Method and apparatus for determining muscle activation

Abstract

The present disclosure relates to a computer-implemented method comprising: receiving sensor data from one or more myogram sensors of a wearable sensor device, where the one or more myogram sensors are configured to monitor muscle activation of a subject; determining whether a first muscle of the subject has been selectively activated based on the sensor data; and providing biofeedback based on a determination of whether the first muscle has been selectively activated.

Description

Technical Field

[0001] This disclosure relates to a method and apparatus for determining muscle activation. Background Technology

[0002] Lower back pain (LBP) is a leading cause of chronic disability and pain. Secondary consequences of lower back pain include depression, pain medication abuse, sick leave at work, or unemployment.

[0003] Evidence-based guidelines recommend back exercises as a first-line treatment. These can be implemented through self-managed programs (e.g., using leaflets or videos) or through individualized supervision by a therapist. However, adherence to self-managed exercise programs is often poor, and wait times for therapist visits can be long and sessions with a therapist can be short.

[0004] It has been shown that the intensity of back exercises is directly proportional to the degree of back pain relief. Therefore, improving methods for back pain self-management may prove to be highly cost-effective compared to, for example, increasing the number of therapists. Two major barriers to back pain self-management are that people find it difficult to exercise properly (e.g., activating deep abdominal muscles rather than superficial ones) and have low motivation for such exercises, especially given the higher prevalence of depression and apathy in this population.

[0005] Back exercises involving specific activation of pelvic stabilizing muscles are more effective than non-specific exercises for treating lower back pain. However, instructing patients to selectively activate trunk muscles is challenging and has so far required manual palpation by a physical therapist or expensive, laboratory-based instruments such as ultrasound or standard EMG (i.e., involving wires connected to an amplifier not worn on the body).

[0006] Therefore, there is a need for a convenient and practical method and device for improving self-management of back pain, particularly for encouraging and supervising users to perform back exercises properly (e.g., at home). This includes helping users actively activate specific core muscle patterns associated with relieving back pain and disability. Summary of the Invention

[0007] This invention provides a more detailed description of concepts in specific embodiments. It should not be used to identify essential features of the claimed subject matter, nor to limit the scope of the claimed subject matter.

[0008] According to a first aspect of this disclosure, a computer-implemented method is provided, comprising: receiving sensor data from one or more myographic sensors of a wearable sensor device, wherein the one or more myographic sensors are configured to monitor muscle activation of a subject; determining, based on the sensor data, whether a first muscle of the subject has been selectively activated; and providing biofeedback based on the determination that the first muscle has been selectively activated.

[0009] By determining whether muscles are selectively activated, it is possible to identify whether the exercise a subject is performing is activating the intended muscle group (i.e., including the primary muscle). By providing subjects with feedback on the selective activation of the primary muscle, they can adjust how they perform their exercise to increase the activation of the intended muscle group, thereby enabling them to perform specific exercises more effectively (e.g., back exercises for relieving LBP). The behavioral effects of MMG-based feedback on selective muscle activation have shown that providing such feedback leads to improved selectivity of muscle activation in subjects.

[0010] Determining whether the subject's first muscle has been selectively activated may include: determining a calibration activation value for the first muscle during the calibration process; and determining the degree of activation of the first muscle relative to the calibration activation value. Determining whether the subject's first muscle has been selectively activated may include determining whether the degree of activation of the first muscle exceeds a threshold percentage of the calibration activation value.

[0011] The one or more myographic sensors may include one or more mechanical myographic (MMG) sensors. Unlike electromyography (EMG), MMG does not require a connection to the skin's electrical activity, avoiding the need for adhesives, gels, or shaving; and it is not sensitive to sweating. Furthermore, fewer sensors are required (EMG requires positive, negative, and ground electrodes), and signal amplification is not necessary. This makes microphone-MMG suitable for low-cost personal devices that can measure muscle activity "anytime, anywhere." In contrast, existing EMG sensors are wired and not wearable, and therefore unsuitable for use in wearable sensor devices. Therefore, the one or more myographic sensors may include one or more wireless myographic sensors. As an alternative to using MMG sensors, the one or more myographic sensors may include one or more wireless, wearable EMG sensors.

[0012] Determining whether the first muscle has been selectively activated based on the sensor data may include classifying the sensor data using one or more trained classifiers. The one or more trained classifiers may be trained using electromyography (EMG) sensor data and mixed myocardial infarction (MMG) sensor data, wherein the EMG and MMG sensor data used to train the one or more trained classifiers are collected simultaneously. Specifically, the one or more trained classifiers may be trained using ground truth data including EMG sensor data and training data including MMG sensor data corresponding to the EMG sensor data, wherein the MMG sensor data in the training data and the EMG sensor data in the ground truth data are collected simultaneously.

[0013] The one or more trained classifiers can be configured to distinguish between sensor data indicating muscle activation and sensor data indicating muscle rest. In this case, the one or more trained classifiers can be trained using EMG sensor data indicating muscle activation and EMG sensor data indicating muscle rest. Specifically, the one or more trained classifiers can be trained using: ground truth data including EMG sensor data indicating muscle activation and EMG sensor data indicating muscle rest; and training data including MMG sensor data corresponding to EMG sensor data indicating muscle activation and MMG sensor data corresponding to EMG sensor data indicating muscle rest, wherein the MMG sensor data in the training data and the EMG sensor data in the ground truth data are collected simultaneously.

[0014] The one or more trained classifiers can be configured to recognize sensor data indicating the degree of activation of the first muscle relative to the subject's second muscle. In this case, the one or more trained classifiers can be trained using EMG sensor data indicating the degree of activation of the first muscle relative to the subject's second muscle. Specifically, the one or more trained classifiers can be trained using: ground truth data including EMG sensor data indicating the degree of activation of the first muscle relative to the subject's second muscle; and training data including MMG sensor data corresponding to the EMG sensor data indicating the degree of activation of the first muscle relative to the subject's second muscle, wherein the MMG sensor data in the training data and the EMG sensor data in the ground truth data are collected simultaneously.

[0015] The one or more trained classifiers can be configured to identify the activation intensity of the first muscle. In this case, the one or more trained classifiers can be trained using EMG sensor data indicating different activation intensities of the first muscle. Specifically, the one or more trained classifiers can be trained using: ground truth data including EMG sensor data indicating different activation intensities of the first muscle; and training data including MMG sensor data corresponding to the EMG sensor data indicating different activation intensities of the first muscle, wherein the MMG sensor data in the training data and the EMG sensor data in the ground truth data are collected simultaneously.

[0016] Determining whether a subject's first muscle has been selectively activated may include determining the degree of activation of the first muscle relative to the subject's second muscle. A first myographic sensor among the one or more myographic sensors may be configured to monitor muscle activation of the first muscle, and a second myographic sensor among the one or more myographic sensors may be configured to monitor muscle activation of the second muscle.

[0017] The first muscle may be a core muscle. The first muscle may be the transverse abdominis. The first myographic sensor of the one or more myographic sensors may be configured to monitor muscle activation of the transverse abdominis. The first muscle may be the internal oblique muscle. The first myographic sensor of the one or more myographic sensors may be configured to monitor muscle activation of the internal oblique muscle.

[0018] The second muscle may be the rectus abdominis. The second myographic sensor in the one or more myographic sensors may be configured to monitor muscle activation of the rectus abdominis.

[0019] The first muscle may be the erector spinae. The first myographic sensor of the one or more myographic sensors may be configured to monitor muscle activation of the erector spinae.

[0020] The method may further include: receiving motion sensor data from one or more motion sensors configured to monitor the subject's movements; and processing the motion sensor data to determine the subject's posture.

[0021] The sensor data can be received from the wearable sensor device including the one or more myographic sensors. Providing biofeedback may include providing a command to cause the wearable sensor device to provide one or more of tactile, audio, and visual feedback to the subject.

[0022] According to a second aspect of this disclosure, a computer-readable medium is provided, including instructions that, when executed by at least one processor of a device, cause the device to perform the method of the first aspect.

[0023] According to a third aspect of this disclosure, an apparatus is provided, comprising: at least one processor; and at least one memory including computer-readable instructions, which, when executed by the at least one processor, cause the apparatus to perform the method of the first aspect. The apparatus may be the wearable sensor device.

[0024] The device may further include at least one electrical stimulation device configured to activate the first muscle of the subject. When executed by the at least one processor, the instruction may cause the device to change the electrical charge applied to the muscle and / or nerve fibers by the electrical stimulation device in response to receiving the biofeedback.

[0025] According to a fourth aspect of this disclosure, a wearable sensor device is provided, configured to be worn by a subject, the wearable sensor device comprising: one or more myographic sensors configured to monitor muscle activation of the subject; and a transmitter for transmitting sensor data from the one or more myographic sensors to at least one processor of a separate device. The one or more myographic sensors may include one or more mechanographic (MMG) sensors. The one or more myographic sensors may be configured to monitor muscle activation of one or more core muscles. The one or more core muscles may include one or more of the following: transverse abdominis, internal oblique, rectus abdominis, and erector spinae.

[0026] According to a fifth aspect of this disclosure, a computer-implemented method is provided, comprising: receiving sensor data from one or more myographic sensors of a wearable sensor device, wherein the one or more myographic sensors are configured to monitor muscle activation of a subject; predicting the subject's posture from a plurality of posture classifications based on the sensor data; and selectively providing biofeedback based on the posture classifications.

[0027] Predicting the subject's posture based on the sensor data may include classifying the sensor data using one or more trained classifiers. The one or more trained classifiers may be trained using ground truth posture data and MMG sensor data, wherein the ground truth posture data and MMG sensor data used to train the one or more trained classifiers are collected simultaneously. Specifically, the posture may be predicted by a classifier trained using ground truth data regarding whether the subject's sitting posture is slouching ("poor") or upright ("good"), in which case the multiple posture classifications consist of slouching and upright sitting postures. The classifier may be trained using training data comprising features extracted from TA / IO and / or ES MMG sensor data associated with the time periods during which each subject was in a slouching or upright posture.

[0028] Selectively providing biofeedback based on the posture classification may include: not providing biofeedback in response to determining that the subject's posture is classified as upright (or "good"); and providing a reminder or nudge to the subject in response to determining that the subject's posture is classified as slouching (or "bad").

[0029] According to a sixth aspect of this disclosure, a computer-implemented method is provided, comprising: receiving sensor data from one or more myographic sensors of a wearable sensor device, wherein the one or more myographic sensors are configured to monitor muscle activation of a subject; predicting, based on the sensor data, a level of pain experienced by the subject during muscle activation; and providing biofeedback based on the predicted level of pain.

[0030] Predicting the pain level experienced by the subject based on the sensor data may include classifying the sensor data using one or more trained classifiers. The one or more trained classifiers may be trained using ground truth pain level data and MMG sensor data, wherein the ground truth pain level data and MMG sensor data used to train the one or more trained classifiers are collected simultaneously. The ground truth pain level data may be a binary pain classification. The binary pain classification may indicate a diagnosis of the presence or absence of back pain. The ground truth pain level data may be a graded pain score provided by the subject. The classifier may be trained using training data comprising features extracted from TA / IO and / or ES MMG sensor data associated with the time period during which each subject experienced a specific pain level. The classifier may also be trained using training data comprising features associated with one or more exercises performed by the subject during muscle activation.

[0031] Providing biofeedback based on the classification of pain levels may include: in response to determining that the subject's pain level is classified as low, providing no biofeedback or providing biofeedback to continue a specific exercise; and in response to determining that the subject's pain level is classified as high, providing the subject with instructions to adjust a specific exercise.

[0032] The method may further include: tracking pain levels input by the subject over time; identifying the correlation between the input pain levels and one or more features extracted from sensor data and / or associated with one or more exercises; and providing the subject with suggestions for performing exercises associated with the reduced pain levels using a reinforcement learning-based model. The method may further include receiving a subjective pain level from the subject after exercising according to the suggestions. If the subjective pain level received from the subject is lower than a previously received subjective pain level from the subject, the method may further include providing a positive reward label to positively weight the suggestions. If the subjective pain level received from the subject is higher than a previously received subjective pain level from the subject, the method may further include providing a negative reward label to negatively weight the suggestions.

[0033] According to a seventh aspect of this disclosure, a computer-readable medium is provided, including instructions that, when executed by at least one processor of a device, cause the device to perform the method of the fifth aspect or the method of the sixth aspect.

[0034] According to an eighth aspect of this disclosure, an apparatus is provided, comprising: at least one processor; and at least one memory including computer-readable instructions, which, when executed by the at least one processor, cause the apparatus to perform the method of the fifth aspect or the method of the sixth aspect. The apparatus may be the wearable sensor device. Attached Figure Description

[0035] The following describes specific embodiments by way of example and with reference to the accompanying drawings, wherein:

[0036] Figure 1 A flowchart is shown for a method to determine muscle selective activation.

[0037] Figure 2 A wearable sensor device is shown.

[0038] Figure 3 The optimal sensor location for myograph sensing of specific core muscles is shown.

[0039] Figure 4 A flowchart is shown for a method to determine whether a subject's primary muscle has been selectively activated.

[0040] Figure 5 An example user interface configured to display selective muscle activation data is shown.

[0041] Figure 6A The sensor locations used to compare MMG and EMG measurements of core muscle activity (i.e., sensor validation experiments) are shown.

[0042] Figure 6B The signal traces corresponding to the signal processing steps performed on the EMG and MMG signals are shown, as well as the signal traces from the inertial measurement unit (IMU).

[0043] Figure 6C The correlation between signals from the torso EMG sensor and two types of torso MMG sensors is shown.

[0044] Figure 6D Box plots of EMG, MMG, and IMU signal strengths are shown during active and rest periods.

[0045] Figure 6E A scatter plot comparing the root mean square (RMS) values ​​of EMG and MMG for a specific muscle during a specific activity is shown.

[0046] Figure 6F Box plots of the MMG-EMG correlation coefficient and the IMU-EMG correlation coefficient are shown.

[0047] Figure 7A EMG and MMG data for maintaining muscle contraction for a specific time period are shown.

[0048] Figure 7B A scatter plot is shown showing trunk MMG as a function of muscle thickness changes from ultrasound images during a specific activity.

[0049] Figure 8A Box plots are shown showing the activity / rest classification accuracy for different thresholds of maximum spontaneous contraction (MVC).

[0050] Figure 8B EMG and MMG data for specific activity and noise conditions are shown.

[0051] Figure 8C Box plots are shown for MMG classification accuracy under different noise conditions in the training test (based on over-threshold EMG).

[0052] Figure 9 A scatter plot of the EMG-MMG coefficient for body mass index (BMI) is shown.

[0053] Figure 10The integration of the MMG sensor in a wearable belt is shown.

[0054] Figure 11A A graph indicating the effect of MMG-based biofeedback on muscle activation (measured by EMG) is shown.

[0055] Figure 11B A graph indicating the effect of MMG-based biofeedback on core muscle selectivity is shown.

[0056] Figure 11C A graph showing the effect of MMG-based biofeedback on core muscle activation compared to placebo feedback is presented.

[0057] Figure 12A A graph indicating the effect of MMG-based biofeedback on the number of repetitions completed during self-guided exercise is shown.

[0058] Figure 12B The figure shows the effect of MMG-based biofeedback on the change in core muscle activation over time (area under the EMG curve) during self-guided exercise compared to no biofeedback.

[0059] Figure 13A Box plots of MMG-EMG correlation coefficients are shown for subjects with lower back pain (LBP) and no back pain (NBP).

[0060] Figure 13B A graph showing the effect of MMG-based biofeedback on core muscle activation in LBP and NBP subjects is presented.

[0061] Figure 14A This shows a common physical therapy-based back exercise.

[0062] Figure 14B It shows the source of the process Figure 14A The signal readings of the IMU channels of the subjects undergoing the exercise are shown.

[0063] Figure 14C The graph shows IMU data for a specific back exercise when performed correctly and incorrectly.

[0064] Figure 15A The diagram shows repeated IMU data for two specific exercises (top), as well as myographic (EMG and MMG) data from the RA (middle) and TA (bottom) muscles.

[0065] Figure 15B Box plots of selective muscle activation, measured with EMG, are shown for physical therapists and non-experts in three exercises.

[0066] Figure 15CThe cumulative confusion matrix is ​​shown for classifying repetitions as non-selective activation (where both RA and TA are <2%), TA-selective activation (where only TA is >5%), or non-selective activation (where both RA and TA are >5%). The decision tree classifier was trained using EMG RMS as ground truth data and features extracted from MMG. F1 (micro): 0.80; F1 (macro): 0.81; Accuracy: 0.81.

[0067] Figure 16 The MMG-EMG correlation coefficients for different sensor locations are shown.

[0068] Figure 17 Box plots of MMG and EMG signal strengths for exercise tests and noise conditions are shown.

[0069] Figure 18 A computing device configured to implement the methods of this disclosure is shown.

[0070] Figure 19 A rear view of another wearable sensor device is shown.

[0071] Figure 20 It shows Figure 19 Front view of the wearable sensor device shown.

[0072] Figure 21 Box plots of MMG signal intensity in the core muscles are shown during different subject postures.

[0073] Figure 22 A flowchart illustrating a method for predicting subject posture is shown.

[0074] Figure 23 A flowchart is shown illustrating a method for predicting the level of pain experienced by a subject. Detailed Implementation

[0075] The embodiments of this disclosure are explained below with particular reference to determining whether certain core muscles of a subject are selectively activated by the subject. However, it should be understood that the embodiments described herein are also applicable to determining whether other muscles of a subject (i.e., muscles that are not core muscles) are selectively activated by the subject. It should also be understood that the embodiments described herein are equally applicable when core muscles are activated by external stimuli (e.g., electrical current transmitted through a peripheral device).

[0076] Overview

[0077] Figure 1This is a flowchart of a computer-implemented method 100 for determining selective muscle activation. As used herein, the term "selective activation" indicates whether a specific muscle of a subject has been activated. For example, an indication of whether a specific muscle has been activated may include the degree of activation of that specific muscle in the subject. In particular, if the degree of activation of a specific muscle exceeds an activation threshold, it can be determined that the specific muscle has been selectively activated, wherein the activation threshold indicates that the muscle is activated rather than at rest. An indication of whether a specific muscle has been activated may also include the degree of activation of a second muscle in the subject, thereby allowing a comparison of the activation of the first and second muscles. By determining whether a muscle has been selectively activated, it is possible to identify whether the exercise being performed by the subject is activating the intended muscle group. By providing the subject with feedback on selective muscle activation, the subject can adjust how they perform their exercise to increase the activation of the intended muscle group, thereby enabling them to perform a specific exercise (e.g., a back exercise for relieving LBP) more effectively. Behavioral effects of MMG-based feedback on selective muscle activation have shown that providing such feedback leads to improved selectivity of muscle activation in the subject.

[0078] Method 100 can be used to determine the selective activation of a subject's core muscles. As used herein, the term "core muscles" refers to muscles located in the subject's trunk. Specifically, core muscles include the transverse abdominis (TA), rectus abdominis (RA), internal oblique (IO), and erector spinae (ES).

[0079] Method 100 is performed by one or more processors of the device (e.g., Figure 18 The processor 1102 of the computing device 1100 shown is implemented. The device can be configured to provide biofeedback to a subject in the form of visual feedback; in this case, the device can be a device with a display screen, such as a smartphone or tablet. Alternatively, the device can be configured to provide biofeedback to a subject in the form of tactile feedback; in this case, the device can be a wearable sensor device (e.g., a belt) including a myocardial sensor and an actuator configured to cause vibration of at least a portion of the wearable sensor device. Therefore, in general, the one or more processors can be located within a wearable device including sensors, or in a device separate from the wearable device including sensors.

[0080] At 110, sensor data is received from one or more myographic sensors of a wearable sensor device. The one or more myographic sensors are configured to monitor muscle activation in the subject.

[0081] The one or more myographic sensors may include one or more mechanographic (MMG) sensors. In one example, each of the one or more myographic sensors is an MMG sensor. MMG refers to various techniques for measuring specific low-frequency vibrations emitted by contracting muscles. For example, an MMG sensor may include a miniaturized microphone small enough to be worn. Unlike wearable electromyography (EMG), MMG does not require a connection to the skin, avoiding the need for adhesives, gels, or shaving; and it is not sensitive to sweat. Furthermore, fewer sensors are required (EMG requires positive, negative, and ground electrodes), and signal amplification is not required. This makes microphone-MMG suitable for low-cost personal devices that can measure muscle activity “anytime, anywhere.”

[0082] At 120, based on the sensor data received at 110, it is determined whether the subject's first muscle has been selectively activated. In one example, determining whether the subject's first muscle has been selectively activated includes determining the degree of activation of the first muscle relative to an activation value of the first muscle achieved during device calibration of method 100 (e.g., the maximum activation value of the first muscle). For example, the degree of activation of the first muscle may be determined relative to the maximum voluntary contraction (MVC) of the first muscle during the device calibration.

[0083] Additionally or alternatively, determining whether a first muscle has been selectively activated includes determining whether sensor data exceeds an activation threshold. The activation threshold may be a threshold that distinguishes between a first muscle being at rest (i.e., if sensor data does not exceed the activation threshold) and a first muscle being activated (i.e., if sensor data exceeds the activation threshold). The activation threshold may additionally or alternatively be a threshold indicating the degree of activation of the first muscle relative to the degree of activation of a second muscle (i.e., where activation of each of the first and second muscles is monitored using different myographic sensors). For example, sensor data may be processed to identify the degree of activation of the first muscle and the degree of activation of the subject's second muscle, thereby allowing a comparison of the activation of the first and second muscles to determine whether the first muscle has been selectively activated. For example, if the degree of activation of the first muscle (e.g., relative to the MVC of the first muscle) exceeds the degree of activation of the second muscle (e.g., relative to the MVC of the second muscle), it can be determined that the first muscle is selectively activated. Alternatively, if the ratio of the degree of activation of the first muscle to the degree of activation of the second muscle exceeds a threshold (e.g., 1.5 times), it can be determined that the first muscle is selectively activated.

[0084] In each of the above scenarios, determining whether sensor data exceeds an activation threshold may include classifying the sensor data using one or more trained classifiers. Specifically, one or more classifiers may be trained to distinguish between sensor data indicating that a first muscle is at rest and sensor data indicating that a first muscle is activated. Additionally or alternatively, one or more classifiers may be trained to distinguish between sensor data indicating superficial activation of core muscles (e.g., activation of the RA muscle) and sensor data indicating deep activation of core muscles (e.g., activation of the TA or IO muscles). Furthermore, one or more classifiers may be trained to identify the degree of activation of core muscles.

[0085] At 130, biofeedback is provided based on the determination of whether the primary muscle has been selectively activated. Biofeedback can be provided to the subject and can be provided in various forms. For example, biofeedback may include visual feedback, audio feedback, and / or tactile feedback.

[0086] When method 100 is implemented by a processor as part of a wearable sensor device, providing biofeedback may include providing audio from a speaker of the wearable sensor device and / or activating an actuator configured to cause vibration in at least a portion of the wearable sensor device. When method 100 is implemented by a processor as part of a device separate from the wearable sensor device, providing biofeedback may include providing visual biofeedback on a display screen of the device, providing audio biofeedback from a speaker of the device, and / or providing commands to the wearable sensor device to provide audio and / or tactile feedback.

[0087] When method 100 is implemented via an external electrical stimulator, a part of a device separate from the wearable sensor device, the electrical stimulator is used to guide muscle activation. Feedback from a biofeedback device is used to modulate the level of stimulation that induces muscle activation. The level of stimulation, such as the electrical current directed to the muscle, is modulated using a processor that receives the same biofeedback signal to control muscle activation.

[0088] Sensor arrays and fixed (MMG / IMU)

[0089] As described above, in method 100, 110, sensor data can be received from multiple MMG sensors. The MMG sensors can be part of a wearable sensor device such as a belt. Specifically, as... Figure 2As shown, the wearable sensor device 200 may include at least two MMG sensors 210 applied along the width of a single neoprene / Lycra (RTM) waistband 220, which is approximately 10 cm wide. Each MMG sensor 210 is adapted into a 3D-printed housing 212 made of PEBA (polyether block amide) to provide flexibility. An alternative material for the housing 212 includes elastic thermoplastic elastomer filaments. The housing 212 is attached to the waistband 220 via a Velcro patch 230, allowing for easy adjustment of the sensor 210's placement and easy removal of the sensor 210 when the waistband 220 needs cleaning.

[0090] The MMG sensor 210 can be used to monitor muscle activation in three core muscle groups. Specifically, the MMG sensor 210 can be used to monitor muscle activation in the RA, TA or IO, and ES muscles. The optimal location of the MMG sensor 210 can be confirmed based on anatomical landmarks identified in previous EMG studies, and as... Figure 3 As shown. Specifically, for RA, the MMG sensor 210 can be placed 3 cm lateral to and 3 cm superior to the umbilicus to avoid the thickest layer of adipose tissue (e.g., as shown). Figure 6A (As shown). For TA / IO, the MMG sensor 210 can be placed in the middle of the muscle, 2-3 cm medial to the anterior superior iliac spine (ASIS), and 2 cm below the line connecting the two ASIS (also, for example, as shown). Figure 6A (As shown). For ES, the MMG sensor 210 can be placed 2 cm outside the interspinous space of L4-L5 (also as shown). Figure 6A (As shown). The sensor arrangement used to monitor these muscles can be bilateral or unilateral, because the accuracy of detecting muscle activation in each muscle using a single MMG sensor 210 is approximately 80%.

[0091] To be implemented in the belt 220, the RA MMG sensor 210 can be shifted downwards to be collinear with the TA, while the TA MMG sensor 210 is moved upwards approximately 1 cm toward the ASIS line (e.g., Figure 6A (As shown). The ES MMG sensor 210 is adjusted to be collinear with the other two sensors 210. Therefore, the wearable sensor device 200, which includes three MMG sensors 210, can be used to monitor muscle activation in RA, TA, or IO muscles as well as ES muscles.

[0092] Figure 2Each MMG sensor 210 shown includes a microphone (e.g., an SPU1410LR5H-QB microphone available from Knowles, Inc., Itasca, Illinois, USA), housed within a chamber of a sealed plastic housing 212, which is 2 cm in diameter and 1 cm high. A Mylar (RTM) membrane at one end of the chamber is positioned to contact the skin covering each muscle of interest. Mechanical perturbations to the membrane, including those caused by muscle fiber oscillations, result in pressure changes within the chamber, which are detected and recorded by the microphone. The MMG sensor 210, worn on the torso, is connected via wires to a single analog-to-digital converter, which can be packaged together with an inertial measurement unit (IMU), a Bluetooth transmitter, a rechargeable battery, and a switch. These components can be packaged in a 4.5 x 2.5 x 1 cm box and attached to the subject's waist via a pocket on a belt 220. By using the MMG sensor 210 instead of an EMG sensor, an amplifier is not required. The IMU can sample at 500 Hz, providing 3D acceleration (using a 16g triaxial accelerometer, such as the ADXL345 accelerometer available from Analog Devices in Wilmington, Massachusetts) and 3D gyroscope data. The battery can have sufficient capacity to provide 4 hours of continuous recording. A Bluetooth transmitter can be configured to transmit data from the MMG sensor 210 to one or more processors implementing method 100.

[0093] Figure 19 and 20 Another example of a wearable sensor device 500 is shown, which includes a belt 520 and multiple MMG sensors 510. (See attached image.) Figure 19 As shown, the RA MMG sensor 510a is centered 20 mm to each side of the anterior sagittal plane, while the TA MMG sensor 510b is centered 80 mm to each side of the anterior sagittal plane. The ES MMG sensors 510c are spaced 40 mm apart and are connected to an adjustable D-ring 522, which allows the ES MMG sensors 510c to move independently of the tightness of the belt 520. This allows the ES MMG sensors 510c to be correctly positioned on the ES muscles, with their centers centered 20 mm to each side of the posterior sagittal plane. The belt 520 includes a male buckle member 524, the position of which is adjustable to tighten or loosen the belt 520. Additionally, the belt 520 includes a female buckle member 526, which provides a snap-fit ​​engagement with the male buckle member 524. A housing 528, including a PCB unit (not shown), is coupled to the female buckle member 526 to minimize the profile of the belt 520. RA MMG sensor 510a and TA MMG sensor 510b are connected to housing 528.

[0094] Calibration using MVC

[0095] To calibrate the wearable sensor device 200 or 500, each participant performed two maximum voluntary contractions (MVCs) (5 seconds each) for each muscle group, with a 30-second rest between them. For the RA MVC, the subject first lay supine with knees bent at 90 degrees and arms crossed over the chest, each hand touching the opposite shoulder (“Rest”). Then, when instructed (“Active”), the subject performed a partial sit-up, lifting the shoulder blades approximately 30 degrees off the ground. Manual resistance was applied at the elbows by a physical therapist. For the TA / IO MVC, the subject assumed the same rest position as the RA MVC and performed an incline sit-up, attempting to move one shoulder toward the opposite bent knee. Manual resistance was applied to the lifted shoulder and the opposite bent knee. For the ES MVC, the subject first lay prone with hands behind the head (“Rest”) and, upon instruction, attempted to slowly lift the torso, performing a trunk extension in the prone position. Manual resistance was applied at the ankles and shoulder blades. The MMG signal for each muscle was then normalized to an average value over a 500 ms window around the subject's maximum value during the corresponding MVC. This allows all subsequent workouts to be displayed as percentage effort values ​​relative to the MVC. For classifier training (described in further detail below), the same workout was used to obtain the subject's EMG signal values ​​for the MVC of each muscle.

[0096] Alternatively, other calibration methods can be implemented (i.e., as alternatives to determining a subject's MVC). As an example, a subject may be instructed to perform specific exercises (e.g., abdominal crunches, bridge exercises, etc.), and the subject's maximum muscle activation values ​​during these exercises can be used to compare muscle activation at a baseline during use (e.g., using a threshold value based on the baseline). As another example, a subject may be instructed to assume a specific posture (e.g., sitting upright or standing) or rest, and the subject's muscle activation values ​​under that specific posture or rest condition can be used as a baseline for comparing muscle activation during use (e.g., using a threshold value based on the baseline).

[0097] Signal processing

[0098] In one example, determining whether a subject's primary muscle had been selectively activated included 120. Figure 4 The signal processing method 300 shown is illustrated.

[0099] In 310, method 300 includes preprocessing or denoising the raw MMG signal data. In one example, preprocessing the raw signal data includes removing noise from the signal by: (i) bandpass filtering the raw signal from 10 Hz to 100 Hz using a zero-phase fifth-order Butterworth filter; (ii) full-wave rectification of the filtered signal; and (iii) applying a Hamper filter with a 1000-sample window size to remove outliers. The output of the preprocessed raw signal data is a clean signal with noise removed.

[0100] Alternatively, in step 310, the raw signal data (or signal data that has been bandpass filtered according to step (i) above) can be processed by a trained denoising model. This denoising model is trained using ground truth data from 70-80 participants for whom binary time-series data is available, indicating when each participant was instructed to exercise or rest. To generate ground truth data, the preprocessing steps of 310 (or, if the signal data has been bandpass filtered, steps (ii) and (iii) of 310) are first applied to the participants' MMG signals. Any signals occurring during rest periods are then attenuated from the signal. Attenuating signals occurring during rest periods is possible because rest periods can be identified from the time-series data. To attenuate the rest period portion of the signal, the signal occurring during the rest period is multiplied by a low value, such as 0.1, to reduce the amplitude of the rest period portion relative to the rest of the signal. The denoising model is trained using this clean ground truth data with a mean squared error loss function and an Adam optimizer to learn and adjust the model. Therefore, the output from the trained denoising model is a clean signal with noise removed.

[0101] In 330, method 300 includes normalizing the clean signal by referencing the subject's MVC for the muscle associated with the MMG signal data being processed. This means that the signal value is expressed as a percentage of the subject's MVC for that muscle.

[0102] In method 300, method 340 includes extracting features from the MMG signal. The following 15 features can be extracted from each MMG channel: mean absolute value, standard deviation, maximum value, range, kurtosis, skewness, root mean square (RMS), variance, waveform length, mean amplitude variation, absolute standard deviation of the difference, logarithmic detector, corrected mean absolute value 1 and 2, and simple square integral. Features are extracted from the signal for each time period during which the subject attempts to selectively activate the first muscle. This time period may correspond to the time between the start and stop times set by the software application used by the subject for exercising (e.g., displayed on a user interface) (see below). Figure 5 (Describe it).

[0103] In method 300 (350), scaling involves subtracting the minimum value present in each feature column and then dividing that value by the range of each feature column to generate input for one or more classifiers. Other scaling methods that can be used include removing the median and scaling the data according to the quantile range.

[0104] At 360, method 300 includes classifying the input generated at 350 using one or more trained classifiers. Specifically, classifying the input may include applying one or more trained classifiers to the input: a first trained classifier 362 configured to determine whether a muscle is at rest or in an activated state; a second trained classifier 364 configured to determine selective activation of the muscle; and a third trained classifier 366 configured to determine the intensity of muscle activation.

[0105] Each of these classifiers is trained using features extracted from specific regions of MMG signal data and corresponding EMG signal data (recorded simultaneously with the MMG signal data) as ground truth. Specifically, each classifier uses ground truth data generated from EMG signal data received from a surface EMG sensor to monitor muscle activity in multiple participants. Surface EMG data is highly accurate and therefore suitable as ground truth. However, EMG sensors are not convenient for the reasons discussed above. Each classifier uses features identified from 340 corresponding simultaneously recorded MMG sensor data (i.e., using features similar to...) Figure 6A The sensor arrangement shown is used for training. Features of the MMG sensor data used to train the classifier are extracted for each region of the MMG signal, which corresponds to the region where the EMG signal exceeds a specific threshold (e.g., 2.5%, 5%, etc. of the EMG MVC) within a specific time period (e.g., at least 400 ms). The specific threshold of the EMG signal may correspond to the muscle activation period, rest period, activation period of one muscle relative to another, or muscle activation intensity, as illustrated in the examples below.

[0106] For the first trained classifier 362 (i.e., the rest and activation classifier), the calibration steps described above and the preprocessing (in 310) and normalization to MVC (in 330) described above for MMG data were performed on the EMG data collected from each participant to generate EMG signal values ​​corresponding to each participant's MVC for each muscle. An activation threshold (relative to individual MVC) was set. For an epoch to be classified as "beyond threshold," the EMG signal must remain above that threshold for at least 400 ms; in this case, the epoch is labeled "1." An epoch comprises the activity repetitions for each subject throughout all exercises and their corresponding rest periods. A rest epoch was determined as a 2-second interval around the minimum EMG point within each rest epoch; in this case, the epoch is labeled "0." Therefore, the ground truth data for the first classifier 362 includes EMG activity epochs (labeled "1": muscle activation) and rest epochs (labeled "0": muscle rest). Features associated with the MMG sensor data for their corresponding epochs (i.e., features listed in 340 and scaled in 350) were used to train classifier 362. The impact of the target EMG threshold on classifier accuracy was evaluated for EMG MVC values ​​of 2.5, 5, 7.5, 10, 12.5, and 15%. Therefore, the threshold used for predefined active or resting states can be modified to produce different difficulty patterns.

[0107] The first classifier, 362, is a C-Support Vector Machine (SVM) classifier. For each muscle, the MMG input generated in 350 (i.e., scaled features extracted from real-time MMG data) is fed as a predictor into the C-SVM classifier to predict exercise repetitions exceeding the EMG target threshold. In one example, the SVM is prepared in Python 3.7 using the scikit-learn, imblearn, and scikit-multilearn packages; parameters are: C=1000, gamma=0.001, kernel='rbf', probability=True. In this example, leave-one-out (subject) cross-validation is used to evaluate performance, reporting the mean accuracy and the area under the receiver operating characteristic curve (AUROCC). Furthermore, in this example, the Synthetic Minority Oversampling Technique (SMOTE) from imblearn is used to train the model with equal proportions of classes.

[0108] For the second trained classifier 364 (i.e., the selective activation classifier), the EMG signal data is preprocessed, denoised, and normalized (using procedures 310 to 330 described above for MMG data) to generate EMG signal data for each muscle relative to the MVC of each participant. This classifier can predict whether activation involves overall non-specific muscle activation or whether activation is specific to deeper (TA / IO) muscles. The ground truth data for the second classifier is generated using preprocessed, denoised, and normalized EMG signals, with a thresholding of the ratio between TA / IO and RA activations in a given repetition. If the TA / IO:RA ratio exceeds a predetermined threshold (e.g., 1.5 times), the activation is classified as TA-specific activation, resulting in binary ground truth data of 1 or 0 (selective or non-selective activation). The training and testing datasets used to train the second classifier 364 consist of MMG features extracted and scaled from each window of MMG data collected simultaneously, corresponding to selective or non-selective activation. Alternatively, the absolute difference between RA and TA values ​​can be used instead of the ratio to create ground truth data, as this also indicates core muscle selectivity. Depending on the exercise being performed, other muscle activation rates may also be used.

[0109] The second classifier, 364, is an Extremely Random Tree (ERT) classifier. MMG inputs generated for TA / IO and RA muscles at 350 (i.e., scaled features extracted from real-time MMG data) are fed into the ERT classifier as predictors to predict exercise repetitions exceeding a ratio threshold. The ERT classifier is prepared in Python 3.7 using the scikit-learn, imblearn, and scikit-multilearn packages; parameters are: No. estimators=100, Split criterion=Gini. Performance is evaluated using leave-one-out (subject) cross-validation, reporting mean accuracy and area under the receiver operating characteristic curve (AUROCC). The model is trained with equal proportions of classes using the Synthetic Minority Oversampling Technique (SMOTE) from imblearn.

[0110] For the third trained classifier 366 (i.e., the activation strength classifier), EMG RMS values ​​are extracted from the EMG signal data. The EMG RMS values ​​are then graded into multiple levels (e.g., 1-5) using thresholds for the EMG MVC values ​​(e.g., 0-10% = 1; 10-20% = 2) to indicate the intensity of activation. The graded EMG RMS values ​​are the ground truth thresholds for the third classifier. For each graded EMG value, the same aforementioned features (i.e., those mentioned in 340) are extracted from the corresponding MMG signal to construct the training / test dataset for training the third classifier 366.

[0111] The third classifier, 366, is a regressor-based model in the form of a Voting Regressor, consisting of three models: a Bayesian Ridge Linear Regressor, a Huber Regressor, and a Support Vector Regressor, implemented using scikit-learn. The MMG features generated in 350 (i.e., scaled features extracted from real-time MMG data) are fed into the third classifier, which classifies the MMG signals from each muscle into graded activations.

[0112] In some examples, the third classifier 366 can be used as an alternative to the second classifier 364 to provide hierarchical activation ground truth data instead of binary selective versus non-selective ground truth data.

[0113] Instead of using machine learning models for feature extraction (i.e., steps 340 and 350 of method 300), deep learning models, such as recurrent or fully convolutional neural networks, that use EMG data as ground truth are alternative methods for predicting EMG RMS data. Therefore, these methods can be used alternatively to implement the three classifiers 362, 364, and 366. Furthermore, those skilled in the art will understand that the specific classifiers and parameters mentioned above are merely illustrative examples, and that other classifiers and parameters can be used to implement the three classifiers 362, 364, and 366.

[0114] Certain signal processing steps can be implemented using the belt firmware instead of the one or more processors implementing method 100. In other words, processing can be distributed between the belt firmware and the one or more processors. Specifically, initial signal preprocessing (e.g., noise reduction via adaptive filtering and data decimation) can be performed using the belt firmware to minimize data transmission requirements (i.e., in embodiments where one or more processors are part of a separate device from the wearable sensor device). This approach enables compatibility with most 32-bit controllers, thus achieving compatibility on common hardware platforms.

[0115] Specifically, the firmware can be used to implement noise reduction algorithms to process inputs from multiple MMG sensors 210 and 510 at high-fidelity data rates (i.e., 512 Hz - 1024 Hz). Specifically, to provide noise reduction through adaptive filtering, the conventional Least Mean Square (LMS) algorithm is enhanced by integrating a Recursive Least Squares (RLS) method, which provides faster convergence and better adaptability in high data rate environments for real-time filtering. Furthermore, the firmware can be used to implement a dynamic buffering strategy tailored for the Nordic nRF5340 microcontroller unit (MCU), available from Nordic Semiconductor in Trondheim, Norway. The dynamic buffering strategy implements a circular buffer combined with selective data retention to manage the high data throughput from the MMG sensors 210 and 510. This approach minimizes memory usage while ensuring that the required data is processed and transmitted without delay. In addition, decimation techniques intelligently reduce the data rate based on contextual analysis of signal importance, using gradient descent over a sliding window to retain critical signal information while reducing processor load and power consumption.

[0116] IMU is used to segment motion and identify posture.

[0117] Method 100 may also include receiving motion sensor data (e.g., accelerometer data from an IMU of wearable sensor device 200). In one example, the IMU accelerometer signal is bandpass filtered and full-wave rectified using a zero-phase fourth-order IIR filter (half-power frequency 1: 1 Hz; half-power frequency 2: 20 Hz) to calculate the three-plane acceleration amplitude.

[0118] The classifiers 362, 364, and 366 described above can also be used with IMU data. For example, for IMU data periods corresponding to activity and rest periods determined from EMG sensor data, the same 15 features described in 340 can be extracted (i.e., in the same manner as extracting features for MMG sensor data). The classifier can then be trained using the features extracted from the IMU sensor data in the same manner as those extracted from the MMG sensor data.

[0119] The IMU provides gyroscope velocity, acceleration, and orientation data, which can be used to identify a person's posture while performing certain exercises (such as abdominal crunches while lying down, standing, or sitting) and during standing / sitting postures. The activity window can be categorized into different postures. Therefore, IMU data can be used to identify the correct position for a subject to begin a specified exercise. For example, if the exercise involves lying down, then by processing the IMU data, the software application the user uses to perform the exercise (see below) Figure 5(Description) can be configured to wait until the user lies down before starting a workout cycle. IMU data can also be used to identify excessive movement (e.g., jerking or swaying while trying to maintain a posture) in the subject during exercise. In response to the identification of excessive movement, the subject can be given additional cues to correct their posture.

[0120] To classify IMU data, a broad, first-step classification involves categorizing the active window as dynamic or static based on accelerometer, gyroscope, and magnetometer data exceeding predetermined thresholds. A second step further classifies the window as sitting, lying down, standing, noisy, walking, and running by using additional thresholds applied to a combination of IMU and MMG data. Therefore, method 100 may include processing motion sensor data to determine the subject's posture.

[0121] Furthermore, IMU data provides motion information, which can be used to determine when a subject is moving rather than exercising. For example, motion windows in MMG data can be identified from IMU data and then multiplied by 0.1 to reduce the influence of motion on the MMG signal. This allows for the removal of large spikes in the MMG data caused by the subject's movement, so that the feedback provided only shows static isometric muscle activation.

[0122] Visual and tactile feedback for exercise

[0123] As described above, the biofeedback provided at 130 may include visual feedback, audio feedback, and / or tactile feedback, and may be provided by wearable sensor devices or devices such as smartphones or tablets. In one example, such as Figure 5 As shown, the software application can be configured to provide a user interface 400 that displays the subject's MMG for RA, TA / IO, and ES using real-time filtered MMG signals. To remove... Figure 5 In the example shown, the original MMG signal is first bandpass filtered at 10-100 Hz, then rectified, and finally low-pass filtered at 1 Hz. Alternatively, the original MMG signal can be processed by a trained denoising model (e.g., such as...). Figure 4 (as described in 310).

[0124] The user interface 400 provides processed MMG sensor data in real time in the form of three rings 402 (one for each muscle). The MMG signal values, after being normalized to the user's MVC, are displayed as 100% full value. Although Figure 5 The example demonstrates how to determine selective muscle activation without using one or more trained classifiers, but it is also possible to determine selective muscle activation using one or more trained classifiers, as explained in detail above.

[0125] Figure 5 Each muscle in the example is color-coded, and the corresponding sensor spatial layout image is displayed on the screen. Figure 5 The partial user interface shown on the right displays the user's activity and rest cycles. Figure 5 The software application shown also includes a timer 404 with start and stop sound signals. This software application allows users to adjust the timer 404 according to how long they wish to rest between repetitions and how long they wish to maintain each repetition. Furthermore, the software application provides functionality that allows users to modify the number of repetitions they wish to perform. Figure 5 The software application shown also includes images (406 images) or short videos of each exercise, which remain on the screen while the user exercises. For example... Figure 5 As shown, the software application includes a pause / play button 408, allowing the workout to be paused and resumed at any time, and a repetition counter 410 on the screen. Furthermore, the software application allows the user to change the difficulty level by defining 100% activation (e.g., in easy mode, ring 402 can display 0.5x MVC as a 100% value, while a more difficult mode can use the actual MVC value as 100%).

[0126] As an alternative to the ring 402 displaying the activation of each muscle, the user interface 400 may display the ratio of TA to RA activation as a single value. In this alternative example, displaying the TA:RA ratio may increase the ease with which the subject focuses on a “good” activation pattern (i.e., relatively high TA activation).

[0127] In one example, the software application includes built-in motivational feedback that can be adjusted in real time based on the user's performance. Furthermore, at the end of each session, the user can use the functionality to view their progress and compare it to previous sessions, including the number of repetitions performed correctly and the average level of effort exerted. The software application can be configured to command actuators on a wearable sensor device (e.g., via a Bluetooth interface) to provide tactile feedback. Tactile feedback from the actuators on the wearable sensor device, whether on the front or back of the torso, signals to the user that they need to further engage that muscle during exercise. Additionally, tactile feedback can also be used to reduce slouching, even when the user is wearing a support belt but not exercising, or when the user is at a desk or standing.

[0128] In alternative examples, devices such as smartphones or tablets may provide only audio feedback, rather than visual feedback. In further alternative examples, biofeedback may be provided in the form of audio and / or haptic feedback from speakers and / or actuators of the wearable sensor device. In such examples, one or more processors of a separate device (e.g., a smartphone or tablet) may command the speakers and / or actuators of the wearable sensor device to provide audio and / or haptic feedback. Alternatively, one or more processors of the wearable sensor device may command the speakers and / or actuators of the wearable sensor device to provide audio and / or haptic feedback.

[0129] Using electrical stimulation to activate muscles and using biofeedback for control.

[0130] In another example, the device can be coupled to an external device configured to stimulate muscles or nerves in the trunk. One form of this example includes a wearable or external electrical stimulator configured to apply a charge to one or more muscles in the trunk to induce activation of those muscles. Functional electrical stimulation (FES) is an example of this form. Muscle activation from the stimulus can be measured via biofeedback in a manner similar to voluntary muscle activation; however, the biofeedback will be used by one or more processors to modify the stimulus level to guide the activation to produce a response. For MMG, biofeedback is based on non-electrical measurements of muscle activity and can therefore be used for biofeedback during stimulation (while EMG sensing may not be suitable for use with electrical stimulation due to electrical interference). Another manifestation of this example is neuromodulation achieved by applying a charge to nerve fibers via a wearable or external device. Transcutaneous electrical nerve stimulation (TENS) is an example of this form. The same biofeedback can be used by one or more processors to control the stimulus level to produce a pain-inhibiting response. The electrical stimulator can also be incorporated into a wearable sensor device that includes sensors (e.g., Figure 2 (The belt shown).

[0131] Support data

[0132] The effectiveness of using a trunk MMG sensor to predict core muscle activity was tested using the gold standard laboratory EMG. Eighty-nine participants performed three standard back exercises (i.e., tuck exercises, core support exercises, and bridge exercises) under the supervision of a physical therapist. MMG and EMG sensors were placed at conventional recording sites on three trunk muscles that were preferentially activated during the three exercises: TA / IO, RA, and ES. For comparison, an IMU was also mounted on the trunk (above RA), similar to the strategy used in current technology for home back exercise monitoring. Electrode sites are shown below. Figure 6AAs shown, based on the optimal muscle location for surface EMG sampling described in previous publications, each MMG sensor is positioned between two EMG sensors for each muscle (TA / IO, RA, and ES). Figure 6A The electrode locations shown are for validating the MMG at the previously described standard locations for TA / IO, RA, and ES sampling using EMG. In contrast, in an MMG belt used for biofeedback (i.e., as...) Figure 2 As shown in the figure, the RA sensor is moved directly downwards to the subumbilical position, so that all three sensor positions are approximately coplanar.

[0133] Figure 6B The signal traces corresponding to the various signal processing steps performed on the signal are shown. These steps are used to remove noise and motion artifacts and focus on specific frequency bands (2 Hz to 200 Hz) activated by muscle fiber contraction. Figure 6B As shown, the processed MMG signal corresponds very well with the processed EMG signal. In contrast, the IMU data and EMG data shown in the bottom signal trace have a very poor correlation.

[0134] Figure 6C The study shows how signals from two different types of trunk MMG sensors (diaphragm-MMG and silicone-MMG) from representative subjects (i.e., those whose MMG-EMG r values ​​were closest to the median) correlate with signals from trunk EMG sensors for three exercise types and three corresponding trunk muscles. Figure 6C The two types of MMG sensors shown provide MMG sensor data that corresponds well to EMG sensor data.

[0135] Figure 6D Box plots are shown of EMG and MMG signal intensities (root mean square (RMS), as a percentage of maximal voluntary contraction (MVC, further described below)) during active and rest periods. The changes in MMG signal between active and rest periods were significant and similar to the EMG signal amplitude across all exercises and core muscles. Figure 6D Only the abdominal contraction movement is shown, but all percentage changes between the active and resting periods are greater than 100%. In contrast, for abdominal contraction or core support movements, the trunk IMU signal did not differ significantly between the active and resting periods, although it was sensitive to the bridge movement due to trunk movement.

[0136] Figure 6E A scatter plot comparing the EMG and MMG RMS values ​​of the RA and TA muscles during core support and contraction exercises is shown. Figure 6E The data shown corresponds to the subjects whose r-values ​​are closest to the group midpoint.

[0137] Figure 6F (Top left panel) shows a box plot of the relationship between the MMG-EMG correlation coefficient (r) and the peak experimental EMG amplitude. For TA and RA, the mean correlation coefficients between MMG and EMG RMS are: r = 0.80 and r = 0.81 for diaphragm-MMG; and r = 0.83 and r = 0.82 for silica-MMG. For ES, the mean correlation coefficients between MMG and EMG RMS values ​​are: r = 0.73 for diaphragm-MMG; and r = 0.79 for silica-MMG. The equivalent r value for IMU-EMG (bottom left panel) is approximately 50% lower (mean: r = 0.41). The right panel shows the relationship between the MMG-EMG correlation coefficient and the peak experimental EMG amplitude for RA, TA, and ES, adjusted for time delay (grey box plots show a 500 ms time correction, black box plots show an unrestricted time correction). Time correction brings all median coefficients to >0.8.

[0138] therefore, Figures 6B to 6F The results showed that during exercise, trunk MMG corresponded well with laboratory EMG, while trunk IMU did not.

[0139] Figure 7A The percentage of muscle contraction for holding an isometric contraction for 4 seconds (top left) and 8 seconds (bottom left) is shown in the TAMVC EMG and MMG data. The duration of hold, defined as the signal being at least 5% of the subject's MVC, was significantly correlated between EMG and MMG across all exercises (RA: r = 0.76, as shown in the top right; TA: r = 0.72, as shown in the middle right; ES: r = 0.61, as shown in the bottom right (p < 0.001)). This indicates that MMG detects muscle activity itself, not just transient motion / sound associated with the start and end. This contrasts with trunk IMU, which was not correlated with the duration of isometric contractions of any core muscles (p > 0.05).

[0140] Figure 7B Scatter plots of trunk MMG as a function of corresponding muscle thickness changes during tuck and core support movements are shown (median r = 0.70 (TA); r = 0.73 (RA); p < 0.01; N = 10). Muscle thickness was measured using ultrasound.

[0141] therefore, Figure 7A and 7B MMG can predict dynamic changes in isometric contraction duration and core muscle thickness.

[0142] As described above, a machine learning classifier for MMG signals was trained to predict exercise trials that exceeded a predetermined EMG threshold. Figure 8A Box plots show the percentage accuracy applied to various percentage MVC thresholds for TA (right panel). For thresholds above 5%, RA, TA, and ES achieved accuracies between 80% and 87% (area under the receiver operating characteristic curve 0.89 to 0.93). In contrast, IMU-based classification accuracy ranged from 63% to 68%. The classifier was trained using balanced classes and validated using leave-one-out cross-validation. The classifier was found to be more accurate than simple signal thresholding.

[0143] Among the 20 participants, after completing a series of back exercises, they underwent three purposeful noise conditions: loud counting (by...). Figure 8B and 8C (in the speech bubble indicator), deep breathing (by...) Figure 8B and 8C (head indicator in the image) and rubbing clothing back and forth on the electrodes (by...) Figure 8B and 8C (The shirt instructions are in the image). Subjects remained lying down throughout the session to minimize actual trunk muscle activation. Figure 8B Example MMG and EMG signal traces are shown. In most cases, the noise intensity is significantly lower than that of standard exercise repetitions (core support and abdominal contraction movements are shown for comparison).

[0144] To improve accuracy beyond the threshold, a machine learning classifier was used to distinguish between exercise repetitions and rest, and / or one of three types of noise (loud counting, deep breathing, and rubbing clothing on the electrodes). This classifier was identified as the first trained classifier described above. Figure 8C Box plots are shown for the MMG classification accuracy (based on overthreshold EMG) for each of the three types of noise mentioned above, targeting the exercise test versus noise. Figure 8C As shown, these classifiers all achieve a validation accuracy of over 80% for RA, TA, and ES.

[0145] therefore, Figures 8A to 8C The results show that core muscle contractions are accurately classified by MMG and can be distinguished from ambient noise.

[0146] Figure 9 This study demonstrates how the EMG-MMG correlation coefficient values ​​for TA and RA change with the subject's body mass index (BMI) during abdominal contraction and core support exercises. BMI does not significantly affect MMG signal amplitude, correlation with EMG, or contraction classification accuracy. Therefore, MMG accuracy does not decrease with increasing BMI. Figure 9 The central shading in each image indicates overweight subjects, while Figure 9The shaded area on the right side of each figure identifies obese subjects. Figure 9 The results showed that overweight and obese subjects exhibited similar EMG-MMG correlations to subjects with lower BMI.

[0147] Figure 10 The integration of MMG sensors within a neoprene waist belt is demonstrated, enabling convenient self-application and a comfortable fit. This places three MMG sensors at the same level while still targeting three separate core muscles. The MMG-EMG correlation is comparable in magnitude to that found with tape-fixed MMG (r > 0.8 for TA and RA, r = 0.73 for ES during corresponding exercises, showing the median for subjects). Therefore, Figure 10 The display shows that the MMG sensor can be integrated into a practical belt, thus verifying that... Figure 2 The effectiveness of the wearable sensor device 200 shown.

[0148] Behavioral experiments were also conducted to determine whether the MMG-biofeedback belt could make back exercises more effective. Forty-six participants (non-physiotherapists) performed three exercise groups, each consisting of repetitions of ab crunches, core support exercises, and bridge exercises (x6), each lasting 4 and 8 seconds respectively. No feedback was provided in group #1; MMG feedback was provided in groups #2 or #3, with the choice of feedback group balanced among the participants. Exercise performance was assessed using accompanying EMG, which was concealed from the participants.

[0149] Figure 11A The effects of MMG biofeedback on muscle activation in three exercises and three core muscles were shown. Greater activation occurred in the MMG feedback groups compared to the "no feedback" group, regardless of the order in which feedback was provided (p < 0.01 for most muscle / exercise / duration combinations). This can be seen from the peak "Ʌ" observed in the group receiving feedback in group #2 and the "V" observed in the group receiving feedback in group #3, indicating a decline in performance after group #1, but feedback increased effort relative to no feedback. Muscle enhancement associated with feedback was greatest for the trunk stabilizer TA (59%) compared to RA (29% increase).

[0150] The impact of feedback on workout quality can also be assessed by measuring core muscle selectivity, i.e., the difference between TA and RA activation levels, which is associated with LBP reduction. Figure 11BThe charts in the figure show the %RMS difference associated with MMG feedback comparing RA and TA. Under all six conditions, feedback resulted in greater relative activation of TA than RA (measured using EMG) (all p < 0.05 except for Hollow 8s). The significance of this was evident in a separate experiment where 11 physical therapists selectively activated TA by more than 100% more than non-physiotherapists during abdominal contractions, in both groups without biofeedback (p < 0.01). Therefore, the MMG-belt can bring non-experts closer to the “appropriate” activation pattern.

[0151] To assess whether MMG biofeedback drove the behavioral effect, rather than app-based encouragement, 16 different subjects were tested using different versions of the app in which the placebo “feedback” consisted of an app dial that increased from 0 to 100% in each trial. Specifically, the placebo “feedback” included subjects visualizing a similar dynamic app display, except that in each trial, visual activity moved alternately from 0 to 100% independently of actual muscle activity. Figure 11C In the image, data from different versions of the application is displayed at the bottom of the scale. In this case, the placebo "feedback" from different versions of the application did not enhance muscle activation or selectivity, indicating that accurate muscle feedback is crucial for improving workout performance.

[0152] therefore, Figures 11A to 11C MMG biofeedback results in greater and more selective core muscle activation.

[0153] In Figures 11A to 11C In the corresponding behavioral experiments, participants were expected to repeat the exercises when prompted, so these results may not reflect the overall effectiveness of the belt in real-life situations where participants choose how much exercise they want to do. To assess the impact of MMG biofeedback on self-guided exercise, another experiment was conducted in which participants were instructed to perform abdominal contractions "as many times as possible" in two 3-minute groups without receiving any prompts (n=50). In one group, MMG feedback was provided; in the other group, no feedback was provided, with the order balanced between participants. Figure 12A The data showed a significant increase in the number of repetitions performed by each subject, averaging 20.8% (left panel; p < 0.05). A significant correlation was observed between the number of repetitions achieved during the no-feedback rounds and the percentage increase in the number of repetitions after feedback (right panel; r: -0.76; p < 0.001). Specifically, Figure 12A The scatter plot illustrates how the increase in the number of repetitions associated with MMG is negatively correlated with the baseline number of repetitions.

[0154] Figure 12BThe results showed that, compared to no feedback, MMG feedback increased uncued TA activation (EMG integral over time; p < 0.05), regardless of group order (line color indicates order, where black = feedback presented first; gray = feedback presented second). These findings are consistent with previous experiments ( Figures 11A to 11C The consistency between the results indicates that the feedback improved the quality of the workout. Figures 11A to 12B The effectiveness of MMG feedback was validated using EMG measurements, which are considered the gold standard for measuring muscle activity. Therefore, the EMG data showed that the embodiments described herein increased activation by providing MMG-based biofeedback.

[0155] The above experiment was re-analyzed ( Figures 11A to 12B This study compared subjects with chronic lower back pain (LBP) to those without back pain (NBP). In all trials using MVC as the threshold for the three muscles of interest in at least 2.5% of the subjects, the MMG-EMG correlation coefficients between LBP and NBP were not statistically significant. Figure 13A As shown. Regarding the behavioral advantages of MMG-biofeedback, both the NBP and LBP groups showed significantly greater muscle activation with feedback compared to no feedback (i.e., Figure 13B (Above the dotted line); Among all exercises, the trunk stabilizer TA showed the greatest enhancement, indicating that MMG-feedback improved the quality of back exercises for both user groups.

[0156] As described above, the wearable sensor device 200 may include an inertial measurement unit (IMU) to detect trunk posture and motion measurements. IMU information can add value to the MMG sensing of muscle activation by: (i) automatically recording the type of exercise and number of repetitions (for dynamic exercises); and (ii) ensuring that the exercise is performed as intended with appropriate movement or by remaining still.

[0157] To demonstrate these abilities, 16 subjects were tested while performing six common back exercises, including dynamic and static trunk movements, using one of three postures. These exercises included... Figure 14A As shown. (The validation results of the MMG-EMG Spearman correlation coefficient for all exercises are: r≥0.8 for RA and TA, and r≥0.7 for ES, which is consistent with the results of previous experiments).

[0158] Figure 14B The display shows the signal readings from the four channels of the belt IMU during a standard workout set (indicated by black circles with numbers 1-6), representing accelerations in three linear (x, y, z) planes and one rotational plane. Figure 14BThe IMU curves showed consistency across all subjects, allowing for the identification of exercise type; and for dynamic exercise, they allowed for the identification of repetition count. Figure 14B The graphs show the first eight subjects, but the graphs for the remaining eight subjects are equally discriminative. (For example, a positive y-signal (orange) indicates a supine exercise, while a negative z-signal (red) indicates an upright posture.) Dynamic exercise can be distinguished from static exercise by the fact that the signal variability in any channel is at least ten times greater than that of static exercise.

[0159] Although Figure 14B This demonstrates that IMUs can distinguish between postural and dynamic exercises, but it also reveals the limitations of using IMUs alone for biofeedback: it records static exercise repetitions poorly and cannot differentiate between static exercises such as core support movements and abdominal contractions; or between bird-dog pose and simply resting on all fours. Myographs are needed to make these distinctions.

[0160] IMU data can also indicate whether the exercise achieves the full range of motion required for stretching, or only incompletely. To demonstrate this, nine physical therapists were asked to both perform the exercise correctly and incorrectly, in ways they typically observe in patients. The IMU data from this experiment showed that... Figure 14C The six images on the left. (For example...) Figure 14C As shown in the figure on the right, the bridge pose (lumbar extension) and rotational exercises showed significant reductions in both linear and rotational acceleration, respectively, at both poor and good performance levels (p<0.01). Conversely, exercises that require the spine to remain static while only the limbs move, such as the bird-dog pose, showed excessive trunk acceleration due to trunk tilt at poor performance levels (p<0.05).

[0161] Compared to existing wearable devices that only use IMU for back exercises, the advantages of integrating a myograph (MMG) sensor with an IMU into a wearable sensor device are: (i) distinguishing static trunk exercises and helping patients learn exercises such as abdominal contractions; and (ii) encouraging selective activation of deep trunk stabilizing muscles, such as the TA muscles.

[0162] Figure 15A This demonstrates the value of adding myographic data to IMU data. IMU motion data (top) showed no significant variation between tuck and core support movements, although small repetitions were distinguishable. In contrast, muscle activation showed a huge difference in magnitude when measured using belt-guided MMG (shallower trace) or laboratory EMG (deeper trace, used for MMG validation) (middle and bottom), particularly in the rectus abdominis (RA).

[0163] Figure 15AThe middle and lower figures also show how the MMG belt provides physiological biomarkers that can distinguish between “good” and “bad” repetitions. This can be seen from the relative activation of RA and TA in each trial. In all core support movement trials (right side), RA was highly activated; but for TA, only in some repetitions (r3-r6 marked) was TA equally activated. Since a key goal of therapeutic back exercises is to activate trunk stabilizers such as TA, the relative MMG output of RA and TA can be compared and fed back to the user. In this case, the core support movement repetitions r3-r6 “hit the target” (by… Figure 15A (Indicated by the checkmark in the image below), while other core support actions repeatedly "missed target" (by...). Figure 15A (The cross in the image below indicates this). Previously, this ability to provide selective activation data for training exercises was limited to the laboratory.

[0164] therefore, Figure 15A The results showed that myograph amplitude changes were more effective than IMU in distinguishing between the two exercises and that the relative variability of TA:RA activation could be assessed on a trial-by-trial basis.

[0165] The necessity of distinguishing different core muscle activities, such as Figure 15B As shown, 23 physical therapists or subjects who had received proper training in each exercise were compared with 63 inexperienced subjects during their exercises. Figure 15B As shown, inexperienced subjects exhibited difficulty selectively activating trunk muscles during abdominal contraction and core support movements, consistent with previous studies. Training individuals to selectively activate trunk muscles can alleviate LBP.

[0166] Figure 15C A cumulative confusion matrix is ​​shown, which demonstrates the number of instances where the MMG-based classifier correctly or incorrectly identifies exercise repetitions as non-selective activation (where both RA and TA are <2%), TA-selective activation (where only TA is >5%), or non-selective activation (where both RA and TA are >5%).

[0167] Figure 15CThe MMG-based classifier was confirmed to distinguish between TA and RA activations, and it was shown that MMG can be used to identify selective muscle activation. A decision tree classifier was trained using features extracted from MMG to distinguish repetitions based on EMGRMS during repeated activity exercises. The thresholds used to distinguish between low and high activation repetitions were determined using 1% and 5% MVC. F1 scores of 80.7% (micro) and 76.6% (macro) were obtained on the test dataset. The mean performance metrics were calculated as follows: F1 (micro): 0.80 (95% CI: 0.80–0.81); F1 (macro): 0.81 (0.80–0.81); Accuracy: 0.81 (0.80–0.81). The cumulative confusion matrix was generated through bootstrapping with replacement (1000 iterations) on the test dataset.

[0168] Figure 5 The user interface shown displays real-time feedback and can be configured to indicate to subjects whether they are performing abdominal contractions correctly by visualizing the degree of TA activation. If a subject "hit the target," a virtual reward (e.g., points) can be awarded. Conversely, high RA activation will tell subjects they are actually performing core support, not abdominal contractions. Performing core support instead of abdominal contractions is a common mistake patients make, and currently can only be verified by an experienced physical therapist placing their hand on the patient's abdominal wall.

[0169] For bridging movements, non-physiotherapists tend to have underactivated extremities (ES) compared to physical therapists, so the stimulus will be redirected to ensure ES activity "hits the target." This will be an additional requirement beyond the user needing to achieve the movement target (measured from the IMU), as non-physiotherapists typically do not provide adequate stretching. Figure 15B Software applications can be used to progressively guide users toward achievable muscle activation goals (e.g., “Today I want 5 good TA activations”), which physical therapists can then adjust.

[0170] As from Figures 6B to 15C The robustness of the trunk MMG signal-to-noise ratio was demonstrated by its high correlation with EMG (r > 0.80 for RA and TA; r > 0.70 for ES); and MMG biofeedback improved exercise performance in 100 subjects. High performance was observed in two MMG types, two MMG fixation methods (tape / belt), two anatomical systems (different MMG latitudes, compared to coplanar fixation for belt convenience), four postures, and eight exercises (i.e.,...). Figure 14A The exercises shown, as well as seated crunches and squats, and a wide range of BMIs were maintained.

[0171] In addition, four alternative sensor locations (e.g., for each muscle of interest, RA, TA, and ES) Figure 16 As shown in the figure, the MMG-EMG correlation coefficient (Spearman's r) was determined when ten subjects performed three exercises. The values ​​for the four sites were similar and consistent with those in earlier experiments: for RA and TA, r ≥ 0.8 in the 16 / 24 site-exercise combination; for ES, r ≥ 0.7 in the 7 / 12 combination. This provides confidence that a good signal can still be established even with an inaccuracy of approximately 6 cm in MMG location. Inevitably, some mislocations can lead to poor signal (e.g., located above the bone). To mitigate this, an initial signal testing phase can be implemented where users self-check signal strength while performing test movements such as core support exercises. If the signal is poor, the software application can provide instructions to adjust the belt position (using belt markers) or tightness.

[0172] In addition, three purposeful noise control conditions were tested on 20 subjects: vigorous rubbing of clothing covering the sensors, loud counting by the user, and deep breathing. None of these activated core muscles (as observed in EMG). The MMG signal intensity in the standard exercise test (5%–15% of maximal effort) was significantly higher in the 11 / 12 muscle / control combinations than in all control conditions. Figure 17 As shown. In Figure 17 In the middle, the order of each series corresponds to Figure 17 The order is indicated in the top legend. To further improve the ability to distinguish exercised muscle contractions from noise, a machine learning classifier for MMG signals was trained. This classifier achieved 80%–100% accuracy in detecting at least 5% of MVCs in the presence of three types of noise: RA, TA, and ES (see [link to diagram]). Figure 8B and Figure 8C (and the accompanying discussion).

[0173] Figure 18 A schematic and simplified representation of a computing device 1100 that can be used to perform the methods described herein is shown.

[0174] The computing device 1100 includes various data processing resources, such as a processor 1102 (particularly a hardware processor) coupled to a central bus architecture. Further data processing resources, such as memory 1104, are also connected to the bus architecture. A display adapter 1106 connects a display device 1108 to the bus architecture. One or more user input device adapters 1110 connect user input devices 1112, such as a keyboard and / or touchscreen, to the bus architecture. One or more communication adapters 1114 are also connected to the bus architecture to provide data connectivity to other devices (e.g., via Bluetooth or other suitable interfaces to wearable sensor devices).

[0175] In operation, the processor 1102 of computer system 1100 executes a computer program comprising computer-executable instructions, which may be stored in memory 1104. When executed, the computer-executable instructions cause computer system 1100 to perform one or more methods described herein. The results of the executed processing can be displayed to the user via display adapter 1106 and display device 1108. User input for controlling the operation of computer system 1100 can be received from user input device 1112 via user input device adapter 1110.

[0176] Furthermore, the computer system 1100 may further include an electrical stimulator 1116 configured to apply an electrical current to the user's muscles to activate them. The processor 1102 may be configured to control the stimulation level in response to determining whether a particular muscle has been selectively activated (e.g., as detected by one or more sensors such as an MMG sensor and optionally an IMU sensor).

[0177] Obviously, in some cases, Figure 18 Some features of the computer system 1100 shown may be absent. For example, one or more of the computer devices 1100 may not require a display adapter 1106 or a display device 1108. This may be the case, for example, for a particular server-side computer device 1100 that is only used for its processing capabilities and does not need to display information to a user. Similarly, user input device adapter 1110 and user input device 1112 may not be required. In its simplest form, the computer device 1100 includes a processor 1102 and memory 1104.

[0178] Changes or modifications to the systems and methods described herein are described in the following paragraphs.

[0179] While MMG sensors are better suited for wearable sensor devices than current wired EMG sensors (for reasons stated above), it should be understood that EMG sensors can also be implemented in wearable sensor devices alternatively. In particular, future EMG sensor development may lead to the easier implementation of wireless EMG sensors in wearable sensor devices, and this disclosure is intended to cover such future myographic sensor developments used in the wearable sensor devices described herein. Therefore, myographic sensors implemented in wearable sensor devices may include wireless, wearable EMG sensors.

[0180] Furthermore, while this disclosure aims to provide selective activation feedback to improve the quality of exercise designed to relieve lower back pain, the methods and apparatus described herein can be used for other purposes. For example, muscle spasms that cause pain occur nonselectively in all trunk muscle groups and are temporally nonspecific, and can occur during both rest and exercise. The methods and apparatus described herein can be used to determine whether trunk muscles have been activated (e.g., activated by muscle spasms) and to provide biofeedback to the subject showing how to relax or inhibit excessive muscle activation. This process involves the same signal processing steps as described above, except that the output to the user promotes a reduction in activation rather than an increase in activation.

[0181] Furthermore, although this disclosure focuses on the selective activation of core muscles, it should be understood that the methods and apparatus described herein can also be applied to determine the selective activation of muscles in other parts of the body.

[0182] The wearable sensor devices 200 and 500 described above can also be used to classify trunk postures based on the activation of specific trunk muscles. Specifically, a classifier can be implemented to predict whether a subject's posture is "good" or "poor" based on MMG sensor data. If a "poor" posture is predicted, feedback can be provided to the subject. For example, the subject can be given a reminder to improve their posture.

[0183] Figure 21 Box plots show the MMG RMS values ​​of TA / IO and ES muscles when subjects were in a slouching ("poor" posture) and an upright ("good" posture) posture. From Figure 21 As can be seen, a significant increase in TA / IO activation was observed during the upright posture using the MMG sensor (p = 0.0396), and ES also showed a significant difference between the upright and relaxed postures (p = 0.0033).

[0184] Therefore, one or more of the TA / IO MMG sensor and ES MMG sensor can be used to determine whether a subject wearing wearable sensor devices 200, 500 is in a slouching posture (“poor” posture) or an upright sitting posture (“good” posture). “Good” and “poor” standing postures result in similar differences in TA / IO and ES activation, meaning that one or more of the TA / IO MMG sensor and ES MMG sensor can also be used to determine whether a subject wearing wearable sensor devices 200, 500 is in a slouching standing posture (“poor” posture) or an upright standing posture (“good” posture). Therefore, a reminder or nudge can be given to the user to return to a good posture and generate behavioral change in the user. The reminder or nudge can be given using any biofeedback method discussed at 130 of method 100.

[0185] Figure 22 A flowchart of a method 600 for classifying subject postures is shown. This method can be performed by one or more processors of a device (e.g., Figure 18 The processor 1102 of the computing device 1100 shown is implemented.

[0186] At 610, sensor data is received from one or more myographic sensors of the wearable sensor device. The one or more myographic sensors are configured to monitor muscle activation in the subject (e.g., TA / IO activation and / or ES activation). The sensor data may be received in the same manner as at 110 of method 100.

[0187] At 620, the subject's posture is predicted from multiple posture classifications. This posture can be predicted by a classifier trained using ground truth data on whether the subject's sitting posture is slouching ("poor") or upright ("good"), in which case the multiple posture classifications consist of slouching and upright sitting postures. Therefore, the multiple posture classifications from which the subject's posture is predicted are the same as those used as ground truth data during classifier training. The classifier can be trained using training data comprising features extracted from TA / IO and / or ES MMG sensor data associated with the time periods during which each subject is in a slouching or upright posture. After training, the classifier predicts whether the subject is in a slouching or upright posture based on the features extracted from the TA / IO and / or ES MMG data collected for the subject. The features extracted from the MMG sensor data can be the same features discussed at 340 of method 300.

[0188] It should be understood that the classifier can alternatively be trained using additional ground truth poses, in which case the classifier can predict the subject's pose from a larger set of pose classifications. As an alternative to the 620 examples mentioned above, upright and relaxed standing postures can be used instead of upright and relaxed sitting postures.

[0189] At 630, based on the posture classification at 620, biofeedback is selectively provided to the subject. If the subject's posture is classified as upright (or "good") at 620, no biofeedback may be provided at 630. However, if the subject's posture is classified as slouching (or "poor") at 620, a reminder or nudge may be provided at 630 using the methods for providing biofeedback discussed at 130 of method 100.

[0190] The wearable sensor devices 200 and 500 described above can also be used to classify pain levels based on the activation of specific trunk muscles. Specifically, a classifier can be implemented to predict whether a subject is experiencing pain, or whether the subject is experiencing a specific level of pain, based on MMG sensor data and / or exercise-related data. If the subject is predicted to experience pain or a threshold pain level, feedback can be provided to the subject. For example, instructions to adjust the ongoing exercise can be given. Pain levels can also be tracked over time to provide exercise recommendations related to a reduction in pain levels.

[0191] Figure 23 A flowchart of a method 700 for classifying the pain level of a subject is shown. This method can be performed by one or more processors of a device (e.g., Figure 18 The processor 1102 of the computing device 1100 shown is implemented.

[0192] At 710, sensor data is received from one or more myographic sensors of the wearable sensor device. The one or more myographic sensors are configured to monitor muscle activation in the subject (e.g., TA / IO activation and / or ES activation). The sensor data may be received in the same manner as at 110 of method 100.

[0193] In 720, the subject's pain level is predicted from multiple pain level categories. These multiple pain level categories can be binary categories (e.g., lower back pain, no lower back pain) or hierarchical categories using multiple subjective pain scores (e.g., a scale of 1 to 10, where 1 represents mild or very mild pain, 5 represents moderate pain, and 10 represents severe pain). Subjective pain scores can be entered by the subject via a user interface (e.g., user interface 400) after one or more workouts.

[0194] Pain levels can be predicted by a classifier trained using ground truth data and training data. For binary classification, ground truth data can include a label "1" indicating a medical diagnosis of back pain and a label "0" indicating the absence of a medical diagnosis of back pain. For graded classification, ground truth data can include pain scores provided by subjects during exercise while wearing wearable sensor devices 200 and 500. For example, subjects can perform a specific exercise and input a subjective pain score, indicating the level of pain experienced while performing that exercise. Subjective pain scores can also be associated with a set of exercises.

[0195] The classifier can be trained using training data comprising features extracted from TA / IO and / or ES MMG sensor data associated with periods of one or more exercises for each subject that are associated with a specific ground truth pain classification. The extracted MMG sensor data features can be the same features discussed at 340 of method 300. Specifically, the extracted features can include RMS, mean frequency, and maximum activation during the exercise period (i.e., the contraction period). The classifier can also be trained using features associated with one or more exercises that are associated with the ground truth pain classification. For example, features regarding exercise duration, exercise type, and exercise intensity can also be used as training data. A neural network (e.g., a recurrent neural network, a gated recurrent unit, or a long short-term memory network) can be trained using the ground truth pain classification and training data to predict future pain classifications based on the extracted MMG features and / or exercise duration features.

[0196] The purpose of categorizing the pain levels of 720 subjects was to determine whether subjects with back pain could be encouraged to change their muscle activation patterns over time to mimic those of subjects who did not experience back pain. The study showed that asymmetry between muscle groups and different muscle activation patterns are associated with the presence of back pain.

[0197] At 730, biofeedback is selectively provided to the subject based on the pain level classification at 720. If the subject's pain level is classified as low at 720, no biofeedback may be provided at 730. However, if the subject's pain level is classified as high (or exceeds a certain threshold) at 720, a cue may be provided at 730 using the methods for providing biofeedback discussed at 130 of method 100. The cue provided at 730 may instruct the subject to adjust the ongoing exercise. In particular, at 730, the subject may be instructed to exercise using exercise characteristics (e.g., duration, intensity, exercise type) associated with the low pain level classification.

[0198] Subjective pain scores input by subjects can be tracked over time using a reinforcement learning-based recommendation model. By tracking subjective pain scores over time, correlations between subjective pain scores and extracted MMG features and / or exercise features can be identified. For example, by tracking subjective pain scores over a period of time, one or more features (i.e., extracted MMG features and / or exercise features) can be identified as being associated with a reduction in subjective pain scores. By identifying the correlation between one or more features and a reduction in subjective pain scores, personalized recommendations can be provided to subjects based on real-time MMG and / or exercise data (e.g., in 730). For example, the model can recommend different actions to subjects based on incoming MMG / exercise data, such as increasing or decreasing exercise intensity, increasing or decreasing exercise difficulty and / or contraction duration, increasing or decreasing rest periods, or changing the exercise.

[0199] The goal of a reinforcement learning-based recommendation model is to increase long-term health benefits for participants and minimize pain experienced by them. A "positive reward tag" can be provided to the model if a participant's pain score decreases or remains unchanged after following the model's recommendations (based on the participant's subjective pain score after following the model's recommendations), and a "negative reward tag" can be provided if a participant's pain score increases after following the model's recommendations. Positive reward tags can also be weighted by improvements in muscle strength or endurance observed based on MMG data. For example, a positive reward tag can be weighted by improvements in muscle activation (e.g., symmetrical muscle activation or increased endurance) observed in mimicking healthy participants, identified from MMG data. Positive reward tags can be used for recommendations proposed by a positively weighted model, while negative reward tags can be used for recommendations proposed by a negatively weighted model.

[0200] Therefore, reinforcement learning-based recommendation models can process incoming MMG data and update the model's state (i.e., MMG data associated with increased or decreased pain). Based on the model's state, the reinforcement learning-based recommendation model can recommend actions. For example, if the MMG data is associated with decreased pain, the model might recommend continuing the exercise. Or, if the MMG data is associated with increased pain, the model might recommend the subject take a break. In either case, the model might recommend increasing the intensity (depending on whether increasing the intensity of a particular exercise is associated with an increase or decrease in subjective pain). Based on reward tags associated with the pain level experienced by the subject after following the model's recommendations, the recommendations provided by the model can be adjusted to prevent pain attacks before they occur, while allowing the subject to safely continue their rehabilitation journey.

[0201] The method described can be implemented using computer-executable instructions. A computer program product or computer-readable medium may include or store computer-executable instructions. A computer program product or computer-readable medium may include hard disk drives, flash memory, read-only memory (ROM), CDs, DVDs, caches, random access memory (RAM), and / or any other storage medium for storing information for an unlimited duration (e.g., long time periods, permanent, brief moments, temporary buffers, and / or information caches). A computer program may include computer-executable instructions. A computer-readable medium may be a tangible or non-transitory computer-readable medium. The term "computer-readable" includes "machine-readable."

[0202] The singular terms “one” and “a” should not be interpreted as meaning “only one”. Rather, unless otherwise stated, they should be interpreted as meaning “at least one” or “one or more”. The word “including” and its derivatives, including “include (third person)” and “contain”, include each of the stated features, but do not exclude the inclusion of one or more additional features.

[0203] The above embodiments are described by way of example only, and the described embodiments should be considered illustrative rather than restrictive in all respects. It should be understood that variations can be made to the described embodiments without departing from the scope of the invention. It will also be apparent that many variations not described exist, but these variations fall within the scope of the appended claims.

Claims

1. A computer-implemented method, comprising: Sensor data is received from one or more myographic sensors of a wearable sensor device, wherein the one or more myographic sensors are configured to monitor muscle activation in a subject; Based on the sensor data, it is determined whether the subject's first muscle has been selectively activated; as well as Based on the determination of whether the first muscle has been selectively activated, biofeedback is provided.

2. The method according to claim 1, wherein, Determining whether the subject's first muscle has been selectively activated includes: During the calibration process, a calibration activation value for the first muscle is determined; and Determine the degree of activation of the first muscle relative to the calibrated activation value.

3. The method according to claim 2, wherein, Determining whether the subject's first muscle has been selectively activated includes determining whether the degree of activation of the first muscle exceeds a threshold percentage of the calibrated activation value.

4. The method according to any one of claims 1 to 3, wherein, The one or more myograph sensors include one or more mechanical myograph (MMG) sensors.

5. The method according to any one of claims 1 to 4, wherein, Determining whether the first muscle has been selectively activated based on the sensor data includes classifying the sensor data using one or more trained classifiers.

6. The method according to claim 5, wherein, The one or more trained classifiers are configured to distinguish between sensor data indicating muscle activation and sensor data indicating muscle rest.

7. The method according to claim 6, wherein, The one or more trained classifiers are trained using electromyography (EMG) sensor data indicating muscle activation and EMG sensor data indicating muscle rest.

8. The method according to any one of claims 5 to 7, wherein, The one or more trained classifiers are configured to identify sensor data indicating the degree of activation of the first muscle relative to the subject's second muscle.

9. The method according to claim 8, wherein, The one or more trained classifiers are trained using EMG sensor data indicating the degree of activation of the first muscle relative to the subject's second muscle.

10. The method according to any one of claims 5 to 9, wherein, The one or more trained classifiers are configured to identify the activation intensity of the first muscle.

11. The method according to claim 10, wherein, The one or more trained classifiers are trained using EMG sensor data that indicates different activation intensities of the first muscle.

12. The method according to any one of claims 1 to 11, wherein, Determining whether the subject's first muscle has been selectively activated includes determining the degree of activation of the first muscle relative to the subject's second muscle.

13. The method according to claim 12, wherein, A first myographic sensor of the one or more myographic sensors is configured to monitor muscle activation of the first muscle, and a second myographic sensor of the one or more myographic sensors is configured to monitor muscle activation of the second muscle.

14. The method according to any one of claims 1 to 13, wherein, The first muscle is the core muscle.

15. The method according to any one of claims 1 to 14, wherein: The first muscle is the transverse abdominis, and the first myographic sensor of the one or more myographic sensors is configured to monitor muscle activation of the transverse abdominis; or The first muscle is the internal oblique muscle, and the first myograph sensor of the one or more myograph sensors is configured to monitor muscle activation of the internal oblique muscle.

16. The method of claim 15, wherein, when subordinate to claim 12, the second muscle is the rectus abdominis, and the second myographic sensor of the one or more myographic sensors is configured to monitor muscle activation of the rectus abdominis.

17. The method according to any one of claims 1 to 14, wherein, The first muscle is the erector spinae, and the first myograph sensor of the one or more myograph sensors is configured to monitor muscle activation of the erector spinae.

18. The method according to any one of claims 1 to 17, further comprising: Motion sensor data is received from one or more motion sensors configured to monitor the subject's movements; as well as The motion sensor data is processed to determine the subject's posture.

19. The method according to any one of claims 1 to 18, wherein: The sensor data is received from the wearable sensor device including the one or more myocardial sensors; and Providing biofeedback includes providing commands to cause the wearable sensor device to provide one or more of tactile, audio, and visual feedback to the subject.

20. A computer-readable medium comprising instructions that, when executed by at least one processor of a device, cause the device to perform the method according to any one of claims 1 to 19.

21. An apparatus comprising: At least one processor; as well as At least one memory includes computer-readable instructions that, when executed by the at least one processor, cause the apparatus to perform the method according to any one of claims 1 to 19.

22. The apparatus according to claim 21, wherein, The device further includes at least one electrical stimulation device configured to activate the subject's first muscle.

23. The apparatus according to claim 22, wherein, When the instruction is executed by the at least one processor, the device, in response to receiving the biofeedback, changes the charge applied to the muscle and / or nerve fibers by the electrical stimulation device.

24. A wearable sensor device configured to be worn by a subject, the wearable sensor device comprising: One or more myograph sensors are configured to monitor muscle activation in the subject; as well as A transmitter for transmitting sensor data from the one or more myograph sensors to at least one processor in a separate device.

Images