Estimation model generation device, estimation device, estimation model generation method, estimation method, and program

Abstract

This estimation model generation device comprises: a feature amount extraction unit that extracts a plurality of types of feature amounts from a potential detected from a subject; a feature amount recording unit that records the plurality of types of feature amounts for each hand motion the subject is instructed to perform; a model generation unit that generates, from a combination of one or more types of feature amounts among the plurality of types of feature amounts, a plurality of estimation models for outputting a hand motion intended by the subject; and a model output unit that outputs, from among the plurality of generated estimation models, an estimation model having a higher estimation accuracy than a predetermined criterion, or an estimation model included in a predetermined number or a predetermined ratio from the top when the estimation models are arranged in descending order of estimation accuracy.

Description

Estimation model generation device, estimation device, estimation model generation method, estimation method, and program

[0001] The present invention relates to an estimation model generation device, an estimation device, an estimation model generation method, an estimation method, and a program. This application claims priority to Japanese Patent Application No. 2024-113911, filed on July 17, 2024, the contents of which are incorporated herein by reference.

[0002] Illnesses such as stroke can leave people with residual paralysis. Patients with residual paralysis undergo rehabilitation to treat the paralysis. Research is being conducted to improve the effectiveness of rehabilitation by analyzing movements and providing feedback.

[0003] For example, Patent Document 1 discloses a method for estimating the movement that a subject wants to make with their hands based on myoelectric potentials measured from the upper limbs.

[0004] JP 2024-67915 A

[0005] However, further improvement in estimation accuracy is necessary. An object of the present invention is to provide an estimation model generation device, an estimation device, an estimation model generation method, an estimation method, and a program that solve the above-mentioned problems.

[0006] One aspect of the present invention is an estimation model generation device that includes a feature extraction unit that extracts multiple types of feature values ​​from an electric potential detected from a subject; a feature recording unit that records the multiple types of feature values ​​for each hand movement instructed by the subject; a model generation unit that generates multiple estimation models that output the hand movement that the subject wants the subject to perform from each combination of feature values ​​based on all combinations of the multiple types of feature values; and a model output unit that outputs an estimation model with high estimation accuracy.

[0007] According to the present invention, it is possible to improve the accuracy of estimating the motion that the subject wants his or her hand to perform.

[0008] 1 is a diagram showing the configuration of the estimation model generating device 1 according to the first embodiment. FIG. 2 is an example showing myoelectric potential signals measured for different indicated hand movements. FIG. 3 is an example showing a combination of types of feature quantities used to generate a BFS stored in the combination storage unit 192, and a ranking of estimation accuracy of an estimation model generated based on the same number of types of feature quantities of the estimation model. FIG. 4 is a flowchart showing the operation of the estimation model generating device 1 according to the first embodiment. FIG. 5 is a diagram showing the configuration of the estimation device 4 according to the first embodiment. FIG. 6 is a flowchart showing the operation of the estimation device 4 according to the first embodiment. FIG. 7 is a diagram showing the estimation accuracy of an estimation model generated by the estimation model generating device 1. FIG. 8 is a diagram comparing the estimation accuracy of estimation models (BFS, SBFS) generated by the estimation model generating device 1 with the estimation accuracy of a conventional estimation model (MFS). FIG. 9 is a diagram showing combinations of feature quantities whose estimation accuracy falls in the top 20% for each of the generated BFSs.

[0009] (Estimation Model Generating Device) Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Fig. 1 is a diagram showing the configuration of an estimation model generating device 1 according to a first embodiment. The estimation model generating device 1 according to the first embodiment includes a measurement signal receiving unit 11, a feature extracting unit 12, a feature recording unit 13, a model generating unit 14, an estimation accuracy calculating unit 15, a combination recording unit 16, a model output unit 17, and a storage unit 19.

[0010] The measurement signal receiving unit 11 receives electromyograms (EMG) measured from the upper limb 2 of the subject by the measuring device 3. The measuring device 3 is attached to, for example, the skin surface of the upper limb 2, and measures the myoelectric potential signals.

[0011] The measurement device 3 may measure myoelectric potential signals at multiple positions of the upper limb 2. For example, the measurement device 3 measures myoelectric potential signals in, for example, the flexor carpi radialis (FCR), the flexor carpi ulnaris (FCU), the abductor pollicis brevis (APB), and the extensor digitorum communis (EDC).

[0012] The measuring device 3 measures the myoelectric potential signal for each different hand movement of the upper limb 2 instructed to the subject. The different hand movements include, for example, thumbs up, pinch, fist, extension, spread, flexion, etc. For example, the measuring device 3 measures the myoelectric potential signal for specific hand movements by having the subject perform different hand movements in sequence for a certain period of time.

[0013] The measurement device 3 may measure myoelectric potential signals of two upper limbs of the same person, one of which may be paralyzed due to an illness such as acute stroke.

[0014] The feature extraction unit 12 extracts features for each hand movement from the EMG signal. The EMG signals from which features are extracted are signals obtained by performing preprocessing on EMG signals measured for each position. The features are numerical values ​​or vectors that indicate the characteristics of the extracted raw data.

[0015] The extracted features include various types of features. The extracted features include, for example, time domain features. The time domain features include, for example, the following values:・integrated EMG parameter (IEMG) ・average amplitude value (AAV) ・mean absolute value (MAV) ・logarithmic mean absolute value (LMAV) ・non-linear scaled value (NSV) ・6 types of modified mean absolute value (MMAV1-6) in which features from 3 to 6 are novel implementations of the signal fractioning ・mean absolute value slope (MAVS) ・enhanced mean absolute value (EMAV) ・simple square integral (SSI) ・logarithmic version of SSI (LSSI) ・Hjorth parameters including activity (HPA) or variance (VAR), model mobility (HPM), and complexity (HPC) ・absolute value of 3rd, 4th and 5th temporal moments (TM3-5) ・modified skewness (MSKEW1-3) ・kurtosis (KT) ・modified kurtosis (MKURT) ・root mean square of 2nd order (RMSV2) ・root mean square of 3rd order (RMSV3) ・logarithmic implementations of RMSV2 and RMSV3 (LRMSV2-3) ・root squared zero order moment (RSM0) ・waveform length ratio (WLR) ・irregularity factor (IRF) ・first and secondorder root squared normalized descriptors (RSD1-2) ・the mean value of square root (MSR) ・mean value of the absolute value of summation of exponent root (ASM) ・mean absolute value of Napier’s constant value (MANC) ・standard deviation (SD) ・log detector (LOG) ・root mean squared normalized value of the LOG (ROG) ・median differential value (MDPV) ・median power difference value ・difference absolute standard deviation value (DASDV) ・maximum fractal length (MFL) ・wavelength (WL) ・average amplitude change (AAC) ・enhanced wavelength (EWL) ・signal’s zero-crossing (ZC) ・slope sign change (SSC) ・Wilson’s amplitude (WAMP) ・myopulse percentage rate (MYOP) ・multi-channel energy ratio (ER) of the sEMG ・log of the ER (LER) ・mean of energy ratio (MER), ・log of MER (LMER) ・max energy ratio of the sEMG (MXR) ・log of MXR (LMXR) ・multiple hamming windows (MHW) ・multiple trapezoidal windows (MTW) ・histogram of the EMG (HIST) ・simple square root of the amplitude histogram (SAHT) ・2nd to 6th orderautoregressive coefficients (AR2-6) ・cepstral coefficients (CCAR) ・linear predictive coefficients from 2nd to 6th order (LPC2-6) ・logarithmic cardinality (LCARD) ・percentiles of the signal with 75% and 50% (PERC1-2) ・sample entropy (SEN) ・approximate entropy (AEN) ・fuzzy entropy (FEN) ・permutation entropy (PEN)

[0016] The extracted features include, for example, frequency domain features, such as the following: modified amplitude spectrum (MASP), logarithmic version of MASP (MLASP), mean frequency (MNF), modified mean frequency (MMNF), median frequency (MDF), modified median frequency (MMDF), total power (TTP), mean power (MNP), peak frequency (PKF), frequency ratio (FRT), 1st, 2nd, and 3rd spectral moments (SM1-3), median power spectrum (MDS), maximum power spectrum (MXS), power spectrum ratio (PSR), logarithmic power spectrum ratio (LPSR), spectral mean density (SND), spectral median density (SMD), fundamental frequency (FDD), and histogram of the EMG frequency with 2 and 4 bins (FTHT2,4).

[0017] The extracted features include, for example, features in the time-frequency domain. The features in the time-frequency domain include, for example, the following values: ・short-time Fourier transform (STFT) ・fractional Fourier transform (FRFT) ・wavelet packet entropy (WENT) ・discrete wavelet transforms (DWT) ・the energy of wavelet coefficient (EWT), median value energy of wavelet packet coefficients (EWP) ・zero-crossing of wavelet coefficient (ZCWC) ・Hilbert-Huang transform (HHT) ・the Stockwell transform (SWT)

[0018] The extracted features include, for example, spatial domain features. Spatial domain features estimate the spatial correlation of sEMG signals in multi-channel applications. Spatial domain features include, for example, the following pattern recognition spatial domain features: scaled mean absolute value (SMAV), modified scaled mean absolute value (MSMAV), EMG channel correlation coefficients of normalized values ​​using median value, square root value, and root mean square value (CC-D, CC-S, CC-R), the general ratio of the four muscle electrodes (RT4), and the biomechanical ratio of the mean absolute values ​​obtained from the forearm flexors and extensors muscles in 4-channel EMG (FER-4).

[0019] The extracted features include, for example, fractal domain features. Fractal domain features describe the geometric shape of sEMG signals by analyzing the fractal dimension of the time-series signal waveform to monitor temporal events. Fractal domain features include, for example, the following values: ・fractal dimension with the 4th time-step (FR4) ・detrended fluctuation analysis (DFA) ・Higuchi fractal dimension (HFD)

[0020] The feature recording unit 13 records the extracted multiple types of feature amounts for each different hand movement in the feature storage unit 191 of the storage unit 19. Fig. 2 shows an example of feature amounts corresponding to one subject and one hand movement recorded in the feature storage unit 191. Feature amounts IEMG, AAV, MAV, etc. are recorded for each measurement position FCR, FCU, APB, EDC.

[0021] The model generation unit 14 generates an estimation model based on the recorded feature amounts. The generated estimation model is a model that receives feature amounts extracted from the EMG signal as input and outputs hand movements of the upper limb. The model generation unit 14 generates a model that classifies the feature amounts by applying, for example, a support vector machine (SVM) to the feature amounts for each different hand movement.

[0022] The model generation unit 14 generates estimation models using different numbers of feature quantities, for example, by using one or more types of feature quantities from the recorded feature quantities. More specifically, for example, the model generation unit 14 generates an estimation model using one type of feature quantity from the recorded feature quantities, and generates an estimation model using two types of feature quantities. The model generation unit 14 generates an estimation model using, for example, a combination of a predetermined number of feature quantities from the recorded feature quantities. In other words, the model generation unit 14 arbitrarily selects, for example, a predetermined number of feature quantities from the recorded feature quantities, and generates an estimation model using the selected feature quantities. The model generation unit 14 generates an estimation model using, for example, all combinations of the predetermined number of feature quantities from the recorded feature quantities. When the number of recorded feature quantities is p and the model generation unit 14 generates estimation models based on q feature quantities, the model generation unit 14 generates the number of estimation models indicated by Equation (1).

[0023] That is, the model generation unit 14 performs a brute force search and generates an estimation model based on all combinations of types of feature quantities. Here, q may be determined to be a single value or a range. For example, when q is determined to be 1 to 4, the model generation unit 14 generates an estimation model based on one type of feature quantity, generates an estimation model based on two types of feature quantities, generates an estimation model based on three types of feature quantities, and generates an estimation model based on four types of feature quantities. Hereinafter, an estimation model generated by a brute force search is referred to as a "BFS," and in particular, an estimation model generated based on n types of feature quantities and estimating hand movements based on the n types of feature quantities is referred to as a "BFS." ｎ It should be noted that in the brute force search, it is not necessary to generate an estimation model based on all combinations, and some combinations may not be used to generate an estimation model. When the recorded features include time domain features and other features that are not time domain features, the model generation unit 14 can generate an estimation model using a combination of the time domain features and other features that are not time domain features.

[0024] The estimation accuracy calculation unit 15 verifies the generated BFS and calculates the estimation accuracy of the BFS. For example, the features recorded in the feature storage unit 191 are classified in advance into training data and test data according to the subject, the model generation unit 14 generates a BFS based on the training data, and the estimation accuracy calculation unit 15 verifies the BFS using the test data. Here, the method of classifying the features into training data and test data may be arbitrary, and may be a classification method that suits the estimation purpose of the estimation model.

[0025] The combination recording unit 16 records the combination of feature types used to generate the BFS and the ranking of the estimation accuracy of the BFS generated based on the same number of feature types in the combination storage unit 192 of the storage unit 19. Figure 3 shows an example of the combination of feature types used to generate the BFS stored in the combination storage unit 192 and the ranking of the estimation accuracy of the estimation model generated based on the same number of feature types. The combination recording unit 16 may record the combination of feature types used to generate the BFS and the estimation accuracy of the BFS in the combination storage unit 192.

[0026] The model generation unit 14 generates an estimation model that uses a larger number of types of feature quantities based on the generated BFS. Based on a combination of s types of feature quantities used by the first estimation model and t types of non-overlapping feature quantities used by the second estimation model, the model generation unit 14 generates a new estimation model that inputs the (s + t) types of feature quantities used by the first estimation model or the second estimation model and outputs hand movements of the upper limbs.

[0027] The model generation unit 14 generates a new estimation model by referring to the ranking of estimation accuracy recorded in the combination storage unit 192. In the above example, for the combination of s types of feature quantities and the combination of t types of feature quantities, a combination that results in a higher ranking of estimation accuracy for the generated first estimation model and second estimation model is preferentially selected, and a new estimation model is generated.

[0028] Hereinafter, the method of generating an estimation model using a larger number of types of feature quantities based on the combinations of the types of feature quantities stored in the combination storage unit 192 will be referred to as a method of generating an estimation model by a "semi-brute force search." In addition, an estimation model generated by a semi-brute force search will be referred to as an "SBFS." In particular, an estimation model that is generated based on n types of feature quantities and that estimates hand movements based on the n types of feature quantities will be referred to as an "SBFS." ｎ " is called.

[0029] The estimation accuracy calculation unit 15 verifies the SBFS in the same way as the BFS and calculates the estimation accuracy. The combination recording unit 16 records, in the combination storage unit 192 of the storage unit 19, the combination of feature types used to generate the SBFS and the ranking of the estimation accuracy of the SBFS generated based on the same number of feature types.

[0030] The number of SBFSs generated by the model generation unit 14 and verified by the estimation accuracy calculation unit 15 is a preset number. The number set here is set to a value smaller than the number of estimation models generated when a brute force search is performed. The number set here is set based on, for example, the computing capacity of the estimation model generation device 1. ｎ Number of SBFS ｎ＋１ The number of SBFSs generated by the model generation unit 14 may be set so that the ratio of the numbers of SBFSs to SBFSs is a predetermined number. For example, ｎ Number of SBFS ｎ＋１ The number of SBFSs generated by the model generation unit 14 may be set so that the ratio of the numbers of SBFSs is close to the golden ratio (1:(1+√5) / 2).

[0031] The model generating unit 14 generates an estimation model that uses a larger number of types of feature quantities based on the combinations of feature quantities stored in the combination storage unit 192 .

[0032] The model generation unit 14 may store the generated estimation model in the storage unit 19. In generating an estimation model by semi-brute force search, an SBFS may be generated by improving an estimation model that uses a small number of feature quantities. As a method for estimating hand movements based on n types of feature quantities, BFS may be used.１ From BFS ４ , SBFS ｘ (x is an integer of 5 or more) can be used for integration processing. In this case, SBFS ５ is BFS １ and BFS ４ A combination of or BFS ２ and BFS ３ It is composed of a combination of SBFS. ６ is BFS １ and SBFS ５ Combination of BFS ２ and BFS ４ A combination of or BFS ３ and BFS ３ It is composed of a combination of the following: ７ is BFS １ and SBFS ６ Combination of BFS ２ and SBFS ５ A combination of or BFS ３ and BFS ４ It is composed of a combination of the following.

[0033] Also, BFS １ From BFS ３ , SBFS ｘ (x is an integer of 4 or more) can be used for integration processing. In this case, SBFS ４ is BFS １ and BFS ３ A combination of or BFS ２ and BFS ２ It is composed of a combination of SBFS. ５ is BFS １ and SBFS ４ A combination of or BFS ２ and BFS ３ It is composed of a combination of SBFS. ６ is BFS １ and SBFS ５ Combination of BFS ２ and SBFS ４ A combination of or BFS ３ and BFS ３ It is composed of a combination of the following.

[0034] Also, BFS １ , BFS ２ , SBFS ｘ (x is an integer of 3 or more) can be used for integration processing. In this case, SBFS ３ is BFS １ and BFS ２ It is composed of a combination of SBFS. ４ is BFS １ and SBFS ３ A combination of or BFS ２ and BFS ２ It is composed of a combination of SBFS. ５ is BFS １ and SBFS ４ Combination of BFS ２ and SBFS ３ It is composed of a combination of the following.

[0035] The model generation unit 14 generates an estimation model by brute force search and semi-brute force search. The model generation unit 14 generates an estimation model (BFS) based on all combinations of feature types up to a predetermined type of feature type by brute force search. The model generation unit 14 generates a BFS based on, for example, one type of feature type. １ and BFS is generated based on two types of features. ２ and BFS is generated based on three types of features. ３ and BFS is generated based on the four types of features. ４ The estimation accuracy calculation unit 15 generates the generated BFS １ , BFS ２ , BFS ３ , BFS ４ The combination recording unit 16 calculates the estimation accuracy of BFS. １ , BFS ２ , BFS ３ , BFS ４ The combination of feature types used to generate the １ , BFS ２ , BFS ３ , BFS ４ The combination is recorded in the combination storage unit 192 for each combination.

[0036] Thereafter, the model generation unit 14 generates an estimation model that uses more types of feature quantities than the types of feature quantities used by the already generated estimation model by a semi-brute force search. １ , BFS ２ , BFS ３ , BFS ４ When ５ At this time, the model generation unit 14 generates an SBFS based on the feature quantity of one non-overlapping combination stored in the combination storage unit 192 and the five feature quantities included in the combination of the four feature quantities. ５ The model generation unit 14 generates an SBFS based on the five types of feature quantities included in the combination of two types of feature quantities and the combination of three types of feature quantities that are stored in the combination storage unit 192 and do not overlap. ５ Here, the feature combinations stored in the combination storage unit 192 are selected in descending order of rank. ５ The estimation accuracy calculation unit 15 calculates the generated SBFS ５ The combination recording unit 16 calculates the estimation accuracy of SBFS. ５ The combination of the types of feature amounts used to generate the feature values ​​and the ranking of the estimation accuracy are recorded in the combination storage unit 192.

[0037] Although the model generating unit 14 generates an SBFS from feature quantities included in two combinations of feature quantities, it may generate an SBFS from feature quantities included in three or more combinations.

[0038] The model generator 14 then ６ The model generation unit 14 generates an SBFS based on the feature quantity of one unique combination stored in the combination storage unit 192 and six feature quantities included in the combination of five feature quantities. ６ The model generation unit 14 generates an SBFS based on the six types of feature quantities included in the combinations of the two types of feature quantities that do not overlap and the four types of feature quantities that are stored in the combination storage unit 192. ６The model generation unit 14 generates an SBFS based on six types of feature amounts stored in the combination storage unit 192 and included in two non-overlapping three-type combinations. ６ Here, the combinations of feature amounts stored in the combination storage unit 192 are selected in descending order of rank.

[0039] The model generation unit 14 generates SBFSs and the estimation accuracy calculation unit 15 calculates the estimation accuracy until the estimation accuracy of an SBFS using a predetermined number of types of feature quantities is calculated. This makes it possible to identify a combination of types of feature quantities that can generate an estimation model with high estimation accuracy.

[0040] The combination recording unit 16 may record in the combination storage unit 192 only combinations of feature types that can generate estimation models with high estimation accuracy. Here, "high estimation accuracy" refers to estimation accuracy that is higher than an arbitrarily set standard. The standard may be a number or a percentage. For example, among estimation models that use the same number of feature types, estimation models with the top 100 estimation accuracies are considered to have "high estimation accuracy." For example, among estimation models that use the same number of feature types, estimation models with estimation accuracies that fall in the top 20% are considered to have "high estimation accuracy."

[0041] The model output unit 17 outputs the estimation model with the highest estimation accuracy among the generated estimation models to the estimation device 4 described later. Here, the model output unit 17 may output other generated estimation models, such as an estimation model with high estimation accuracy, to an external device.

[0042] Here, the "estimation model with the highest estimation accuracy" may be the SBFS with the highest ranking of estimation accuracy among the SBFSs that use the largest number of types of feature quantities. Furthermore, if an estimation model is not stored in the storage unit 19, the model generation unit 14 may regenerate an estimation model based on a combination of types of feature quantities that can generate the estimation model with the highest estimation accuracy, which is stored in the combination storage unit 192.

[0043] 4 is a flowchart showing the operation of the estimation model generating device 1 according to the first embodiment. First, the measurement signal receiving unit 11 receives the myoelectric potential signal measured by the measurement device 3 (step S11). Then, the feature extracting unit 12 extracts multiple types of feature values ​​from the myoelectric potential signal (step S12). The feature recording unit 13 records the multiple types of extracted feature values ​​in the feature storage unit 191. At this time, the feature recording unit 13 may record, together with the feature values, the corresponding hand movements, subject information, whether the measured upper limb is paralyzed, and the like.

[0044] The model generation unit 14 performs a brute force search to generate a BFS １ From BFS ａ (a is a predetermined integer equal to or greater than 2) (step S14). １ From BFS ａ The combination recording unit 16 calculates the estimation accuracy of BFS (step S15). １ From BFS ａ The combination of feature types used to generate BFS １ From BFS ａ The model generating unit 14 generates an SBFS that uses one more feature quantity than the maximum number of feature quantities used by the already generated estimation models (step S17). ａ＋１ is generated.

[0045] Thereafter, the estimation accuracy calculation unit 15 calculates the estimation accuracy of the generated SBFS (step S18). The combination recording unit 16 records the combination of the types of feature quantities used to generate the SBFS and the ranking of the estimation accuracy in the SBFS in the combination storage unit 192 of the storage unit 19 (step S19).

[0046] If the SBFS using the predetermined number of types has not been generated (step S20: NO), the operation from step S17 is executed again. ａ＋１ After the generation of SBFS, in step S18 ａ＋１The estimation accuracy of SBFS is calculated in step S19. ａ＋１ The combination of feature types used to generate SBFS ａ＋１ The ranking of the estimation accuracy in the combination storage unit 192 is recorded in the combination storage unit 192. After that, if the result of step S20 is NO, step S17 is performed to ａ＋２ is generated.

[0047] If an SBFS using a predetermined number of types has been generated (step S20: YES), the model output unit 17 outputs the estimation model with the highest estimation accuracy among the generated estimation models (step S21). The model output unit 17 can output the estimation model with the highest estimation accuracy by referring to the estimation accuracy rankings recorded in the combination storage unit 192 and identifying the feature combination with the highest estimation accuracy. The model generation unit 14 may record the generated estimation model in the storage unit 19 along with information on the combination of feature types used. In this case, the model output unit 17 can search the storage unit 19 for an estimation model corresponding to the identified feature combination with the highest estimation accuracy ranking, and output the estimation model with the highest estimation accuracy. If no generated estimation model is recorded in the storage unit 19, the model output unit 17 may identify the feature combination with the highest estimation accuracy, and then the model generation unit 14 may generate an estimation model using the identified feature combination, and the model output unit 17 may output the generated estimation model.

[0048] The combination recording unit 16 is a BFS １ From BFS ａ The combination of feature types used to generate BFS １ From BFS ａThe combination recording unit 16 may record the combination of feature types used to generate the SBFS and the estimation accuracy in the SBFS in the combination storage unit 192 of the storage unit 19. In this case, when the model generation unit 14 generates the SBFS, it determines the combination of feature quantities to use by comparing the estimation accuracies recorded in the combination storage unit 192. In this case, the model output unit 17 identifies the combination of feature quantities that results in the highest estimation accuracy by comparing the estimation accuracies recorded in the combination storage unit 192, and outputs the estimation model generated using the combination of feature quantities that results in the highest estimation accuracy.

[0049] The model generation unit 14 may terminate generation of the estimation model when an SBFS with an estimation accuracy equal to or greater than a predetermined value is generated, and the model output unit 17 may output the estimation model with the highest estimation accuracy. Alternatively, the model generation unit 14 may terminate generation of the estimation model when an improvement in estimation accuracy cannot be achieved even if the number of types of feature quantities is increased, and the model output unit 17 may output the estimation model with the highest estimation accuracy.

[0050] As described above, the estimation model generation device 1 can identify combinations of feature types that can generate estimation models with high estimation accuracy and generate estimation models with a large number of feature types. As shown in formula (1), as the number of feature types (q) used to generate estimation models increases, the number of combinations of feature types increases explosively. In the first embodiment, the model generation unit 14 performs a semi-brute-force search and generates an estimation model (SBFS) by referring to combinations of feature types that generate estimation models with high estimation accuracy, thereby reducing the amount of calculation.

[0051] The feature amount recording unit 13 may record the feature amount separately depending on whether the upper limb of the measurement subject is paralyzed or not. Also, the feature amount recording unit 13 may record the feature amount separately for each person to be measured.

[0052] The model generation unit 14 may generate different estimation models depending on whether the upper limb of the measurement subject is paralyzed or not, or may generate an estimation model for each person to be measured.

[0053] 5 is a diagram showing the configuration of an estimation device 4 according to the first embodiment. The estimation device 4 includes a measurement signal receiving unit 41, a feature extraction unit 42, an estimation unit 43, an estimation result output unit 44, and a storage unit 51. The storage unit 51 stores the estimation model output by the estimation model generation device 1.

[0054] The measurement signal receiving unit 41, like the measurement signal receiving unit 11, receives the myoelectric potential measured by the measuring device 3 from the upper limb 5. The feature extracting unit 42 extracts feature amounts from the myoelectric potential signal. The feature amounts extracted by the feature extracting unit 42 are feature amounts used by the estimation model stored in the storage unit 51. The feature amounts extracted by the feature extracting unit 42 are specified by the estimation model stored in the storage unit 51.

[0055] The estimation unit 43 inputs the feature quantities extracted by the feature quantity extraction unit 42 into the estimation model, thereby outputting the estimated hand movements of the upper limbs. The estimation result output unit 44 outputs the hand movements of the upper limbs as the estimated results. In this way, the estimation device 4 can estimate the movements of the upper limbs that the hands are about to perform. The outputted hand movements of the upper limbs are displayed, for example, on a VR headset or AR headset worn by the person being measured. This provides visual assistance when the person being measured is undergoing rehabilitation of their paralyzed upper limbs.

[0056] 6 is a flowchart showing the operation of the estimation device 4 according to the first embodiment. The measurement signal receiving unit 41 receives the myoelectric potential signal measured by the measurement device 3 (step S31). Thereafter, the feature extracting unit 42 extracts feature values ​​from the myoelectric potential signal (step S32). The estimation unit 43 estimates hand movements from the feature values ​​using an estimation model (step S33). The estimation result output unit 44 outputs the estimation result obtained by the estimation unit 43 (step S34).

[0057] (Experimental Results) Estimation models were generated using the estimation model generation device 1 according to the first embodiment, and their estimation accuracy was calculated. Electromyogram signals were measured for different hand movements from 19 subjects, and 127 types of feature quantities were extracted from the electromyogram signals. To calculate the accuracy of the estimation models, 10-fold cross-validation was performed, and data from 17 randomly selected subjects was used as training data, and data from two subjects was used as test data. The accuracy of the hand movements output by the estimation models was calculated as the accuracy of the estimation models. This was performed 100 times, and the average value was used as the accuracy of the estimation models.

[0058] FIG. 7 shows the estimation accuracy of the estimation model generated by the estimation model generation device 1. Estimation models using 1 to 20 types of feature quantities were generated, and the maximum estimation accuracy was shown for each number of types of feature quantities used. (a) of FIG. 7 shows the maximum estimation accuracy of the estimation model generated by the BFS using brute force search. １ From BFS ４ Generate an estimated model up to ５ From SBFS ２０ The maximum estimation accuracy of each estimation model is shown in (b) of Fig. 7. １ From BFS ３ Generate an estimated model up to ４ From SBFS ２０ The maximum estimation accuracy of each estimation model when generating estimation models up to １ From BFS ２ Generate an estimated model up to ３ From SBFS ２０ The maximum estimation accuracy of each estimation model when generating estimation models up to

[0059] In FIG. 7(a), BFS ４ The accuracy rate of SBFS is 71.8571%. ５ The accuracy rate of SBFS is 71.0000%. ２０ The accuracy rate of BFS was 75.2857%. ３ The accuracy rate of SBFS is 70.7857%.４ The accuracy rate of SBFS is 69.5000%. ２０ The accuracy rate of BFS was 76.0714%. ２ The accuracy rate of SBFS is 68.2857%. ３ The accuracy rate of SBFS is 68.7857%. ２０ The estimated accuracy was 73.1429%.

[0060] FIG. 8 is a diagram comparing the estimation accuracy of the estimation models (BFS, SBFS) generated by the estimation model generating device 1 with the estimation accuracy of a conventional estimation model (MFS). １ From MFS ８ The estimation accuracy of the eight estimation models was calculated. １ is an estimation model that inputs MAV, WL, ZC, and SSC extracted from EMG signals and outputs the hand movements of the upper limbs. ２ is an estimation model that inputs RMS and AR6 extracted from EMG signals and outputs hand movements. ３ is an estimation model that inputs LMAV and NSV extracted from EMG signals and outputs hand movements. ４ is an estimation model that inputs IEMG, VAR, WL, ZC, SSC, and WAMP extracted from EMG signals and outputs hand movements. ５ is an estimation model that inputs SSI, RSD1, RSD2, MSR, ASM, and ROG extracted from EMG signals and outputs hand movements. ６ is an estimation model that inputs ZC, SSC, AR4, PSR, DFA, and HFD extracted from EMG signals and outputs hand movements. ７ is an estimation model that inputs MAV, RMS, WL, VAR, ZC, SSC, WAMP, MMAV1, MMAV2, PSP, AR2, AR5, MDF, and MNF extracted from EMG signals and outputs upper limb hand movements. ８is an estimation model that inputs MAV, VAR, AR-4, ZC, MNF, and MDF extracted from EMG signals and outputs upper limb hand movements.

[0061] In FIG. 8, the estimation accuracy of SBFS is shown as the SBFS shown in FIG. ２０ The accuracy rate of the SBFS shown in (a) of FIG. ２０ The accuracy rate of MFS is 71.8571%. １ From MFS ８ The accuracy rate of the eight estimation models is MFS ５ From the above, it can be said that the BFS and SBFS generated in the first embodiment have improved accuracy in estimating hand movements.

[0062] Figure 9 shows the combinations of features that fall into the top 20% of estimation accuracy in each generated BFS. Among the combinations of features that fall into the top 20% of estimation accuracy, we investigated the proportions of combinations that include only time-domain features (TD), combinations that include time-domain features and other features (TD-Other), and combinations that include only non-time-domain features (Other). １ In this case, the feature combination contains only one type of feature, so the proportion of TD-Other is 0.

[0063] BFS ２ From BFS ４ Therefore, by combining time domain features with other features, it is possible to generate an estimation model with high estimation accuracy.

[0064] Second Embodiment An estimation model generation device 1 and an estimation device 4 according to a second embodiment will be described below, but descriptions of similar features to those of the estimation model generation device 1 and the estimation device 4 according to the first embodiment will be omitted. A feature recording unit 13 according to the second embodiment records extracted feature values ​​and hand movement evaluation indices for the feature values ​​in a storage unit 21. The hand movement evaluation indices are indices for evaluating the degree of paralysis, such as the Fugl-Meyer Assessment (FMA), Brunnstrom stage, Stroke Impairment Assessment Set (SIAS), and Modified Ashworth Scale (MAS). A model generation unit 14 according to the second embodiment receives feature values ​​extracted from an electromyography signal as input and generates a model that outputs a hand movement evaluation indices.

[0065] The estimation unit 43 according to the second embodiment inputs feature quantities into an estimation model to output an evaluation index of hand movement. The estimation result output unit 44 according to the second embodiment outputs the evaluation index of hand movement as an estimation result. The output evaluation index of hand movement is displayed, for example, on an external display device. This makes it possible to estimate an index of paralysis of the upper limbs of the person being measured.

[0066] When one of the two upper limbs of the person being measured is paralyzed and the other is not, the feature recorded by the feature recording unit 13 according to the second embodiment and the feature extracted by the feature extraction unit 42 according to the second embodiment may be a feature extracted from the myoelectric potential signals measured from the two upper limbs.

[0067] Third Embodiment An estimation model generation device 1 and an estimation device 4 according to a third embodiment will be described below, but descriptions of similar aspects to those of the estimation model generation device 1 and the estimation device 4 according to the first embodiment will be omitted. A feature recording unit 13 according to the third embodiment records extracted feature values ​​and values ​​indicating the level of cognitive function for the feature values ​​in a storage unit 21. The level of cognitive function is, for example, MMSE (Mini-Mental State Examination). A model generation unit 14 according to the third embodiment receives feature values ​​extracted from a myoelectric potential signal as input and generates a model that outputs a value indicating the level of cognitive function.

[0068] The estimation unit 43 according to the third embodiment inputs feature quantities into an estimation model to output a value indicating the level of cognitive function. The estimation result output unit 44 according to the third embodiment outputs a value indicating the level of cognitive function as an estimation result. The value indicating the level of cognitive function is displayed, for example, on an external display device. This makes it possible to estimate the level of cognitive function of the person being measured.

[0069] Other Embodiments One embodiment of the present invention has been described in detail above with reference to the drawings, but the specific configuration is not limited to that described above, and various design changes and the like can be made within the scope that does not deviate from the gist of the present invention.

[0070] In the above embodiment, the model generation unit 14 generates a BFS and an SBFS, but it is not necessary to generate an SBFS. In a case where an SBFS is not generated, the estimation model generation device 1 outputs the BFS with the highest estimation accuracy from among the generated BFSs to the estimation device 4.

[0071] In the above embodiment, the measurement device 3 measures myoelectric potential signals from the subject, but the measurement device 3 is not limited to myoelectric potential signals and may measure any potential detected from the subject. For example, the measurement device 3 may measure electroencephalogram signals from the subject. It is believed that hand movements can also be estimated by extracting features from the electroencephalogram signals and generating an estimation model based on the features.

[0072] The estimation model generating device 1 and the estimation device 4 in the above-described embodiments may be partially or entirely implemented by a computer. In this case, a program for implementing these functions may be recorded on a computer-readable recording medium, and the program may be loaded into a computer system and executed. Note that the term "computer system" as used herein includes an operating system (OS) and peripheral hardware. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, as well as storage devices such as hard disks built into computer systems. Furthermore, the term "computer-readable recording medium" may also include media that dynamically store programs for a short period of time, such as communication lines used when transmitting programs via networks such as the Internet or telephone lines, or media that store programs for a fixed period of time, such as volatile memory within a computer system serving as a server or client. The programs may be programs that implement some of the above-described functions, or may be programs that can be implemented in combination with programs already stored in the computer system. Furthermore, the estimation model generating device 1 and the estimation device 4 may be partially or entirely implemented using programmable logic devices such as field programmable gate arrays (FPGAs).

[0073] According to the present invention, it is possible to improve the accuracy of estimating the motion that the subject wants his or her hand to perform.

[0074] REFERENCE SIGNS LIST 1 Estimation model generation device, 11 Measurement signal receiving unit, 12 Feature extraction unit, 13 Feature recording unit, 14 Model generation unit, 15 Estimation accuracy calculation unit, 16 Combination recording unit, 17 Model output unit, 19 Memory unit, 2 Upper limb, 3 Measurement device, 4 Estimation device, 41 Measurement signal receiving unit, 42 Feature extraction unit, 43 Estimation unit, 44 Estimation result output unit, 5 Upper limb, 51 Memory unit

Claims

1. An estimation model generation device comprising: a feature extraction unit that extracts multiple types of feature values ​​from an electric potential detected from a subject; a feature recording unit that records the multiple types of feature values ​​for each hand movement instructed by the subject; a model generation unit that generates multiple estimation models that output the hand movement that the subject wants the subject to perform, from one or a combination of multiple types of feature values; and a model output unit that outputs, from the multiple generated estimation models, an estimation model with higher estimation accuracy than a predetermined standard, or an estimation model that is included in a predetermined number or a predetermined percentage of the top estimates when sorted in order of highest estimation accuracy.

2. The estimation model generation device according to claim 1, wherein the model generation unit generates the plurality of estimation models based on all combinations of a predetermined number of types of feature quantities, and generates an estimation model that uses more types of feature quantities than the predetermined number based on combinations of types of feature quantities used by an estimation model that has already been generated and has high estimation accuracy.

3. The estimation model generation device according to claim 2, wherein the model generation unit generates an estimation model that uses one more type of feature than the maximum number of types of feature used by already generated estimation models, based on a combination of types of feature used by already generated estimation models with high estimation accuracy.

4. The estimation model generating device according to any one of claims 1 to 3, wherein the feature extraction unit extracts a plurality of types of feature from a myoelectric potential signal measured from the upper limb of the subject.

5. The estimation model generating device according to claim 1 , wherein the model output unit outputs an estimation model with the highest estimation accuracy.

6. An estimation device that estimates a movement that a subject wants to make with their hand by inputting a current generated by the subject into an estimation model that takes the current generated by the subject as input and outputs the movement that the subject wants the hand to make, wherein the estimation model is generated by an estimation model generation device described in any one of claims 1 to 3.

7. An estimation model generation method comprising: a feature extraction step of extracting multiple types of feature values ​​from an electric potential detected from a subject; a feature recording step of recording the multiple types of feature values ​​for each hand movement instructed by the subject; a model generation step of generating multiple estimation models that output the hand movement that the subject wants the subject to perform from each combination of feature values ​​based on all combinations of the multiple types of feature values; and a model output step of outputting an estimation model with high estimation accuracy.

8. An estimation method for estimating an action that a subject wishes to have his or her hand performed by inputting the potential detected from the subject into an estimation model that uses the potential detected from the subject as input and outputs the action that the subject wishes the hand to perform, wherein the estimation model is generated by the estimation model generation method described in claim 7.

9. A program that causes a computer to operate as the estimation model generating device according to any one of claims 1 to 3.

10. A program for causing a computer to operate as the estimation device according to claim 6.

11. The estimation model generating device according to claim 1, wherein the model generating unit generates a plurality of the estimation models from all combinations of feature quantities for a predetermined number of types of feature quantities.

12. The estimation model generating device according to claim 11, wherein the predetermined number is an integer between 1 and 4.

13. The estimation model generating device according to claim 1, wherein the combination of features includes a combination of a feature in the time domain and another feature.

14. The estimation model generating device according to claim 1, wherein the model generating unit generates the plurality of estimation models from each combination of the plurality of types of feature quantities based on all combinations of the plurality of types of feature quantities.

Images