Prediction device, prediction method, and prediction program for predicting manic state or depressive state of subject

Abstract

A prediction device 3 comprises an acquisition unit 31 that acquires objective information of a subject, and a prediction unit 33 that predicts a manic state or a depressive state of the subject on the basis of the objective information.

Description

Predictive device, prediction method, and prediction program for predicting manic or depressive states in a subject.

[0001] The present invention relates to a prediction device, prediction method, and prediction program for predicting manic or depressive states in a subject.

[0002] In treating bipolar disorder, which involves alternating depressive and manic episodes, it is crucial for the patient to be aware of their own mood swings and to take care to avoid irregular sleep, activity, and lifestyle. However, it has been difficult for patients with this disorder to recognize even mild manic episodes. Furthermore, attending physicians often cannot accurately assess the patient's condition and adequately provide treatment and lifestyle guidance to prevent symptom relapse during short consultations that occur every few weeks to several months. Therefore, efforts are being made to prevent symptom relapse by encouraging bipolar disorder patients to self-monitor, record their activity and sleep, and promote a consistent lifestyle (for example, Non-Patent Document 1).

[0003] "The Importance of Self-Monitoring," Tomoshibi Counseling Office, December 11, 2021, Internet <URL: https: / / kokoro-onayami.com / self-monitoring / 1330 / > Bjorn Schuller et al., “The INTERSPEECH 2013 Computational Paralinguistics Challenge: Social Signals, Conflict, Emotion, Autism,” in Proc. Interspeech 2013. doi: 10.21437 / Interspeech.2013-56 Bjorn Schuller et al., “The INTERSPEECH 2012 Speaker Trait Challenge,” in Proc. Interspeech 2012, 254-257, doi: 10.21437 / Interspeech.2012-86

[0004] However, conventional self-monitoring methods require affected individuals to record their activity and sleep daily, which is time-consuming, and activity levels and sleep content are easily subjective and therefore prone to inaccuracy. Furthermore, they do not function as a means of preventing relapse unless conscious recording is continued over a long period, and it is difficult to feel the effects of such efforts in the short term. As a result, maintaining motivation and continuing these methods is difficult, and they are not yet fully implemented in society.

[0005] Furthermore, while a certain degree of mood fluctuation is observed even in the general population who do not suffer from mental illness, monitoring moods is expected to improve mental health, social life, and interpersonal relationships for the general population as well. It is also desirable that monitoring enables the early detection of depression and bipolar disorder.

[0006] This invention has been made in view of the above problems, and aims to predict manic or depressive states while reducing the burden on the subject.

[0007] To solve the above problems, the present invention includes the following embodiments: 1. A prediction device comprising: an acquisition unit for acquiring objective information of a subject; and a prediction unit for predicting the manic or depressive state of the subject based on the objective information. 2. The prediction device according to 1, wherein the subject's target disease is bipolar disorder. 3. The prediction device according to 1 or 2, wherein the objective information includes any of (1) to (4) below: (1) Activity information including at least one of the following: number of steps, calories burned, distance traveled, sitting time, high-intensity activity time, moderate-intensity activity time, light-intensity activity time and elevation gain, walking speed, step frequency, stride length, step width, step width instability index, floor climbing speed, rate of altitude change, climbing efficiency, metabolic equivalent, metabolic efficiency, and energy efficiency. (2) Sleep information including at least one of the following: time to get into bed, time to fall asleep, number of consecutive sleeps, and sleep stages. (3) Heart rate information. (4) Voice information. 4. The prediction device according to 3, wherein the objective information includes (1) and (2) above. 5. The prediction device according to item 3, wherein the objective information includes (4) above. 6. The prediction device according to item 3, wherein the activity level information includes at least one of steps, calories burned, and distance traveled. 7. The prediction device according to any one of items 1 to 6, wherein the prediction unit predicts the manic or depressive state of the subject using a prediction model that has been machine-trained to output a manic or depressive state when the objective information is input. 8. The prediction device according to item 7, wherein the prediction model is a random forest (RF), a gradient boosting tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), a fully convolutional network (FCN), a recurrent neural network (RNN), long short-term memory (LSTM), or a gated regressive unit (GRU). 9. The prediction device according to item 8, wherein the prediction model is a random forest or long short-term memory (LSTM). 10. A prediction device according to any one of items 1 to 9, further comprising an alert unit that issues an alert in accordance with the prediction result of the prediction unit. Item 11. A prediction system comprising: a prediction device according to any one of items 1 to 10; and a terminal device that is communicably connected to the prediction device and records the objective information.Item 12. The prediction system according to Item 11, wherein the terminal device is a wearable device or smartphone application for accumulating at least one of activity level information, sleep information, heart rate information, and voice information in daily life. Item 13. A prediction method comprising: an acquisition step of acquiring objective information of a subject; and a prediction step of predicting a manic or depressive state of the subject based on the objective information. Item 14. A prediction program that causes a computer to operate as an acquisition unit for acquiring objective information of a subject, and a prediction unit for predicting a manic or depressive state of the subject based on the objective information.

[0008] According to the present invention, it is possible to reduce the burden on the subject and more accurately predict whether they are in a manic or depressive state, and whether they are likely to become manic or depressive in the future, based on objective information recorded in their daily life.

[0009] This is a block diagram illustrating the schematic configuration of a prediction system according to one embodiment of the present invention. This is a graph showing the relationship between the number of features in voice information and the cumulative HSIC score. This is an example of an objective information database. This is a flowchart showing the processing procedure for a method of predicting the manic and depressive states of a subject. (a) is a graph showing the percentage of time points in which a depressive state was predicted based on activity level, sleep, heart rate variability, and voice in individuals with bipolar disorder and healthy controls, and (b) is a graph showing the percentage of time points in which a manic state was predicted based on activity level, sleep, heart rate variability, and voice in individuals with bipolar disorder and healthy controls.

[0010] Embodiments of the present invention will be described below with reference to the accompanying drawings. However, the present invention is not limited to the embodiments described below, and various modifications are possible without departing from its spirit.

[0011] (Overall Configuration) Figure 1 is a block diagram showing the schematic configuration of a prediction system 1 according to one embodiment of the present invention. The prediction system 1 is a system for predicting a manic or depressive state in a subject. In this embodiment, the target disease is bipolar disorder, and the subject is a patient who is receiving treatment at a medical institution for bipolar disorder, but it may also be a general person who does not suffer from the target disease.

[0012] The prediction system 1 comprises a terminal device 2 and a prediction device 3. The terminal device 2 and the prediction device 3 are connected to each other via a network N such as the Internet.

[0013] (Terminal device) Terminal device 2 is a smartphone, smartwatch, fitness tracker, etc., managed by the subject. Terminal device 2 may be owned by the patient or lent to the patient, as long as it is something the subject can use on a daily basis.

[0014] The subject can use terminal device 2 to record one to several types of objective information recorded during their daily life. The objective information recorded during daily life refers to any of the following (1) to (4), or a combination thereof: (1) Activity information including at least one of the following: number of steps, calories burned, distance traveled, sitting time, high-intensity activity time, moderate-intensity activity time, light-intensity activity time, elevation gain, walking speed, step rate, stride length, step width, step width instability index, floor climbing speed, rate of elevation change, metabolic equivalent, and metabolic efficiency (2) Sleep information including at least one of the following: time to go to bed, time to fall asleep, number of consecutive sleep cycles, and sleep stage (3) Heart rate information (4) Voice information

[0015] Furthermore, due to the ease of obtaining the information, it is preferable that the objective information includes (1) activity level information and (2) sleep information. In addition, since audio information can be obtained from smartphones, etc., and contributes greatly to prediction when used in combination with other information, it is preferable that the objective information includes (4) audio information.

[0016] The terminal device 2 includes a sensor 21, an objective information recording unit 22, and an objective information transmission unit 23.

[0017] Sensor 21 is a motion sensor that detects the subject's body movement, or a sensor that detects heart rate, or both. Sensor 21 does not need to be built into the terminal device 2 and may be a separate unit from the terminal device 2.

[0018] The objective information recording unit 22 records objective information, including activity level information and sleep information calculated from body movements detected by the sensor 21, and heart rate information detected by the heart rate sensor, in a memory (not shown), and also records voice information using a voice recording function. The objective information transmission unit 23 periodically (for example, every day) reads the recorded objective information from the memory and transmits it to the prediction device 3. The objective information transmission unit 23 may also transmit the objective information to the prediction device 3 in real time. As an example of a terminal device 2 having the functions of detecting, recording, and transmitting the activity level information, sleep information, and heart rate information described above, the Charge 4 manufactured by Fitbit is given, and as an example of an end device 2 that records and transmits voice information, a smartphone and application having such functions are given.

[0019] Among the objective information recorded in daily life, activity level information preferably includes at least one of the following: steps, calories burned, distance traveled, sedentary minutes, very active minutes, fairly active minutes, lightly active minutes, floor elevation, walking speed, step rate, stride length, width of gait, stability proxy index, floor rate, elevation change rate, metabolic equivalent of task (MET), and metabolic efficiency. In particular, it is preferable to include at least one of the steps, calories burned, and distance traveled.

[0020] Distance traveled refers to the distance traveled by the subject on their own and does not include distance traveled by vehicle or wheelchair. Seating time refers to the time spent sitting in a chair or similar position. Light-load activity time, moderate-load activity time, and high-load activity time refer to the time (minutes) spent moving slightly, moderately, and significantly, respectively. Ascent / descend amount refers to the number of floors ascended or descended in a day.

[0021] Walking speed refers to the distance traveled per unit time and is calculated by dividing the distance traveled by the time taken. Step rate indicates the number of steps per unit time and is calculated by dividing the number of steps by the time taken. Stride length refers to the average stride length obtained by dividing the distance traveled by the number of steps. Step width is an index that shows the distance between the left and right feet in contact with the ground during walking and is estimated by mathematical models (Donelan type, Hak type, etc.) based on walking speed and height. The step width instability index is an index for evaluating walking stability and is calculated using statistical methods based on walking speed variability and heart rate variability (HRV). Floor ascent rate indicates the number of floors climbed per unit time and is calculated by dividing the number of floors climbed by the time taken. Elevation change rate indicates the amount of elevation change per unit time and is calculated by dividing the elevation change by the time taken. Metabolic equivalent is a metabolic index that represents the intensity of physical activity. Based on standards set by the American College of Sports Medicine and the World Health Organization (WHO), activities are classified into five levels (static, light, moderate, high, and very high intensity). Energy efficiency refers to the distance traveled per calorie expenditure and is calculated by dividing the distance traveled by the calories burned. It is used to evaluate the exercise efficiency of physical activity. Climbing efficiency indicates the amount of vertical movement per calorie expenditure and is calculated by dividing the amount of ascent / descent by the calories burned. When distance travel data is unavailable, metabolic efficiency is calculated by dividing the metabolic equivalent by the heart rate (HR) and is used as an indicator of metabolic intensity relative to physiological load.

[0022] Among the objective information recorded in daily life, sleep information preferably includes at least one of the following: total time in bed, total minutes asleep, total sleep records, and sleep stage.

[0023] Bedtime refers to the total time (in minutes) spent in bed during the night, sleep duration refers to the total time (in minutes) spent asleep within that bedtime, and consecutive sleep count refers to the number of consecutive sleep periods recorded during the day.

[0024] In the heart rate information recorded during daily life, it is preferable that the heart rate variability information includes at least one of the following: coefficient of variation of RR intervals (CVRR), standard deviation of normal and normal intervals (SDNN), root mean square of successive differences (RMSSD), the number of adjacent NN intervals that differ by more than 50 milliseconds (NN50), the percentage of differences between adjacent normal RR intervals that are greater than 50 milliseconds (pNN50), low frequency component (LF), high frequency component (HF), LF / HF, or LF / (LF+HF).

[0025] In recent years, various wearable devices have become commonplace. Some provide real heart rate variability data, while others only provide information in the form of heart rate data every 5 seconds. In the former case, heart rate variability indices such as CVRR, SDNN, RMSSD, NN50, pNN50, LF, HF, LF / HF, and LF / (LF+HF) can be calculated using a predetermined method. The present invention includes a method described in the following sections that enables the estimation of approximate values ​​for each heart rate variability index calculated from real heart rate variability data with a certain degree of accuracy, even from information in the form of heart rate data every 5 seconds. This makes it possible to utilize many commercially available wearable devices that only provide information in the form of heart rate data every 5 seconds in conjunction with this system.

[0026] The calculation of the approximate value of the specific heart rate variability index SDNN is performed as follows. First, the average heart rate information at 5-second intervals during the observation period, provided by Fitbit's Charge 4, is used. The series at the sampling point is to, t i =5 i [s] is defined as (i=0,…,N-1).

[0027] At this time, the heart rate series HR at each point in time ａｖ teeth It is given as follows: Here h ｉ This represents the average heart rate for the i-th 5-second interval.

[0028] This HR ａｖ Based on this, the approximate value of the RR interval (RRI) is calculated using the following formula.

[0029] Here, the constant 60000 represents the conversion from bpm to milliseconds. Each r ｉ Since it is based on a 5-second average heart rate, it can be said to be a stable estimated RRI that smooths out instantaneous fluctuations.

[0030] Next, we will remove outliers. The conditions for a valid RRI are... This is how it is defined.

[0031] According to the American Heart Association, the lower limit of the normal range of resting heart rate is 40 bpm, and the maximum normal heart rate during exercise is usually 220 bpm. Therefore, in this embodiment, r ｍｉｎ is defined as 40 bpm, and r ｍａｘ is defined as 220 bpm.

[0032] r ｍｉｎ and r ｍａｘ are defined as follows, respectively:

[0033] That is, RRIs less than 272.7 ms or exceeding 1500 ms are to be excluded.

[0034] When valid(r i ) = false, the valid time points pri(i) and post(i) before and after are determined under the conditions pri(i) < i < post(i) respectively, and linear interpolation is performed as follows to reconstruct the RRI:

[0035]

[0036] Here, θ = {a, b, c} is estimated by minimizing the following equation by the least squares method:

[0037] The final RRI after removal (estimated_RRI) is calculated by the following equation:

[0038] Further, based on the obtained estimated_RRI, the difference between consecutive heartbeat intervals is calculated by the following equation. Here, estimated_RRI i+1 represents the (i + 1)-th heartbeat interval, and estimated_RRI i represents the i-th heartbeat interval:

[0039] Based on the obtained estimated_RRI, the average value of estimated_RRI is calculated. Here, "N" represents the total number of estimated_RRIs, and "estimated_RRI i " represents the length of the i-th estimated_RRI.

[0040] Based on the estimated_RRI and bar estimated_RRI obtained from the above calculations, the standard deviation of the RR intervals (SDNN), which is the coefficient of variation of the RR interval, is calculated. Here, "N" represents the total number of heart rate intervals during the measurement period, and "estimated_RRI i " represents the length of the i-th normal heart rate interval, and "bar estimated_RRI" represents the average time of all heart rate intervals.

[0041] The approximate value of the specific heart rate variability index (CVRR) is calculated as follows. Based on the estimated_RRI and estimated_SDNN bars obtained from the above calculation, the coefficient of variation of RR intervals (CVRR) is calculated.

[0042] The approximate value of the specific heart rate variability index RMSSD is calculated as follows: Based on the difference from the heart rate interval obtained in the above calculation, the root mean square of successive differences between normal heartbeats (RMSSD) is calculated. Here, "N" represents the total number of heart rate intervals during the measurement period. RMSSD is considered an indicator of vagal tone.

[0043] The approximate value of the specific heart rate variability index NN50 is calculated as follows: Based on the difference from the heart rate interval obtained in the above calculation, the total number of adjacent NN intervals that differ by more than 50 milliseconds (NN50) is calculated. Here, "N" represents the total number of heart rate intervals within the measurement period. If the difference exceeds 50 milliseconds, the index function I is set to 1; otherwise, it is set to 0. NN50 is also considered an index of vagal nerve tone.

[0044] The approximate value of the specific heart rate variability index pNN50 is calculated as follows. Based on the estimated_NN50 obtained from the above calculation, the percentage of heart rates where the difference between consecutive adjacent RRIs exceeds 50 ms (the percentage of differences between adjacent normal RR intervals that are greater than 50 milliseconds: pNN50) is calculated. Here, "N" represents the total number of heart rate intervals during the measurement period. pNN50 is also considered an indicator of vagal nerve tone.

[0045] The calculation of approximate values ​​for the specific heart rate variability indices LF and HF is performed as follows. Based on the estimated_RRI obtained from the above calculation, the window is calculated as an array generated by the Hanning window function. hann(L) generates a window of length L. In the second equation, windowed_estimated_RRI is calculated as the estimated_RRI data after applying the Hanning window. This is an array of L elements obtained by multiplying the estimated_RRI array element by element with the window array.

[0046] Based on the windowed_estimated_RRI obtained from the above calculation, the time series data is converted to frequency domain data by applying the FFT. The FFT is a common algorithm for efficiently calculating the Discrete Fourier Transform (DFT) and its inverse. The DFT is defined by the following equation:

[0047]

[0048] The following formula generates an array of frequency components of the FFT result, denoted as f. L is the number of samples and represents the length of the windowed_estimated_RRI array. △t indicates the time interval between consecutive samples of the original time-domain data. The fftfreq function returns an array of frequency values ​​from 0 to the Nyquist frequency, and from negative Nyquist frequencies to less than 0.

[0049] Based on the FFT results obtained from the above calculations, the power spectral density is obtained by squaring the FFT amplitude and then calculating the power spectral density. In the first equation, the squared amplitude of the FFT result is |X(f)| 2 It is calculated as follows: |X(f)| 2 This represents the power of each frequency component in the frequency domain. This is X(f) and its complex conjugate X * It is calculated by multiplying by (f). In this context, X(f) is the result of applying FFT to windowed average RRI data, and includes both amplitude and phase information of the frequency components. * (f) represents the complex conjugate of X(f). Product |X(f)| 2 is a real number representing the power of each frequency component. The power spectral density, PSD(f), is calculated as the power distribution between the frequency components that make up the signal using the following formula. Here, L represents the number of samples in the time-domain data. PSD is obtained by dividing the squared amplitude of the FFT by L, and provides a measurement of the signal power per unit frequency.

[0050] Based on the FFT results obtained from the above calculations, the LF and HF frequency bands are identified. The LF band represents the low frequency band and includes frequencies in the range of 0.04 to 0.15 Hz. This band typically reflects the activity of both the sympathetic and parasympathetic nervous systems. The HF band represents the high frequency band and includes frequencies in the range of 0.15 to 0.40 Hz. This band is primarily associated with the activity of the parasympathetic nervous system (vagus nerve).

[0051] Based on the FFT and PSD results obtained from the above calculations, the LF and HF values ​​are calculated. The LF power is obtained by integrating the PSD over a frequency range of 0.04 Hz to 0.15 Hz. The HF power is obtained by integrating the PSD over a frequency range of 0.15 Hz to 0.40 Hz.

[0052] The approximate values ​​of the specific heart rate variability indices LF / HF and LF / (LF+HF) are calculated as follows. Based on the LF and HF obtained from the above calculations, LF / HF and LF / (LF+HF) are calculated.

[0053] To validate the accuracy of approximations of heart rate variability (HRV) indices based on average heart rate data collected every 5 seconds, we examined the correlation between measured and approximate heart rate variability values ​​using data from a cohort of 154 pregnant women for whom measured heart rate variability data was available. The correlation matrix results for heart rate variability indices and their approximations showed high Pearson correlation coefficients for most heart rate variability (HRV) indices, indicating a high degree of similarity between each index and its mean approximation. In particular, SDNN showed a correlation coefficient of 0.921. Similarly, RMSSD and pNN50 also showed correlations of 0.913 and 0.918, respectively, demonstrating the reliability of their approximations. On the other hand, the LF / HF ratio, which is commonly used to measure autonomic nervous system balance, showed a relatively low correlation coefficient of 0.714.

[0054] Furthermore, to evaluate the similarity of the distributions of approximate and observed values ​​and the stability of the estimation error, in addition to correlation analysis, equivalence testing using TOST (Two One-Sided Test), principal component analysis (PCA), Bland-Altman plot, and confidence interval analysis based on bootstrap were performed. In all of these methods, a high degree of agreement was confirmed between the estimated and observed values. In TOST, the p-values ​​were less than 0.05 for all major indices, and the effect size (Cohen's d) was also extremely small at less than 0.1, indicating statistical equivalence. In principal component analysis, the similarity index in the principal component space between approximate and observed values ​​was 0.95, confirming the agreement in shape. Furthermore, the generalizability and reproducibility of this algorithm were confirmed in validation using high-resolution and low-resolution heart rate data from multiple publicly available databases.

[0055] Among the objective information recorded in daily life, audio information is collected using various forms of smartphone applications. In verifying the content of this invention, at the end of each day, audio information recorded on a smartphone application regarding (1) the date of the day, (2) good experiences of the day, and (3) bad experiences of the day was quantified as 6,837 types of features extracted using the Interspeech 2013 Paralinguistic Challenge Feature (ComParE2013) method (Non-Patent Literature 2, Non-Patent Literature 3).

[0056] This feature calculates statistics such as the mean, standard deviation, peak position, mean of peak values, and first to third quartiles for 64 types of low-level features (LLDs) (e.g., auditory spectrum, MFCC, speech power per frequency band, fundamental frequency, harmonic component to noise ratio, etc.) and their time differences (ΔLLD), thereby calculating a feature of the same length for any speech.

[0057] For these features, we used HSIC Lasso to verify the importance of features that contribute to mood prediction. The graph in Figure 2 shows the number of features on the X axis and the change in cumulative HSIC score on the Y axis.

[0058] Forty-five features were extracted as features that reached the point where the cumulative score reached 95% of the total score. The following 45 features were extracted from the features extracted by CompParE2013 that particularly contribute to the prediction of manic or depressive states: 'pcm_fftMag_fband250-650_sma_de_posamean', 'pcm_fftMag_spectralKurtosis_sma_de_lpc3', 'audSpec_Rfilt_sma_de

[0020] _quartile2', 'pcm_RMSenergy_sma_de_risetime', 'pcm_fftMag_spectralRollOff90'.0_sma_de_upleveltime75', 'pcm_fftMag_spectralSlope_sma_segLenStddev', 'mfcc_sma

[0010] _lpc3', 'audSpec_Rfilt_sma

[0020] _segLenStddev', 'pcm_fftMag_fband1000-4000_sma_de_minPos', 'mfcc_sma[5]_qregerrQ', 'pcm_fftMag_spectralSkewness_sma_de_maxPos', 'audSpec_Rfilt_sma[9]_peakMeanRel', 'audSpec_Rfilt_sma_de[6]_meanSegLen', 'audSpec_Rfilt_sma_de

[0023] _lpc2', 'mfcc_sma[2]_lpc1', 'pcm_fftMag_spectralFlux_sma_de_pctlrange0-1', 'audSpec_Rfilt_sma_de[8]_peakMeanAbs', 'mfcc_sma_de[3]_meanFallingSlope', 'audSpec_Rfilt_sma[6]_peakMeanMeanDist', 'pcm_fftMag_spectralSlope_sma_upleveltime50', 'pcm_fftMag_fband1000-4000_sma_qregerrQ', 'audSpec_Rfilt_sma

[0023] _upleveltime90', 'jitterDDP_sma_de_centroid', 'pcm_zcr_sma_de_lpc4', 'pcm_fftMag_spectralSlope_sma_stddevRisingSlope', 'pcm_fftMag_spectralSkewness_sma_linregc1', 'audSpec_Rfilt_sma

[0021] _stddev', 'pcm_fftMag_spectralHarmonicity_sma_leftctime', 'pcm_fftMag_spectralCentroid_sma_percentile99.0', 'audSpec_Rfilt_sma

[0015] _stddevRisingSlope', 'mfcc_sma[2]_stddevFallingSlope', 'audSpec_Rfilt_sma[3]_qregerrQ', 'audSpec_Rfilt_sma

[0020] _iqr2-3', 'audSpec_Rfilt_sma_de[9]_lpc3', 'audSpec_Rfilt_sma

[0023] _range', 'audSpec_Rfilt_sma

[0012] _upleveltime25', 'pcm_fftMag_fband1000-4000_sma_de_pctlrange0-1', 'mfcc_sma_de[8]_peakMeanMeanDist', 'mfcc_sma[8]_minSegLen', 'audspecRasta_lengthL1norm_sma_upleveltime25', 'audspec_lengthL1norm_sma_iqr2-3', 'pcm_fftMag_spectralKurtosis_sma_stddevRisingSlope', 'audSpec_Rfilt_sma

[0011] _upleveltime90', 'mfcc_sma_de

[0014] _posamean', 'mfcc_sma

[0010] _peakMeanRel'.

[0059] However, in this invention, the method for acquiring speech information, the content of speech, the method for extracting features, the method for evaluating the contribution to mood prediction, and the types of features used for prediction are merely examples, and the method for acquiring speech information, the content of speech, the method for extracting features, the method for evaluating the contribution to mood prediction, and the types of features used for prediction in this invention are not limited to the content of this example.

[0060] (Prediction device) The prediction device 3 is located on the cloud and can be configured using a general-purpose computer. The hardware configuration of the prediction device 3 includes a processor such as a CPU or GPU (not shown), a main memory such as DRAM or SRAM (not shown), and an auxiliary storage device 30 such as an HDD or SSD. The auxiliary storage device 30 stores the prediction program P, prediction models M1 and M2, and an objective information database D.

[0061] The prediction device 3 comprises, as functional blocks, an acquisition unit 31, a storage unit 32, a prediction unit 33, and an alert unit 34. These functional blocks can be implemented in software by the processor of the prediction device 3. In this case, each of the above units can be implemented by the processor reading the prediction program P stored in the auxiliary storage device 30 into the main memory and executing it. The prediction program P may be downloaded to the prediction device 3 via a communication network such as the Internet, or it may be installed in the prediction device 3 via a computer-readable non-temporary recording medium such as a CD-ROM on which the prediction program P is recorded.

[0062] The acquisition unit 31 acquires objective information of the subject transmitted from the terminal device 2. In this embodiment, the objective information recorded in daily life is any of the following (1) to (4), or a combination thereof: (1) Activity level information including at least one of the following: number of steps, calories burned, distance traveled, sitting time, high-intensity activity time, moderate-intensity activity time, light-intensity activity time, and elevation change; (2) Sleep information including at least one of the following: time to go to bed, time to fall asleep, number of consecutive sleeps, and sleep stage; (3) Heart rate information; (4) Voice information

[0063] The storage unit 32 stores objective information recorded in daily life, acquired by the acquisition unit 31, in the objective information database D. An example of the objective information database D is shown in Figure 3. The objective information database D stores objective information of the subject for at least the past week.

[0064] The prediction unit 33 predicts the subject's manic or depressive state based on objective information stored in the objective information database D. In this embodiment, the prediction unit 33 predicts the subject's manic or depressive state using prediction model M1 and / or prediction model M2. Prediction models M1 and M2 use machine learning or deep learning to predict whether they output a manic state, a depressive state, or a normal state when objective information is input. Specifically, prediction model M1 is a model for short-term prediction, and in this embodiment, when objective information is input, machine learning or deep learning is performed to output whether the subject is manic, depressive, or normal at the same time on the recording date of the objective information. Prediction model M2 is a model for long-term prediction, and when objective information is input, machine learning or deep learning is performed to output whether the subject is manic or depressive one week after the recording date, using four days of past information from the recording date of the objective information as a retrospective window. Studies have shown that predictive performance deteriorates with retrospective periods shorter or longer than four days, such as two to three days or five to seven days, indicating that four days of information is the most effective temporal context. However, the period from the most recent record date to the prediction target date is not limited to the above.

[0065] The predictive models M1 and M2 are preferably, for example, Random Forest (RF), Gradient Boosting Tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), Fully Convolutional Network (FCN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Regressive Unit (GRU).

[0066] For example, when predicting a manic or depressive state at the same time on the latest recording date of objective information stored in the objective information database D (let's say February 3rd), the prediction unit 33 inputs the objective information for February 3rd into the prediction model M1 to predict the manic or depressive state at that time. Similarly, when predicting a manic or depressive state one week later (February 10th) from the latest recording date of objective information stored in the objective information database D (February 3rd), the prediction unit 33 inputs the objective information for the four days prior to the latest recording date (January 30th to February 2nd) into the prediction model M2 to predict the manic or depressive state on February 10th. As shown in the embodiment described later, both prediction models M1 and M2 can predict manic or depressive states with high accuracy.

[0067] The alert unit 34 issues an alert in accordance with the prediction result of the prediction unit 33. The criteria for whether or not to issue an alert are not particularly limited, but for example, if it is predicted that the person will be in a manic or depressive state one week after the latest recording date of objective information that the person is not currently in a manic or depressive state, the alert unit 34 will send a notification to the terminal device 2 recommending a review of lifestyle habits, including sleep and exercise, and a visit to the hospital. The specific form of the alert is not particularly limited, and for example, the prediction result may be displayed on a display (not shown) of the prediction device 3, or if the person wishes, the information may be sent as a notification to terminals used by family members or attending physicians for sharing.

[0068] (Processing Procedure) Figure 4 is a flowchart showing the processing procedure for predicting a manic or depressive state in a subject.

[0069] In step S1, the subject uses terminal device 2 to record objective information about their daily life. In this embodiment, the objective information recorded is one of the subject's activity level information, sleep information, heart rate information, or voice information, or a combination of these.

[0070] In step S2, the recorded objective information is transmitted from the terminal device 2 to the prediction device 3.

[0071] In step S3 (acquisition step), the acquisition unit 31 of the prediction device 3 acquires objective information about the subject transmitted from the terminal device 2.

[0072] In step S4, the storage unit 32 stores the objective information in the objective information database D.

[0073] In step S5 (prediction step), the prediction unit 33 predicts the subject's manic or depressive state at a given point in time, as well as their manic or depressive state in the near future, such as one week later, based on objective information stored in the objective information database D.

[0074] As a result, if it is predicted that the person will be in a manic or depressive state at the same time or in the near future, such as one week later (YES in step S6), the alert unit 34 will issue an alert.

[0075] (Summary) In this embodiment, the subject records objective information using the terminal device 2 in their daily life, and the prediction device 3 predicts the subject's manic or depressive state based on the objective information. Furthermore, by issuing alerts based on the prediction results, self-monitoring, which is considered useful for preventing relapses of manic or depressive states in bipolar disorder, can be effectively performed. Since objective information is automatically recorded by the terminal device 2 and the manic or depressive state is predicted based on that information, the subject does not need to evaluate and record their own mood changes. Therefore, it is possible to predict manic or depressive states while reducing the burden on the subject.

[0076] (Additional Notes) The present invention is not limited to the embodiments described above, and various modifications are possible without departing from the spirit of the invention. Forms obtained by appropriately combining the technical means disclosed in the embodiments are also included in the technical scope of the present invention.

[0077] In the above embodiment, the indicators used by the prediction unit 33 to predict the manic or depressive state of the subject were based solely on objective information about the subject. However, other information, such as medication information, may also be included as indicators.

[0078] Furthermore, in the above embodiment, the prediction unit 33 performs predictions using a machine learning-based prediction model (artificial intelligence algorithm), but predictions may also be performed using a rule-based algorithm.

[0079] Furthermore, the terminal device 2 may include at least some of the functions of the prediction device 3.

[0080] The following describes examples of the present invention, but the present invention is not limited to the following examples.

[0081] In this example, objective information and information on whether the subjects were in a manic or depressive state were collected from subjects (113 individuals with bipolar disorder and 97 control subjects (healthy individuals)) over a period of 6 or 9 months, and predictive models M1 and M2 were generated using the collected data. Specifically, each participant recorded the following objective information daily: (1) activity data including at least one of steps taken, calories burned, distance traveled, sitting time, high-intensity activity time, moderate-intensity activity time, light-intensity activity time, and elevation changes; (2) sleep data including at least one of time to go to bed, time to fall asleep, number of consecutive sleep cycles, and sleep stages; (3) heart rate data; and (4) voice data. They also self-assessed their mood daily on an 11-point scale from -5 (bad) to +5 (good). Furthermore, they assessed the degree of mania or depression once a week using the Altman Self-Rating Mania Scale (ASRM) and the Quick Inventory of Depressive Symptomatology (QIDS-J). In the self-assessment of mood, -5 to -2 was rated as depressive, -1 to 1 as normal, and 2 to 5 as hypomanic. In ASRM, the cutoff value was set to 11, while in QIDS-J, the cutoff value was set to 20.

[0082] A training dataset was generated by labeling each subject's objective information with their self-assessment of their mood at the time the objective information was recorded. Predictive model M1 was then generated by predicting this dataset using machine learning or deep learning. Furthermore, a training dataset was generated by dividing each subject's objective information over the entire period into one-week segments and labeling the ASRM and QIDS-J evaluation results recorded one week after the end of each segment's objective information. Predictive model M2 was then generated by applying machine learning to this dataset. Predictive models M1 and M2 were predicted using Random Forest (RF), Gradient Boosting Tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), Fully Convolutional Network (FCN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Regressive Unit (GRU), and the algorithm with the highest prediction accuracy was selected.

[0083] In this embodiment, an architecture based on RNN, LSTM, and GRU is adopted as a deep learning model to predict the manic or depressive state of a subject, using objective information as input. Furthermore, an enhanced configuration combining the following two mechanisms is introduced to capture the structural characteristics of time-series information with higher accuracy.

[0084] First, the architecture incorporates a Temporal Multi-Scale Architecture as its primary mechanism. This architecture provides multiple input sequences in parallel, each composed of different time scales, such as "past 2 days," "past 3 days," "past 4 days," "past 5 days," and "past 7 days," and inputs each into an independent LSTM or GRU subnetwork. This makes it possible to simultaneously extract and learn information with different periodicities, such as short-term rapid fluctuations (e.g., a sudden decrease in sleep) and long-term trend changes. The outputs from each scale are ultimately combined and input as an overall representation vector to a higher layer (fully connected layer or attention layer).

[0085] Next, the second mechanism is a Dynamic Temporal Feature Selection mechanism. In this mechanism, a module (e.g., a linear layer or self-attention) is placed at each time step in the time series to calculate the importance score of the features, and the temporal weights are dynamically adjusted based on these scores. This allows for assigning high weights to points in time when, for example, rapid sleep changes or abnormal heart rate fluctuations occur, while applying lower weights to redundant data in normal states, thereby contributing to the optimization of prediction accuracy. Evaluation experiments described later confirmed that the configuration incorporating these two mechanisms improves the accuracy of identifying manic or depressive states and the predictive performance of future states compared to conventional single-window and statically weighted models.

[0086] In this embodiment, for example, the following formula is used to determine the weight coefficient α based on the importance of the feature quantities at each time step. ｔ Calculate and use a weighted vector h to highlight time series information that has a high contribution to the prediction. ｗｅｉｇｈｔｅｄ It is also possible to obtain the above. The objective information input may include various derived features such as walking speed, step width, step width instability index, MET, and energy efficiency.

[0087]

[0088] This vector is input to a higher-level fully connected layer or attention mechanism layer and transformed into the final prediction result.

[0089]

[0090] Here, W ｘ , v, and b are learnable parameters.

[0091] Next, the accuracy of predictions by predictive model M1 (manic and depressive states at the same point in time) and predictive model M2 (manic and depressive states one week later) was verified by cross-validation. As predictive model M1, random forest (RF), gradient boosting tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), fully convolutional network (FCN), recurrent neural network (RNN), long short-term memory (LSTM), and gated regressive unit (GRU) were used. For each model, the accuracy of simultaneous prediction of manic and depressive states based on the Altman Self-Rated Mania Scale (ASRM) and the Quick Indicator of Depressive Symptoms (QIDS-J) was evaluated using seven combinations of input features: (1) activity level and sleep information only, (2) heart rate information only, (3) voice information only, (4) combination of activity level, sleep information, and heart rate information, (5) combination of activity level, sleep information, and voice information, (6) combination of heart rate information and voice information, and (7) combination of activity level, sleep information, heart rate information, and voice information. The results of the prediction accuracy (accuracy, precision, recall, F-score, and area under the ROC curve (AUC)) for each input condition are shown in Table 1. In particular, RNN, LSTM, and GRU showed generally higher performance than the other models. For example, when only activity level and sleep information was input, LSTM and GRU showed high prediction accuracy with an accuracy of 0.89, an AUC of 0.88, and an F-score of 0.87. (2) Using only heart rate variability information, LSTM and GRU achieved an accuracy of 0.87, an AUC of 0.87, and an F-value of 0.86. (3) Even with only voice information input, RNN, LSTM, and GRU showed stable performance with an F-value of 0.84 to 0.85 and an AUC of 0.84 to 0.85. Furthermore, even higher performance was confirmed with multimodal feature combinations. (4) In combination of activity level / sleep information and heart rate variability information, LSTM and GRU showed an F-value of 0.87 and an AUC of 0.87. (5) In combination of activity level / sleep information and voice information, LSTM recorded excellent results with an accuracy of 0.91, an AUC of 0.88, and an F-value of 0.88. (6) When combining heart rate variability information and voice information, LSTM and GRU maintained high performance with an F value of 0.86 and an AUC of 0.86.The highest prediction accuracy was achieved in (7) a model that integrated all activity level / sleep information, heart rate variability information, and voice information, with both LSTM and GRU achieving accuracy of 0.95, AUC of 0.95, and F-score of 0.91.

[0092]

[0093] Next, Table 2 shows the accuracy, precision, recall, F-score, and area under the ROC curve (AUC) for predicting manic and depressive states based on daily subjective manic / depressive state assessments from +5 to -5 points, for the following prediction models M2: (1) activity level and sleep information only, (2) heart rate variability information only, (3) voice information only, (4) combination of activity level and sleep information and heart rate variability information, (5) combination of activity level and sleep information and voice information, (6) combination of heart rate variability information and voice information, and (7) all combinations of activity level and sleep information, heart rate variability information and voice information. RNN, LSTM, and GRU consistently demonstrated higher performance than other models in all feature combinations. In particular, (7) when all features (activity level and sleep information, heart rate variability information, and voice information) were integrated, LSTM and GRU achieved AUC = 0.97, accuracy = 0.96, and F-score = 0.93, confirming that they were the best configuration in prediction model M2. Furthermore, (1) when only activity level and sleep information was used as input, LSTM and GRU showed high accuracy with AUC = 0.88, accuracy = 0.89, and F-score = 0.88, and (2) even with only heart rate variability information, they maintained AUC = 0.88 and F-score = 0.86. (3) Even under the condition of using only voice information, RNN, LSTM, and GRU all recorded AUC = 0.87 and F-score = 0.85, demonstrating that even voice data alone, which was previously considered to be noisy, can achieve practical accuracy. Furthermore, the prediction accuracy remained stable even with two modalities: (4) the combination of activity level and sleep information with heart rate variability information recorded an AUC of 0.93 and an F-value of 0.92 for LSTM and GRU; (5) the combination of activity level and sleep information with voice information recorded an AUC of 0.94 and an F-value of 0.92; and (6) the combination of heart rate variability information with voice information also recorded an AUC of 0.91 and an F-value of 0.91.For comparison, Random Forest (RF) showed stable performance for all single features ((1) to (3)), achieving AUC = 0.85, 0.83, and 0.85 respectively. For the two-feature combinations (4) to (6), RF's AUC improved to 0.87 to 0.88, and for the all-feature combination (7), it achieved an AUC of 0.89, accuracy of 0.90, and F-score of 0.88, demonstrating a good balance between stability and accuracy. On the other hand, LGBM performed similarly to or slightly worse than RF for single information, but its accuracy improved to AUC = 0.87 to 0.88 for the combinations (4) to (6), and for the all-feature combination (7), it recorded an AUC of 0.93, accuracy of 0.95, and F-score of 0.92, confirming that it is the best configuration for the prediction model M2.

[0094]

[0095] Next, Table 3 shows the accuracy, precision, recall, F-score, and area under the ROC curve (AUC) for predictions based on a daily subjective assessment of manic and depressive states from +5 to -5 points after one week, for the following prediction models M2: Random Forest (RF), Gradient Boosting Tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), Fully Convolutional Network (FCN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Regressive Unit (GRU). This is done using (1) activity level and sleep information only, (2) heart rate variability information only, (3) voice information only, (4) a combination of activity level and sleep information and heart rate variability information, (5) a combination of activity level and sleep information and voice information, (6) a combination of heart rate variability information and voice information, and (7) a combination of all three types of activity level and sleep information, heart rate variability information and voice information. The prediction model M2 is configured to predict the state one week later, using four days' worth of objective information from the most recent recording date as input. Random Forest (RF) also demonstrated stable predictive performance in predicting manic and depressive states one week later, recording AUC = 0.80, 0.75, and 0.76 for each single information input: (1) activity level and sleep information, (2) heart rate variability information, and (3) voice information. Furthermore, for all combinations of two types of information—(4) activity level and sleep information and heart rate variability information, (5) activity level and sleep information and voice information, and (6) heart rate variability information and voice information—RF recorded AUC = 0.83 to 0.86, achieving higher accuracy than single information input. In addition, when all three types of information—(7) activity level and sleep information, heart rate variability information, and voice information—were integrated and inputted, RF achieved AUC = 0.90, accuracy = 0.91, and F-score = 0.88. In particular, LSTM and GRU showed the highest prediction performance in all configurations. Especially in the all-feature input configuration (7), both LSTM and GRU recorded AUC = 0.96, accuracy = 0.96, and F-score = 0.91 to 0.92. Furthermore, in all input conditions (1) to (6), LSTM and GRU consistently achieved AUC = 0.84 or higher and F-score = 0.83 or higher, demonstrating the effectiveness of deep learning time series models in long-term prediction.From these results, it was confirmed that the predictive model M2 can also accurately predict future fluctuations in manic or depressive states by utilizing a multimodal feature configuration that combines multiple physiological and behavioral information, as well as time series models such as LSTM and GRU.

[0096]

[0097] In paragraph 0077, when predicting a manic or depressive state one week later, if the subject is already in a manic or depressive state at the time the objective information used for prediction is collected, predicting the state one week later may simply represent a continuation of the state at that time. Therefore, in this verification, we limited the analysis to cases where the subject is in a normal state, neither manic nor depressive, at the time the objective information used for prediction is collected, and evaluated the accuracy of predicting a manic or depressive state one week later from that state. Table 4 shows the predictive performance (accuracy, precision, recall, F-score, AUC) of manic and depressive states based on subjective evaluations after one week, for the following cases where the predictive model M2 is Random Forest (RF), Gradient Boosting Tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), Fully Convolutional Network (FCN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Regressive Unit (GRU): (1) activity level and sleep information only, (2) heart rate variability information only, (3) voice information only, (4) combination of activity level and sleep information and heart rate variability information, (5) combination of activity level and sleep information and voice information, (6) combination of heart rate variability information and voice information, and (7) combination of three types of activity level and sleep information, heart rate variability information and voice information. Predicting future abnormalities (mania or depression) from a normal state is difficult to classify and tends to result in slightly lower prediction accuracy. However, Random Forest (RF) showed stable prediction performance compared to other algorithms, and even when using only (1) activity level and sleep information, (2) heart rate variability information, and (3) voice information as inputs, it was possible to make predictions with an accuracy of AUC = 0.74, 0.73, and 0.72. Furthermore, in any configuration combining two types of information, such as (4) activity level and sleep information and heart rate variability information, (5) activity level and sleep information and voice information, and (6) heart rate variability information and voice information, a clear improvement in performance was observed, with an AUC of 0.80 to 0.85. (7) When all features were integrated, RF recorded an AUC of 0.89, accuracy of 0.90, and F-score of 0.86, demonstrating that it is possible to predict future manic and depressive states from a normal state with high accuracy.Furthermore, RNN, LSTM, and GRU consistently maintained high accuracy in all configurations. In particular, in the all-feature configuration (7), LSTM recorded an AUC of 0.92, accuracy of 0.93, and F-score of 0.92, while GRU recorded an AUC of 0.91, accuracy of 0.92, and F-score of 0.91. These models were confirmed to have high practicality even under more stringent conditions, such as predicting anomalies from normal states.

[0098]

[0099] Next, we integrated all four types of objective information—activity level, sleep information, heart rate variability, and voice information—and examined the prediction accuracy (accuracy, precision, recall, F-score, AUC) of manic or depressive states one week later when the length of past information input to the predictive model M2 was changed (2 days, 3 days, 4 days, 5 days, 7 days). The results are shown in Table 5. Even when the input period was 2 or 3 days, a consistent prediction accuracy of AUC of 0.82 to 0.88 was obtained in all models. However, the highest overall prediction accuracy was obtained when 4 days of information was input. Specifically, in all models—Random Forest (RF), LGBM, RNN, LSTM, and GRU—the AUC when 4 days of information was input was recorded as 0.91, 0.89, 0.92, 0.96, and 0.94, respectively, showing results that surpassed other input periods. Furthermore, the accuracy, recall, and F-score were highest when using 4 days of data, and in particular, the LSTM achieved very high predictive performance with an accuracy of 0.96 and an F-score of 0.96. On the other hand, when using 5 days and 7 days of data as input, the overall accuracy tended to be the same as or slightly lower than that of 4 days. This is thought to be because the historical data is too long, potentially including noise and irrelevant fluctuations. Therefore, it is suggested that using 4 days of objective data as input is the most effective and optimal memory length (input window) for predicting manic or depressive states one week later.

[0100]

[0101] While mood swings similar to manic and depressive states are expected to some extent in the general population who do not suffer from bipolar disorder, the present invention has identified similar hypomanic and depressive states in the general population who do not suffer from bipolar disorder, albeit at a significantly lower frequency than those observed in individuals with bipolar disorder. Figure 5(a) shows a comparison of the proportion of time points predicted to be depressive based on activity levels, sleep, heart rate variability, and voice in individuals with bipolar disorder and healthy controls (RF model), and Figure 5(b) shows a comparison of the proportion of time points predicted to be manic (RF model).

[0102] This invention is applicable not only to predicting the severity of bipolar disorder, but also to predicting manic or depressive states in depressive disorders, substance-related disorders, and anxiety disorders. Differentiating between depression, which presents only as depressive episodes, and bipolar disorder is often difficult, but this invention can be used to detect depressive episodes in depression and to differentiate between depression and bipolar disorder. Furthermore, it is expected that even people without mental illness can use this invention to monitor their daily mood changes, thereby improving their mental health, social life, and interpersonal relationships, and contributing to the early detection of bipolar disorder.

[0103] 1 Prediction system 2 Terminal device 21 Sensor 22 Objective information recording unit 23 Objective information transmission unit 3 Prediction device 30 Auxiliary storage device 31 Acquisition unit 32 Storage unit 33 Prediction unit 34 Alert unit D Objective information database M1 Prediction model M2 Prediction model N Network P Prediction program

Claims

1. A prediction device comprising: an acquisition unit for acquiring objective information about a subject; and a prediction unit for predicting whether the subject is in a manic or depressive state based on the objective information.

2. The prediction device according to claim 1, wherein the target disease of the subject is bipolar disorder.

3. The prediction device according to claim 1 or 2, wherein the objective information includes any of the following (1) to (4): (1) Activity information including at least one of the following: number of steps, calories burned, distance traveled, sitting time, high-intensity activity time, moderate-intensity activity time, light-intensity activity time and elevation gain, walking speed, step frequency, stride length, step width, step width instability index, floor climbing speed, rate of altitude change, climbing efficiency, metabolic equivalent, metabolic efficiency, and energy efficiency; (2) Sleep information including at least one of the following: time to get into bed, time to fall asleep, number of consecutive sleep cycles, and sleep stages; (3) Heart rate information; (4) Voice information 4. The prediction device according to claim 3, wherein the objective information includes (1) and (2) above.

5. The prediction device according to claim 3, wherein the objective information includes (4) above.

6. The prediction device according to claim 3, wherein the activity information includes at least one of the number of steps, calories burned, and distance traveled.

7. The prediction device according to claim 1 or 2, wherein the prediction unit predicts the manic or depressive state of the subject using a prediction model that has been trained by machine learning to output a manic or depressive state when the objective information is input.

8. The prediction device according to claim 7, wherein the prediction model is a random forest (RF), a gradient boosting tree (GBT), eXtreme Gradient Boosting (XGBoost), LightGBT (LGBM), a fully convolutional network (FCN), a recurrent neural network (RNN), long short-term memory (LSTM), or a gated regressive unit (GRU).

9. The prediction device according to claim 8, wherein the prediction model is a random forest or long-term short-term memory.

10. The prediction device according to claim 1, further comprising an alert unit that issues an alert in accordance with the prediction result of the prediction unit.

11. A prediction system comprising: a prediction device according to claim 1; and a terminal device that is communicatively connected to the prediction device and records the objective information.

12. The prediction system according to claim 11, wherein the terminal device is a wearable device or smartphone application for accumulating at least one of the activity level information, sleep information, heart rate information, and voice information in daily life.

13. A prediction method comprising: an acquisition step of obtaining objective information about a subject; and a prediction step of predicting whether the subject is in a manic or depressive state based on the objective information.

14. A prediction program that operates a computer as an acquisition unit for acquiring objective information about a subject, and a prediction unit for predicting whether the subject is in a manic or depressive state based on the objective information.

Images