Motion analysis system

Abstract

To provide an analysis time section suitable for analyzing a specific motion by dividing a complicated motion with mathematically rational division time.SOLUTION: A storage part of a motion analysis system stores surface myoelectric potential time series data. A timing determination plane generation part applies a first dimension reduction method to the surface myoelectric potential time series data, reduces the dimension to a predetermined dimension number of the dimension number of the surface myoelectric potential time series data or less, and creates timing determination planes of the number of the determined dimension number with the respective acquired dimensions as a variable axis and with other axis as a temporal axis. A selection part has a function to select at least one shape feature from a shape of a function on at least one timing determination plane. An analysis part determines at least one division time with a shape feature occurrence time as a reference, acquires an analysis time section by dividing a temporal axis of the surface myoelectric potential time series data by the division time, and executes appropriate analysis for the analysis time section.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present invention relates to a motion analysis system that measures and analyzes muscle activity during exercise using surface myoelectric potential measuring electrodes. [Background technology]

[0002] Swallowing is a series of movements that send food from the mouth to the stomach, and is divided into three stages: the oral stage, the pharyngeal stage, and the esophageal stage. The swallowing mechanism is shown in Figure 1 (a) and (b). The food bolus formed by chewing is sent to the pharynx by pressing the tongue against the palate. Then, the swallowing reflex is induced, and the hyoid bone and larynx are elevated forward and upward due to contraction of the suprahyoid muscles. Furthermore, as a reflex movement associated with the contraction of the suprahyoid muscles, the infrahyoid muscles (thyrohyoid muscles) contract, causing the larynx to reach its highest position. In conjunction with this elevation of the hyoid bone and larynx, the epiglottis is inverted to close the larynx. This allows the food bolus to pass through the esophagus without aspiration or choking. This is the mechanism by which safe swallowing occurs.

[0003] In this way, swallowing is achieved by a complex mechanism in which "voluntary movement" and "reflex involuntary movement" coexist. Voluntary movement mainly refers to the movement in the mouth where the tongue moves to send the food bolus formed by chewing food into the pharynx (oral phase), while involuntary movement refers to the movement of the food bolus passing through the pharynx by the swallowing reflex (pharyngeal phase). The swallowing reflex is a highly reproducible and extremely precise pattern movement programmed by the central pattern generator (CPG) in the medulla oblongata. The applicant has previously proposed techniques for analyzing this movement as disclosed in Patent Documents 1 to 4.

[0004] Eating and swallowing are complex movements that involve the coordinated movement of many organs, and are a mixture of voluntary and involuntary movements. Ingestion and swallowing is a series of actions in which food is recognized, brought to the mouth, chewed in the oral cavity, and transported through the pharynx and esophagus to the stomach, and its basic mechanism is explained in five stages (anticipatory phase, preparatory phase, oral phase, pharyngeal phase, and esophageal phase) known as the "five-stage model." There are also other models, such as the "four-stage model" which explains it in four stages (oral preparatory phase, oral transport phase, pharyngeal phase, and esophageal phase), and the "process model" which explains the process of chewing and swallowing in four stages (stage I transport, processing, stage II transport, and swallowing). The fact that various models have been proposed symbolizes the difficulty of analyzing eating and swallowing movements.

[0005] For example, in the conventional model that divides speech into three phases (oral, pharyngeal, and esophageal), due to age and individual differences, the division of movements is not precise, and for example, the boundary between the oral and pharyngeal phases cannot be said to be accurately divided, which causes problems in analysis. In particular, because the suprahyoid muscles are involved in both the oral and pharyngeal phases, it is difficult to directly determine the boundary between the oral and pharyngeal phases from their surface myoelectric potential signals.

[0006] Furthermore, whether or not swallowing is possible, as both "voluntary movement" and "reflex involuntary movement" coexist, is heavily influenced by the presence or absence of oral motor problems and the breakdown of the swallowing reflex.

[0007] The swallowing movement induced by the swallowing reflex shows a highly reproducible movement pattern, but it is known that the swallowing movement pattern (swallowing pattern) such as the distance and speed of hyoid bone movement, the timing and time of laryngeal closure, and the opening time of the esophageal entrance change depending on the difference in the amount of food swallowed at one time (amount of food) and the physical properties of the food (viscosity, hardness, etc.). In this way, the ability to fine-tune the movement of the swallowing-related organs and change the swallowing pattern according to the swallowing conditions is considered to be swallowing reserve. This is the swallowing response to food so as not to cause choking or aspiration. However, swallowing reserve decreases due to cerebrovascular disorders, neuromuscular diseases, muscle weakness due to aging, changes in the position of the swallowing organs, delayed swallowing reflex, etc. Furthermore, if any problem occurs in swallowing due to systemic diseases, etc., food bolus may remain in the pharynx, enter the larynx, or be aspiration, which will increase the risk of aspiration pneumonia and choking. As such, analysis of swallowing movements is complicated by a mixture of difficulties, including the difficulty of accurately classifying swallowing movements, the fact that there are periods in which voluntary and involuntary movements (swallowing reflexes) occur simultaneously, making it difficult to make a precise distinction, and the difficulty of analysis due to differences in foods. Conventionally, there was a problem in that the analysis time interval was not accurately determined.

[0008] The above explanation explains the difficulty in analyzing swallowing movements. The inventor realized that if it is possible to set an analysis time interval for analyzing swallowing movements, which are very complicated movements, in detail, it can be applied to setting an analysis time in the analysis of any movement. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] JP 2019-208629 A [Patent Document 2] JP 2022-41316 A [Patent Document 3] JP 2021-142087 A [Patent Document 4] Patent Publication No. 2022-27304 Summary of the Invention [Problem to be solved by the invention]

[0010] An object of the present invention is to divide a complex movement into mathematically rational division times and to provide an analysis time interval suitable for analyzing a specific movement included in the complex movement. [Means for solving the problem]

[0011] In view of the above problems, the present invention has the following configuration. The device includes a surface myoelectric potential measuring electrode, a memory unit, a timing determination unit, and an analysis unit, the surface myoelectric potential measuring electrodes measure time-series data of a surface myoelectric potential generated in association with muscle movement; The storage unit is storing the surface myoelectric potential time series data by at least a predetermined movement included in an exercise period; the timing decision unit includes a timing decision plane generation unit and a selection unit; the timing decision plane generating unit applies a first dimension reduction method to the surface myoelectric potential time series data to reduce the dimensions to a predetermined number of dimensions equal to or less than the number of dimensions of the surface myoelectric potential time series data, and creates timing decision planes with each of the obtained dimensions as one variable axis and the other axis as a time axis, the number of which corresponds to the predetermined number of dimensions; the selection unit has a function of selecting at least one geometric feature from a shape of a function on at least one of the timing determination planes, and acquiring at least one geometric feature occurrence time at which the geometric feature occurs; The analysis unit determines at least one division time based on at least one selected geometric feature occurrence time, divides the time axis of the surface myoelectric potential time series data by the division time to obtain an analysis time interval, and performs an appropriate analysis on the analysis time interval. Effect of the Invention

[0012] The present invention provides an advantage in that an analysis time interval used for motion analysis can be determined based on a criterion having a mathematical background. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1(a) is an explanatory diagram of the mechanism of swallowing, and FIG. 1(b) is an explanatory diagram showing the suprahyoid and infrahyoid muscles. [Diagram 2] Figure 2 is an explanatory diagram of the surface myoelectric potential measurement electrodes. Figure 2(a) is an external view of the suprahyoid muscle group surface myoelectric potential measurement electrodes, and Figure 2(b) is an external view of the infrahyoid muscle group surface myoelectric potential measurement electrodes. Figure 2(c) is an external view of the surface myoelectric potential measurement electrodes attached to a subject. [Diagram 3] FIG. 3 is an explanatory diagram for explaining how surface myoelectric potential time series data is stored in the memory unit 2 from the surface myoelectric potential measuring electrodes. [Figure 4] FIG. 4 is an illustration of the image constructor used to construct the image used for analysis. [Diagram 5] The number of dimensions determined by nonnegative matrix factorization (NMF), which is a first dimensionality reduction method, is set to 2, and a process is described in which a muscle activity matrix M is created from a matrix (RMS) of a subject to be processed. Fig. 5 is an explanatory diagram illustrating the process in which nonnegative matrix factorization (NMF) is applied to the muscle activity matrix M to create a first timing determination plane and a second timing determination plane of two dimensions, which is a predetermined number of dimensions. [Figure 6] FIG. 6 is an explanatory diagram for selecting a geometric feature from the first timing decision plane and the second timing decision plane obtained in FIG. [Figure 7] FIG. 7 is an explanatory diagram for creating data including only the analysis time interval performed by the analysis unit. [Figure 8] FIG. 8 is an explanatory diagram showing the positions of the divided times 51 in the surface myoelectric potential time series data. [Figure 9] FIG. 9 is an explanatory diagram of the overall configuration of the embodiment. [Figure 10] FIG. 10 is a schematic diagram of an Alex Net type convolutional neural network. [Figure 11] FIG. 11 is an explanatory diagram showing an outline of the experiment of the embodiment. [Figure 12] FIG. 12 is an explanatory diagram of the Euclidean distance measured by the distance measurement unit. [Figure 13] Figure 13 shows the results of an experiment in which a fatigue task was performed on a young group of 12 young men with no history of dysphagia and an elderly group of 11 elderly men, in which normalized Euclidean distance was calculated. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the following description, the same reference numerals in different drawings indicate parts having the same function, and duplicated descriptions in each drawing will be omitted as appropriate. EXAMPLES

[0015] The motion analysis system A of the embodiment is a system for analyzing data related to swallowing-related muscles. The specification will explain the wide range of application of the present invention, which can be applied to motions other than swallowing motion, by explaining modified examples of performing motion analysis during other eating and swallowing while explaining the embodiment.

[0016] (Surface myoelectric potential measurement electrodes) Fig. 2 is an explanatory diagram of the surface myoelectric potential measuring electrode 1. Fig. 2(a) is an external view of the suprahyoid muscle group surface myoelectric potential measuring electrode 11, and Fig. 2(b) is an external view of the infrahyoid muscle group surface myoelectric potential measuring electrode 12. A ground electrode (GND electrode) and an insensitive electrode, which is a bipolar electrode (not shown), were placed on the left and right earlobes, and an RLD electrode (not shown) was placed on the seventh cervical vertebra. FIG. 2(c) is an external view of the subject wearing the surface myoelectric potential measurement electrode 1, with the suprahyoid muscle group surface myoelectric potential measurement electrode 11 and the infrahyoid muscle group surface myoelectric potential measurement electrode 12 attached under the chin and throat, respectively.

[0017] (Modification) The surface myoelectric potential measurement electrode 1 may measure the masseter muscle and temporalis muscle during mastication. The surface myoelectric potential measurement electrode 1 may be added to the suprahyoid muscle group surface myoelectric potential measurement electrode 11 and the infrahyoid muscle group surface myoelectric potential measurement electrode 12 to measure the masseter muscle, temporalis muscle, and sternocleidomastoid muscle that stabilizes the head. The surface myoelectric potential measurement electrode 1 may be a surface myoelectric potential measurement electrode 1 that measures only specific muscle groups, such as the suprahyoid muscle group surface myoelectric potential measurement electrode 11, the infrahyoid muscle group surface myoelectric potential measurement electrode 12, the masseter muscle surface myoelectric potential measurement electrode, the temporalis muscle surface myoelectric potential measurement electrode, and the sternocleidomastoid muscle surface myoelectric potential measurement electrode. The muscle group to be measured and the combination of the surface myoelectric potential measurement electrodes 1 vary depending on the purpose of the analysis. In the present invention, the surface myoelectric potential measuring electrodes 1 can be used for various muscle groups depending on the purpose of analysis.

[0018] (Data storage) FIG. 3 is an explanatory diagram for explaining how surface myoelectric potential time series data 21 is stored in storage unit 2 from surface myoelectric potential measuring electrodes 1. As shown in FIG. The suprahyoid muscle group surface myoelectric potential measuring electrode 11 and the infrahyoid muscle group surface myoelectric potential measuring electrode 12 are each multi-channel electrodes with 22 channels. The sampling device 13 amplifies the weak sEMG (surface myoelectric potential) of each channel by 125 times and samples it at 2000 Hz. This allows the surface myoelectric potential time series data 21, which are the suprahyoid muscle group surface myoelectric potential time series data 211 and the infrahyoid muscle group surface myoelectric potential time series data 212, to be acquired.

[0019] (Surface myoelectric potential time series data) The surface myoelectric potential time series data 21 may be not only raw potential data acquired from the surface myoelectric potential measurement electrodes 1 but also variously processed data. For example, the surface myoelectric potential time series data 21 may be wavelet transformed data or data of the RMS signal 552. The surface myoelectric potential time series data 21 means any time series data acquired from the surface myoelectric potential measurement electrodes 1. In the embodiment, a muscle activity matrix M is used in which the rows represent the RMS signal 552 and the columns represent the time axis (number of frames). The row data of the matrix can be changed depending on the research purpose of the researcher. For example, the median frequency (MDPF) may be acquired as the surface myoelectric potential time series data 21.

[0020] In the embodiment, time series data of the root mean square (RMS) RMS signal 552 is used as the rectified surface myoelectric potential time series data 21, but the surface myoelectric potential time series data 21 may be subjected to processing such as smoothing. In addition to the root mean square (RMS), rectification may be performed by rectification smoothing (ARV: Average Rectified Value) or integrated myoelectric potential (iEMG: Integrated EMG). The surface myoelectric potential time series data 21 may be normalized by the maximum value of the rectified muscle activity during maximum voluntary contraction or the amplitude of the rectified muscle activity during a specific movement to obtain the surface myoelectric potential time series data 21. The surface myoelectric potential time series data 21 in the present invention may be time series data derived from the raw surface myoelectric potential time series data 21, and may include data that has been subjected to some processing.

[0021] (Timeline) In the present invention, the "time axis" may refer to any time series, and for example, the concept of a time axis also includes cases where a movement is expressed as one cycle (for example, one cycle is defined as from standing up from a chair to the first step of walking, or from the start to the end of swallowing), and one cycle is normalized to 100%. In this embodiment, the time axis is represented by the number of frames.

[0022] (Preprocessing section) FIG. 4 is an explanatory diagram of the image constructing unit 6 used to construct an image to be used for analysis. (1) Root Mean Square (RMS) The pre-processing section 55 is a processing section for obtaining analytical values ​​to be used in the analysis. The pre-processing unit 55 of the embodiment includes an RMS calculation unit 551 that determines an RMS signal 552 . The RMS signal 552 is a feature quantity in the time domain, and includes amplitude information of the sEMG signal (surface myoelectric potential signal). The RMS calculation unit 551 obtains the RMS signal 552 (Root Mean Square), which is a feature component in the time domain, for the sEMG signal of each channel while shifting the 128 ms frame at a cycle of 16 ms. The RMS calculation unit 551 calculates the RMS signal 552 from both the suprahyoid muscle group surface myoelectric potential time series data 211 and the infrahyoid muscle group surface myoelectric potential time series data 212.

[0023] (2) Cepstrum coefficients (CC) The preprocessing unit 55 includes a cepstrum coefficient calculation unit 555. Cepstrum coefficients are features in the frequency domain, and include information on the power spectrum obtained by Fourier transforming the sEMG signal. Hereinafter, the cepstrum coefficients are abbreviated as CC values. The CC values ​​are obtained by taking logarithmic values ​​of the absolute values ​​of data obtained by discrete Fourier transforming the sEMG signal, and then performing inverse Fourier transform. A low-order CC value corresponds to a gentle fluctuation (envelope shape), and a high-order CC value corresponds to a fine fluctuation (fine structure). In the embodiment, attention is paid to the envelope shape of the power spectrum, and low-order (1st to 3rd) CC values ​​are adopted as features. The cepstrum coefficient calculation unit 555 of the embodiment calculates a 1st order CC value 556, a 2nd order CC value 557, and a 3rd order CC value 558 from both the suprahyoid muscle group surface myoelectric potential time series data 211 and the infrahyoid muscle group surface myoelectric potential time series data 212. In the embodiment, the order of the cepstrum coefficients of the surface myoelectric potential time series data 21 related to swallowing becomes flat and loses characteristics as it increases, so it is set to first to third order, but depending on the purpose of the analysis, higher order CC values ​​can be obtained and used for the analysis.

[0024] (Modification) Depending on the purpose of the analysis, the analysis value acquired by the pre-processing unit 55 may be an analysis value other than the RMS signal 552 or the CC value. For example, the analysis value may be a median frequency (MDPM). The embodiment is merely an example.

[0025] (Image Composition Department) The RMS signal 552, the tertiary CC value 558, the secondary CC value 557, and the primary CC value 556 obtained from the suprahyoid muscle group surface myoelectric potential time-series data 211 are sent to the image construction unit 6. The image construction unit 6 creates a total of four images, with the horizontal axis being the time series (182 frames) and the vertical axis being the intensities of the RMS signal 552, the tertiary CC value 558, the secondary CC value 557, and the primary CC value 556 for each channel, that is, an RMS image 61, a primary CC image 62, a secondary CC image 63, and a tertiary CC image 64. These four images are linked vertically by the image construction unit 6 to create a suprahyoid muscle group composite image 65. Similarly, the RMS signal 552, the tertiary CC value 558, the secondary CC value 557, and the primary CC value 556 obtained from the infrahyoid muscle group surface myoelectric potential time-series data 212 are sent to the image construction unit 6. The image construction unit 6 creates a total of four images, with the horizontal axis being the time series (182 frames) and the vertical axis being the intensities of the RMS signal 552, the tertiary CC value 558, the secondary CC value 557, and the primary CC value 556 for each channel, that is, an RMS image 61, a primary CC image 62, a secondary CC image 63, and a tertiary CC image 64. These four images are linked vertically by the image construction unit 6 to create an infrahyoid muscle group composite image 66.

[0026] In this embodiment, the RMS signal 552, the tertiary CC value 558, the secondary CC value 557, and the primary CC value 556 of the suprahyoid muscle group composite image 65 and the infrahyoid muscle group composite image 66 are arranged in the order shown in Fig. 4. In the analysis by the neural network 7 described later, the order is not important, so the arrangement can be arbitrary.

[0027] The suprahyoid muscle group composite image 65 and the infrahyoid muscle group composite image 66 include all data during the eating and swallowing period. However, with all data during the eating and swallowing period in its original form, it is difficult to analyze only the muscle groups related to voluntary movement or only the involuntary movement (swallowing reflex). When analyzing voluntary movement, it is necessary to select a time interval during which only voluntary movement is performed or a time interval during which voluntary movement is mainly performed as an analysis time interval for analysis. Swallowing is not uniform depending on age and personality, from children to the elderly, whether the food is soft or hard, large or small, and some people swallow slowly and some quickly. The inventor has discovered a mathematically rational method for dividing the analysis time interval for swallowing-related movements.

[0028] (The first dimensionality reduction method used for timing decisions) The first dimensionality reduction method converts data from a high-dimensional space to a low-dimensional space while attempting to obtain some meaningful characteristics of the original data from the low-dimensional space. There are various first dimensionality reduction methods, such as principal component analysis (PCA), nonnegative matrix factorization (NMF), singular value decomposition (SVD), and independent component analysis (ICA), and an appropriate method can be selected depending on the object to be analyzed. Among them, since muscles exert force only in the contraction direction, non-negative matrix factorization (NMF), which handles only positive signals and associates them with the contraction direction of muscles, meets the analysis objective of the present invention. Therefore, the inventors selected non-negative matrix factorization (NMF) as the first dimensionality reduction method in the embodiment. Note that non-negative matrix factorization (NMF) is known for its ease of interpretation due to non-negativity constraints and for its theoretical support as an unsupervised learning model.

[0029] Nonnegative matrix factorization (NMF) can predetermine the number of dimensions to be ultimately reduced. Fig. 5 illustrates a process for creating a muscle activity matrix M from a matrix (RMS6111) of a subject to be processed, with the number of dimensions set to 2 by nonnegative matrix factorization (NMF), which is a first dimensionality reduction method. Fig. 5 is an explanatory diagram illustrating a process for applying nonnegative matrix factorization (NMF) to muscle activity matrix M to create a first timing determination plane 321 and a second timing determination plane 322 of two dimensions, which is a predetermined number of dimensions.

[0030] First, the following matrices (1) and (2) are created. (1) The 22-channel data included in the suprahyoid muscle group surface myoelectric potential time-series data 211 is sent to the RMS calculation unit 551 in the pre-processing unit 55. Then, the RMS calculation unit 551 outputs a matrix in which the RMS signals 552 of the suprahyoid muscles for 22 channels are arranged in chronological order. (2) The 22-channel data included in the infrahyoid muscle group surface myoelectric potential time-series data 212 is sent to an RMS calculation unit 551 in the pre-processing unit 55 . Then, the RMS calculation unit 551 outputs a matrix in which the RMS signals 552 of the infrahyoid muscles for 22 channels are arranged in chronological order.

[0031] In reality, since they are treated as matrices, they do not need to be imaged; however, in FIG. 5, for the purpose of visualizing the explanation, matrix (1) is represented as suprahyoid muscle group RMS image 611, and matrix (2) is represented as infrahyoid muscle group RMS image 612. The matrix to which the first dimensionality reduction method, nonnegative matrix factorization (NMF), is applied is the muscle activity matrix M, which is a total of 44 channels × 182 frames, obtained by concatenating matrices (1) and (2). For the sake of explanation, Figure 5 visualizes the muscle activity matrix M as RMS6111.

[0032] (Timing decision unit) The timing determination unit 3 applies non-negative matrix factorization (NMF) with a predetermined number of dimensions set to 2 to the muscle activity matrix M, thereby compressing the dimensions to two. The predetermined number of dimensions can be appropriately determined taking into consideration the complexity of the movement, etc. Then, the timing decision plane generator 31 creates a timing decision plane 32 with each of the compressed two dimensions as the vertical axis. The two created timing decision planes 32 are a first timing decision plane 321 (32) and a second timing decision plane 322 (32).

[0033] FIG. 6 is an explanatory diagram of selecting a geometric feature from the first timing decision plane 321 and the second timing decision plane 322 obtained in FIG. 6, the predetermined number of dimensions is set to two, so two timing decision planes 32 are created: a first timing decision plane 321 and a second timing decision plane 322. The selection unit 4 displays the two timing decision planes 32 in a manner that allows comparison. A first timing decision function 33 is shown on the first timing decision plane 321, which has the first dimension as the vertical axis and the time axis as the horizontal axis. A second timing decision function 34 is shown on the second timing decision plane 322, which has the second dimension as the vertical axis and the time axis as the horizontal axis. It should be noted that the time axis of the timing determination plane 32 shown in FIG. 6 is not in units of seconds, so that it is referred to as time axis number 40, for example, according to the numbers assigned to the time axis. As will be explained in the section (shape feature occurrence time) below, the time axis need only be in chronological order and does not have to be in general units of time.

[0034] (predetermined number of dimensions) The number of predetermined dimensions is appropriately determined by the researcher using the functions of the timing determination unit 3 and the selection unit 4, taking into consideration the purpose of the analysis. Since the maximum number of dimensions is the number of electrodes, the researcher can appropriately determine the predetermined number of dimensions within the range of 2 to the number of electrodes. Taking into consideration ease of analysis, etc., the predetermined number of dimensions will generally be 2 to several dimensions.

[0035] The function on the timing decision plane 32 obtained by the first dimension reduction method contains information resulting from some muscle activity. Since the first timing decision plane 321(32) and the second timing decision plane 322(32) have different dimensions, the information contained therein is also different. Furthermore, the shape of the first timing decision function 33 on the first timing decision plane 321 and the shape features 41 of the second timing decision function 34 on the second timing decision plane 322 contain information resulting from any changes in muscle activity. It is up to the researcher to clarify what muscle activity and changes are associated with these dimensions and the geometric feature 41. For example, the first timing decision function 33 of the first timing decision plane 321 has a geometric feature 41 (shoulder) whose slope changes around time axis number 80. If the researcher wants to analyze this part, he or she can use the selection unit 4 to select the beginning and end of the shoulder as the geometric feature 41, thereby obtaining the geometric feature occurrence time 42 (around time axis number 70) at which the geometric feature 41 begins and the geometric feature occurrence time 42 (around time axis number 83) at which the geometric feature 41 ends. The geometric feature occurrence time 42 can be used as it is as the division time 51 for dividing the analysis time interval. The researcher can select the range from time axis number 70 to time axis number 83 as the analysis time interval. The researchers then examine the relationship between the analysis time interval and the meaning of muscle movement. Such an analysis time interval is set by the analysis unit 5, which will be described later, and it is possible to analyze what kind of change in muscle activity has occurred between the occurrence of two adjacent geometric features 41.

[0036] (Time to appearance of shape features) As explained in the above section (Time axis), the "time" in the present invention may be any time series, for example, in rehabilitation, when the movement of standing up is decomposed into several phases and the movement is expressed as one cycle and normalized for analysis, each phase becomes an individual time. Each phase is still time series data and is included in the concept of time in the present invention. The geometric feature occurrence time 42 is similar to this, and if the time axis is configured in phase units, the phase in which the geometric feature 41 occurs becomes the geometric feature occurrence time 42.

[0037] Furthermore, the researcher can use the selection unit 4 to select a particular geometric feature 41 from among the geometric features 41, such as a peak 411, a start point, an end point, a rise, a fall, and including the first, second and third derivatives of a function. The first timing determination plane 321 and the second timing determination plane 322 are displayed on the display, and the researcher can specify an appropriate position with a mouse using the function of the selection unit 4 described later to specify the division time 51. The researcher can also specify an arbitrary position between two geometric features 41, such as between a specific geometric feature 41 and another geometric feature 41, as the division time 51 as necessary. For example, a researcher can select two division times 51 to divide the analysis time interval into three analysis time intervals, early, middle, and late, and compare and analyze the changes that occur in these three analysis time intervals. The division time 51 can be set to 1 or more depending on the purpose of the experiment.

[0038] Researchers can then study muscle activity by comparing the activity before and after the occurrence of the selected shape feature 41. In addition, muscle activity can be studied by setting an analysis time interval of several frames around the selected shape feature 41 (division time 51) as the center. The present invention makes it possible to analyze the meaning and changes of specific muscle activity from within a complex swallowing movement.

[0039] The inventor's previous research has revealed that the time interval up to the peak 411 of the first timing decision function 33 reduced to two dimensions by nonnegative matrix factorization (NMF) is the time interval during which voluntary movement is predominant.

[0040] As described above, the timing decision unit 3 creates two timing decision planes 32, and the selection unit 4 displays the two timing decision planes 32, prompting the researcher to select a geometric feature 41. In the experiment described below, a peak 411 was selected as the geometric feature 41. The peak 411 on the first timing decision plane 321 occurs earlier than the peak 411 on the second timing decision plane 322. In order to select the geometric feature occurrence time 42 using the selection unit 4, the researcher displayed the first timing decision plane 321 and the second timing decision plane 322 so that they could be compared, and determined that the "earliest peak time 421" was the division time 51 that divides the analysis time interval.

[0041] The section from time axis 0 to time axis 63, which is the earliest peak time 421, is called a previous analysis time interval 512. The section from time axis 63 to time axis 82 is called a later analysis time interval 513.

[0042] Furthermore, previous research has revealed that a small amount of involuntary movement (reflex swallowing) is mixed in at the earliest peak time 421, which is the time when the first timing decision function 33 reaches its peak 411. Therefore, if one wishes to analyze purely voluntary movement, one can analyze purely voluntary movement by defining a time slightly before the earliest peak time 421 as the divided time 51 and setting it as the analysis time interval. In this manner, the researcher can use the timing determination unit 3 and the selection unit 4 to determine a division time 51 that divides the analysis time interval around one shape feature 41 according to the purpose of the analysis. A later analysis time interval 513 can be used to analyze involuntary movements (reflex swallowing).

[0043] Conventionally, the determination of the analysis time interval has been largely based on the researcher's experience, and it has not been possible to determine an analysis time interval such as one in which the above-mentioned two adjacent geometric features 41 occur. The present invention can determine an analysis time interval with a mathematical background, and therefore can contribute to the analysis of complex swallowing movements.

[0044] (Analysis Department) The analysis unit 5 of the embodiment includes a preprocessing unit 55, an image construction unit 6, a neural network 7, and a distance measurement unit 8. The analysis unit 5 determines an analysis time interval, extracts the analysis time interval from the time series data to be processed, and outputs a final analysis result.

[0045] (Creating data for the analysis time interval) FIG. 7 is an explanatory diagram for creating data including only the analysis time interval performed by the analysis unit 5. In FIG. FIG. 7 is also an explanatory diagram of a process of increasing analysis data by inversion processing 67 in order to compensate for the muscle activity data that has been reduced by using only the previous analysis time interval 512 in divided time 51 as the analysis time interval. As previously mentioned, the earliest peak time 421 selected using the timing decision plane 32 was selected as the split time 51 . The division time 51 is approximately 63 on the time axis. Using the division time 51, the suprahyoid muscle group combined image 65 and the infrahyoid muscle group combined image 66 are cropped to an image of only the previous analysis time interval 512. To increase the analysis data, the image of only the previous analysis time interval 512 is inverted. At that time, the division time 51 is adjusted so that it is exactly located in the center.

[0046] The surface myoelectric potential time series data 21 was sampled for a sufficiently long time so that the number of samples included from the start point of the previous analysis time interval 512 to the divided time 51 is 3000. If the number of samples included in the previous analysis time interval 512 is less than 3000, a no-signal interval may be supplemented toward the start point. In particular, when the analysis time interval to be analyzed is short, such as between two shape features 41, the no-signal interval is supplemented to improve the accuracy of the subsequent analysis by the neural network 7.

[0047] Next, a combining process 68 is performed, in which the inverted inversion processed image 671 and the image only from the previous analysis time interval 512 are linked to generate a combined processed image 681. This creates a combined processed image 681 that is symmetrical with respect to the axis of the division time 51. In the subsequent process, the combined processed image 681 of the suprahyoid muscles and the combined processed image 681 of the infrahyoid muscles are subjected to extraction of feature vectors by the neural network 7. The combining process 68 is a process performed to increase the amount of analysis data, and is not essential. Subsequent analyses are possible without the combining process 68, so the inversion process 67 and the combining process 68 are one of the preferred modes.

[0048] 8 is an explanatory diagram showing the position of the division time 51 in the surface myoelectric potential time series data 21. The upper graph is the first timing decision function 33 plotted on the first timing decision plane 321, and the lower graph shows the surface myoelectric potential time series data 21 (sEMG signal). FIG. 8 shows how a previous analysis time interval 512 and a subsequent analysis time interval 513 are divided by setting a peak 411 of the first timing decision function 33 as the division time 51. FIG. 8 shows that the surface myoelectric potential time series data 21 (sEMG signal) is complex and has no particular features at the point corresponding to the division time 51, making it difficult to set the division time 51 using only the surface myoelectric potential time series data 21 (sEMG signal). In this way, the present invention makes it possible to set an analysis time interval with a mathematical background.

[0049] (Summary of the process so far and overall structure) FIG. 9 is an explanatory diagram of the overall configuration of the embodiment. The inventor will summarize the configurations of the embodiments described so far and provide an overview of the analysis to be described below.

[0050] The researcher prepares a surface EMG measuring electrode 1 and measures a surface EMG (sEMG) signal using a suprahyoid muscle group surface EMG measuring electrode 11 and an infrahyoid muscle group surface EMG measuring electrode 12 . The measured surface myoelectric potential (sEMG) signal is sent to a preprocessing unit 55, which calculates an RMS signal 552, a first order CC value 556, a second order CC value 557, and a third order CC value 558. The image construction unit 6 creates a total of four images, with the horizontal axis representing the time axis (number of frames) and the vertical axes representing the RMS signal 552, the first order CC value 556, the second order CC value 557, and the third order CC value 558, respectively.

[0051] These four images are aligned side by side to create a suprahyoid muscle group composite image 65 and an infrahyoid muscle group composite image 66 . 9, the timing determination unit 3 then uses the RMS image 61 to reduce the dimensions with a first dimension reduction method to create a timing determination plane 32. Using the selection unit 4, the researcher can select a geometric feature 41 from the shape of the function on the timing determination plane 32 and obtain the geometric feature occurrence time 42. The time selected as a result of this selection is the division time 51.

[0052] The suprahyoid muscle group composite image 65 and the infrahyoid muscle group composite image 66 are divided by the division time 51, and in the embodiment, the previous analysis time interval 512 is selected as the analysis time interval. The suprahyoid muscle group composite image 65 and the infrahyoid muscle group composite image 66 are cut leaving the previous analysis time interval 512, and an image in which the previous analysis time interval 512 is horizontally inverted is created in an inversion process 67. The original image and the inverted image, which has been flipped left and right, of the previous analysis time interval 512 are subjected to a combining process 68 to generate a combined processed image 681 . The above is an outline of the embodiment described so far.

[0053] The analysis beyond this will be described later, but the motion analysis system A of the embodiment uses the combined processed image 681, extracts feature vectors by the neural network 7, and has a distance measurement unit 8 that measures the distance of the feature vectors.

[0054] (Neural Networks) The neural network 7 used in the embodiment is an Alex Net type convolutional neural network 7. The embodiment will be described focusing on the Alex Net type convolutional neural network 7 as an analysis method, but the feature of the present invention is the determination of the analysis time interval, and the analysis method is arbitrary.

[0055] (Alex Net type convolutional neural network) FIG. 10 is a schematic explanatory diagram of an Alex Net type convolutional neural network 7. The embodiment is an example in which a combined processed image 681 created using data from the surface myoelectric potential measuring electrodes 1 is subjected to extraction of features using an Alex Net type convolutional neural network 7. The Alex Net type neural network 7 of the embodiment is pre-trained using more than one million natural images stored in the Image Net database. No images such as the combined processed image 681 derived from the data of the surface myoelectric potential measuring electrode 1 are used in the training. If analysis is possible using an Alex Net type convolutional neural network 7 that does not use images derived from the data of the surface myoelectric potential measurement electrodes 1 for training, this means that the accuracy of analysis can be further improved by using a neural network 7 that has been trained with a sufficient number of images derived from the data of the surface myoelectric potential measurement electrodes 1. In the embodiment, an experiment was performed using an Alex Net type convolutional neural network 7 that was trained only on natural images.

[0056] As shown in Figure 10, the Alex Net type convolutional neural network7 consists of five convolutional layers, two pooling layers, two normalization layers, and three fully connected layers. Finally, the fully connected layer fc8 can classify images into 1,000 categories, such as mice, coffee cups, pencils, and dogs.

[0057] In the embodiment, two combined images 681 were created by performing an inversion process 67 and a combination process 68 on the suprahyoid muscle group combined image 65 and the infrahyoid muscle group combined image 66 created in FIG. In the embodiment, the feature vector is acquired from fc6. When these two combined images 681 are input, the Alex Net type convolutional neural network 7 outputs a 4096-dimensional suprahyoid muscle group feature vector 75 and a 4096-dimensional infrahyoid muscle group feature vector 76 from fc6, which is the fully connected layer. The 4096-dimensional suprahyoid muscle group feature vector 75 and the 4096-dimensional infrahyoid muscle group feature vector 76 are concatenated to generate an 8192-dimensional total feature vector 77.

[0058] The previous analysis time interval 512 is a time interval in which voluntary movements are predominant. The total feature vector 77 extracted from the combined processed image 681 created using the previous analysis time interval 512 includes features resulting from voluntary movements during swallowing.

[0059] (analysis) The analysis is performed by examining the difference between the total feature vector 77 during a swallowing movement to be compared and the total feature vector 77 during a swallowing movement to be analyzed. One example of an analytical method for examining differences is to quantify the similarity between the 8,192-dimensional total feature vector77 of the comparison target and the analysis target.

[0060] The swallowing function evaluation method using the similarity of swallowing pattern images according to the embodiment calculates all feature vectors 77, calculates angles and distances of all feature vectors 77 under different swallowing conditions, and quantifies the similarity of the images. The similarity can be cosine similarity, Euclidean distance, Hamming distance, Mahalanobis distance, Manhattan distance, Pearson's correlation coefficient, Chebyshev distance, Minkowski distance, etc. In addition, there are also image similarity evaluation methods such as MSE (Mean Square Error), PSNR (Peak Signal to Noise Ratio), and SSIM (structural similarity index measure). The evaluation process is a process of creating an index for evaluating swallowing function from the image similarity and evaluating the swallowing function of each individual. For this reason, it is sufficient to measure the subject several times under a specified swallowing condition, and a safe and simple swallowing function evaluation method can be provided that can quantitatively and routinely grasp changes in swallowing reserve without using machine learning.

[0061] (Second dimensionality reduction method) Too many features can lead to poor accuracy, so it is preferable to use the second dimensionality reduction method to reduce the number of dimensions of all feature vectors 77 and to quantify the similarity. In the embodiment, kernel principal component analysis (KPCA) is adopted as the second dimensionality reduction method. The kernel principal component analysis (KPCA) adopted in the embodiment maps all 8192-dimensional feature vectors77 into a linearly separable high-dimensional space, and then performs principal component analysis in that space to compress the dimensions into an easily handled three-dimensional feature vector. Unlike principal component analysis (PCA), kernel principal component analysis (KPCA) is capable of handling nonlinear features.

[0062] All feature vectors 77 of the comparison target and analysis target in the embodiment are converted into dimensionality-reduced feature vectors compressed to three dimensions by the second dimensionality reduction method, kernel principal component analysis (KPCA). Then, the normalized Euclidean distance between the dimensionality-reduced feature vectors of the comparison target and analysis target is calculated and treated as the similarity. Note that the kernel principal component analysis (KPCA) is merely one example of the present invention for calculating the similarity. Depending on the purpose of analysis, there are suitable similarity measures, and researchers can select an appropriate similarity measure.

[0063] As with the first dimension reduction method, various dimension reduction methods can be selected for the second dimension reduction method according to the purpose of analysis. Both the first and second dimension reduction methods can be various, such as PCA: principal component analysis, FA: factor analysis, ICA: independent component analysis, NMF: nonnegative matrix factorization, ICAPCA: ICA applied to the subspace defined by PCA, pICA: probabilistic ICA with nonnegativity constraints, etc. Furthermore, the dimensionality reduction method included in the present invention also includes a method of compressing the dimensionality using an artificial neural network such as an autoencoder.

[0064] (experiment) FIG. 11 is an explanatory diagram showing an outline of the experiment of the embodiment. The inventor conducted an experiment to investigate fatigue during swallowing from all of the feature vectors 77 described above. Fatigue was investigated by focusing on the tongue's movement of the food bolus during swallowing. The force with which the tongue presses against the palate during swallowing is an important factor in efficiently and safely moving the food bolus into the pharynx. The fatigue task performed below is a task that artificially tires the tongue. The experiment performed the fatigue task using a tongue pressure measuring device, which will be described later.

[0065] (Compare and contrast) Subjects were instructed to consciously swallow 10 ml of cold water as forcefully as possible. Hereinafter, this "act of swallowing as hard as possible" will be referred to as "effortful swallowing." The effortful swallowing test was initiated by injecting 10 ml of cold water sample into the floor of the mouth, and the subject was asked to remain silent for 3 seconds and then swallow within 2 seconds. The comparison controls are the time series data 211 of the surface myoelectric potential of the suprahyoid muscles and the time series data 212 of the surface myoelectric potential of the infrahyoid muscles during this effortful swallowing.

[0066] (Data normalization) Since the total feature vector 77 differs depending on the subject, the inventor performed the following normalization in the embodiment. Before the actual measurement, the subject performs a standard swallowing action (in the embodiment, swallowing 1 ml of cold water as usual). Then, the total feature vector 77 to be evaluated for each individual is normalized so that the magnitude (norm) of the total feature vector 77 extracted from the muscle activity at that time becomes 1. Normalization allows comparison across subjects with different anterior neck sizes and different muscle locations relative to electrode locations.

[0067] (Tongue pressure measuring device) The tongue pressure measuring device (JMS Corporation TPM-01) is a device that measures the pressure when a balloon inserted in the oral cavity is pressed by the tongue. In this example, the tongue pressure measuring device was used not only to measure tongue pressure, but also in a fatigue task that fatigues the tongue.

[0068] (Measurement of maximum tongue pressure) Because maximum tongue pressure differs from person to person, the subjects were asked to press a balloon with their tongue against the palate as hard as they could using a tongue pressure measuring device three times, and the maximum value was recorded as the maximum tongue pressure. For example, if the first measurement was 39.8 kPa, the second measurement was 40.5 kPa, and the third measurement was 39.1 kPa, the maximum tongue pressure would be 40.5 kPa, which was measured the second time.

[0069] (Fatigue Task) The fatigue task required subjects to maintain a tongue pressure of 60% of their maximum tongue pressure for 30 seconds. In the above example, 40.5kPa x 0.6 = 24.3kPa The subjects were instructed by the researchers to maintain the tongue pressure of 24.3 kPa for 30 seconds. The fatigue task consisted of the subjects pushing the balloon with their tongue and trying to maintain the tongue pressure of 24.3 kPa output by the tongue pressure measuring device for 30 seconds.

[0070] After that, the subjects were instructed to perform a fatigue task and then to swallow 10 ml of cold water with effort. That is, the subject was instructed to swallow 10 ml of cold water as a sample as much as possible within 2 seconds after the fatigue task, after which 10 ml of cold water was injected into the floor of the mouth and the subject was asked to remain still for 3 seconds. At that time, time series data 211 of the surface myoelectric potential of the suprahyoid muscles and time series data 212 of the surface myoelectric potential of the infrahyoid muscles are acquired and become experimental data during swallowing after the fatigue task.

[0071] (Measurement of electrical potential and creation of composite images of suprahyoid muscles and infrahyoid muscles) 11, the measured suprahyoid muscle group surface myoelectric potential time series data 211 and infrahyoid muscle group surface myoelectric potential time series data 212 are subjected to the above-mentioned processing in the pre-processing unit 55 to generate an RMS image 61, a primary CC image 62, a secondary CC image 63, and a tertiary CC image 64, which are then linked to generate a suprahyoid muscle group composite image 65 and an infrahyoid muscle group composite image 66. A total of four images are generated: a suprahyoid muscle group composite image 65 and an infrahyoid muscle group composite image 66 during normal swallowing as a control, and a suprahyoid muscle group composite image 65 and an infrahyoid muscle group composite image 66 during swallowing after a fatigue task.

[0072] (Timing decision unit) Details of the timing determination unit 3 are not shown in FIG. 11, but as described above, the timing determination unit 3 derives the division time 51 using nonnegative matrix factorization (NMF), which is a first dimensionality reduction method. The effortful swallow before the fatigue task to be compared and the effortful swallow after the fatigue task in the experiment are different swallowing movements. The timing determination unit 3 performs nonnegative matrix factorization (NMF) for each effortful swallow to derive the division time 51.

[0073] (Determining the analysis time interval) Although not shown, the inventor can use the selection unit 4 to select the shape feature occurrence time 42 by comparing the times of the peak 411 on the first timing decision plane 321 and the second timing decision plane 322, and selecting the peak 411 on the first timing decision plane 321 as the earliest peak time 421. Then, the earliest peak time 421 was used as the division time 51 for dividing the analysis time interval. As described above, it is known that voluntary movement is the main period during the time period from the start of swallowing to the earliest peak time 421. Therefore, in the experiment, the previous analysis time period 512 was used as the analysis time period suitable for analyzing voluntary movement.

[0074] (Inversion and Combination) 11, the suprahyoid muscle group composite image 65 and the infrahyoid muscle group composite image 66 were subjected to an inversion process 67 and a combination process 68. This process created a bilaterally symmetrical combined processed image 681 with the division time 51 as the center of the horizontal axis.

[0075] The combined processed image 681 is subjected to calculation of total feature vectors 77 for both the comparison subject and the analysis subject using an Alex Net type convolutional neural network 7.

[0076] (Detection of fatigue) Fatigue detection is possible by comparing the total feature vector77 (8192 dimensions) of the control before the fatigue task with the total feature vector77 (8192 dimensions) of the subject after the fatigue task.

[0077] (Second dimensionality reduction method) In the embodiment, kernel principal component analysis (KPCA) is adopted as the second dimensionality reduction method. The 8192-dimensional full feature vector 77 was converted into a dimension-reduced full feature vector compressed to an easier-to-manage three dimensions by kernel principal component analysis (KPCA).

[0078] (Distance measurement section) FIG. 12 is an explanatory diagram of the Euclidean distance measured by the distance measurement unit 8. As shown in FIG. The distance measurement unit 8 calculates the Euclidean distance between the swallowing dimension compressed total feature vector 781 before the fatigue task and the swallowing dimension compressed total feature vector 782 after the fatigue task, and obtains the similarity. In this experiment to examine fatigue, when the similarity is low, i.e., when the distance is large, it is interpreted that the influence of fatigue is appearing in voluntary movement. Conversely, when the similarity is high, i.e., when the distance is small, it is interpreted that the influence of fatigue is not appearing in voluntary movement.

[0079] (Experimental Results) Figure 13 shows the results of an experiment in which a fatigue task was performed on a young group of 12 young men with no history of dysphagia and an elderly group of 11 elderly men, in which normalized Euclidean distance was calculated. This experiment was conducted with the approval of the Iwate University Medical Research Ethics Committee for Human Subjects (No. 202016).

[0080] (Statistical processing) A t test (significance; * P < 0.05) was used to compare results between the younger and older groups.

[0081] As a result, a significant difference was found between the young and elderly groups. As mentioned above, it was found that the elderly group was more susceptible to the effects of fatigue tasks than the young group. The neural network 7 used in this experiment is an Alex Net type convolutional neural network 7 that does not use images derived from the data of the surface myoelectric potential measurement electrodes 1 for learning. It is clear that more detailed analysis results can be obtained by using an Alex Net type convolutional neural network 7 that uses images derived from the data of the surface myoelectric potential measurement electrodes 1 for learning.

[0082] Although the experimental data is not shown, the analysis was performed in the post-analysis time period 513, during which involuntary movements (swallowing reflex) were predominant. The distance between the analysis results before and after the fatigue task was measured as the degree of similarity. The difference in the degree of similarity between the young and elderly groups was small and not significant. It is premature to interpret that the tongue-fatiguing task does not affect involuntary movements (swallowing reflex). It should be noted that the experiment was conducted using an Alex Net-type convolutional neural network7 trained only with natural images. It is clear that the neural network7 can output a more accurate total feature vector77 if it is trained using potential images during swallowing movements. Research on involuntary movements (swallowing reflex) will progress as the accuracy of neural networks 7 improves.

[0083] (Examples of other swallowing applications) It has been found that the latter analysis time section 513 is a section in which involuntary movement (reflex swallowing) is predominant. A decline in the function of involuntary movement (reflex swallowing) increases the risk of aspiration and choking. If a decline in the function of involuntary movement (reflex swallowing) is detected during a normal meal, the caregiver needs to change to an dysphagia diet that is easier to swallow. Analyzing the latter analysis time section 513 in which involuntary movement (reflex swallowing) is predominant is useful for knowing the timing to change to an dysphagia diet that is easier to swallow. In addition, by comparing all feature vectors77 of elderly people before, during, and after a meal, it is also possible to analyze how fatigue accumulates in the elderly.

[0084] (Examples of applications other than swallowing movements) As mentioned above, swallowing involves a large number of muscle groups, and is a mixture of voluntary and involuntary movements (reflex swallowing). The ability to analyze complex swallowing movements can also be applied to the analysis of various other movements, such as arm movements and cardiac movements.

[0085] Although the embodiments and modifications of the present invention have been described in detail above with reference to the drawings, the specific configuration is not limited to these embodiments and modifications, and the present invention also includes design changes and the like that do not deviate from the gist of the present invention. Furthermore, the above-described embodiments and modifications can be combined by utilizing each other's technologies, so long as there are no particular contradictions or problems in the purpose, configuration, or the like. [Explanation of symbols]

[0086] A. Motion Analysis System 1. Surface EMG Measurement Electrodes 11 Suprahyoid muscle group surface myoelectric potential measurement electrode 12 Infrahyoid muscle group surface myoelectric potential measurement electrode 13 Sampling Device 2 Storage section 21 Surface EMG Time Series Data 211 Time series data of surface myoelectric potential of suprahyoid muscles 212 Time series data of surface myoelectric potential of the infrahyoid muscles 3 Timing decision section 31 Timing decision plane generation unit 32 Timing Decision Plane 321 First Timing Decision Plane 322 Second Timing Decision Plane 33 First Timing Decision Function 34 Second Timing Decision Function 4 Selection section 411 Peak 41 Shape characteristics 42 Shape feature occurrence time 421 Earliest Peak Hour 5 Analysis section 51 Division Time 512 Previous analysis time interval Analysis time interval after 513 55 Pretreatment section 551 RMS calculation section 552 RMS signal 555 Cepstrum Coefficient Calculation Unit 556 1st CC value 557 Secondary CC value 558 3rd CC value 6 Image Composition Section 61 RMS images 611 Suprahyoid Muscles RMS Image 612 Infrahyoid muscles RMS image 6111 RMS 62 1st CC image 63 Secondary CC Images 64 3rd CC images 65 Suprahyoid Muscles Composite Image 66 Composite image of the infrahyoid muscles 67 Reverse Processing 671 Inverted image 68 Joining Process 681 Combined Processing Images 7. Neural Networks 75 Suprahyoid muscles feature vector 76 Infrahyoid muscles feature vector 77 total feature vectors 781 Swallowing dimension reduction feature vector before fatigue task 782 Swallowing dimensionality reduction feature vector after fatigue task 8 Distance measurement section

Claims

1. The device includes a surface myoelectric potential measuring electrode, a memory unit, a timing determination unit, and an analysis unit, the surface myoelectric potential measuring electrodes measure time-series data of a surface myoelectric potential generated in association with muscle movement; The storage unit is storing the surface myoelectric potential time series data by at least a predetermined movement included in an exercise period; the timing decision unit includes a timing decision plane generation unit and a selection unit; the timing decision plane generating unit applies a first dimension reduction method to the surface myoelectric potential time series data to reduce the dimensions to a predetermined number of dimensions equal to or less than the number of dimensions of the surface myoelectric potential time series data, and creates timing decision planes in which each of the obtained dimensions is a variable axis and the other axis is a time axis, the number of which corresponds to the predetermined number of dimensions; the selection unit has a function of selecting at least one geometric feature from a shape of a function on at least one of the timing determination planes, and acquiring at least one geometric feature occurrence time at which the geometric feature occurs; The analysis unit determines at least one division time based on at least one selected geometric feature occurrence time, divides the time axis of the surface myoelectric potential time series data by the division time to obtain an analysis time interval, and performs an appropriate analysis on the analysis time interval.

2. The motion analysis system according to claim 1 , wherein the timing decision plane generating unit applies the first dimensionality reduction method to time series data rectified as the surface myoelectric potential time series data.

3. 2. The motion analysis system according to claim 1, wherein the surface myoelectric potential measuring electrodes are multi-channel electrodes.

4. The motion analysis system of claim 1 , wherein the first dimensionality reduction method is non-negative matrix factorization.

5. The motion analysis system according to claim 1, characterized in that the analysis unit defines the division time as a time before a predetermined time or a predetermined time percentage from the selected geometric feature occurrence time based on the selected geometric feature occurrence time, divides the division time to obtain the analysis time interval, and performs an appropriate analysis on the analysis time interval.

6. 2. The movement analysis system according to claim 1, wherein the surface myoelectric potential measuring electrodes include at least a suprahyoid muscle group surface myoelectric potential measuring electrode and / or an infrahyoid muscle group surface myoelectric potential measuring electrode.

Images