the depression risk obtained using the second evaluation model. The evaluation result is expressed as a first evaluation value V ｐ and the second evaluation value V ｑ The output unit 35 may output the first evaluation value V ｐ When outputting the second evaluation value V, the output unit 35 may further output the range of the evaluation value for each depression risk set in the first evaluation model. ｑ When outputting the evaluation result, the output unit 35 may output the range of evaluation values ​​for each severity of depression set in the second evaluation model. The output unit 35 may display the evaluation result on a display (not shown) of the display unit 36 ​​or the EEG-measuring device 4, or may output the evaluation result by voice from a speaker (not shown) provided in the depression risk assessment device 3 or the EEG-measuring device 4. The output unit 35 may also transmit the evaluation result to another device connected to the depression risk assessment device 3, or may write the evaluation result to a recording medium. The display unit 36 ​​is a display device such as an LCD (Liquid Crystal Display). The display unit 36 ​​displays various information.

[0087] The EEG measuring device 4 measures the EEG of one or more channels of the subject and outputs EEG data indicating the measured EEG to the depression risk assessment device 3. Any existing device can be used as the EEG measuring device 4. For example, the EEG measuring device 4 may be a simple portable electroencephalograph with a sampling frequency of 128 Hz.

[0088] The depression risk assessment model generation device 2 may be realized by multiple computer devices connected to a network. In this case, it is possible to arbitrarily determine which of these multiple computer devices each functional unit of the depression risk assessment model generation device 2 is realized by. For example, the memory unit 21, or the memory unit 21 and acquisition unit 22, and other functional units may be realized by different computers connected by a network. Furthermore, the same functional unit of the depression risk assessment model generation device 2 may be realized by multiple computer devices.

[0089] Similarly, the depression risk assessment device 3 may be realized by multiple computer devices connected to a network. In this case, each functional unit of the depression risk assessment device 3 may be realized by any of the multiple computer devices. For example, the memory unit 31 of the depression risk assessment device 3 may be realized by a server computer on a network such as the Internet, and the other functional units of the depression risk assessment device 3 may be realized by the subject's personal computer, tablet, smartphone, etc. Furthermore, the memory unit 31, feature calculation unit 33, evaluation unit 34, and output unit 35 of the depression risk assessment device 3 may be realized by a server computer on a network such as the Internet, and the acquisition unit 32 and display unit 36 ​​of the depression risk assessment device 3 may be realized by the subject's personal computer, smartphone, tablet, etc. Furthermore, the display unit 36 ​​of the depression risk assessment device 3 may be provided in the EEG measurement device 4.

[0090] The depression risk assessment meter may also be an integrated device in which the depression risk assessment device 3 and the EEG measuring device 4 are combined. For example, the depression risk assessment meter may be an EEG measuring device 4 used as a portable EEG meter that is equipped with the functions of the depression risk assessment device 3. The functions of the depression risk assessment device 3 provided in the depression risk assessment meter may also be realized by a plurality of computer devices connected to a network as described above. The depression risk assessment model generation device 2 and the depression risk assessment device 3 may also be integrated into a device, or the depression risk assessment model generation device 2, the depression risk assessment device 3, and the EEG measuring device 4 may also be integrated into a device. In this case, too, the functions of the depression risk assessment model generation device 2 and the functions of the depression risk assessment device 3 may each be realized by a plurality of computer devices connected to a network as described above.

[0091] Next, we will explain the operation of the depression risk assessment system 1. Fig. 15 is a flow diagram showing the processing of the depression risk assessment model generation device 2. First, using Fig. 15, we will explain the processing of the depression risk assessment model generation device 2 when the types of first evaluation feature quantities and second evaluation feature quantities are feature quantities obtained based on the frequency components of the electroencephalogram of one channel.

[0092] First, the acquisition unit 22 of the depression risk assessment model generation device 2 acquires a plurality of training data (step S11). The training data includes, for example, electroencephalogram (EEG) data of each channel c of the subject measured with the eyes closed and in a resting state, and attributes of the subject. The electroencephalogram data EEG of channel c is ｃ For example, channel c is AF3, AF4, F3, F4, F7, F8, FC5, FC6, T7, T8, P7, P8, O1, and O2, but it may be a part of these, or other channels may be used instead of or in addition to these channels. Furthermore, the learning data indicates electroencephalogram (EEG) data measured under similar conditions, such as using the same type of electroencephalogram measuring device. The acquiring unit 22 writes the acquired learning data to the storage unit 21.

[0093] The feature calculation unit 23 selects unselected learning data from the storage unit 21 stored in the storage unit 21 as the processing target (step S12). The selected learning data is referred to as processing target learning data. The feature calculation unit 23 selects an unselected channel c of the processing target learning data as the processing target (step S13). The selected channel c is referred to as processing target channel c. The type of channel c is determined in advance, and there may be one or more types of channel c.

[0094] The feature calculation unit 23 calculates electroencephalogram data EEG of the processing target channel c from the processing target learning data. ｃ are read out from the storage unit 21 and passed through a band-pass filter (step S14). ｃ The power supply noise and DC components contained in the signal are cut off. Note that the process in step S14 does not have to be performed.

[0095] In step S14, the feature calculation unit 23 calculates the electroencephalogram data EEG of the processing target channel c through the band-pass filter. ｃ The electroencephalogram data EEG is decomposed by wavelet transformation to obtain frequency components of 2 to 32 Hz (step S15). Note that the feature calculation unit 23 may calculate each frequency component by other time-frequency analysis. ｃ Frequency f ｊ The frequency components of E ｃｊ For example, the feature amount calculation unit 23 calculates the electroencephalogram data EEG ｃ The feature calculation unit 23 decomposes the frequency component E obtained in step S15 into frequency components with a frequency interval of 20 points per octave. ｃｊ Each of the amplitudes is normalized so that the average amplitude becomes 1 (step S16).

[0096] The feature calculation unit 23 calculates the frequency components E of the processing target channel c normalized in step S16. ｃｊA first evaluation feature and a second evaluation feature are obtained from the first evaluation model and the second evaluation feature are obtained from the second evaluation model (step S17). The type of the first evaluation feature is read from the first evaluation model, and the type of the second evaluation feature is read from the second evaluation model. The types of the first evaluation feature and the second evaluation feature include PRR increase rates such as PRR, normalized PRR, PRR difference, and PRR frequency difference, PRR standard deviation and variation coefficient, PRR temporal fluctuations such as VPRI and normalized VPRI, mean phase velocity increase rates such as mean phase velocity, mean phase velocity difference, and mean phase velocity frequency difference, mean phase velocity temporal fluctuations such as mean phase velocity standard deviation and variation coefficient, normalized amplitude increase rates such as normalized amplitude, normalized amplitude difference, and normalized amplitude frequency difference, and normalized amplitude temporal fluctuations such as normalized amplitude standard deviation and variation coefficient. The types of the first evaluation feature and the second evaluation feature may be the same or different.

[0097] The feature amount calculation unit 23 determines whether all channels have been selected (step S18). If there is an unselected channel c (step S18: NO), the feature amount calculation unit 23 repeats the process from step S13.

[0098] When it is determined that all channels have been selected (step S18: YES), the feature calculation unit 23 calculates the measurement subject attributes read out from the processing target learning data and the frequency f of each channel c. ｊ The first evaluation feature quantity and the second evaluation feature quantity for each image are written to the storage unit 21 (step S19).

[0099] The feature calculation unit 23 determines whether all the learning data have been selected (step S20). If there is unselected learning data (step S20: NO), the feature calculation unit 23 repeats the process from step S12. If it determines that all the learning data have been selected (step S20: YES), the feature calculation unit 23 instructs the selection unit 24 to perform the process of step S21.

[0100] The selection unit 24 selects a channel c and a frequency f for the HC group in which HC is set as the category of the learning data. ｊ First evaluation feature average A, which is the average of the first evaluation feature for eachｃｊ（ＨＣ） Furthermore, the selection unit 24 calculates the channel c and frequency f for the MDD group in which MDD is set as the category of the learning data. ｊ First evaluation feature average A, which is the average of the first evaluation feature for each ｃｊ（ＭＤＤ） (Step S21). Note that instead of the average, a median or the like may be used.

[0101] The selection unit 24 selects the first evaluation feature average A of the HC group for each channel c. ｃｊ（ＨＣ） and the first evaluation feature average A of the MDD group ｃｊ（ＭＤＤ） The first evaluation frequency f is a frequency at which the difference with ｃｐ For example, the selector 24 selects the frequency f ｊ Every A ｃｊ（ＨＣ） -A ｃｊ（ＭＤＤ） The selection unit 24 selects n channels from the channel with the largest t-test result for each channel c. ｐ Number of frequencies f ｃｊ’ The selection unit 24 selects n ｐ Number of frequencies f ｃｊ’ The channel c with the largest sum of t values ​​in ｐ and the selected channel c ｐ n selected from ｐ Number of frequencies f ｃｊ’ The first evaluation frequency f ｃｐ (Step S22). The selector 24 selects n channels from all channels c in descending order of t value. ｐ First evaluation frequencies f ｃｐ In this case, n ｐ First evaluation frequencies f ｃｐ Every channel c ｐ is selected.

[0102] In addition, when the first evaluation feature is normalized PRR, for the HC group, the frequency f ｊ The first evaluation feature average A is calculated by averaging the normalized PRRs of all channels for each ｊ（ＨＣ） For the MDD group, the frequency f ｊ The first evaluation feature average A is calculated by averaging the normalized PRRs of all channels for eachｊ（ＨＣ） The selector 24 may calculate the frequency f ｊ Every A ｊ（ＨＣ） -A ｊ（ＭＤＤ） The selection unit 24 selects n values ​​from the largest result of the t-test. ｐ Number of frequencies f ｊ’ The selector 24 selects the selected frequency f ｊ’ For each channel c ｐ Select the selected channel c ｐ Frequency f ｊ’ The first evaluation frequency f ｃｐ In this case, the selection unit 24 selects the selected n ｐ First evaluation frequencies f ｃｐ Channel C ｐ and frequency f ｊ’ When the first evaluation feature is the normalized VPRI, the selection unit 24 may select the first evaluation frequency f ｃｐ may be selected.

[0103] The model generating unit 25 generates a first evaluation model (step S23). That is, the model generating unit 25 generates a first evaluation model using the first evaluation frequency f ｃｐ The evaluation feature quantity is input to the calculation of the first evaluation model, and the first evaluation value V ｐ（ＨＣ） Furthermore, the model generating unit 25 calculates the first evaluation frequency f obtained from each learning data of the MDD group. ｃｐ The evaluation feature quantity is input to the calculation used in the first evaluation model, and the first evaluation value V ｐ（ＭＤＤ） For example, the calculation uses the evaluation feature as the first evaluation value (where n ｐ = 1), each first evaluation frequency f ｃｐ Calculate the average of the evaluation features of each first evaluation frequency f ｃｐ The model generation unit 25 calculates a linear sum of the evaluation feature quantities V ｐ（ＨＣ） and the first evaluation value V of the MDD group ｐ（ＭＤＤ）Based on the above, the model generation unit 25 determines a range of first evaluation values ​​containing many HCs and a range of first evaluation values ​​containing many MDDs so that these ranges do not overlap. The model generation unit 25 determines a range containing many HCs as a range of low depression risk, and a range containing many MDDs as a range of high depression risk. The model generation unit 25 generates the first evaluation model using the selected channel c ｐ and the first evaluation frequency f ｃｐ and the ranges of the first evaluation value for each of the high depression risk and low depression risk are set.

[0104] Next, the selector 24 selects, for each channel c, the frequency f ｊ The selection unit 24 calculates the correlation between the second evaluation feature amount and the depression scale for each channel c. ｑ Number of frequencies f ｃｊ” The selection unit 24 selects n ｑ Number of frequencies f ｃｊ” The channel c with the largest sum of correlations in ｑ and the selected channel c ｑ n selected for ｑ Number of frequencies f ｃｊ” The second evaluation frequency f ｃｑ (Step S25). The selector 24 selects n channels from all channels c in descending order of correlation. ｑ Number of second evaluation frequencies f ｃｑ In this case, n ｑ Number of second evaluation frequencies f ｃｑ Every channel c ｑ is selected.

[0105] In addition, when the second evaluation feature is normalized PRR, the selection unit 24 selects the frequency f ｊ The normalized PRR average of all channels is calculated for each frequency f ｊ The selector 24 may calculate either or both of the normalized PRR average by region, which is an average of the normalized PRRs by region for each frequency f ｊ The correlation between the normalized PRR average and the normalized PRR average by region and the depression scale was calculated, and n was selected from the highest correlation.ｑ Number of frequencies f ｊ” The selector 24 selects the selected frequency f ｊ” For each channel c ｑ Select the selected channel c ｑ Frequency f ｊ” The second evaluation frequency f ｃｑ In this case, the selection unit 24 selects the selected n ｑ Number of second evaluation frequencies f ｃｑ Channel C ｑ and frequency f ｊ” When the second evaluation feature is the normalized VPRI, the selection unit 24 may select the second evaluation frequency f ｃｑ may be selected.

[0106] The model generating unit 25 generates a second evaluation model (step S26). That is, the model generating unit 25 generates a second evaluation model using the second evaluation frequency f ｃｑ The evaluation feature quantity is input to the calculation of the second evaluation model, and the second evaluation value V ｑ For example, the calculation is performed by calculating the second evaluation frequency f ｃｑ Calculate the average of the evaluation features of each second evaluation frequency f ｃｑ The model generation unit 25 calculates a linear sum of the evaluation feature quantities V ｑ and the depression scale set in the learning data, the model generation unit 25 calculates a range of the second evaluation value corresponding to each depression scale so that the ranges do not overlap. ｑ and the second evaluation frequency f ｃｑ and a range of the second evaluation value according to the severity of depression represented by the depression scale.

[0107] The model generation unit 25 stores the first evaluation model generated in step S23 and the second evaluation model generated in step S26 in the storage unit 21 (step S27). The output unit 26 reads out and outputs a depression risk assessment model including the first evaluation model and the second evaluation model generated by the model generation unit 25 from the storage unit 21 (step S28). The acquisition unit 32 of the depression risk assessment device 3 acquires the depression risk assessment model output from the depression risk assessment model generation device 2 and stores it in the storage unit 31.

[0108] Next, we will explain the depression risk assessment model generation device 2 in the case where the type of first evaluation feature and the type of second evaluation feature are features obtained based on the frequency components of electroencephalograms of multiple channels (hereinafter referred to as multi-channel features). Examples of multi-channel features include normalized PRR average, normalized PRR average by region, normalized PRR average difference by region, normalized PRR average minimum difference, normalized PRR average maximum difference, average normalized VPRI across all channels (hereinafter referred to as normalized VPRI average), average normalized VPRI for each electrode region group (hereinafter referred to as region-specific normalized VPRI average), difference in average normalized VPRI by region between different electrode region groups (hereinafter referred to as region-specific normalized VPRI average difference), difference between minimum normalized VPRI across all channels and normalized VPRI average (hereinafter referred to as normalized VPRI average minimum difference), difference between maximum normalized VPRI across all channels and normalized VPRI average (hereinafter referred to as normalized VPRI average maximum difference), etc. When the type of the first evaluation feature quantity and the type of the second evaluation feature quantity are multi-channel features, the depression risk assessment model generation device 2 performs the same processing as that shown in Figure 15 described above, except for the following processing.

[0109] When the type of the first evaluation feature quantity or the type of the second evaluation feature quantity is a multi-channel feature quantity such as a normalized PRR average, a region-specific normalized PRR average, a region-specific normalized PRR average difference, a normalized PRR average minimum value difference, or a normalized PRR average maximum value difference, in step S17, the feature quantity calculation unit 23 calculates the frequency component E ｃｊ Using the PRR obtained from the above, each frequency f of the processing target channel c is calculated by equation (3). ｊThen, in the process of step S17 in which all channels are selected as processing target channels c, or after determining YES in step S18, the feature calculation unit 23 uses the normalized PRRs of each frequency that have been calculated for all channels to calculate the normalized PRR average for each frequency, the normalized PRR average by region, the normalized PRR average difference by region, the normalized PRR average minimum value difference, or the normalized PRR average maximum value difference.

[0110] Similarly, when the type of the first evaluation feature quantity or the type of the second evaluation feature quantity is a multi-channel feature quantity such as a normalized VPRI average, a region-specific normalized VPRI average, a region-specific normalized VPRI average difference, a normalized VPRI average minimum value difference, or a normalized VPRI average maximum value difference, in step S17, the feature quantity calculation unit 23 calculates, in the same manner as above, each frequency component E ｃｊ Using the VPRI obtained from the above, each frequency f of the processing target channel c is calculated by equation (3). ｊ Then, in the process of step S17 in which all channels are selected as processing target channels c, or after determining YES in step S18, the feature calculation unit 23 uses the normalized VPRIs of each frequency that have been calculated for all channels to calculate the normalized VPRI average for each frequency, the normalized VPRI average by region, the normalized VPRI average difference by region, the normalized VPRI average minimum value difference, or the normalized VPRI average maximum value difference.

[0111] In step S21, the selection unit 24 selects the frequency f ｊ First evaluation feature average A, which is the average of the first evaluation feature for each ｊ（ＨＣ） Furthermore, the selection unit 24 calculates the frequency f ｊ First evaluation feature average A, which is the average of the first evaluation feature for each ｊ（ＭＤＤ） Instead of the average, the median or the like may be used.

[0112] In step S22, the selection unit 24 selects the first evaluation feature average A of the HC group.ｊ（ＨＣ） and the first evaluation feature average A of the MDD group ｊ（ＭＤＤ） The first evaluation frequency f is a frequency at which the difference with ｐ For example, the selector 24 selects the frequency f ｊ Every A ｊ（ＨＣ） -A ｊ（ＭＤＤ） The selection unit 24 selects n values ​​from the largest result of the t-test. ｐ Number of frequencies f ｊ’ The selection unit 24 selects the selected n ｐ Number of frequencies f ｊ’ The first evaluation frequency f ｐ Let's say.

[0113] In step S23, the model generating unit 25 calculates the first evaluation frequency f obtained from each learning data of the HC group. ｐ The evaluation feature quantity is input to the calculation of the first evaluation model, and the first evaluation value V ｐ（ＨＣ） Furthermore, the model generating unit 25 calculates the first evaluation frequency f obtained from each learning data of the MDD group. ｐ The evaluation feature quantity is input to the calculation used in the first evaluation model, and the first evaluation value V ｐ（ＭＤＤ） The model generation unit 25 calculates the first evaluation value V of the HC group. ｐ（ＨＣ） and the first evaluation value V of the MDD group ｐ（ＭＤＤ） Based on this, a range of low depression risk and a range of high depression risk are determined.

[0114] The model generation unit 25 adds a channel c used to calculate the evaluation feature quantity to the first evaluation model. ｐ and the first evaluation frequency f ｐ For example, when the types of the first evaluation feature are a normalized PRR average, a normalized PRR average minimum difference, a normalized PRR average maximum difference, a normalized VPRI average, a normalized VPRI average minimum difference, and a normalized VPRI average maximum difference, the first evaluation value ranges for channel c are set. ｐ In this case, the type of the first evaluation feature is the channel c ｐIn addition, when the types of the first evaluation feature are the normalized PRR average by region, the normalized PRR mean difference by region, the normalized VPRI average by region, and the normalized VPRI mean difference by region, the channel c ｐ is a channel included in the region group. In this case, channel c ｐ may be represented by a site group.

[0115] In step S25, the selector 24 selects the frequency f ｊ The selection unit 24 calculates the correlation between the second evaluation feature amount and the depression scale. ｑ Number of frequencies f ｊ” The selection unit 24 selects the selected n ｑ Number of frequencies f ｊ” The second evaluation frequency f ｑ Let's say.

[0116] In step S26, the model generating unit 25 generates the second evaluation frequency f obtained from each learning data of the MDD group. ｑ The evaluation feature quantity is input to the calculation of the second evaluation model, and the second evaluation value V ｑ The model generation unit 25 calculates the second evaluation value V obtained from the training data. ｑ and the depression scale set in the learning data, the model generation unit 25 calculates a range of the second evaluation value corresponding to each depression scale so that the ranges do not overlap. ｑ and the second evaluation frequency f ｑ and the range of the second evaluation value according to the severity of depression represented by the depression scale. For example, ｑ is all channels, channels belonging to the region group.

[0117] In addition, when machine learning is used for the first evaluation model, in step S23 of FIG. 15, the model generation unit 25 calculates the first evaluation frequency f obtained from each learning data of the HC group. ｃｐ or first evaluation frequency f ｐThe first evaluation value, which is an output obtained by inputting the evaluation feature quantity of the first evaluation model, indicates a value of low depression risk, and the first evaluation frequency f ｃｐ or first evaluation frequency f ｐ The first evaluation model is trained so that the first evaluation value, which is the output obtained by inputting the evaluation features into the first evaluation model, indicates a high depression risk. In this case, the low depression risk value and the high depression risk value are determined in advance.

[0118] In addition, when machine learning is used for the second evaluation model, in step S26 of FIG. 15, the model generation unit 25 generates the second evaluation frequency f ｃｑ or second evaluation frequency f ｑ The evaluation features are input into the second evaluation model, and learning is performed so that the second evaluation value obtained as the output is the depression scale value indicated by the training data or a value corresponding to that depression scale.

[0119] The depression risk assessment model generation device 2 may generate multiple types of first assessment models. In this case, the depression risk assessment model generation device 2 performs the processes of step S17 and steps S21 to S23 for each first assessment model, but only needs to perform the processes common to multiple first assessment models once. Similarly, the depression risk assessment model generation device 2 may generate multiple types of second assessment models. In this case, the depression risk assessment model generation device 2 performs the processes of step S17 and steps S24 to S26 for each second assessment model, but only needs to perform the processes common to multiple second assessment models once.

[0120] 16 is a flow diagram showing an example of the processing of the depression risk assessment device 3. The EEG measuring device 4 measures the EEG of the subject in a resting state with their eyes closed, and transmits the results to the depression risk assessment device 3. The EEG of the subject is measured under the same conditions as those under which the EEG of the learning data was measured, for example, using the same type of EEG measuring device 4 as that used to measure the EEG of the learning data. The acquisition unit 32 of the depression risk assessment device 3 receives the EEG data measured from the EEG measuring device 4 and writes it to the storage unit 31 (step S31). The EEG data EEG is EEG, which is the brain wave of each channel c ｃ Including EEG ｃ to which data of the measurement time is added by the EEG measurement device 4. Note that, as long as the channels used in the depression risk assessment model are included, the number of electrodes (measurement locations) may be, for example, 1 to 4. The measurement time may be about 1 minute, but when the first evaluation feature quantity or the second evaluation feature quantity is the PRR, the average phase velocity, or the temporal fluctuation of the normalized amplitude, the measurement time is set to be long enough to calculate these.

[0121] The feature calculation unit 33 receives from the storage unit 31 the type of the first evaluation feature added to the first evaluation model and the channel c ｐ and the first evaluation frequency f ｃｐ or first evaluation frequency f ｐ and the range of the first evaluation value for each depression risk. If the type of the first evaluation feature is not a multi-channel feature, the first evaluation frequency f ｃｐ One first evaluation frequency f ｃｐ Channel c corresponding to ｐ When the type of the first evaluation feature is a multi-channel feature, the first evaluation frequency f ｐ One first evaluation frequency f ｐ Channel c corresponding to ｐ is multiple (all channels or channels of a site group).

[0122] The feature amount calculation unit 33 calculates the channel c from the electroencephalogram data written in the storage unit 31 in step S31. ｐ Brain wave data EEG ｃｐ(Step S32). The feature amount calculation unit 33 obtains the read electroencephalogram data EEG ｃｐ is passed through a band-pass filter of 2 to 32 Hz (step S33). Note that if the process of step S14 was not performed when generating the depression risk assessment model, the process of step S33 does not need to be performed.

[0123] In step S33, the feature calculation unit 33 calculates the electroencephalogram data EEG ｃｐ The feature amount calculation unit 33 calculates the frequency components by decomposing the electroencephalogram data EEG using the time-frequency analysis used in step S15. ｃｐ Among the frequency components calculated from ｐ Each first evaluation frequency f ｃｐ The first frequency component E ｃｐ Or each channel c ｐ First evaluation frequency f ｐ The first frequency component E ｐ (Step S34). The feature calculation unit 33 calculates each of the first frequency components E obtained in Step S34. ｃｐ or each first frequency component E ｐ The feature amount calculation unit 33 normalizes the amplitude of each first frequency component E normalized in step S35 so that the average of the amplitudes of the first frequency components E becomes 1. ｃｐ The first evaluation feature quantity w of the type read out in step S32 is ｃｐ Alternatively, the feature calculation unit 33 may obtain the normalized multiple channels c ｐ The first frequency component E ｐ Based on the above, the first evaluation feature quantity w of the type read out in step S32 is ｐ is acquired (step S36).

[0124] The evaluation unit 34 evaluates each of the first evaluation frequencies f ｃｐ The first evaluation feature w ｃｐ or each first evaluation frequency f ｐ The first evaluation feature w ｐis input into the calculation of the first evaluation model, and the first evaluation value V ｐ (Step S37). The evaluation unit 34 calculates the calculated first evaluation value V ｐ The evaluation unit 34 determines the depression risk based on whether the first evaluation value V is included in the high depression risk range or the low depression risk range read from the first evaluation model. ｐ If the risk of depression is within the low depression risk range, the risk of depression is determined to be low (step S38: YES).

[0125] If the risk of depression is low, the output unit 35 outputs the first evaluation result indicating the low risk of depression (step S39). The first evaluation result may be a first evaluation value V ｐ and a range of evaluation values ​​for a low depression risk and a range of evaluation values ​​for a high depression risk. For example, the output unit 35 displays the first evaluation result on the display unit 36 ​​or a display unit (not shown) included in the EEG-measuring device 4. At this time, the output unit 35 may display the first evaluation result using a color corresponding to a low depression risk. Furthermore, the output unit 35 may output the first evaluation result by sound from a speaker (not shown) included in the depression risk assessment device 3, the EEG-measuring device 4, or another device connected to the depression risk assessment device 3. For example, the output unit 35 may read out the first evaluation result aloud, or may output a sound corresponding to a low depression risk.

[0126] The evaluation unit 34 calculates a first evaluation value V ｐ is included in the range of high depression risk (step S39: NO), the first evaluation result is determined to be high depression risk. ｑ and the second evaluation frequency f ｃｑ or second evaluation frequency f ｑ and the second evaluation value V for each severity of depression ｑ If the type of the second evaluation feature is not a multi-channel feature, the range of the second evaluation frequency f ｃｑ One second evaluation frequency f ｃｑ Channel c corresponding to ｑWhen the type of the second evaluation feature is a multi-channel feature, the second evaluation frequency f ｑ One second evaluation frequency f ｑ Channel c corresponding to ｑ is multiple (all channels or channels of a site group).

[0127] The feature amount calculation unit 33 calculates the channel c from the electroencephalogram data written in the storage unit 31 in step S31. ｑ Brain wave data EEG ｃｑ (Step S41). The feature amount calculation unit 33 acquires the read electroencephalogram data EEG ｃｑ is passed through a band-pass filter of 2 to 32 Hz (step S42). Note that if the process of step S14 was not performed when generating the depression risk assessment model, the process of step S42 does not need to be performed.

[0128] In step S42, the feature calculation unit 33 calculates the electroencephalogram data EEG ｃｑ The feature amount calculation unit 33 calculates the frequency components by decomposing the electroencephalogram data EEG using the time-frequency analysis used in step S15. ｃｑ Among the frequency components calculated from ｑ Each second evaluation frequency f ｃｑ The second frequency component E ｃｑ Or each channel c ｑ The second evaluation frequency f ｑ The second frequency component E ｑ (Step S43). The feature calculation unit 33 calculates the second frequency components E obtained in Step S43. ｃｑ or each second frequency component E ｑ The amplitude of the channel c is normalized so that the average of the amplitudes becomes 1 (step S44). ｐ and Channel C ｑ If they are the same, the electroencephalogram data EEG is ｃｑ The process of acquiring the value of the parameter and the processes of steps S42 to S44 can be omitted.

[0129] The feature calculation unit 33 calculates the normalized second frequency components E ｃｑ From the second evaluation feature quantity w of the type read out in step S41 ｃｑ Alternatively, the feature amount calculation unit 33 may obtain the normalized multiple channels c ｑ The second frequency component E ｃｑ Based on the above, the second evaluation feature quantity w of the type read out in step S41 is ｑ The evaluation unit 34 acquires the second evaluation frequencies f ｃｑ The second evaluation feature w ｃｑ or each second evaluation frequency f ｑ The second evaluation feature w ｑ is input into the calculation of the second evaluation model, and the second evaluation value V ｑ The evaluation unit 34 calculates the calculated second evaluation value V ｑ The evaluation unit 34 determines which of the ranges of severity read from the second evaluation model the second evaluation value V is included in (step S47). ｑ The severity of the range including the above is the depression assessment risk of the subject.

[0130] The output unit 35 outputs the evaluation result (step S48). The evaluation result includes a first evaluation result indicating a high risk of depression and a second evaluation result indicating the severity of depression. The first evaluation result may include a first evaluation value V instead of or in addition to the high risk of depression. ｐ The second evaluation result may include a range of evaluation values ​​for a low risk of depression and a range of evaluation values ​​for a high risk of depression. ｑand ranges of evaluation values ​​corresponding to each severity level. Furthermore, the evaluation result does not have to include the first evaluation result. For example, the output unit 35 displays the evaluation result on the display unit 36 ​​or a display unit (not shown) included in the EEG-measuring device 4. At this time, the output unit 35 may display the evaluation result using a color corresponding to the severity level. Furthermore, the output unit 35 may output the evaluation result by sound from a speaker (not shown) included in the depression risk assessment device 3, the EEG-measuring device 4, or another device connected to the depression risk assessment device 3. For example, the output unit 35 may read out the evaluation result aloud, or may output a sound corresponding to the severity level of depression.

[0131] If the depression risk is low, the depression risk assessment device 3 may perform the processes of steps S41 to S48 without performing step S39. However, the first evaluation result output in step S48 is the same information as the first evaluation result output in step S39. Furthermore, if only the severity of depression is to be assessed, the depression risk assessment device 3 may not need to perform the processes of steps S34 to S39.

[0132] The first evaluation model is based on the calculated first evaluation value V ｐ In this case, in step S37, the evaluation unit 34 may perform a calculation to output a depression risk that includes the range of each first evaluation frequency f ｃｐ The first evaluation feature w ｃｐ or each first evaluation frequency f ｐ The first evaluation feature w ｐ By inputting this into the calculation of the first evaluation model, an output is obtained indicating whether the risk of depression is low or high.

[0133] Similarly, the second evaluation model calculates the second evaluation value V ｑ In this case, in step S47, the evaluation unit 34 may perform a calculation to output a depression risk that includes the range of each second evaluation frequency f ｃｑ The second evaluation feature w ｃｑ or each second evaluation frequency f ｑ The second evaluation feature w ｑBy inputting the above into the calculation of the second evaluation model, the depression risk, which indicates the severity of depression, is obtained.

[0134] If there are multiple first evaluation models, the depression risk assessment device 3 may perform the processes of steps S32 to S37 for each first evaluation model and combine the depression risks obtained for each first evaluation model to obtain a first evaluation result. If there are multiple second evaluation models, the depression risk assessment device 3 may perform the processes of steps S41 to S47 for each second evaluation model and combine the depression severity levels obtained for each second evaluation model to obtain a second evaluation result.

[0135] For example, the depression risk assessment device 3 may be provided on a network such as the Internet, and a user device such as a personal computer or smartphone may transmit EEG data measured by the EEG-measuring device 4 to the depression risk assessment device 3 via the network. The depression risk assessment device 3 transmits the assessment results to the user device that is the source of the EEG data. This allows the assessment results of the possibility or risk of depression to be obtained almost instantaneously.

[0136] As described above, it has previously been discovered that PRR reflects depressed mood in individuals, including healthy individuals. According to this embodiment, this PRR can be used not only to identify depression patients but also to identify healthy individuals who have a strong tendency toward depression or are at high risk of depression. Furthermore, this embodiment can estimate the severity of MDD. The depression risk assessment system 1 described above can screen individuals at high risk of depression using electroencephalograms measured for approximately one minute to several minutes with eyes closed and in a resting state using a portable electroencephalograph or the like. It can also be used to determine whether a person is merely temporarily depressed but not actually depressed. It may also be possible to identify factors that indicate recovery from depression, making it useful for treating depression patients. Furthermore, this embodiment can assess the risk of depression using an inexpensive device and a single measurement over a short period of time, making it suitable for group health checks at workplaces and schools.

[0137] Next, the phase resetting phenomenon detection algorithm will be described. Figure 21 is a flow diagram showing the process of detecting the phase resetting phenomenon by the depression risk assessment model generation device 2 and the depression risk assessment device 3. Here, the case where the feature calculation unit 23 of the depression risk assessment model generation device 2 detects the phase resetting phenomenon will be described as an example, but similar processing is also performed when the feature calculation unit 33 of the depression risk assessment device 3 detects the phase resetting phenomenon. The feature calculation unit 23 performs the processing of Figure 21 for each combination of channel c and frequency f that are targets for phase resetting phenomenon detection, to detect the phase resetting phenomenon.

[0138] The feature calculation unit 23 obtains a frequency component E (electroencephalogram component) of frequency f from the electroencephalogram data EEG of the channel c to be processed by wavelet transformation. The feature calculation unit 23 obtains a time-series signal φ ｆ (t) is acquired, and A, which is a time series signal of the amplitude of the frequency component E of the frequency f, is obtained. ｆ (t) is acquired (step S51).

[0139] The feature calculation unit 23 calculates the phase time-series signal φ ｆ Based on (t), the relative phase velocity RAS is calculated using the following equation (4): ｆ (t) is calculated (step S52).

[0140]

[0141] For simplicity, the phase is not limited to 0 to 2π, but is assumed to increase continuously with time, and Δt is the sampling interval.

[0142] The feature calculation unit 23 calculates the relative phase velocity RAS at the time t when the amplitude is locally minimum and ｆ (t) is the relative phase velocity RAS ｆ Time average value of MRAS ｆ If the difference is equal to or greater than a certain value h from A, it is estimated that the phase resetting phenomenon occurred at that time t (step S53). ｆ (t) < A ｆ (t-Δt) and A ｆ (t) < Aｆ When (t+Δt) holds, it is determined that the amplitude is minimal at time t. ｆ (t)-MRAS ｆ |>h, the relative phase velocity RAS ｆ (t) is the time average value MRAS ｆ The feature calculation unit 23 determines that the phase resetting phenomenon occurred at a time t greater than or equal to a certain value h. For example, if the certain value h is 5% of 2πf, then h=0.05. The feature calculation unit 23 records the time t at which the phase resetting phenomenon is estimated to have occurred.

[0143] The feature amount calculation unit 23 calculates the feature amount based on the time t estimated in step S53. For example, the feature amount calculation unit 23 counts the number of occurrences of the phase resetting phenomenon in s seconds, and multiplies the counted number of occurrences by 2πf ｓ By dividing by f, the PRR, which is the occurrence frequency per period at frequency f, is obtained.

[0144] We will now explain an example of the hardware configuration of the depression risk assessment model generation device 2 and the depression risk assessment device 3. Fig. 17 is a device configuration diagram showing an example of the hardware configuration of the depression risk assessment model generation device 2 and the depression risk assessment device 3. The depression risk assessment model generation device 2 and the depression risk assessment device 3 each include a processor 71, a storage unit 72, a communication interface 73, and a user interface 74.

[0145] The processor 71 is a central processing unit that performs calculations and control. The processor 71 is, for example, a CPU. The processor 71 reads and executes programs from the storage unit 72. The storage unit 72 further has a work area and the like when the processor 71 executes various programs. The communication interface 73 is connected to other devices so as to be able to communicate with them. The user interface 74 is an input device such as a keyboard, a pointing device (a mouse, a tablet, etc.), a button, a touch panel, etc., a display device such as a display, and an audio output device such as a speaker or headphones.

[0146] The functions of the acquisition unit 22, feature calculation unit 23, selection unit 24, and model generation unit 25 of the depression risk assessment model generation device 2, and at least some of the functions of the acquisition unit 32, feature calculation unit 33, evaluation unit 34, and output unit 35 of the depression risk assessment device 3, are realized by the processor 71 reading and executing a program from the storage unit 72. The program may be transmitted via a telecommunications line or recorded on a portable recording medium. The program may also be for realizing some of the above-described functions, or may be capable of realizing the above-described functions in combination with a program already stored in a computer system. All or some of the above functions may be realized using hardware such as an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a field-programmable gate array (FPGA). The memory unit 21 of the depression risk assessment model generation device 2 and the memory unit 31 of the depression risk assessment device 3 are realized by the memory unit 72.

[0147] According to the above-described embodiment, the depression risk assessment device includes a storage unit, a feature calculation unit, and an evaluation unit. The storage unit stores a depression risk assessment model for calculating an assessment value representing the risk of depression using feature values ​​calculated from predetermined frequency components obtained by decomposing electroencephalograms using time-frequency analysis and whose values ​​change with changes in the frequency of occurrence of the phase resetting phenomenon. The feature calculation unit calculates feature values ​​of the same type as the feature values ​​used in the depression risk assessment model from the predetermined frequency components obtained by decomposing the measured electroencephalograms of the subject using time-frequency analysis. The evaluation unit calculates an assessment value representing the risk of depression of the subject using the feature values ​​calculated by the feature calculation unit as input to the depression risk assessment model.

[0148] The evaluation value indicates whether the risk of depression is low or high, or the severity of depression. The risk of depression includes the possibility of already being depressed, the possibility of being in a pre-depressed state, the possibility of not currently being depressed but becoming depressed in the future, or the possibility of not currently being depressed but being prone to developing severe depression if it does occur. The possibility of already being depressed includes the possibility of being depressed but not yet diagnosed.

[0149] The feature quantity is any one of the occurrence frequency of the phase resetting phenomenon, the normalized occurrence frequency, the average phase velocity, the normalized amplitude, the rate of increase of the occurrence frequency, the average phase velocity, or the normalized amplitude increase rate, and the temporal fluctuation of the occurrence frequency, the average phase velocity, or the normalized amplitude increase rate. The temporal fluctuation of the occurrence frequency is any one of the standard deviation of the occurrence frequency of the phase resetting phenomenon, the coefficient of variation of the occurrence frequency of the phase resetting phenomenon, the degree of variation in the occurrence interval of the phase resetting phenomenon, and the normalized degree of variation in the occurrence interval of the phase resetting phenomenon.

[0150] The depression risk assessment model is generated based on a pair of feature amounts obtained from each frequency component of the electroencephalogram of the subject and the presence or absence of depression in the subject or the severity of depression in the subject.

[0151] The depression risk assessment model generation device includes a feature calculation unit, a selection unit, and a model generation unit. The feature calculation unit calculates feature values ​​that change in accordance with changes in the frequency of occurrence of the phase resetting phenomenon from frequency components of each frequency obtained by decomposing the electroencephalograms of each of multiple subjects using time-frequency analysis. The selection unit selects frequency components that are highly correlated with the risk of depression based on the feature values ​​obtained from each frequency component of the electroencephalograms of each of the multiple subjects and the presence or absence of depression or the severity of depression in the subjects. The model generation unit generates a depression risk assessment model for calculating an assessment value representing the risk of depression, using the feature values ​​obtained from the frequency components selected by the selection unit and the presence or absence of depression or the severity of depression in the subjects.

[0152] The depression risk assessment meter has an electroencephalogram (EEG) measurement unit, the components of the depression risk assessment device described above, and an output unit. The EEG measurement unit measures the electroencephalograms of the person being assessed. The output unit outputs the assessment results obtained by the assessment unit. The depression risk assessment system has the components of the assessment model generation device described above and the components of the depression risk assessment device. The depression risk assessment system may further have an EEG measurement unit.

[0153] Although an embodiment of the present invention has been described above in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and includes designs within the scope of the gist of the present invention.

[0154] It can be used in the medical field as well as in healthcare-related industries.

[0155] REFERENCE SIGNS LIST 1 Depression risk assessment system 2 Depression risk assessment model generation device 3 Depression risk assessment device 4 Electroencephalogram measurement device 21 Memory unit 22 Acquisition unit 23 Feature amount calculation unit 24 Selection unit 25 Model generation unit 26 Output unit 31 Memory unit 32 Acquisition unit 33 Feature amount calculation unit 34 Evaluation unit 35 Output unit 36 ​​Display unit 71 Processor 72 Memory unit 73 Communication interface 74 User interface

Claims

1. A depression risk assessment device comprising: a memory unit that stores a depression risk assessment model for calculating an assessment value representing the risk of depression using features calculated from predetermined frequency components obtained by decomposing electroencephalograms using time-frequency analysis and whose values ​​change in accordance with changes in the frequency of occurrence of the phase resetting phenomenon; a feature calculation unit that calculates features of the same type as the features used in the depression risk assessment model from the predetermined frequency components obtained by decomposing the measured electroencephalograms of a subject using time-frequency analysis; and an evaluation unit that calculates an assessment value representing the risk of depression of the subject using the features calculated by the feature calculation unit as input to the depression risk assessment model.

2. The depression risk assessment device of claim 1, wherein the evaluation value represents whether the risk of depression is low or high, or the severity of depression, and the risk of depression includes the possibility of already being depressed, the possibility of being in a pre-depressive state, the possibility of not currently being depressed but becoming depressed in the future, or the possibility of not currently being depressed but being likely to become depressed more severely if depressed.

3. The depression risk assessment device of claim 1, wherein the feature quantity is any one of the occurrence frequency of the phase resetting phenomenon, the normalized occurrence frequency, the average phase velocity, the normalized amplitude, the increase rate of the occurrence frequency, the average phase velocity or the normalized amplitude increase rate, and the temporal fluctuation of the occurrence frequency, the average phase velocity or the normalized amplitude increase rate.

4. The depression risk assessment device of claim 1, wherein the depression risk assessment model is generated based on a pair of the features obtained from each frequency component of the electroencephalogram of the subject and the presence or absence of depression in the subject or the severity of depression in the subject.

5. A depression risk assessment model generation device comprising: a feature calculation unit that calculates a feature whose value changes in accordance with changes in the frequency of occurrence of the phase resetting phenomenon from the frequency components of each frequency obtained by decomposing the electroencephalograms of each of a plurality of subjects by time-frequency analysis; a selection unit that selects frequency components that have a high correlation with the risk of depression based on the feature obtained from each frequency component of the electroencephalograms of each of the plurality of subjects and the presence or absence of depression or the severity of depression of the subject; and a model generation unit that generates a depression risk assessment model for calculating an assessment value representing the risk of depression using the feature obtained from the frequency components selected by the selection unit and the presence or absence of depression or the severity of depression of the subject.

6. A depression risk assessment device comprising: a memory unit that stores a depression risk assessment model for calculating an assessment value representing the risk of depression using features calculated from predetermined frequency components obtained by decomposing electroencephalograms by time-frequency analysis and whose values ​​change in accordance with changes in the frequency of occurrence of the phase resetting phenomenon; an electroencephalogram measurement unit that measures the electroencephalograms of a subject to be assessed; a feature calculation unit that calculates feature values ​​of the same type as the feature values ​​used in the depression risk assessment model from the predetermined frequency components obtained by decomposing the measured electroencephalograms by time-frequency analysis; an evaluation unit that calculates an assessment value representing the risk of depression of the subject to be assessed using the feature values ​​calculated by the feature calculation unit as input to the depression risk assessment model; and an output unit that outputs the evaluation result by the evaluation unit.

7. A depression risk assessment system comprising: a first feature calculation unit that calculates a feature, the value of which changes in accordance with a change in the frequency of occurrence of the phase resetting phenomenon, from the frequency components of each frequency obtained by decomposing the electroencephalograms of each of a plurality of subjects by time-frequency analysis; a selection unit that selects frequency components that have a high correlation with a risk of depression, based on the feature obtained from each frequency component of the electroencephalograms of each of the plurality of subjects and the presence or absence of depression or the severity of depression of the subject; a model generation unit that generates a depression risk assessment model for calculating an assessment value representing the risk of depression, using the feature obtained from the selected frequency components that are the frequency components selected by the selection unit and the presence or absence of depression or the severity of depression of the subject; a second feature calculation unit that calculates feature values ​​of the same type as the feature values ​​used in the depression risk assessment model, from the selected frequency components obtained by decomposing the measured electroencephalograms of the subject by time-frequency analysis; and an evaluation unit that calculates an assessment value representing the risk of depression of the subject by using the feature calculated by the second feature calculation unit as an input to the depression risk assessment model.

8. A depression risk assessment method comprising: a reading step of reading from a memory unit a depression risk assessment model for calculating an assessment value representing the risk of depression using features calculated from predetermined frequency components obtained by decomposing electroencephalograms by time-frequency analysis and whose values ​​change in accordance with changes in the frequency of occurrence of the phase resetting phenomenon; a feature calculation step of calculating features of the same type as the features used in the depression risk assessment model from the predetermined frequency components obtained by decomposing the measured electroencephalograms of the subject by time-frequency analysis; and an evaluation step of calculating an assessment value representing the risk of depression of the subject by using the features calculated in the feature calculation step as input to the depression risk assessment model.

9. A depression risk assessment model generation method comprising: a feature calculation step of calculating a feature whose value changes in accordance with changes in the frequency of occurrence of the phase resetting phenomenon from the frequency components of each frequency obtained by decomposing the electroencephalograms of each of a plurality of subjects by time-frequency analysis; a selection step of selecting frequency components having a high correlation with the risk of depression based on the feature obtained from each frequency component of the electroencephalograms of each of the plurality of subjects and the presence or absence of depression or the severity of depression of the subjects; and a model generation unit that generates a depression risk assessment model for calculating an assessment value representing the risk of depression using the feature obtained from the frequency components selected in the selection step and the presence or absence of depression or the severity of depression of the subjects.

10. A program for causing a computer to function as the depression risk assessment device according to claim 1.

11. A program for causing a computer to function as the depression risk assessment model generating device according to claim 5.

Images