Gait assistance systems and methods of control thereof

Abstract

Determining predicted torque values for a powered gait assistance system are described. The method comprises: obtaining a training dataset comprising training sensor data associated with measurements of a plurality of first individuals using one or more first powered gait assistance systems; training a multilayer perceptron (MLP) using the training sensor data; obtaining, from one or more sensors of a powered gait assistance system, sensor data associated with a second individual using the powered gait assistance system; processing the sensor data associated with the second individual using the trained MLP to generate predicted torque values for a second powered gait assistance system; and applying the predicted torque values to a motor of the second powered gait assistance system to cause the motor to apply assistive torque, that corresponds to the predicted torque values, between the first and second arms of the second powered gait assistance system.

Description

GAIT ASSISTANCE SYSTEMS AND METHODS OF CONTROL THEREOFSTATEMENT OF GOVERNMENT SUPPORT

[0001] This invention was made with Government support under project number Z01 CL090099 by the National Institutes of Health, National Cancer Institute Clinical Center. The Government has certain rights in the invention.CROSS-REFERENCE TO PRIOR APPLICATION

[0002] Priority is claimed to U.S. Provisional Application Serial No. 63 / 539,898 filed on September 22, 2023, the entire contents of which is hereby incorporated by reference herein.FIELD

[0003] This application is related to the field of powered systems for assisting human ambulation and other activities.BACKGROUND

[0004] Cerebral palsy (CP) is a group of motor disabilities caused by a perinatal nonprogressive injury to the central nervous system. CP is the most common pediatric motor disorder, affecting 1 to 4 per 1000 live births. Individuals with CP present a variety of disabilities including spasticity, rigidity and diminished coordination and motor control. Individuals with CP are often partitioned into subgroups based on the type and distribution of motor impairments. One of the most common subgroups is spastic diplegia, in which pathologic gait patterns are frequently observed. Crouch gait, which is characterized by excessive stance phase knee flexion and may be accompanied by other deficits at the hip and / or ankle, is one of the most frequently observed gait deviations. Crouch gait can lead to joint pain and degenerative arthritis due to elevated joint contact forces, bony deformities, and increased risk for falls from inadequate foot clearance. Additionally, crouch gait is less efficient than normal walking patterns and many experience deterioration of walking ability with age, leading to a loss of ambulation in a large portion of the population.

[0005] The most common treatments for crouch gait include surgery, botulinum toxin injections, physical therapy and strengthening, and orthotic interventions. Outcomes from surgical interventions, which typically target the hamstrings, are variable. Distal femoral extensionosteotomy procedures have shown improvement of knee extension in short-term follow up, particularly when combined with patellar tendon advancement, offering a potentially effective but more invasive solution. One reason for the mixed surgical outcomes may be heterogeneity underlying the crouch gait pattern in diplegic CP. In addition to hamstring spasticity, anti -gravity (extensor) weakness contributes to the deterioration of walking capabilities as a function of age. However, like surgery, muscle strengthening programs have inconsistent outcomes. Achieving and maintaining adequate strength levels in individuals with motor disabilities is a challenging problem that is exacerbated because ambulatory children with CP are less active than children without disabilities. Orthotic bracing, aimed to increase mobility of individuals with crouch gait, may block or restrict motion at the ankle and or the knee joint to provide passive weight support or to suppress unwanted motions. Traditional and floor reaction ankle-foot orthoses (AFO and FRAFO, respectively) have been shown to temporarily improve knee extension and spatiotemporal gait parameters while worn. However, long term use can lead to greater weakness in the restricted muscle groups over time. Treadmill based robotic assisted therapy offers the advantage of repetitive practice. Improved function has been reported in pilot studies of treadmill based training in CP, but their overall effectiveness compared to traditional therapies of equal intensity is equivalent. Furthermore, these strategies are limited in the environment and duration in which they can be applied and have not been explicitly studied for remediation of crouch gait. Functional electrical stimulation (FES) devices have been applied to treat crouch gait in children with CP with limited success using percutaneous and surface electrodes to stimulate extensor muscles; however, the intensity of stimulation required to extend the flexed knee can be difficult to reach.

[0006] Thus, there is a clear need for more effective interventions which can preserve and / or augment strength on a continuous basis for those with crouch gait from CP.SUMMARY

[0007] In an exemplary embodiment, the present disclosure provides a method, comprising: obtaining a training dataset comprising training sensor data associated with measurements of a plurality of first individuals using one or more first powered gait assistance systems; training a multilayer perceptron (MLP) using the training sensor data; obtaining, from one or more sensors of a powered gait assistance system, sensor data associated with a second individual using the powered gait assistance system; processing the sensor data associated with the second individualusing the trained MLP to generate predicted torque values for a second powered gait assistance system; and applying the predicted torque values to a motor of the second powered gait assistance system to cause the motor to apply assistive torque, that corresponds to the predicted torque values, between first and second arms of the second powered gait assistance system to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

[0008] In some instances, processing the sensor data comprises: inputting the sensor data into the trained MLP to determine a predicted gait phase from a plurality of gait phases of the gait cycle; and determining the predicted torque values for the powered gait assistance system based on the predicted gait phase.

[0009] In some examples, the one or more sensors comprises a torque sensor, an angle sensor, a foot sensor, and / or an inertial based measurement (IMU) sensor, wherein the measurements from the training sensor data and the sensor data includes kinematic measurements and dynamic measurements that are obtained by the one or more sensors of the first and the second powered gait assistance systems.

[0010] In some variations, a transition between gait phases of the gait cycle is determined by the MLP using the sensor data or transition thresholds for the gait phases are determined by the MLP using the sensor data, and wherein each gait phase of the gait phases includes a certain assistive torque applied by the motor.

[0011] In some instances, obtaining the sensor data associated with the second individual comprises: obtaining actual sensor data from the one or more sensors of the powered gait assistance system; obtaining surface electromyography (sEMG) data from one or more sEMG sensors of the powered gait assistance system; and processing the sEMG data using an attention-based recurrent neural network (RNN) to generate predicted joint angles, wherein processing the sensor data associated with the second individual comprises processing, using the trained MLP, the actual sensor data and the predicted joint angles that are generated using the attention-based RNN to generate the predicted torque values.

[0012] In some examples, the method further comprises processing the sensor data associated with the second individual using the trained MLP to generate a binary value that indicates whether to provide extension assistance to the second individual.

[0013] In some variations, the method further comprises training the attention-based RNN by: receiving sEMG training data from a data source, wherein the attention-based RNN comprises anencoder, a decoder, and an attention mechanism; training, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input to the decoder; and subsequent to training the encoder and the decoder, using a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided to an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

[0014] In some instances, the one or more sEMG sensors are located in an area that corresponds to a foot print of the powered gait assistance system.

[0015] In some examples, the actual sensor data from the one or more sensors is used as ground truth data, wherein training the attention mechanism includes using the ground truth data.

[0016] In another exemplary embodiment, the present disclosure provides a computing environment, comprising: a training computing system configured to: obtain a training dataset comprising training sensor data associated with measurements of a plurality of first individuals using one or more first powered gait assistance systems; train a multilayer perceptron (MLP) using the training sensor data; and provide the trained MLP to the powered gait assistance device; the powered gait assistance system, comprising: a first arm and a second arm coupled together to allow relative pivoting between the first and second arms, wherein the first arm is configured to be coupled to an upper leg portion of a second individual and the second arm is configured to be coupled to a lower leg portion of the second individual such that relative pivoting between the first and second arms is about an axis proximate the second individual’s native knee pivot axis; a motor coupled to the first and second arms and operable to apply torque between the first and second arms to urge relative pivoting between the first and second arms; one or more sensors that are configured to provide sensor data; and a controller programmed to: obtain the sensor data associated with the second individual, the sensor data obtained by the one or more sensors as the second individual uses the powered gait assistance system; process the sensor data associated with the second individual using the trained MLP to generate predicted torque values for the powered gait assistance system; and apply the predicted torque values to the motor to apply assistive torque, that corresponds to the predicted torque values, between the first and second arms to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

[0017] In some instances, processing the sensor data comprises: inputting the sensor data into the trained MLP to determine a predicted gait phase from a plurality of gait phases of the gait cycle; and determining the predicted torque values for the powered gait assistance system based on the gait phase.

[0018] In some examples, the one or more sensors comprises a torque sensor, an angle sensor, a foot sensor, and / or an inertial based measurement (IMU) sensor, wherein the measurements from the training sensor data and the sensor data includes the kinematic measurements and dynamic measurements that are obtained by the one or more sensors of the powered gait assistance system.

[0019] In some variations, a transition between gait phases of the gait cycle is determined by the MLP using the sensor data or transition thresholds for the gait phases are determined by the MLP using the sensor data, and wherein each gait phase of the gait phases includes a certain assistive torque applied by the motor.

[0020] In some instances, obtaining the sensor data associated with the second individual comprises: obtaining actual sensor data from the one or more sensors; obtaining surface electromyography (sEMG) data from one or more sEMG sensors of the powered gait assistance system; and processing the sEMG data using an attention-based recurrent neural network (RNN) to generate predicted joint angles, wherein processing the sensor data associated with the second individual comprises processing, using the trained MLP, the actual sensor data and the predicted joint angles that are generated using the attention -based RNN to generate the predicted torque values.

[0021] In some examples, processing the sensor data associated with the second individual using the trained MLP includes generating a binary value that indicates whether to provide extension assistance to the second individual.

[0022] In some variations, the training computing system is further configured to train the attention-based RNN by: receiving sEMG training data from a data source, wherein the attentionbased RNN comprises an encoder, a decoder, and an attention mechanism; training, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input to the decoder; and subsequent to training the encoder and the decoder, using a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided to an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

[0023] In yet another exemplary embodiment, the present disclosure provides a method, comprising: obtaining a training dataset comprising surface electromyography (sEMG) training data from a data source, wherein the sEMG training data is associated with a plurality of first individuals; training an attention-based model using the sEMG training data, the attention-based model comprising an encoder, a decoder, and an attention mechanism; obtaining sEMG data of a second individual using a powered gait assistance system; processing the sEMG data associated with the second individual using the trained attention-based model to generate predicted joint angles for the second individual; and based on the predicted joint angles, instructing a motor of the powered gait assistance system to apply assistive torque between a first arm and second arm of the powered gait assistance system, the assistive torque corresponding to a stage of the second individual’s gait cycle, to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

[0024] In some instances, training the attend on -based model comprises: training, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input to the decoder; and subsequent to training the encoder and the decoder, using a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided as an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

[0025] In some examples, the method further comprises: providing the predicted joint angles and actual sensor data from one or more sensors of the powered gait assistance system to a trained multilayer perceptron (MLP), the actual sensor data captured by the one or more sensors as the second individual uses the powered gait assistance system; processing the actual sensor data and the predicted joint angles using the trained MLP to generate predicted torque values for the powered gait assistance system; and applying the predicted torque values to the motor of the powered gait assistance system to cause the motor to apply the assistive torque that corresponds to the predicted torque values.

[0026] In some variations, processing the actual sensor data comprises: inputting the actual sensor data into the trained MLP to determine a gait phase from a plurality of gait phases of the gait cycle; and determining the predicted torque values for the powered gait assistance system based on the gait phase.

[0027] All examples and features mentioned above may be combined in any technically possible way.BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present application will be described in even greater detail below based on the exemplary figures. The application is not limited to the examples described below. All features described and / or illustrated herein can be used alone or combined in different combinations in examples of the application. The features and advantages of various examples of the present application will become apparent by reading the following detailed description with reference to the attached drawings which illustrate the following:

[0029] FIG. 1 shows a child with crouch gait wearing an exemplary gait assistance system comprising orthotic braces and motorized devices that provide assistive torque at the knees during selected portions of the gait cycle.

[0030] FIG. 2 is a lateral-anterior view of another exemplary gait assistance system comprising thigh, shank, and foot orthotics and a modular powered knee portion that provides assistive torque at the knee joint.

[0031] FIG. 3 is a medial-anterior view of the system of FIG. 2.

[0032] FIG. 4 is a lateral -posterior view of the modular powered knee portion of FIG. 2.

[0033] FIG. 5 is a medial -posterior view of the modular powered knee portion of FIG. 2.

[0034] FIG. 6 is a table showing comparative data for kinematic, muscle activity, and spatiotemporal parameters measured during clinical evaluation of an exemplary gait assistance system.

[0035] FIG. 7 is a diagram of an exemplary finite state machine used to specify desired knee extensor assistance at different phases of the gait cycle.

[0036] FIG. 8 is a diagram illustrating the gait cycle and exemplary portions of the gait cycle when knee extensor assistance can be provided and not provided.

[0037] FIG. 9 is a schematic illustrating an exemplary control architecture for a gait assistance system.

[0038] FIG. 10 is a graph showing mean angular compliance between a patient and an exemplary gait assistance system throughout the gait cycle.

[0039] FIGS. 11A and 11B show mean knee extensor torques applied by an exemplary gait assistance system across the gait cycle, as measured by an onboard torque sensor at the knee.FIGS. 11C-11H show mean sagittal plane hip, knee, and ankle joint angles across the gait cycle during baseline, free, and powered system conditions. Shaded regions represent plus-minus one standard deviation from the mean.

[0040] FIGS. 12A-12H show mean EMG linear envelopes (in mV) across the gait cycle for various leg muscles during free and powered system conditions. Shaded regions represent plusminus one standard deviation from the mean.

[0041] FIG. 13 shows an exemplary gait assistance system comprising motors oriented along lateral aspects of the thighs and coupled to knee joints via right-angle drives.

[0042] FIG. 14 shows an exemplary gait assistance system comprising torso mounted cabledrive motors coupled to knee joints via tensioned cables.

[0043] FIG. 15 shows an exemplary gait assistance system comprising a linear actuator coupled between a thigh brace and a shank brace to provide torque at the knee joint.

[0044] FIG. 16 shows an exemplary gait assistance system comprising a thigh mounted motor that drives a cable coupled to a shank portion along an arcuate path to provide torque about the knee joint.

[0045] FIG. 17 shows an exemplary gait assistance system comprising a body mounted system controller that is wirelessly coupled to a remote control system and / or to a cloud based computing environment / data base.

[0046] FIGS. 18A and 18B illustrate an exemplary computer game system that can be used to help train patients using disclosed powered gait assistance systems.

[0047] FIGS. 19 and 20 are screen shots from an exemplary computer game usable to help train patients using disclosed powered gait assistance systems.

[0048] FIG. 21 is a simplified block diagram depicting an exemplary computing environment in accordance with one or more examples of the present application.

[0049] FIG. 22 is a simplified block diagram of one or more devices or systems within the exemplary environment of FIG. 21.

[0050] FIG. 23 is an exemplary process for training a multilayer perceptron (MLP) using training sensor data and generating predicted torque values for the disclosed powered gait assistance system using the trained MLP and sensor data in accordance with one or more examples of the present application.

[0051] FIG. 24 shows an exemplary first machine learning - artificial intelligence (ML - Al) model that is used to generate the predicted torque values in accordance with one or more examples of the present application.

[0052] FIG. 25 is an exemplary process for obtaining surface electromyography (sEMG) data and using a trained attention-based recurrent neural network to generate predicted joint angles as well as using the predicted joint angles and actual sensor data from sensors of the powered gait assistance system to generate the predicted torque values by the trained MLP in accordance with one or more examples of the present application.

[0053] FIG. 26 shows an exemplary second ML - Al model that is used to generate predicted joint angles in accordance with one or more examples of the present application.

[0054] FIG. 27 is an exemplary process for training an attention-based recurrent neural network in accordance with one or more examples of the present application.DETAILED DESCRIPTION

[0055] Examples of the presented application will now be described more fully hereinafter with reference to the accompanying exemplary figures, in which some, but not all, examples of the application are shown. Indeed, the application may be embodied in any different forms and should not be construed as limited to the examples set forth herein; rather, these examples are provided so that the disclosure will satisfy applicable legal requirements. Where possible, any terms expressed in the singular form herein are meant to also include the plural form and vice versa, unless explicitly stated otherwise. Also, as used herein, the term “a” and / or “an” shall mean “one or more” even though the phrase “one or more” is also used herein. Furthermore, when it is said herein that something is “based on” something else, it may be based on one or more other things as well. In other words, unless expressly indicated otherwise, as used herein “based on” means “based at least in part on” or “based at least partially on”.

[0056] Disclosed herein are powered gait assistance systems (sometimes alternatively referred to as robotic systems or exoskeletons) that can function as mobility aids and as training devices. The disclosed systems can include a programmed controller, various sensors, and a motor coupled to a patient’s leg (e.g., at the upper leg / thigh and at the lower leg / shank) and operable to apply assistive and / or resistive torque to the patient’s knee joint to assist or resist the patient’s volitional knee extensor muscle output during selected stages of the patient’s gait cycle. The sensors caninclude a torque sensor that measures torque applied by the motor and / or the user at the knee joint, muscle output sensors to track volitional effort exerted by the patient, a knee angle sensor, a foot- located sensor that measures ground contact of the patient’s foot, and / or other sensors. The controller can determine what stage of the patient’s gait cycle the patient’s leg is in based on sensor inputs and cause the motor to apply corresponding assistive or resistive torque to the knee based on the gait cycle stage, sensor inputs, known patient characteristics, and / or other factors.

[0057] In an embodiment, the controller can incorporate artificial intelligence (Al) algorithms to predict joint angles of the leg during different stages of the gait cycle. In particular, an attentionbased model (e.g., an attention-based recurrent neural network (RNN)) can be utilized to predict the joint angles (e.g., knee joint angles) based on surface electromyography (sEMG). One potential application of this technology is to enable more ubiquitous use of an exoskeleton beyond rote overground walking by enabling real-time prediction of joint movements. The controller can include a gated recurrent unit (GRU) neural network. The GRU neural network may also enable transfer learning, such that the model can be trained on a particular data set from one individual or a group of (similar) individuals and then updated when applied to data collected from a new individual.

[0058] In another embodiment, the controller may implement a finite state machine (FSM), which is a closed-loop control algorithm. The FSM can include a variable number of states corresponding to the different stages of the gait cycle. The FSM serves as the highest level of a hierarchical control system. Within each state of the FSM, a separate lower-level real-time control algorithm can be activated to provide assistive or resistive torques to the knee to provide mobility assistance, gait training, strength training, or the like.

[0059] Transitions between states are managed in accordance with transition thresholds, which may include pre-tuned values for one or more sensor inputs. For example, to transition from a state corresponding to a late stance stage of the gait cycle to another state corresponding to an early swing stage of the gait cycle, the foot pressure sensor reading must exceed a pre-set threshold. A key process for successful exoskeleton operation is to tune these transition thresholds, which is conventionally done manually based on visual inspection of sensor data, in concert with other gait kinematics.

[0060] In an embodiment, a multi-layer perceptron (MLP) network can be implemented in the controller for tuning these threshold parameters automatically. The user may first wear theexoskeleton to collect sensor data particular to the user. During this stage, the exoskeleton is not activated to provide any assistive or resistive torque. This data is then labeled to provide the ground-truth finite states for the collected sensor data. The MLP is then trained using the collected data and the ground-truth labels. Once trained, the MLP can be used during inference to automatically identify the correct sensor variables to maximize the accuracy of the state determination. During use, sensor data can be continually collected and used to update the transition thresholds based on the determined states, and the relative accuracy of the output of the MLP.

[0061] Disclosed systems and methods can be particularly beneficial as a treatment for crouch gait, such as in in children with cerebral palsy (CP). The disclosed systems can be based on the architecture of a conventional orthotic device (e.g., a KAFO), can be lightweight, adjustable over time, and modular, making them suitable for growing and developing children. Moreover, the disclosed technology can help improve the patient’s native ability to walk over time (e.g., by strengthening muscles, improving native muscle output, correcting posture, and / or improving coordination) such that the patient may eventually no longer need assistive devices to walk with a sufficiently “normal” and efficient gait pattern.

[0062] In some embodiments, the system includes a transmission system coupling the motor to the knee joint. The transmission can include a chain or cable transmission system for example. A chain or cable transmission can allow placement of the motor farther from the knee joint. A transmission system can also be used to increase the torque output of the motor or increase the angular velocity output of the motor, as needed for a particular application.

[0063] In some embodiments, the system can assist patient knee joint extension during a late swing phase of the gait cycle prior to foot touch down, can assist patient knee joint extension during a ground contact phase of the gait cycle prior to toe take off, and / or can assist patient knee joint flexion during an early swing phase of the gait cycle after toe take off. In some embodiments, the motor applies no assistive torque or resistance during one or more portions of the gait cycle, such as an early swing phase.

[0064] In various embodiments, the motor can be located in several different locations and still be operable to apply assistive torque to the patient. For example, in some embodiments, the motor is positioned anterior to the patient’s knee, in some embodiments, the motor is positionedlateral to the patient’s thigh, and in some embodiments, the motor is positioned on the patient’s torso or hips.

[0065] In some embodiments, the controller is programmed to wirelessly communicate with a remote computing system and / or a cloud based communication system. This can allow for remote communication with a clinician and / or active adjustments from a clinician or operator, for example. This can also allow the system to feedback data to a database and receive software and firmware updates. This can also allow heavy computational work to be performed remotely, minimizing the size and cost of the disclosed systems.

[0066] The powered gait assistance systems can also be used with or as part of a computerized gaming / training system. For example, the patient’s leg motions can control action in a computer game or other displayed environment, such that the patient is encouraged to bend and extend his knees to make an object in the game / display move. The patient may also wear a virtual reality headset as part of the gaming / training system.Exemplary Powered Gait Assistance Devices

[0067] The varying postures and physical deformities present in children with CP, combined with the heterogeneous causes of crouch gait, create a need for versatile, adjustable, and adjustable assistance devices, having human-machine interfaces that can be customized to each individual. Some disclosed powered gait assistance devices can be used with existing leg orthotics, such as knee-ankle-foot orthoses (KAFOs) for example. Conventional KAFOs can comprise custom molded shells for the foot, shank, and / or thigh segments connected by rigid uprights mounted along the side of the leg (see, e.g., FIG. 1). KAFOs can employ joint mechanisms in parallel with anatomical joint centers to allow (or control) segment motion, but typically do not include any powered assistance.

[0068] To generate additional assistive torque at the knee joint, the disclosed devices can comprise a powered motor that is coupled via a power transmission to a pivot joint that is positioned adjacent to the patient’s knee (e.g., the pivot joint can be part of a KAFO or coupled to the knee joint of a KAFO). The motor and transmission can be positioned in various locations and the transmission can take various configurations, such as a chain drive, cable drive, right-angle drive, or direct drive, for example. The motor and transmission are desirably light weight, low- profile, and capable of providing a sufficient level of torque assistance to the patient. Of coursethe torque output, and therefore the size and weight, of the device that is sufficient depends on the size and pathology of each particular patient.

[0069] Internal knee extensor moment output scales with the degree of crouch and body mass, but rarely exceeds and is often much less than 50 Nm for children with mild to moderate crouch. As such, in powered gait assistance systems for such patients, peak torque output need not be very high, which can help reduce the size and power consumption of the device and increase compatibility of the patient’s daily life activities. For example, in some embodiments, the powered device can be configured to provide about one third of the internal demand on knee extensors for patients with mild-moderate crouch (e.g., up to about 17 Nm of torque assistance). To provide such levels of torque at the knee joint, a transmission system can be utilized to increase a lower level of torque output from the motor, allowing for smaller motors. However, using a transmission to increase torque output can result in reduced angular speed capabilities. Further, in some exemplary embodiments, the motor-transmission system can achieve a no-load angular velocity similar to peak angular velocity during swing phase knee extension, which can exceed 300 degrees per second.

[0070] In some embodiments, the sensors and controller can estimate the internal joint moment applied at the knee (i.e., the volitional muscle effort exerted by the user across the knee joint) and provide assistance based on that effort. In some methods, the assistance may be adjusted using a feedback controller which tracks volitional knee extensor moment (effort) and provides assistance as a percentage of instantaneous internal knee moment, thereby encouraging volitional muscle use while also providing knee extension assistance.

[0071] In some methods, the assistive torque may be adjusted in anticipation of an upcoming shortfall in volitional knee extension effort. The target for this approach can be to proactively adjust the posture of the limb so that the user’s own muscles are able to create more torque around the joint. In this manner, the powered system can facilitate and / or encourage the user to activate their own muscles, producing the desired training effect. After repeated exposure to these adjustments, the resulting increased muscle activation can be adopted as a new walking strategy, enabling the assistance required to be slowly decreased over time. Examples of this can include anticipatory knee extension assistance during late swing to enable improved extension torque during loading response, and anticipatory knee extension during mid-stance to facilitate passiveknee flexion and active ankle plantarflexion during push-off. Other anticipatory adjustments can be applied as well.

[0072] In some other embodiments, real-time biofeedback can be incorporated into the control strategy such that the user is made aware of their limb posture during walking for the purpose of eliciting greater volitional effort of knee extension. The biofeedback can be provided in the form of electro-tactile, vibro-tactile, auditory, and / or visual feedback, among others. Incorporating biofeedback with the exoskeleton provides a mechanism for which to incentivize the training of improved gait postures with the aid of robotic assistance. One possible embodiment of this incentive is to utilize the feedback to create a video (or audio) based game (i.e., gamification), as illustrated in FIGS. 18-20, to encourage training and reward retention of the improved gait posture or walking strategy.

[0073] FIG. 1 illustrates an exemplary powered gait assistance system 10 that was constructed and tested (as discussed below) with a child having crouch gait from cerebral palsy. The system 10 includes orthotic braces 12 of an existing KAFO and modular powered torque assist system 14 including rigid members 16 that couple to the KAFO braces 12 at the thigh and shank and a motortransmission assembly 18. As shown in the inset of FIG. 1, the motor-transmission assembly 18 includes an encoder 21, electrical motor 22, gear box 24, transmission comprising a chain 26 and sprockets 28 and 30, chain tensioner 32, chain guard 34, and torque sensor 36. The system 10 can also include a ground contact sensor 38 mounted below the foot to detect when the patient’s foot is in contact with the ground and when the leg is in the swing phase.

[0074] FIGS. 2-5 show various views a modular system 100 similar to the system 10 in FIG. 1. The system 100 includes thigh brace 102, shank brace 104, foot brace 106, ankle linkage 108, rigid shank member 110, rigid thigh member 112, rigid torque members 114, 116 coupling the members 110, 112 to the powered knee mechanism, encoder 117, motor 118, gear box 120, sprockets 122, 126 coupled via chain transmission 124, torque sensor 128, rigid transmission frame members 130, 132, and ground contact sensor 134. The motor 118 can comprise a back-drivable 24 V, 90 W brushless motor with a 3-stage, for example. The gear box 120 can comprise an 89: 1 reduction planetary gear head, for example. The encoder 117 can comprise an embedded quadrature encoder, for example. The chain sprocket transmission can provide a 3.5:1 speed reduction, for example, and can transmit torque from the inline shaft of the motor to the knee center of rotation. Other speed reductions ratios can be provided in other embodiments based on the motorand the patient’s needs. The maximum rated output torque of the tested assembly in one embodiment is approximately 16.1 Nm with a 3.82 amp maximum continuous current draw. The torque sensor 128 can be mounted on the drive shaft at the knee and provide a feedback signal indicating the currently applied assistive torque. The powered gait assistance assembly can be modular such that it can be attached to the lateral uprights on the thigh and shank, and / or other portions of existing KAFOs.

[0075] To facilitate both powered assistance and the allowance of unrestricted knee motion during various phases of the gait cycle, the motor can be mounted anterior and superior to the knee joint, as shown in the exemplary systems 10 and 100, for example. The motor-transmission assembly can be attached to both the lower (shank) and the upper (thigh) uprights of a KAFO such that the torque output causes the lower upright to pivot relative to the upper upright, assisting the patient’s knee motion. Each upright can connect to and disconnect from a custom molded shell or other portion of the KAFO via quick release attachment mechanisms to facilitate simple and speedy removal and attachment even after the KAFO has been donned by the user.

[0076] In some embodiments, an adjustable ankle joint mechanism, such as an adjustable dynamic response (ADR) ankle joint device, can be mounted on the lateral upright and can connect the foot and shank components of the system. The adjustable ankle joint mechanism can be located adjacent the angle linkage 108 in the system 100, for example. The adjustable ankle joint mechanism can be freed to allow unencumbered ankle motion, locked to restrict motion, or adjusted to provide dynamic assistance. The system can also include a force sensitive sensor 134, such as force sensitive resistor (FSR) mounted on the foot portion 106 of the device to detect and provide information regarding foot-ground contact. The foot portion 106 of the system can include a foot plate that goes under the patient’s foot and is worn inside the shoe. The force sensitive sensor 134 can be coupled to the foot plate and located under the patient’s foot to detect when the patient’s weight is applied down on the foot plate, indicating when the foot is on the ground and the leg is in the stance phase of the gait.

[0077] The system 100 can also include a controller (not shown) comprising circuitry that provides power and signal conditioning for sensors (e.g., torque sensors, knee angle sensors, ground contact sensors, etc.) as well as feedback control of the motors. The controller can be positioned anywhere, such as mounted near the motor, on the side of the leg, on the upper body, or in a remote device that is coupled to the rest of the system via wires or wireless means. Thecontroller can include or be coupled to servocontrollers that enable closed loop current control of the device motors (e.g., to control torque application at the knee) and knee joint position feedback based on encoder quadrature. Sensor signals can be input into a feedback control system implemented in an onboard computer / processor for real-time, autonomous control. The system can powered by an onboard batteries, for example.

[0078] FIGS. 13-17 show additional alternative embodiments of powered gait assistance devices, as described in more detail below.

[0079] Additional details related to control algorithms implemented by the controller are described below.Exemplary Controller and Software

[0080] Rule-based hierarchical control, as implemented in a finite state machine (FSM), can be deployed for control any of the disclosed systems to help assist and / or restore walking ability. Splitting the gait cycle into discrete states based on detected gait events provides enhanced consistency and robustness to an inherently variable process and allows implementation of lower level controllers within each state. In the exemplary FSM shown in FIG. 7, the gait cycle is divided into three discrete phases: stance 202, early swing 206, and late swing 210. FSM thresholds can be selected based on various criteria, such as data obtained from evaluated children with mild- moderate crouch, and can be verified and adjusted (if necessary) during initial walking sessions for each particular patient. As shown at 204, 212 and 214, the foot plate sensor can be used to govern the transition between stance and swing states based on detected foot ground contact. As shown at 208 and 216, knee angle position and velocity (e.g., computed as first derivative) can be used to detect transition between early and late swing phases. A proportional-integral-derivative (PID) control scheme can be used, for example, to achieve the desired torque output at the knee within each state. The PID error value can be computed as the difference between the measured and desired torque. In some embodiments, the PID gains can be adjusted prior to a patient donning the device, so that torque output reaches a stable response with minimal chatter and latency. For example, in an exemplary system, when the set-point is zero, as in early swing, the PID controller can compensate for the inherent friction of the motor-transmission assembly, allowing for the knee joint to be freely articulated (e.g., effectively frictionless).

[0081] Various control strategies can be implemented to provide knee extension assistance at different intervals within the gait cycle. For example, as shown in FIG. 8, the powered device canprovide extension torque at the knee during stance phase to assist body weight support and during the late-swing phase to assist knee extension for foot placement. At the same time, the control strategy can provide for a frictionless (e.g., near zero torque applied) period during early swing to facilitate knee flexion for toe-clearance.

[0082] FIG. 9 is a flow diagram illustrating an exemplary control architecture 300 for the disclosed gait assistance devices. At 302, kinematic measurements are made (e.g., knee joint angles, other joint angles). The kinematic measurements 302 are fed into an estimation model 304 that outputs a predicted moment profile for the patient over the gait cycle. The predicted moment profile can be generated based on the kinematic measurements 302 along with various other input data, such as patent size, weight, muscle strength, crouch severity, etc. The predicted moment profile can then be fed to a motor controller 306 that converts the predicted moment profile into an electrical control signal appropriate for the particular motor / transmission system being used (e g., the assistive device 308). The assistive device 308 then provides an assistive torque profile over the gait cycle based on the control signal and optionally based on real-time sensory feedback during the gait cycle. The control architecture 300 can cycle continuously over the gait cycle to provide real-time adaptive control for the powered system.Exemplary Device Testing and Evaluation

[0083] To validate the functional performance and preliminary efficacy of the exemplary gait assistance system 10 shown in FIG. 1, implementing the disclosed control strategy illustrated in FIGS. 7 amd 8, one male participant diagnosed with crouch gait from CP was recruited and evaluated. He was classified level II on the gross motor function classification system (GMFCS) and was able to walk for at least 30 feet without assistance. At the time of data collection, he was six years old, height of 118 cm, and weight of 20.0 kg.

[0084] The participant completed five sessions of testing and evaluation. The first session included a clinical assessment of lower limb strength and spasticity, over-ground gait assessment of his baseline gait condition while wearing prescribed AFOs on each leg, and casting for fabrication of custom orthotic braces. The initial session was followed by three practice sessions, during which the participant walked with the powered gait assistance system over-ground and on a treadmill, and a final session for data collection. When using the powered gait assistance system, the participant was instructed to walk in a manner similar to their normal walking habits.

[0085] During the practice sessions, the system torque output was adjusted upwards from 1 Nm at 0.5 Nm increments. Final values of 3.5 Nm for stance phase and 2.625 Nm (75% of the stance value) for swing were established based on participant preference and visual feedback of participant comfort and walking stability. Total walking time in the powered gait assistance system was 18 minutes across the four sessions. Frequent short breaks were given to reduce fatigue and maintain participant attention and focus on the walking task. Total walking time does not include experimental setup, device tuning and calibration, or walking in control conditions (i.e., without powered assistance). The motor controller also incorporated safety measures that limited the maximum torque output and prevented knee hyperextension.

[0086] The fifth session was for experimental data collection with the gait assistance system under two conditions: a free knee joint (no motor assist) and the assistive mode with extensor torque set at 3.5 Nm for stance phase and 2.625 Nm (75% of the stance value) for swing (FIG. 8). The purpose of these conditions was to assess the effect of the passive braces and the motorized assembly on gait separately. The ankle joint was set for free rotation in both conditions. The participant completed three successful overground walking bouts along a 5.5 m pathway in each mode. Kinematic data were collected at 100 Hz using 10 motion capture cameras and a custom marker set. Specifically, three markers were placed on the foot, clusters comprised of four noncollinear markers on the shank and thigh segments, four markers on the pelvis, and three on the trunk. Markers were also placed on the medial and lateral aspects of the ankle and knee joints. Three markers were placed on each shank and thigh segment of the orthotic brace. Muscle activity was collected bilaterally from rectus femoris, vastus lateralis, semitendinosus, and medial gastrocnemius using a wireless EMG system recorded at 1000 Hz. EMG data were band-pass filtered at 15-380 Hz, full-wave rectified, and low-pass filtered at 7 Hz to create a linear envelope. Lower-extremity joint angles were computed from marker trajectories. Experimental data were time normalized to each gait cycle and averaged across the gait cycles for each walking condition. Performance Validation

[0087] The latency of the exemplary gait assistance system 10 was measured by attaching the uprights to a solid extruded aluminum frame, creating an isometric condition under which timing of torque application could be quantified. The on-off and off-on latency was measured as the time between when the controller triggered the transition and when torque measured by the onboard sensor reached the target value. The average latency, across 10 trials, was 53 ± 8 ms for activationand 57 ± 7 ms for deactivation. The compliance between the brace and limb was evaluated by comparing sagittal plane knee joint angle computed from the system markers with the biological knee angle computed from markers placed directly on the skin, as shown in FIG. 10. The difference in knee angle was not significant (p > 0.99) indicating minimal motion between the patient’s body and system. As shown in FIGS. 11A and 11B, the torque provided by the powered system was consistent with the design specifications. Swing phase assistance was provided only after volitional knee extension began. Swing phase knee extension was delayed in the left leg compared to the right leg, which resulted in a reduced interval during which swing extension assistance could be applied to the left limb prior to heel-strike.Joint Kinematics and Spatiotemporal Measures

[0088] The experimental biomechanics data obtained from overground walking were screened to ensure that only complete gait cycles free from adverse events (e.g., toe-drag) were included in the subsequent analyses. The total number of analyzed gait cycles for each condition, along with other data obtained, is shown in the table in FIG. 6.

[0089] As illustrated in FIGS. 11C-11H, to assess the effects of motorized extension assistance on the participant’s gait, kinematics (joint angles) and spatiotemporal parameters were calculated under three conditions: shod + AFO (baseline), walking with the powered system with a free knee joint (free), and walking with assistance from the powered system (powered). There was a significant increase in total knee range of motion (ROM) and knee extension during stance phase when walking with the powered system compared to the free knee and baseline conditions (as shown in FIG. 6). Compared to baseline, the powered system increased peak knee extension during stance by 18.1° and 11.5° in the right and left leg, respectively. Maximum knee flexion during stance was decreased in the powered condition by similar levels (18.6° and 11.9°) compared to baseline. There was no significant difference in the hip angle between the baseline, free knee, and powered conditions. However, peak ankle dorsiflexion was increased significantly for both the free knee and powered conditions due to the free ankle joint compared to the rigid AFO at baseline. Yet, there was no significant difference in knee extension during stance or knee ROM between the baseline and free conditions. There were no significant differences in step length or step width between walking conditions. Cadence was significantly reduced in both the free knee and powered conditions compared to baseline.Muscle Activity

[0090] As shown in FIGS. 12A-12H, walking with the powered system affected EMG activity of several muscles compared to walking with the brace with a free knee joint. There was elevated activity during late stance phase in both rectus femoris and semitendinosus in the powered condition. Post-hoc tests showed a significant increase in peak semitendinosus linear envelope during late stance (p < 0.001). Importantly, the tested powered gait assistance system did not have a large effect on the vastus lateralis activity, which maintained similar levels during late-swing and stance phases when assistance was provided. Likewise, no differences in medial gastrocnemius EMG were observed between the free knee and powered conditions.Discussion

[0091] The exemplary powered systems and control methods described herein can significantly help with ambulation, especially as a treatment of crouch gait in children with CP. The testing and evaluation described above examined the effects of relatively straightforward knee extension assistance during stance and late-swing phase. The analysis shows that the exemplary FSM control strategy activated and deactivated the assistance during walking with latencies of similar duration to those of biological processes (e.g., muscle contractions), and therefore are well- suited for adequate human-machine interaction.

[0092] The evaluation also assessed the level of relative motion that was permitted between the patient’s limbs and the exemplary device when assistive torque is applied. The data indicate that the as-tested powered gait assistance system and implementation strategy suitably mitigated this issue as no significant motion of the system was observed relative to the underlying limb at the knee joint in the patient. Owing to the patient’s weight (20.0 kg), the target torque levels were small but sufficient to demonstrate the functional capacity of the device to elicit the desired outcomes (e.g., increased knee extension during stance). The torque set-point during each assistive phase was held constant at 3.5 Nm for stance and 2.6 Nm for swing, respectively. Even though the torque was constant for each state, the resulting kinematic profile maintained relatively normal waveforms, including continuous knee extension for foot placement and an early stance loading response. In other embodiments, more sophisticated control methods can be utilized, to even more effectively inject additional energy into the gait cycle to alleviate crouch.

[0093] The tested device significantly altered lower extremity kinematics and reduced the amount of crouch compared to the baseline condition. As a result, the subject walked in a more extended, upright posture, and showed significant increases in maximum knee extension duringstance phase in both the left and right legs. The improvement in maximum knee extension (12- 19°) was similar to the range of long-term improvement reported from invasive surgical treatment (10-20°). Peak stance phase knee flexion was also significantly reduced in both legs when walking with the powered system. The participant displayed baseline asymmetry with slightly less crouch during stance in the left leg. The testing results illustrate that the assistive torque, set to the same value in both legs, had a larger effect on stance posture in the more affected limb. Other embodiments can employ alternative methods for delivery of assistive torque, such as by varying levels of assistance between limbs and / or adjusting the amount of torque applied within each phase in real-time.

[0094] The overall knee ROM for the entire gait cycle was significantly increased during walking with the powered system, resulting in a trajectory that was closer to that of normal waking. While there were gains in knee extension during stance, an increase in knee flexion during swing phase was also observed with both the powered system and the free knee condition.

[0095] The increased knee ROM provided by the powered system can mitigate the kinematic benefit of extension assistance provided by the motor. A more flexed knee during swing increases the late swing extension requirement as the limb must articulate through a larger ROM. In the testing results, the transition from swing phase flexion to swing phase extension was delayed in the left leg compared to the right, limiting the time available to provide swing phase assistance from the powered system. Peak swing knee extension velocity in the left leg was approximately 350° / sec, resulting in motor shaft velocity of 18,160 revolutions per minute (RPM), which exceeded the no load speed of the motor (16,300 RPM or 314° / sec) and therefore limited the ability to provide extension assistance in this phase (FIGS. 11A and 1 IB). Swing phase knee extension velocities in the right leg were less than 300° / sec. In other embodiments, a higher power motor with a lower gear reduction can be deployed.

[0096] The tested torque set-points were also relatively small and selected according to patient preference and visual observation of walking stability. The applied torque levels were gradually increased during testing over the three practice sessions. Still, the participant walked with reduced cadence while using the powered system compared to his baseline gait, a somewhat expected finding given the novelty of the task that may be improved with additional time for accommodation to the system. Considering the limited time the participant walked with the powered system,continued use would likely result in increasing comfort levels, the opportunity to increase the assistance, and further kinematic improvements.

[0097] Differences in EMG were observed during walking evaluation with the tested powered system. In the left leg, elevated hip and knee flexor activity at the end of stance phase and in early swing (50-75% of the gait cycle) was observed. The knee extensor torque provided by the motor at the end of stance may have triggered this antagonistic flexion response. No differences were observed in knee extensor activity during stance or swing phase extension assistance in either leg. At the end of the stance phase in the left leg, vastus lateralis activity ceased earlier when using the powered system compared to the free knee. This pattern of activity more closely resembles that of normal walking, and illustrates the application of the disclosed technology for long term use of the disclosed assistance devices to help facilitate elimination of the persistent extensor activity during stance phase of crouch.

[0098] Robotic assistance during walking can be employed in neurologically intact populations without diminishing muscle activity, and can actually increase muscle activity. The testing results described herein show that certain patients, such as children with crouch gait from CP, do not significantly reduce their knee extensor activity when walking with powered extension assistance from a powered gait assistance system. For example, it was observed that the test participant did not simply allow the motors to take over control of knee extension responsibilities and that kinematic improvements were observed. The disclosed powered assistance devices can maintain and increase the user’s volitional muscle activity over time, making them suitable as a rehabilitation tool.

[0099] Another challenge for long-term use of powered gait assistance systems, especially in children, is growth of the patient overtime. Accordingly, the disclosed technology can be modular, adjustable in size and power, and can comprise or be used with orthotic devices that do not require custom molding to be individually adaptable. This can have the added benefit of reduced cost as well as longer-term adaptability.

[0100] Another advantage of the disclosed technology as a therapeutic intervention is that the implemented control strategy can be tuned to optimize the assistance for each individual, and can be changed over time as the patient changes. The modular design allows optional integration of additional modes of assistance, such as passive support at the ankle and surface electrical stimulation.Alternative Embodiments of Powered Gait Assistance Systems

[0101] The systems illustrated in FIGS. 1-5 are representative of a wide variety of possible system in which the disclosed technology can be implemented. In alternative embodiments, the system can utilize different types of motors, transmissions, controllers, sensors, quick-connect mechanisms, etc. For example, the motors can comprise electrical motors, pneumatic or hydraulic actuators, artificial muscles, etc. The transmissions can comprise direct drives, belt drives, chain drives, Bowden cables, geared drives, etc. The systems can be employed on one leg of a patient only, or on both legs. The technology can also be employed analogously on other joints, such as the ankle or hip joints. The motor and transmission system can also be positioned in different locations, such as to the lateral side of the leg, below the anterior aspect of the knee, at the hips or torso, etc. In some embodiments, it is desirable to reduce the lateral width of the device, such as when the patient uses a wheel chair, walker, or crutch. In other embodiments, it may be desirable to locate motor components far from the knee, such as at the hip or thigh, and use a chain, belt or cable type transmission to transfer torque to the knee joint. The device-limb interface can also have various structures, such as rigid thermoplastic orthotic braces, compliant restraints such as cloth, webbing, synthetic skin materials, or other interfaces. In some embodiments, the system controller can communicate wirelessly with other remote devices (e.g., via radio, cellular networks, WiFi, Bluetooth, NFC, etc.). For example, the controller may communicate with a cloud based databases or remotes computing systems to exchange information, data logging, remote control and adjustment of system settings, firmware / software updates, etc.

[0102] FIG. 13 shows an exemplary powered gait assistance system 400 that includes a motor 406 mounted along the lateral aspect of the patient’s thigh brace 402. The motor is coupled to the knee joint 410 via a right-angle drive 408 to apply assistive torque to pivot the shank 404 relative to the thigh. In this system, there is less structure anterior to the knee, and the mass of the motor is shift upward closer to the hip joint, which can reduce the added burden on the hip muscles caused by the system. The system 400 also obviates the need for cables or chains to transmit power.

[0103] FIG. 14 shows an exemplary powered gait assistance system 500 that includes torso mounted cable-drive motors 506. The motors 506 include cable drives 508 that apply tension to cables 510 running down the lateral aspect of the thighs to knee pulleys 512, which use the cable tension to apply torque to the knee joints. The system 500 minimizes the structural impedancearound the legs, allowing less obstruction of gait motion between the braces 502 and 504. The cables 510 can be protected by being routed through tubes, for example.

[0104] FIG. 15 shows an exemplary powered gait assistance system 600 that includes a linear actuator 606 that has a first end 608 coupled to a thigh brace 602 and a second end 610 coupled to a shank brace 604. Linear extension and contraction of the actuator 606 causes torque application about the knee joint 612. The system 600 allows there to be minimal structures around the knee joint.

[0105] FIG. 16 shows an exemplary powered gait assistance system 700 that includes a thigh brace 702 coupled to a shank brace 704 via rigid uprights 706, 708 and a pivot joint 726, with a motor 710 and gearbox 712 mounted along the thigh. The motor turns a cable driver 714 that pulls and pushes a cable 716 that follows an arcuate path through a cable guide 718 around the knee joint 726 to a connection point on a lower member 722 fixed relative to the shank upright 706. The cable drive 714 pushes and pulls the cable to cause the member 722 to pivot the shank about the joint 726. A load cell 716 can be mounted at the end of the cable to measure the applied force. A spring or other biasing member 720 can be included in the upper portion and coupled to the cable to help bias the cable, such as to avoid kinking / crimping of the cable and maintain tension in the cable. The system 700 can provide a greater range of motion over which the assistive torque can be applied, such as over knee angle range from 90 degrees to 270 degrees.

[0106] FIG. 17 shows an exemplary powered gait assistance system 800 that includes a system controller 802 configured to control the powered gait assistance system 804 (e.g., any of the disclosed arrangements) wherein the controller 802 is wirelessly coupled to one or more remote systems, such a removed computing or control system 806 and / or a cloud based computing environment or database 808. The controller 802 can include an onboard microcontroller that adjusts torque assistance of the mechanism 804 based on wireless communications with remote systems, such as with researches or clinicians, and can send back data regarding the performance, compliance, etc. of the system. The controller 802 can also communicate wirelessly with the motors and / or sensors of the system 804. The controller 802 can optionally also include memory / data storage to store data until a later time when it can be communicated or used in processing steps.

[0107] The controller may optionally include a mobile application user interface to adjust the type, timing, and level of assistance, to provide user-interaction for bio-feedback based games for training, and to report user performance while using the exoskeleton system.

[0108] FIGS. 18A and 18B show a computer based game system that can be used to help train a patient wearing one of the disclosed powered gait assistance systems. The game can show images that encourage the patient to bend and extend at the knees in a desired sequence, such as to control an avatar or other object in the game. This can help the patient, especially a child, perform the desired amount of repetitions and foster improved training results, while making the training fun and rewarding.

[0109] Gaming mechanisms and immersive / non-immersive forms of virtual reality are showing promising results and rapidly gaining traction for neuro-rehabilitation. The powered gait assistance systems may optionally be used as a peripheral device in this emerging field of neurorehabilitation gaming. Disclosed gait assistance systems can be capable of collecting and transmitting real-time measurements capturing user movement and effort, and as a result, can be integrated as a game controller for a custom application for exercise targeting knee extension and flexion in children with crouch gait from CP. Proper game control resulting in successful completion of repetitive exercises provides strengthening and cardiovascular training to the user. In addition, when coupled with a neuroimaging modality such as electroencephalography (EEG), these games allow for analysis of brain activation during task execution, allowing for further optimization of both game development and treatment strategies to maximize neuroplasticity. FIGS. 18A and 18B demonstrate knee movements measured by the worn device being used as input to drive a game engine capable of rendering real-time feedback to the patient using either immersive (head mounted display) or non-immersive virtual reality technologies. The software integration pipeline can utilize commercial game engines to broaden the scope of use of the powered gait assistance devices as a haptic game controller for neurorehabilitation gaming and tele-rehabilitation applications.

[0110] For example, FIGS. 19 and 20 show screenshots of an exemplary training game, where squatting at the knees causes a spaceship to move down and extending at the knees causes a spaceship to move up. The patient controls the spaceship to try get the spaceship to move up and down in real time to hit certain objects in the spaceship’s flight path, creating a game that the patient can play and enjoy while training.

[0111] FIG. 21 is a simplified block diagram depicting an exemplary computing environment in accordance with one or more examples of the present application. The environment 2100 includes training data sources 2102, training computing system 2104, imaging system 2106, powered gait assistance system 2108, one or more sensors 2110, and control system 2112. Although the entities within environment 2100 may be described herein and / or depicted in the FIGs. as being singular entities, it will be appreciated that the entities and functionalities discussed herein may be implemented by and / or include one or more entities. For example, the training computing system 2104 may be integrated into the powered gait assistance system 2108 or the functionality of the training computing system 2104 may be performed by the powered gait assistance system 2108. The powered gait assistance system 2108 may be an example of the powered gait assistance system 10 of FIG. 1.

[0112] In some instances, the entities within the environment 2100 such as the training data sources 2102, training computing system 2104, imaging system 2106, and powered gait assistance system 2108 may be in communication via a direct line and / or wirelessly. For instance, the entities within the environment 2100 may communicate via a global area network (GAN) such as the Internet, a wide area network (WAN), a local area network (LAN), or any other type of network or combination of networks either wired or wireless.

[0113] The training data sources 2102 may include publically available data sources such as the publicly available biomechanical open source datasets, or other data sets collected for individuals or subjects interacting with or without the powered gait assistance system 2108 and include sensor data from the one or more sensors 2110 of the powered gait assistance system 2108. The imaging system 2106 may include one or more motion capture cameras and / or imaging devices configured to capture data for users along with custom markers as described herein.

[0114] The data sources 2102 include one or more devices, computing platforms, repositories, databases, systems, and / or other entities that generate and / or store training data that is used by the training computing system 2104 to train one or more ML - Al models such as the MLP and / or the attention-based recurrent neural network (RNN).

[0115] The training computing system 2104 includes one or more computing devices and / or systems that are configured to utilize data from the training data sources 2102, data from the imaging system 2106, and / or data from the one or more sensors 2110 to train and implement one or more ML - Al models such as an attention-based recurrent neural network and / or MLP. Inexamples, the training computing system 2104 may provide the trained ML -Al models to the powered gait assistance systems 210 and control system 2112 for predicting torque values and / or predicting joint angles for an individual utilizing the powered gait assistance system 2108. For example, the training computing system 2104 includes one or more computing devices, computing platforms, systems, servers, desktops, laptops, tablets, mobile devices (e.g., smartphone device, or other mobile device), or any other type of computing device that generally comprises one or more communication components, one or more processing components, and one or more memory components.

[0116] The training computing system 2104 is capable of performing tasks, functions, and / or other actions. For example, the training computing system 2104 trained ML - AL models to the powered gait assistance system 2108 and / or control system 2112. The training computing system 2104 may receive real time data from the one or more sensors 2110 to update or adjust the trained ML - AL models. The training computing system 2104 may use subsets of training data from the training data sources 2102, ground truth data or labeled data from the imaging system 2106, and actual data from the one or more sensors 2110 to train and implement the one or more ML - Al models described herein. The training of these models is described below.

[0117] In some variations, the training computing system 2104 may be implemented as engines, software functions, and / or applications. In other words, the functionalities of the training computing system 2104 may be implemented as software instructions stored in storage (e.g., memory) and executed by one or more processors.

[0118] The control system 2112 may implement, apply, or otherwise execute the one or more ML - Al models provided by training computing system 2104. For example, the control system 2112 may include one or more processors such as central processing units (CPUs) and / or memory. The processor may include one or more physical cores (e.g., CPU cores) or processing units, which are pieces of logic capable of independently performing the functions of the processor. The physical cores may include multiple different threads of execution that can execute multiple different tasks at one time. For instance, each of the threads of execution may be executing programming code independent of each other. By using multiple threads that are executed simultaneously (e.g., multi -threaded processing), the processor may perform multiple tasks at the same time. Once the ML - Al models are trained they may be used to predict joint angles or torque values for applying an assistive or resistive torque or determining a gait phase with an associatedassistive or resistive torque for assisting an individual’s volitional knee muscle output while they utilize the powered gait assistance system 2108. Although some of the examples below describe applying an assistive torque it is to be understood that the embodiments of the present disclosure also include applying a resistive torque. The control system 2112 may use the trained ML - Al models to generate an estimation model 304 for predicting moment and generate control signals associated with the estimation model for providing the assistive torque to the powered gait assistance system 2108. For example, as described above with reference to FIG. 3, the control system 2112, which may be an example of the motor controller 306, may convert the predicted moment profde into an electrical control signal appropriate for the powered gait assistance system 2108. The powered gait assistance system 2108 may provide data from the one or more sensors 2110 to the training computing system 2104 for updating the trained ML - Al models or to train the ML - Al models by the training computing system 2104. The training computing system 2104 and the powered gait assistance system 2108 and control system 2112 may provide updated data or ML - Al models to each other to provide more accurate predictions of joint angles or predicted torque values.

[0119] It will be appreciated that the exemplary environment depicted in FIG. 21 is merely an example, and that the principles discussed herein may also be applicable to other situations — for example, including other types of institutions, organizations, devices, systems, and network configurations. As will be described herein, the environment 2100 may be used by health care enterprise organizations.

[0120] FIG. 22 is a simplified block diagram of one or more devices or systems within the exemplary environment of FIG. 21 (e.g., training computing system 2104, powered gait assistance system 2108, and / or control system 2112) within the environment 2100. The device / system 2200 includes one or more processors 2204, such as one or more CPUs, controller, and / or logic, that executes computer executable instructions for performing the functions, processes, and / or methods described herein. In some examples, the computer executable instructions are locally stored and accessed from a non-transitory computer readable medium, such as storage 2210, which may be a hard drive or flash drive. Read Only Memory (ROM) 2206 includes computer executable instructions for initializing the processor 2204, while the random-access memory (RAM) 2208 is the main memory for loading and processing instructions executed by the processor 2204. The network interface 2212 may connect to a wired network or cellular network and to a local areanetwork or wide area network, such as the network 2106. The device / system 2200 may also include a bus 2202 that connects the processor 2204, ROM 2206, RAM 2208, storage 2210, and / or the network interface 2212. The components within the device / system 2200 may use the bus 2202 to communicate with each other. The components within the device / system 2200 are merely exemplary and might not be inclusive of every component, server, device, computing platform, and / or computing apparatus within the device / system 2200. For example, training computing system 2104, powered gait assistance system 2108, and / or control system 2112 may include some of the components within the device / system 2200 and may also include additional and / or alternative components. Additionally, and / or alternatively, the device / system 2200 may further include components that might not be included within every entity of environment 2100.

[0121] FIG. 23 is an exemplary process for training a multilayer perceptron (MLP) using training sensor data and generating predicted torque values for the disclosed powered gait assistance system using the trained MLP and sensor data in accordance with one or more examples of the present application. The process 2300 may be performed by the training computing system 2104 and / or the powered gait assistance system 2108 depicted in FIG. 21. However, it will be recognized that the process 2300 may be performed in any suitable environment and that any of the following blocks may be performed in any suitable order.

[0122] At block 2302, a system (e.g., the training computing system 2104 and / or the powered gait assistance system 2108) obtains a training dataset comprising training sensor data associated with measurements of a plurality of first individuals using one or more powered gait assistance system. The measurements of the first individuals may be obtained by one or more sensors of a powered gait assistance system as each individual uses the powered gait assistance system to execute a mobility task (e.g., walk over a flat surface, walk on a domed surface, climb, jog, etc.,), or by a motion capture system of the imaging system 2106 as the first individuals execute a mobility task, thereby generating the training sensor data. In other words, in some examples, the training data sources 2102 may be and / or include one or more powered gait assistance systems such as the powered gait assistance systems described above. First individuals may use the powered gait assistance systems to perform mobility tasks, and the sensors of the powered gait assistance systems may generate sensor data. For example, as described above, the sensors of the powered gait assistance systems may include, but are not limited to, a torque sensor, an angle sensor, a foot sensor, an inertial measurement unit (IMU) sensor, a pressure sensor, and / or othersensors. Using the sensors, the powered gait assistance systems may obtain sensor data including, but not limited to torque measurements, angular position measurements, angular velocity measurements, force sensitive resistor (FSR) measurements, force measurements, angular rate measurements, orientation measurements, human interaction with the powered gait assistance system related to force and direction obtained by the pressure sensor, and / or additional and / or alternative measurements. For example, pressure sensors may be built into the body or material of the powered gait assistance system and be configured to measure intent or effort (e.g. force and direction of the pressure or force detected by the pressure sensors) as the individual uses the powered gait assistance system to complete tasks. The data from the pressure sensors may be used to determine whether an individual using the powered gait assistance system is starting to walk or stopping from walking or each time the individual takes a step. After obtaining the sensor data, the powered gait assistance systems may store the sensor data in memory such as memory of the powered gait assistance systems and / or external memory (e.g., a database, repository, and / or other data storage elements). At block 2302, the system may obtain the training dataset, which may include the sensor data (e.g., training sensor data) that was obtained from the sensors from the powered gait assistance systems. In other words, in some instances, the training data sources 2102 may include and / or be the powered gait assistance systems. In other instances, the training data sources 2102 may be a database, repository, and / or other memory / data storage elements (e.g., memory from a computing device), and the training data sources 2102 may receive the training dataset (e.g., the sensor data from the powered gait assistance systems), store the training dataset in memory, and provide the training dataset to the system (e.g., the training computing system 2104).

[0123] At block 2304, the system may train an MLP using the training data. For instance, as mentioned above, a powered gait assistance system such as the powered gait assistance system 2108 may use a finite state machine (FSM) to move between different phases of a gait cycle. For instance, the FSM may split the gait cycle into five discrete phases: early stance, mid-stance, late stance, early swing, and late swing. For each phase, a pre-tuned value of knee extension assistance or no assistance may be provided by the powered gait assistance system 2108. In some instances, the FSM may use on-board sensors (e.g., the FSR sensor) and / or a sensor that detects the angular velocity (e.g., an angle sensor) to switch between different gait phases. For instance, based on one or more thresholds (e.g., for each limb, five FSM thresholds such as two FSR and three angularvelocity thresholds), the powered gait assistance system 2108 may switch between the different gait phases. However, each individual may have slightly different thresholds that are used to switch between different gait phases. For example, an individual that is taller or younger may have different thresholds to switch between each gait phase. Furthermore, for even the same individual, as the individual gets older, their gait phase may change, and thus a threshold that is used to switch between a first to a second gait phase may change for the individual as the individual gets older. As such, typically, individuals may attend sessions such that these thresholds are adjusted and / or calibrated to be specific to the individual that is using the powered gait assistance system 2108. However, this is tedious.

[0124] Accordingly, examples of the present disclosure may use training data (e.g., training sensor data such as FSR and / or angular velocity data for a plurality of different individuals) to train an MLP to predict a gait phase for an individual and / or a predicted torque value or indication directly (e.g., torque assistance value or whether to provide torque assistance). For example, as will be described with reference to FIG. 24, a first ML - Al model (e.g., the MLP) may be trained using the training sensor data, and the trained MLP may be used to predict gait phases and / or torque values for another individual. As such, by using the MLP, this may eliminate the requirement to tune these thresholds manually and instead segment the gait cycle - and thereby provide knee extension assistance during overground walking - automatically and accurately.

[0125] FIG. 24 shows an exemplary first ML - Al model in accordance with one or more examples of the present application. For instance, the first ML - Al model 2402 includes an MLP 2414. The MLP 2414 includes the input layer 2408, hidden layers 2410, and an output layer 2412. During the training, the training computing system 2104 may provide the training sensor data (e.g., a portion of the training sensor data) to the first ML - Al model 2402 and the MLP 2414 to generate a training output. Afterwards, the training output may be compared to a groundtruth output to compute one or more losses. Then, using the one or more computed losses, the training computing system 2104 may use one or more optimization algorithms such as A Stochastic Optimization Algorithm (ADAM) to update the parameters of the first ML - Al model 2402 and / or the MLP 2414.

[0126] In some instances, the MLP 2414 may include twenty neurons and two hidden layers 2410 (e.g., four layers total with the input and output layers 2408 and 2412). In other instances,the MLP 2414 may include any number of neurons and / or layers, including, but not limited to, four, six, and / or ten hidden layers 2410.

[0127] In some examples, the output from the first ML - Al model 2402 (e.g., the training output) may be, include, and / or indicate a specific gait phase (e.g., early stance, mid-stance, late stance, early swing, and late swing). In other examples, the output from the first ML - Al model 2402 (e.g., the training output) may be, include, and / or indicate a predicted torque output value.

[0128] In some variations, the training computing system 2104 may determine the groundtruth output based on the training sensor data and an automatic labeling algorithm. For instance, after the training sensor data is collected or obtained, it can be analyzed and labeled to indicate which sensor data corresponds to a particular gait phase the individual was in and / or entered across a plurality of time steps. For example, a user may analyze the various input of the training sensor data and assign which gait phase the controller of the powered gait assistance system should have been in given the input of the training sensor data. An automatic labeling algorithm may automate the manual labeling process of labeling the groundtruth output by using specified, sometimes individual specific, thresholds to assign a gait phase to the training sensor data.

[0129] In some instances, the training computing system 2104 may separate the training sensor data into a plurality of batches of data, and use the batches of data to train the first ML - Al model 2402.

[0130] Returning back to FIG. 23, at block 2306, the system may obtain, from one or more sensors 2110 of the powered gait assistance system 2108, sensor data associated with a second individual using the gait assistance system 2108. The second individual may be an individual that is not part of the plurality of first individuals that provided the training sensor data. In other words, as described above, by using the MLP (e.g., the MLP 2414), these may be helpful to avoid having to tune the thresholds for the FSM for the different gait phases manually. Thus, at block 2306, the system may obtain sensor data for a second individual and as will be described in blocks 2308 and 2310, the system might not need to provide further calibration / tuning of the thresholds for the FSM for the different gait phases. Instead, the system may use the trained first ML - Al model 2402 to determine the gait phases and / or a predicted torque value.

[0131] For example, the one or more sensors 2110 may be configured to capture the sensor data as the second individual uses the gait assistance system 2108. The sensor data associated withthe second individual may be provided by the one or more sensors 2110 to the control system 2112 of the powered gait assistance system 2108 for processing using the trained MLP.

[0132] At block 2308, the system may process the sensor data associated with the second individual using the trained MLP to generate predicted torque values for the powered gait assistance system 2108. For example, referring to FIG. 24, after training the parameters of the first ML - Al model 2402 (e.g., the weights and / or other aspects of the MLP 2414), the system may use the trained first ML - Al model 2402 to generate predicted torque values. In some examples, the system may provide the sensor data for the second individual (e g., from block 2306) into the first ML - Al model 2404 and the first ML - Al model 2404 may provide an output (e.g., an inference output). The output may include and / or indicate a particular gait phase from a plurality of gait phases of a gait cycle, which is described above. Each gait phase may be associated or mapped to predicted torque values that correspond to a motor of the powered gait assistance system applying an assistive torque to aid the mobility of the second individual. The control system 2112 may process the sensor data associated with the second individual using the trained MLP to generate predicted torque values by first determining a gait phase. The control system 2112 may maintain or otherwise access a mapping of gait phases to certain torque values that correspond to assistive torque to be applied by the motor of the powered gait assistance system.

[0133] Additionally, and / or alternatively, the output from the first ML - Al model 2404 may include and / or indicate an indication of whether to provide extension assistance. For example, the output from the first ML - Al model 2404 may indicate a binary value (e.g., a value of “1” if assistive torque is provided and a value of “0” if assistive torque is not provided). Thus, in such examples, the output from the first ML - Al model 2404 may include and / or indicate a timing of when extension assistance is provided, but might not indicate the magnitude, which may allow a clinician and / or the second individual to maintain control over the amount of assistance to provide to the second individual.

[0134] Additionally, and / or alternatively, the output from the first ML - Al model 2404 may include and / or indicate an actual torque value (e.g., a predicted torque value). For example, the output from the first ML - Al model 2404 may include and / or indicate a magnitude of torque to be provided to the second individual by the powered gait assistance system 2108.

[0135] Additionally, and / or alternatively, the output from the first ML - Al model 2404 may include and / or indicate transition thresholds that may be used for transitioning between thedifferent gait phases. For example, the MLP 2414 may be configured to determine the actual transition thresholds for transitioning between gait phases thereby automating a previous manual process that was conducted by a clinician or engineer interacting with the powered gait assistance system 2110 and reviewing the sensor from the one or more sensors 2110. The MLP 2414 may determine the gait phase or when to transition between gait phases based on identifying correlations in the sensor data from the one or more sensors 2110. For example, a portion of the sensor data obtained by the one or more sensors 2110 may correspond to data obtained while an individual takes a few strides while using the powered gait assistance system 2108. The MLP 2414 may identify a correlation between this portion of sensor data that corresponds to a particular gait phase thereby predicting or determining the gait phase that the individual is in and an assistive torque can be determined from the gait phase thereby by passing the need to tune or use specific transition thresholds. The powered gait assistance system 2108 and control system 2112 may maintain and / or store one or more control laws and / or rules where each gait phase has a corresponding control law. Each control law and / or rule may specify the torque (assistive or resistive torque) that may be a specific value of torque applied or could be within a range of torque applied in an assistive manner to aid an individual using the powered gat assistance system. The MLP 2414, once trained, may directly predict a gait phase using the sensor data from the one or more sensors 2110.

[0136] In some examples, after training the first ML - Al model 2402, the training computing system 2104 may provide the trained first ML - Al model 2404 to the powered gait assistance system 2108. At block 2306, the powered gait assistance system 2108 may obtain sensor data from the second individual using the one or more sensors 2110. At block 2308, the powered gait assistance system 2108 may process the obtained sensor data to generate an inference output (e.g., a binary indication of whether to provide extension assistance, a predicted gait phase, and / or a predicted torque value). In other examples, the training computing system 2104 might not provide the trained first ML - Al model 2404 to the powered gait assistance system 2108. In such examples, at block 2306, the training computing system 2104 may receive the sensor data for the second individual from the powered gait assistance system 2108. At block 2308, the training computing system 2104 may process the sensor data using the trained first ML - Al model 2404 to generate an inference output.

[0137] At block 2310, the system may apply the predicted torque values to a motor of the powered gait assistance system 2108 to cause the motor to apply assistive torque, that corresponds to the predicted torque values, between first and second arms of the powered gait assistance system to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

[0138] For instance, in examples where the training computing system 2104 provides the trained first ML - Al model 2404 to the powered gait assistance system 2108, the powered gait assistance system 2108 may use the inference output that it predicted to apply the predicted torque values to the motor. For example, based on a binary indication of whether to provide extension assistance, the powered gait assistance system 2108 may determine whether to apply an additional torque amount or value (e.g., a pre-defined and / or user-defined such as provided by a clinician and / or the second individual as described above) to provide extension assistance. Based on the predicted gait phase, the powered gait assistance system 2108 may determine a torque value associated with the predicted gait phase (e.g., a torque value associated with a late stance or early swing), and apply the determined torque value. Based on directly predicting a torque value, the powered gait assistance system 2108 may apply the torque value that is predicted by the trained first ML - Al model 2404. In other examples where the training computing system 2104 does not provide the trained first ML - Al model 2404 to the powered gait assistance system 2108, the training computing system 2104 may determine the output from the first ML - Al model 2404 and provide the output to the powered gait assistance system 2108. The powered gait assistance system 2108 may then apply the predicted torque values based on the received output.

[0139] For example, the disclosed powered gait assistance devices can be used with existing leg orthotics, such as knee-ankle-foot orthoses (KAFOs) for example. The powered gait assistance devices of the current disclosure can include a motor that is coupled via a power transmission to a pivot joint that is positioned adjacent to an individual’s knee (e.g., (e.g., the pivot joint can be part of a KAFO or coupled to the knee joint of a KAFO). The motor and transmission can be positioned in various locations and the transmission can take various configurations, such as a chain drive, cable drive, right-angle drive, or direct drive, for example. The motor and transmission are desirably light weight, low-profile, and capable of providing a sufficient level of torque assistance to the individual. Of course the torque output, and therefore the size and weight, of the device that is sufficient depends on the size and pathology of each particular individual. The assistive torqueapplied by the motor of the powered gait assistance devices of the current disclose can include applying the assistive torque between a first and second arm of the powered gait assistance system such as the rigid members 16 that are coupled to the KAFO braces 12 at the thigh and shank and a motor-transmission assembly 18 of FIG. 1.

[0140] In some examples, referring back to block 2306, the obtained sensor data may include predicted sensor data that is predicted from another (e.g., second) ML - Al model. For example, as will be described in further detail below, the training computing system 2104 may further train and use a second ML - Al model (e.g., an attention-based RNN that includes an encoder, a decoder, and an attention-based mechanism) to predict sensor values. The sensor values may then be processed by the trained MLP to generate predicted torque values (e.g., the output of the trained MLP may include and / or indicate a binary indication of whether to provide extension assistance, predicted gait phase thresholds, a predicted gait phase, and / or a predicted torque value). Additionally, and / or alternatively, the second ML - Al model may be used separately from the first ML - Al model (e.g., the trained MLP that is that described in FIG. 23 above).

[0141] FIG. 25 is an exemplary process 2500 for obtaining surface electromyography (sEMG) data and using a trained attention-based recurrent neural network to generate predicted joint angles as well as using the predicted joint angles and actual sensor data from sensors of the powered gait assistance system to generate the predicted torque values by the trained MLP in accordance with one or more examples of the present application. The process 2500 may be performed by a system (e.g., the training computing system 2104 and / or powered gait assistance system 2108). However, it will be recognized that the process 2500 may be performed in any suitable environment and that any of the following blocks may be performed in any suitable order. The process 2500 describes the obtaining of additional data (e.g., sEMG data) and processing the sEMG data for use by the MLP of FIG. 23 and 24 to generate the predicted torque values and / or determined gait phase.

[0142] At block 2502, the system receives sEMG data for the second individual from sEMG sensors of the powered gait assistance system. The sEMG sensors may be part of or included in an sEMG system of the powered gait assistance system (e.g., a wireless sEMG system that is included within the powered gait assistance system 2108). The sEMG data may obtained by the sEMG sensors as the second individual uses the powered gait assistance system to execute a mobility task (e g., walk over a flat surface, walk on a domed surface, climb, jog, etc.,) thereby generating the sEMG data. For instance, the sEMG sensors may obtain sEMG data indicating electrical activitygenerated within contracting muscle fibers in response to activation by the nervous system. For instance, the sEMG sensors may include one or more electrodes that are attached to the skin of a target muscle, and the collected sensor data may indicate information about the neural drive (e.g., the number and intensity of the activated motor neurons) and the fiber electrical properties. Thus, when collected during volitional movement (e.g., when the second individual uses the powered gait assistance system 2108 to execute a mobility task), the sEMG data may encode information about the movement intent of the individual. In other words, the powered gait assistance system 2108 may include one or more sEMG sensors (e.g., electrodes) that collect sEMG data while the second individual is in motion. Furthermore, ajoint angle (e.g., a knee joint angle) may be useful for achieving smooth motion control and has broad application prospects, such as, but not limited to, using the joint angle as an input to an MLP to predict gait phase and / or torque (e.g., as an input to the first ML - Al model 2402 that is described above). The sEMG data may have a relationship to the joint angle, but this relationship may vary based on walking speed and / or other factors, which impacts the ability to estimate the joint angles. For instance, sEMG may be consistent during slower walking speeds whereas stride-to-stride variability may increase as speed increases, which may impact estimation accuracy. In addition, sEMG data may also vary across individuals. For instance, as people age, changes occur in the neuromuscular system that affect the relationship between sEMG and joint angle. Furthermore, different muscles may play a more or less significant role in controlling joint movement at different walking speeds. As such, as will be described below, the system may use an attention-based model such as an attention-based recurrent neural network (RNN) that includes an attention mechanism to predict joint angles from the sEMG data from the second individual.

[0143] In some examples, prior to providing the sEMG data to the attention-based model, the system may perform a preprocessing phase. For instance, the system may filter sEMG data to remove noise, artifacts and / or otherwise enhance the quality of the sEMG data prior to providing the filtered / enhanced sEMG data to the attention-based model. This will be described in further detail below.

[0144] At block 2504, the system processes the sEMG data using an attention-based model (e.g., an attention-based recurrent neural network) to generated predicted sensor data (e.g. predicted joint angles). For example, the attention-based model (e.g., a second ML - Al model that is separate from the first ML - Al model 2402) may include an attention mechanism. Attentionmechanisms may be used in deep learning for improving the performance of processing long input sequences. For instance, the attention mechanism assigns a weight or importance score to each element of the input, based on its relevance to the current system output. This allows the model to selectively focus on the most salient parts of the input and ignore irrelevant or noisy information. The attention mechanism may include three main components: a query, a set of key-value pairs, and a scoring function. The query is a vector that represents the current context or state of the model, while the key-value pairs represent the input data. The scoring function computes a similarity score between the query and each of the key vectors, which is used to compute a weight or importance score for each value. The values are then combined using a weighted sum, where the weights are the importance scores computed by the attention mechanism. In some instances, the attention-based model may be and / or include RNNs, convolutional neural networks (CNNs) and / or transformers. For instance, the attention-based mechanism may be integrated into an RNN, CNN, and / or transformers.

[0145] In some examples, the attention-based model may include an encoder, a decoder, and the attention mechanism described above. This attention-based model is shown in FIG. 26 and will be described below. However, the system may use any type of attention-based model that includes an attention mechanism.

[0146] In operation, after the system receives the sEMG data for the second individual, the powered gait assistance system 2108 may process the obtained sEMG data to generate an inference output (e.g., predicted sensor data such as a predicted joint angle). In some examples, the training computing system 2104 may first train the attention-based model, which is described below in FIG. 27, and then provide the trained attention-based model to the powered gait assistance system 2108. The powered gait assistance system 2108 may process the sEMG for the second individual to generate the inference output. The powered gait assistance system 2108 may then use the inference output to assist the movement of the second individual. In other examples, the training computing system 2104 might not provide the trained attention-based model to the powered gait assistance system 2108. In such examples, at block 2502, the training computing system 2104 may receive the sEMG data for the second individual from the powered gait assistance system 2108. At block 2504, the training computing system 2104 may process the sEMG data using the trained attention-based model to generate an inference output (e.g., predicted sensor data such as a predicted joint angle).

[0147] At block 2506, the system obtains actual sensor data from the one or more sensors of the powered gait assistance system. As shown in FIG. 25, process 2500 may further describe block 2306 of process 2300 in more detail. For instance, the sensor data obtained from the one or more sensors of the powered gait assistance system 2108 may include the actual sensor data from the one or more sensors 2100 as well as the predicted sensor data (e g. predicted joint angles) that are determined based on the sEMG sensors (e.g., electrodes). In other words, the one or more sensors 2110 may include the sEMG sensors as well as additional sensors (e.g., the FSR and / or other sensors described above). To put it another way, the sensor data obtained from the one or more sensors of the powered gait assistance system may also include the raw or unprocessed sEMG data from the sEMG sensors. The system may generate the predicted torque values and / or determine the gait phase and corresponding predicted torque values using the trained MLP, the actual sensor data, and the predicted sensor data (e.g., predicted using the attention-based model). For instance, the MLP may utilize the actual and predicted sensor data to determine the gait phase and corresponding predicted torque values, the binary indication as to whether to provide torque assistance, and / or the predicted torque values directly.

[0148] FIG. 26 shows an exemplary second ML - Al model 2606 that is used to generate predicted joint angles in accordance with one or more examples of the present application. For instance, the second ML - Al model 2606 may include an encoder 2608, an attention mechanism 2610, and decoder 2612. As mentioned above, the second ML - Al model (e.g., the attention-based model) may include the elements shown in FIG. 26. However, in other examples, the second ML - Al model may include the attention mechanism 2610 and additional or alternative elements. For instance, the example of the second ML - Al model 2606 shown in FIG. 26 is merely one such example, and the second ML - Al model 2606 may include additional and / or alternative layers and / or Al models / algorithms (e.g., a CNN or a machine learning method that includes an attention mechanism otherwise referred to herein as an attention-based model). The training of the second ML - Al model 2606 will be initially described. During training, the training computing system 2104 may obtain sEMG sensor data for a plurality of individuals (e.g., the same plurality of first individuals described above that was used to train the MLP 2414 and / or a different group / plurality of individuals), and use the sEMG sensor data to train the second ML - Al model 2606.

[0149] For instance, in operation, the training computing system 2104 may receive and / or obtain the sEMG sensor data 2602. For instance, the training computing system 2104 may receivethe sEMG sensor data from a data source such as a training data sources 2102, which is described above. Additionally, and / or alternatively, the training computing system 2104 may receive the sEMG sensor data from the powered gait assistance system 2108. For example, the training computing system 2104 may initially receive and / or obtain the second ML - Al model 2606 and train the second ML - Al model 2606 using the sEMG sensor data from the training data sources 2102 (e.g., the sEMG training sensor data from a first or second group of individuals). After training using the sEMG sensor data from the training data sources 2102, the training computing system 2104 may further train the second ML - Al model 2606 using the sEMG sensor data from the powered gait assistance system 2108 (e.g., sEMG sensor data associated with a second individual). In other words, the training computing system 2104 may perform transfer learning such that initially the second ML - Al model 2606 may be trained from sensor data from a group of individuals (e.g., the plurality of first individuals). Then, the training computing system 2104 may perform additional training of the second ML - Al model 2606 based on sensor data from a specific individual (e.g., the second individual). Following the additional training, the second ML - Al model 2606 may be used to determine assist the specific individual to perform mobility tasks (e.g., by predicting joint angles that are then used to determine torque assistance for the second individual). In some instances, the training computing system 2104 may also perform transfer learning for the MLP 2414. For instance, referring to FIG. 23, prior to performing 2306, the training computing system 2104 may further train the MLP using training sensor data from the second individual.

[0150] After receiving the sEMG sensor data 2602, the training computing system 2104 may perform pre-processing 2604 of the sEMG sensor data (e.g., filtering the sEMG sensor data). For instance, in some examples, the pre-processing / filtering 2604 may include a multi-stage filtration procedure that processes the raw sEMG signals of the sEMG sensor data 2602. For instance, the multi-stage filtration procedure may include first, implementing a band stop filter (e.g., a HAMPEL filter) with a cutoff frequency of 5 hertz (Hz) and 9 Hz. This filter may be used to remove outliers that are present in the raw EMG signals, which reduces noise and / or artifacts that may interfere with accurate signal analysis. The second stage of the multi-stage filtration procedure may include using a second order infinite impulse response (IIR) notch filter with a center frequency at 6 Hz and quality factor of 30. The next stage of the multi-stage filtration procedure may include undergoing rectification of the raw sEMG signal using a sliding windowof 50 samples. Consecutive peak and root mean square envelope detectors may then be applied to the rectified signal. This process may be aimed to emphasize the amplitude and temporal characteristics of the sEMG signal, which enables a more precise analysis of muscle activation patterns. Additionally, and / or alternatively, to further refine the output, another low-pass filter may be applied to the signal to smooth the filtered sEMG data that is output from the detectors. This step ay be performed to facilitate the removal of any remaining high-frequency noise and may result in a cleaner and more reliable sEMG signal for subsequent processing. In some instances, the second stage may be optional. In some examples, the parameters that are used for the multistage filter (e.g., for the pre-processing / filtering 2604) may user-defined and / or fine-tuned using trial and error.

[0151] Afterwards, the training computing system 2104 may provide the filtered / enhanced sEMG data to the second ML - Al model 2606. The training of the second ML - Al Model 2606 may be performed in two stages. In a first stage, the training computing system 2104 may first train the encoder 2608 and the decoder 2612. For example, the output of the encoder 2608 may be provided directly to the decoder 2612. In this stage, the training computing system 2104 may provide the filtered / enhanced sEMG data and / or the raw sEMG data to the encoder 2606. The encoder 2608 may encode the data (e.g., determine a fixed length internal representation such as a latent space representation) and provide the encoded data to the decoder 2612. The decoder 2612 may then decode the encoded data and generate a first stage output. Subsequently, the training computing system 2104 may compare the first stage output with the input to the encoder 2606 to compute a reconstruction loss. In other words, the encoder 2608 and the decoder 2612 may be trained by using the output of the decoder to compute a loss, a reconstruction loss, by comparing the output of the decoder 2612 to ground truth data (e.g., the initial data that is input to the encoder 2608 such as the actual sEMG data). The training computing system 2104 may then update the parameters of the encoder 2606 and / or the decoder 2612 using the computed reconstruction loss (e.g., by using an optimization algorithm such as ADAM described above).

[0152] After training the encoder 2606 and the decoder 2612, the training computing system 2104 may freeze the parameters of the encoder 2606 and the decoder 2612, and then train the attention mechanism 2610, which may be, but is not limited to, a gated recurrent unit (GRU) neural network and / or an long short-term memory (LSTM). For instance, once the system has trained the encoder 2608 and decoder 2612 of the second ML - Al model 2606, the system may proceed totrain the attention mechanism 2610. The attention mechanism 2610 may enable the second ML - Al model 2606 to selectively focus on the most salient parts of the input (e.g., pre-processed / filtered sEMG data 2604) by assigning a weight to each input element based on its relative to the system output.

[0153] The attention mechanism 2610 may include three components: a query, a set of keyvalue pairs, and a scoring function. The query is a vector that represents the current state of the second ML - Al model 2606. The key-value pair may represent the input data, which in this example includes the pre-processed / filtered sEMG data 2604 and a corresponding ground truth joint angle (knee joint angle). The scoring function computes a similarity score between the query and each key vector which is used to compute a weight or importance score for each value. The values are then combined using a weighted sum, where the weights are the importance scores computed by the attention mechanism 2610.

[0154] The decoder 2612 can selectively access the encoded information, using the context vector c(t), which is computed at each time step of the decoder 2612 based on the previous hidden state and all the hidden states of the encoder 2608. The relative importance of each element in the input sequence is determined by trainable weights assigned to these states, with greater emphasis placed on the most significant inputs. The context vector is constructed by combining each time step j of the encoder 2608 with each time step t of the decoder 2612, and the alignment score. The formula for computing the alignment score is given as: score (j, t) = Elan h (Uas (t- 1) + WMj) , where Va, Wa, and Ua are trainable weights that comprise the attention mechanism. Va weights the alignment score, while Ua and Wa weight the hidden states of the decoder 2612and encoder 2608 ( / ?( / )), respectively. To obtain the attention weights a(j, f) for each time step / , the alignment score is normalized using the SoftMax function in conjunction with the time steps j. The attention weights enable the trainable weights to capture the relative importance of the input at time step j in decoding the output at time step t. Once the attention weights are determined, the context vector is computed as the weighted sum of the relationship between all the encoder 2608 hidden values (Z?(z)) and the attention weights.

[0155] By incorporating the context vector into the decoding process (decoder 2612), the attention mechanism 2610 enables more focus on the relevant inputs in the pre-processed / fdtered sEMG data 2604. The context vector (c( / )) is passed through the decoder 2612 to calculate the probability for the next possible output, which is performed for all time steps at the input. After this step, the current hidden state s(t) is computed based on the context vector (c( / )), the previous hidden state s(J - 1), and the output y(t - 1) from the previous time step as follows:where the notation f() indicates a mapping function. The attention mechanism 2610 allows the second ML - Al model 2606 to identify the correlations between different parts of the input sequence an the corresponding parts of the output sequence. At each time step, the decoder 2612 output is computing using the SoftMax function applied to the hidden state, which is in turn computed suing the context vector, previous hidden state, and previous output.

[0156] During training (e.g., at the second stage), the training computing system 2104 may receive imaging data from the imaging system 2106. For instance, the imaging system 2106 may obtain images of the second individual. The training computing system 2104 and / or the imaging system 2106 may determine ground truth data (e.g., sensor data such as actual joint angles) based on the images. Therefore, based on providing an input into the second ML - Al Model 2606, the encoder 2068 may process the input (e.g., the raw and / or pre-processing sEMG data) and generate an encoder output that is provided to the attention mechanism 2610 (e.g., a GRU). Based on the encoder output, the attention mechanism 2610 may generate an attention mechanism output and provide the output to the decoder 2612. Based on the attention mechanism output, the decoder 2612 may generate a decoder output, which may indicate predicted sensor data such as a predicted joint angle. Then, the training computing system 2104 may determine and / or compute a loss (e.g., by using a binary cross-entropy loss function and / or other loss functions). For example, the training computing system 2104 may compare the decoder output (e.g., the predicted joint angle) with the ground truth data (e.g., the sensor data from the imaging system 2106) to compute the loss. Based on the loss, the training computing system 2104 may update the parameters of the attention mechanism 2610 using one or more optimization algorithms such as ADAM described above. In some instances, the training data (e.g., the sEMG sensor data 2602) may include theground truth data. For example, as mentioned previously, for transfer learning, the training computing system 2104 may train the second ML - Al model initially using sensor data from a plurality of first individuals. As such, initially, the training computing system 2104 may use the ground truth data from the sEMG sensor data 2602 to train the sceond ML - Al model 2606 and the attention mechanism 2610. Then, using imaging data from the imaging system 2106, the training computing system 2104 may perform additional training of the sceond ML - Al model 2606 and the attention mechanism 2610.

[0157] Once the attention mechanism 2610 is trained, the system (e.g., the training computing system 2104 and / or the powered gait assistance system 2108) may use the attention-based recurrent neural network that includes the encoder 2608, attention mechanism 2610, and decoder 2612 to predict joint angles using the pre-processed / filtered sEMG data 2604. In some examples, the predicted joint angles may be used, as described above with reference to FIG. 25, to determine a gait phase from a plurality of gait phases of a gait cycle and / or generate predicted torque values using the MLP. The predicted joint angles may be used to generate the predicted moment and corresponding assistive torque to be applied by the motor of the powered gait assistance system to assist an individual’s volitional knee muscle output during stages of the individual’s gait cycle. Additionally, and / or alternatively, the system may use the second ML - Al model 2606 separately from the first ML - Al model 2402. For instance, the system may use the trained second ML - Al model 2606 to predict sensor data such as joint angles and joint moment, which is related directly to the effort exerted by the individual using the powered gait assistance system, and use the joint angles to perform and / or assist an individual to perform one or more mobility tasks. For example, the joint moment can be used in rehabilitation processes or strengthening processes for an individual to maximize the energy or effort that the individual outputs while using the powered gait assistance system. Although some of the examples described herein are related to predicting knee joint angles other joint angles may be predicted based on the trained second ML - Al model using the sEMG data such as hip or ankle joint angles and hip, knee, or ankle joint moments in three dimensions. In some instances, the trained second ML - Al model may use the sEMG data to predict underfoot center of pressure location of an individual. The underfoot center of pressure location may provide information on forward progression of an individual as they use the powered gait assistance system which can be used in gait determination, by the MLP, or for other purposes such as training or rehabilitation as described herein.

[0158] In some examples, during the training and / or the inference phases for using the second ML - Al model 2606, the system may use one or more channels and / or segmentations to separate the sEMG data / sEMG signals. For example, the sEMG sensors may obtain sensor data from a plurality of different muscles of the individual (e.g., eight muscles). The system may further segment the sEMG data from the different muscles into a plurality of segments (e.g., five segments). The system may then use the segmented sEMG data (e.g., forty total segments based on the five segments for each of the eight muscle groups) to predict the sensor data during the training and inference phases described above.

[0159] FIG. 27 is an exemplary process 2700 for training an attention-based recurrent neural network in accordance with one or more examples of the present application. The process 2700 may be performed by a system (e.g., the training computing system 2104 and / or powered gait assistance system 2108). However, it will be recognized that the process 2700 may be performed in any suitable environment and that any of the following blocks may be performed in any suitable order.

[0160] At block 2702, the system receives sEMG training data from a data source for training an attention-based recurrent neural network comprising an encoder, a decoder, and attention mechanism - as depicted and described above with reference to FIG. 26.

[0161] At block 2704, the system trains, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input of the decoder.

[0162] At block 2706, the system, subsequent to training the encoder and the decoder, uses a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided as an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

[0163] As described above with reference to FIG. 26, the attention mechanism, once trained, is integrated into the second ML - Al model such that the second ML - Al model is trained an attention-based recurrent neural network that is configured to process input of sEMG data (pre- processed or filtered) to determine predicted joint angles. The powered gait assistance system and control system may use the predicted joint angles to determine what stage of a gait cycle an individual’s leg is in. A motor may apply an assistive torque between a first arm and second arm of the powered gait assistance system as the determined stage of the gait cycle may be mapped or otherwise correspond to a certain assistive torque. The assistive torque may be applied by themotor via the first arm and second arm of the powered gait assistance system to assist an individual’s volitional knee muscle output during stages of their gait cycle. In some examples, after training the second ML - Al model 2606, the training computing system 2104 may provide the trained second ML - Al model 2604 to the powered gait assistance system 2108. At block 2306, the powered gait assistance system 2108 may obtain or receive sEMG data for the second individual using the one or more sensors 2110 that include the sEMG sensors as well as actual sensor data from the one or more sensors of the powered gait assistance system 2108. Continuing at block 2306, the powered gait assistance system 2108 may process the obtained sEMG data to generate an inference output (e.g., predicted sensor data such as a predicted joint angle). In other examples, the training computing system 2104 might not provide the trained second ML - Al model 2606 to the powered gait assistance system 2108. In such examples, at block 2306, the training computing system 2104 may receive the sEMG data for the second individual from the powered gait assistance system 2108. In such an example, the training computing system 2104 may process the sEMG data using the trained second ML - Al model 2606 to generate an inference output (e.g., predicted sensor data such as a predicted joint angle). At block 2308, the training computing system 2104 may process the sensor data, the predicted sensor data, and the actual sensor data to generate an inference output using the trained first ML - Al model 2404.

[0164] At block 2310, the system may apply the predicted torque values to a motor of the powered gait assistance system 2108 to cause the motor to apply assistive torque, that corresponds to the predicted torque values, between first and second arms of the powered gait assistance system to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle. For instance, in examples where the training computing system 2104 provides the trained first ML - Al model 2404 and the trained second ML - Al model 2606 to the powered gait assistance system 2108, the powered gait assistance system 2108 may use the inference output that it predicted to apply the predicted torque values to the motor. For example, based on a binary indication of whether to provide extension assistance, the powered gait assistance system 2108 may determine whether to apply an additional torque amount or value (e.g., a pre-defined and / or user-defined such as provided by a clinician and / or the second individual as described above) to provide extension assistance. Based on the predicted gait phase, the powered gait assistance system 2108 may determine a torque value associated with the predicted gait phase (e.g., a torque value associated with a late stance or early swing), and apply the determined torquevalue. Based on directly predicting a torque value, the powered gait assistance system 2108 may apply the torque value that is predicted by the trained first ML - Al model 2404. In other examples where the training computing system 2104 does not provide the trained first ML - Al model 2404 and the trained second ML - Al model 2606 to the powered gait assistance system 2108, the training computing system 2104 may use the output from the second ML - Al model 2606 as input to the first ML - Al model 2404, and provide the output of the first ML - Al model 2404 to the powered gait assistance system 2108. The powered gait assistance system 2108 may then apply the predicted torque values based on the received output.

[0165] The following provides further background and description of exemplary embodiments of the present disclosure, which may overlap to some extent with some of the information provided above. To the extent the terminology used to describe the exemplary embodiments may differ from the terminology used to describe the above embodiments, a person having skill in the art would understand that certain terms correspond to one another in the different embodiments. Features described below can be combined with features described above in various embodiments.

[0166] As described above, the powered gait assistance system may utilize onboard sensors to track or capture kinematic measurements of a user (e.g., knee angle, knee angular velocity, footground contact, and resultant torque applied at the knee) to parse a gait cycle of a user into discrete phases or states during walking. This process is termed a finite state machine (FSM) and can include a variable number of states (e g., from 3-5). The FSM may serve as a highest level of a hierarchical control system used by the powered gait assistance system and / or control system. Within each of these gait phases, a separate lower level real-time control law can be activated to provide assistive or resistive torque to the knee to provide mobility assistance, gait training, or both.

[0167] The FSM may work by specifying transition thresholds that govern when the FSM moves from one state to another. In most instances, these thresholds are pre-tuned values for a single (or combinations) of sensor output from the powered gait assistance system. For example, to transition from late stance to early swing, the foot sensor reading needs to exceed a pre-set value. As described herein, the MLP may automate the specification and tuning of these transition thresholds. Once the MLP is trained it may automatically identify correct sensor variables for the individual to maximize the accuracy of state determination. The MLP may continually be updated through the obtaining of new sensor data or continual sensor data that is collected as the userutilizes the powered gait assistance system. The updated or continual sensor data may be used to update the thresholds based on the determined states and relative accuracy of the output of the MLP.

[0168] In embodiments, training data for training the MLP may be obtained by sensors of the powered gait assistance system from walking experiments. The powered gait assistance system may implement a hierarchical control system that has three layers. The supervisory layer consists of a five state finite state machine (FSM) that splits the gait cycle into discrete phases: early stance, mid-stance, late stance, early swing and late swing. Within each phase, a pre-tuned value of knee extension assistance is provided, or alternatively, if no assistance is provided a net zero torque is imposed across the knee to account for the device inertia and friction. At the lowest level a PID based feedback controller combined with a feedforward loop is implemented to output the desired torque level. The FSM uses on-board sensors to transition between phases. The transition between swing phase and stance phase is dictated by a force sensitive resistor (FSR) placed under the foot. Two thresholds are set to account for hysteresis in the FSR. The default mode is stance and when the FSR value falls below the swing threshold the controller transitions to swing; likewise when the FSR reading exceeds the stance threshold it returns to stance phase. Within stance and swing phases, transitions are dictated by knee angular velocity measured by differentiating the encoder position data. The transition from early to mid-stance and from early to late swing are indicated when the knee flexion velocity transitions from flexion to extension, though the threshold value for each of these transitions can be tuned separately. The transition from mid-stance to late stance is dictated by when the knee angular velocity that exceeds a threshold in the flexion direction. Thus, in total, there are 5 FSM thresholds (two FSR and three angular velocity) that must be tuned for each limb. Typically, these thresholds are tuned when the participant uses the powered gait assistance system in the first session and then are adjusted as needed at the beginning of each experimental session.

[0169] The present disclosure describes a multilayer perceptron (MLP) that eliminates the need to tune these thresholds manually and instead segment the gait cycle - and thereby provide knee extension assistance during overground walking - automatically and accurately. This approach combines the clinical utility of the FSM, i.e., the ability to specify that extension assistance should be provided during different phases of stance and swing, without the necessary step of manually tuning thresholds. To establish the initial feasibility of this approach, a study wasconducted where previously collected data from one participant with CP walking with the powered gait assistance system was analyzed. At the time of enrollment in the study, the participant was 13 years old, a female, had a weight of 41 :9 kg, height of 151 cm and was assessed clinically as GMFCS III. In total, there are 10 visits as part of the study, including one initial visit for a baseline gait assessment, 8 intermediate visits for walking practice with the powered gait assistance system, and final gait assessment with and without the powered gait assistance system on the 10th visit. In each visit, the 5- state FSM was deployed to provide knee extension assistance during early stance, mid-stance and late swing phase. The thresholds for the FSM were tuned manually for each limb during the second visit, and these thresholds were manually updated at the beginning of each visit if necessary. Data from the onboard sensors, i.e., knee angle, knee angular velocity, FSR, and motor output torque, of the powered gait assistance system were collected at 100 Hz and examined.

[0170] The output of an automatic labeling algorithm was examined during 8 trials and it is concluded that it has less than 2:7% error with standard deviation of 1 :4% across all trials. Therefore, due to the low error of automatic labeling, its output is used as the ground truth for the subsequent offline analysis of both the FSM and the MLP based methods. An offline analysis was performed to assess the feasibility of using MLP to segment the gait cycle into these phases and provide the desired assistive torque. The MLP model was trained and tested using data from one of the early walking tests with the powered gait assistance system using cross validation procedures. Specifically, 90% of the data was used to train the model and the remaining 10% was used for evaluation. Prior to training, an automatic labeling algorithm was deployed to assign to each data point whether the assistive torque should be provided (a value of 1) or not (a value of 0). This framework was chosen because it isolates the MLP algorithm so that it only indicates the timing of when extension assistance is provided and not the magnitude, thereby allowing a clinician or user to maintain control over the amount of assistance.

[0171] Training of the MLP algorithm was performed using a PYTORCH package in PYTHON environment on a computer with an INTEL core i7-7600 processor with the clock speed of 2.8 gigahertz (GHz) and a 15.9 gigabytes (GB) installed memory. A stochastic optimization algorithm, known as ADAM was used because its implementation is straightforward, it is well suited for problems that are large in terms of parameters and / or data, invariant to diagonal rescaling of the gradients, computationally efficient, has adaptive estimates of lower order moments and needs a low memory. The learning rate was chosen as 0:0005. Batches of data points have beenused for the training to reduce the effect of outliers and increase the speed of training. The size of each batch was chosen as 8. Additionally, the model was chosen to have 20 neurons and two hidden layers (4 layers total) and was trained over 40 epochs. The learning rate was chosen as the highest number that did not lead to accuracy oscillation of more than 5%. Number of layers, neurons, and epochs were chosen as the least possible numbers that change the accuracy less than 1 percent. Furthermore, for avoiding high velocities to dominate, the velocity input is saturated by a maximum velocity of 20 deg . Saturation of velocity data was done also to minimize the effect of outliers on the deep networks. The maximum velocity value was chosen as the least number that doesn’t reduce the accuracy more than 1%. It should also be noted that higher numbers can be chosen for the values of number of layers, neurons, epochs, and the maximum velocity.

[0172] Walking time for participants using the powered gait assistance system ranged from 40 seconds (secs) and 35 secs for the participant with CP and the healthy participant respectively. The study showed that training data that has a record of sensors during walking for more than 40 secs should provide enough data for the training. The study showed that the accuracy of the right and left legs for the participant with CP after 40 secs of training are 83.6% and 84.9% respectively. It is also resulted that the estimation accuracy for the right and left legs of the healthy participant after 40 secs of training are 91.2% and 88.6% respectively.

[0173] After training, the MLP algorithm was evaluated for accuracy by systematic comparison with the manually tuned 5-state FSM. Accuracy was assessed offline as the true positive rate, i.e., the number of data points in which the MLP algorithm correctly identified whether assistance was to be provided (i.e., value of 1) or not provided (i.e., value of 0) using the automatic labeling output data as the ground truth. This evaluation was performed using the data from nine visits. It should be noted that although the FSM threshold was tuned if necessary during each of these additional visits, the MLP model was only trained one time on the data from the initial visits. After the implementation of the developed MLP based method using the pre-collected data and a statistical analysis of the data, it resulted that the MLP method could achieve an accuracy of 87% on the data set. For having a better insight about the performance and superiority of the method, the control method and the FSM based control method were implemented and then compared for the both legs of the participant. During analysis, it was found that the right leg and left leg have very different gait cycles which makes it more challenging for the developed algorithm to label and learn the participant gait. It also makes the tuning of an FSM based controlmethod more challenging. According to the results over all trials for all visits, the developed MLP method had an accuracy of 81.6% and 78.9% for right and left legs respectively while the on board online finite state machine based control method have the accuracy of 72.0% and 64.1% for right and left legs respectively and offline finite state machine based control method have the accuracy of 78.5% and 68.7% for right and left legs respectively. Therefore, the developed MLP method had 1.51 times and 1.70 times more accuracy than the online FSM based control method for right leg and the left leg respectively and it has 1.16 times and 1.48 times more accuracy than the offline FSM based control method for right leg and the left leg respectively. Therefore, the online and offline FSM based method has lower accuracy despite being tuned before and during all trials by the operator. One of the reasons that the FSM tuning was essential for each visit was getting various values of FSR. However, the MLP based method worked with a high accuracy despite these changes without tuning for each visit. Additionally, the study showed that the MLP based method was able to enforce torque in the right gait location, despite various values of torques has been used for the trials across the visits.

[0174] EMG signals offer a unique and direct insight into the neuromuscular activity associated with human movement. By analyzing these signals, one can extract valuable information about muscle activation patterns, which, in turn, can be translated into meaningful kinematic estimates, such as knee angles. The development of a robust and accurate knee kinematic estimator using neural networks can significantly improve the performance and adaptability of robotic systems in response to human movements such as in individuals using powered gait assistance systems.

[0175] To evaluate the performance of the attention-based neural network, a study was performed that followed a two-fold approach. Firstly, as LSTM is a commonly used RNN architecture in the literature, a comparison was made between the attention-based estimation and the estimation using the Gated Recurrent Unit (GRU), an advanced version of LSTM. A dataset was utilized to train and test both models, enabling a comprehensive assessment of their predictive capabilities. Secondly, the intra-subject evaluation involved assessing the model's capability to adapt continuously within a single subject. The attention-based model was tested on the walking data of healthy subjects, allowing insights into its ability to perform consistently during continuous locomotion task. Furthermore, the attention-based model's effectiveness was examined in terms of its adaptability to different subjects, considering both inter-subject and intra-subject scenarios. Forinter-subject assessment, three healthy subjects from a dataset and four subjects from in-house studies were used. Another dataset (the researchers' dataset) consisted of three healthy subjects and one subject with cerebral palsy (CP). By evaluating the attention-based model's performance on data from diverse healthy and CP subjects, its ability to generalize across different individuals was thoroughly investigated. By dividing the assessment into these two folds and leveraging datasets with diverse subject profiles, the study aimed to provide a comprehensive understanding of the attention-based neural network's strengths and limitations in knee angle estimation. The results of these evaluations contribute valuable insights into the model's versatility, adaptability, and overall efficacy, and further advance the field of motion analysis and assistive technologies.

[0176] During this study, two distinct datasets were employed to train and validate the neural network for knee angle estimation. The first dataset utilized in the study is from a publicly available Georgia Tech dataset. This dataset encompasses locomotion trials performed by 22 healthy subjects across four diverse terrain types: treadmill, level ground, ramp, and stairs. During the experiments, the subjects were equipped with wearable sensors, motion capture technology, and force-plates to capture comprehensive motion and force data. To process the motion capture data, the researchers utilized their custom MATLAB toolbox, MOCAP Tools, which incorporates a gap filling method to address missing data and provides a user-friendly application programming interference (API) to interface with C3D files and OPENSIM.

[0177] The second dataset was collected through experimental overground trials conducted in a Human Motion Lab. This dataset comprises data obtained from participants in controlled experimental settings. While specific details of the data collection process may vary from the other dataset described above, it similarly provides valuable motion information related to knee angles during locomotion tasks. The ability estimate knee angle during over-ground walking in two experiments was examined. In the first, four healthy volunteers (1 male, 3 females; age: 14.9 ± 1.1 years; height: 164 ± 7 centimeter (cm); weight: 73 ± 18 kilogram (kg)) traversed a 5.5 meter (m) walkway repeatedly at self-selected pace. Only sEMG from steady state walking was retained for analysis. EMG was recorded at 1000 Hz bilaterally from the tibialis anterior (TA), medial gastrocnemius (MG), soleus (SL), peroneus longus (PL), hallucis longus (HL), rectus femoris (RF), vastus lateralis (VL), and medial hamstrings (MH). Kinematic data was recorded at 100 Hz using motion capture. In the second experiment, data from one participant with cerebral palsy (CP) (age: 7 years; height: 111 cm; 251weight: 18.5 kg) walking overground with and without kneeextension assistance from a powered gait assistance system was used. Like the healthy volunteers, a participant with CP walked overground across a 5.5 m walkway at self-selected pace. Here, two conditions were examined. The first was a baseline condition in which the participant walked without powered gait assistance system assistance. The second condition involved walking overground while the powered gait assistance system provided knee extension assistance during early stance, mid stance, and late swing phases of the gait cycle. In the powered gait assistance system experiment, sEMG was recorded at 2000 Hz bilaterally from tibialis anterior (TA), medial gastrocnemius (MG), vastus lateralis (VL), rectus femoris (RF), and semitendinosus (ST). Custom pockets in the powered gait assistance system were created to accommodate the sEMG electrodes and prevent interference during recording. Motion capture data was recorded at 100 Hz using a custom marker set placed on the pelvis, thighs, shanks and feet. For both experiments, kinematics data was processed offline using Visual 3D software to compute lower limb joint angles. Kinematic data from the foot markers were used to segment the walking data into gait cycles for each limb spanning from one foot-ground contact event to the next. A multiple stage fdtration procedure was applied to process the raw EMG signals.

[0178] In the preprocessing phase, a multi-stage filtration procedure was employed to process the raw electromyographic (EMG) signals. This procedure aimed to enhance the quality and reliability of the EMG data for subsequent knee angle estimation. The first stage of the filtration involved applying a Hampel filter with a window size of 20 samples. This filter was utilized to remove any outliers present in the raw EMG signals, effectively reducing noise and artifacts that could interfere with accurate signal analysis. Subsequently, a bandpass filter was implemented, with cutoff frequencies set at 5 Hz and 9 Hz. This filter served to pass EMG signal components within the desired frequency range, effectively removing unwanted noise and irrelevant frequencies that may have been captured during data acquisition. The next step involved rectification of the filtered EMG signal using a sliding window of 50 samples. Consecutive peak and root mean square (RMS) envelop detectors were then applied to the rectified signal. This process aimed to emphasize the amplitude and temporal characteristics of the EMG signal, enabling a more precise analysis of muscle activation patterns. To further refine the output, a low- pass filter was applied to the signal to smooth the filtered EMG data. This step facilitated the removal of any remaining high-frequency noise and resulted in a cleaner and more reliable EMG signal for subsequent processing. It is worth noting that the parameters of the multi-stage filter,such as window sizes and cutoff frequencies, were fine-tuned through trial and error. This iterative process ensured that the filtration procedure consistently and accurately performed on both datasets, enhancing the robustness and reliability of the knee angle estimation model.

[0179] During the gait cycle, the EMG signals were divided into five segments, resulting in a total of 40 segments across the 8 muscles. MATLAB software was used to preprocess and segment the data. The deep learning attention model was created using the Keras library in PYTHON. ADAM optimization algorithm and binary cross-entropy loss function were used to train the model. The dataset of healthy volunteers was split into 60% for training and 40% for testing.

[0180] The study included a series of comparative assessments of the proposed model in two forms: one with the attention mechanism neural network and one without the attention mechanism. This comparison was first performed on data from overground walking in 4 healthy adolescents. This analysis focused on the dominant (right) limb. The input dataset included the 8 channels of sEMG and the output was the resultant knee joint angle trajectory during walking. The primary outcome measure was the RMSE between the motion capture measured knee joint angle and the one estimated from the models. The correlation coefficient between the estimated and measured joint angles was also computed. To further evaluate the model and its sensitivity to the input muscle set, each sEMG was removed sequentially and the RMSE and correlation coefficient were computed. The study also assessed the feasibility of the deep learning attention-based neural network to estimate knee angle during overground walking in a participant with CP under two conditions: overground walking with and without robotic knee extension assistance from a powered gait assistance system. Like in the evaluations with healthy volunteers the analysis was restricted to a single limb.

[0181] Data from the healthy volunteers was split such that 60% was randomly assigned for training and the remaining 40% was used for testing. The training results showed an average model accuracy, defined as mean absolute percentage error (MAPE), of 92.3% for the GRU with the attention mechanism and 88.7% for the GRU without the attention mechanism. The testing results showed that, on average, the GRU network with attention mechanism results in a lower average RMSE when estimating knee joint angle than the GRU without the attention mechanism. Similarly, the correlation coefficient was higher on average when the attention mechanism was included in the estimator. Representative gait cycles from each of the four healthy participants demonstratedthe ability of the GRU with attention mechanism to estimate the knee angle profile during overground walking in adolescents.

[0182] The GRU with attention mechanism was used to estimate knee angle during overground walking in a participant with crouch gait from CP with and without assistance from a powered gait assistance system. A mean RMSE of 13.7 and 22.8 for each condition, respectively; this is associated with correlation coefficient of 0.86 and 0.67 respectively. These results indicate knee angle estimation improved with the assistance applied, possibly due to the expanded range of knee motion that occurred in this condition.

[0183] The transfer of a pre-trained model, initially trained on data from healthy subjects, was adapted for use on the CP subject’s data collected during different visits. The objective was to assess the model’s performance and generalization capabilities when applied to the CP subject’s gait data in a longitudinal manner. The study also demonstrated the ability of the GRU-RNN with attention mechanism to estimate knee angle in a participant with CP. Importantly, this participant was assessed clinically with mild-to-moderate spasticity of the knee extensor and flexor muscles. The estimator as constructed was able to predict knee angle, although with a lower level of accuracy compared to healthy controls, particularly in the condition without powered gait assistance system assistance. Interestingly, the study observed a similar RMSE magnitude in presence of powered gait assistance system torque of 0.11 newton meter (Nm) / kg being applied during early stance, mid-stance, and late swing phase of walking as was observed in healthy volunteers. Collectively, these results demonstrate the ability to estimate knee angle during overground in an individual with CP while walking with a powered gait assistance system. The results of the study establish the initial feasibility of this approach and suggest that in the future, it can be applied to estimate user intent before movement execution, given the up to 300 millisecond (ms) delay between sEMG measurement and resultant limb motion. This delay means provides the opportunity to execute a closed loop system which adapts to user intent before the movement execution. Such functionality is particularly important for long-term goal of developing robotic exoskeletons such as powered gait assistance systems which can be used across a range of real- word environments in children with neuromuscular disorders.

[0184] Estimating joint angles from sEMG typically involves four steps: signal acquisition, signal processing, feature extraction, and model selection. The study described above focuses on feature extraction and model selection for the purposes of estimating knee angle during walking,with the ultimate goal of creating novel approaches to controlling a wearable knee exoskeleton (e g. powered gait assistance system). To continue building on the experiments described above, another experiment was conducted in which the study evaluated the transfer learning capability of the GRU-AM model by first training it on an open-source data set before training again on the same data set as the first experiment (four typically developing adolescents) and then assessing its predictive performance compared to the first experiment.

[0185] As mentioned above, Attention mechanisms, which are advancements of encoderdecoder models, have emerged as one of the frontiers in deep learning for improving the performance of processing long input sequences. The attention mechanism assigns a weight or importance score to each input element, based on its relevance to the current system output. This allows the model to selectively focus on the most salient parts of the input and ignore irrelevant or noisy information. The attention mechanism involves three main components: a query, a set of key-value pairs, and a scoring function. The query is a vector that represents the current context or state of the model, while the key-value pairs represent the input data. The scoring function computes a similarity score between the query and each of the key-value pairs, which is used to compute a weight or importance score for each value. The values are then combined using a weighted sum, where the weights are the importance scores computed by the attention mechanism. The attention mechanism can be integrated into various neural networks, including RNNs, convolutional neural networks (CNNs), and transformer models. The transformer / attention mechanism model (Figure 1), which was introduced first for natural language processing, is a popular and powerful model architecture that deploys the attention mechanism to achieve excellent performance on a range of tasks.

[0186] The decoder in attention mechanisms can selectively access the encoded information, using the context vendor c(t), which is computed at each time step of the decoder based on the previous hidden state and all the hidden states of the encoder. The relative importance of each element in the input sequence is determined by trainable weights assigned to these states, with greater emphasis placed on the most significant inputs, giving rise to the name “attention mechanisms”. The context vector is constructed by combining each time step t of the decoder, and the alignment score. The formula for computing the alignment score is given as: score(j, t) = vatanhwhere Va, Wa, and Uaare the trainable weights that comprise the attention mechanism. Vaweights the alignment score, while Uaand Waare associated with the hidden states of the decoder (s(t)) and encoder (h(j)), respectively. To obtain the attention weights oc ( / , t) for each time step t, the alignment score may be normalized. This is achieved using the SoftMax function in conjunction with the time steps j, and the resulting attention weights are defined as follows:

[0187] The attention weights enable the trainable weights to capture the relative importance of the input at time step j in decoding the output at time step t. Once the attention weights are determined, the context vector is computed as the weighted sum of the relationship between all the encoder hidden values h(i) and the attention weights. This relationship is given by:

[0188] By incorporating the context vector into the decoding process, attention mechanisms enable more focus on the relevant inputs in the EMG features. The context vector c(t) is passed through the decoder to calculate the probability for the next possible output, which is performed for all time steps at the input. After this step, the current hidden state s(t-l), and the output y (t — 1) from the previous time step as follows: s(t) = / (s(t - 1), y(t - 1), c(t))

[0189] The attention mechanism described above allows the model to identify the correlations between different parts of the input sequence and the corresponding parts of the output sequence. At each time step, the decoder output is computed using the SoftMax function applied to the hidden state, which is in turn computed using the context vector, previous hidden state, and previous output. This approach enables the model to selectively focus on the most relevant parts of the input sequence, leading to improved performance in generating the corresponding output sequence.

[0190] A GRU is a type of RNN used for sequential data processing. It consists of a hidden state that captures information from previous time steps. Two critical components, the update gate and the reset gate, regulate how information is retained or discarded in the hidden state. The update gate determines how much of the previous hidden state should be preserved, allowing the network to adapt its memory usage dynamically. Simultaneously, the reset gate controls the extent to which the previous hidden state is reset, enabling the GRU to respond to changing patterns in the data. The candidate hidden state, generated using the current input and the previous hidden state, isblended with the previous hidden state according to the update gate, producing the new hidden state or output. This design helps the GRU capture long-range dependencies in sequential data while addressing the vanishing gradient problem, making it suitable for various applications like natural language processing and time series analysis.

[0191] In this study, an encoder-decoder architecture with an attention mechanism was employed. Specifically, the encoder-decoder units consist of 100 gated recurrent unit (GRU) cells. To enhance this architecture, a precisely defined Attention mechanism known as the Bahdanau attention mechanism was incorporated. This integration results in the creation of the attentionbased recurrent neural network model described herein, the GRU with an attention-based mechanism (GRU-AM). To evaluate the performance of the GRU-AM estimator model, the root mean squared error (RMSE) and correlation coefficient (CC) are used as a performance metrics.RMSE is computed as:where N is the number of data points in the time series to be estimated, 0, is the time series’ measured value, and 0(is the prediction model’s estimated value at sampling time i. CC is calculated as:where 0, and 0 is the mean of the measured and estimated knee angle vectors, respectively.

[0192] To further gauge estimation quality, the signal-to-noise ratio (SNR) of the estimate knee angle serves as an additional metric and is calculated as:

[0193] Each EMG signal corresponding to a single gait cycle was partitioned into five segments, each consisting of 20 samples. This resulted in a total of 40 segments for the 8-muscle input model and 30 segments for the 6-muscle channel model. MATLAB R2022b software was used to preprocess and segment the data. The deep learning attention model was created using the Keras library in PYTHON 3.11. An ADAM optimization algorithm and binary cross-entropy loss function were used to train the model. The dataset was split into 60% for training and 40% fortesting. Regarding the training session of the attention-based mechanism, the session concludes either upon reaching an RMSE of less than 15 degrees or after a maximum of 10,000 epochs. Notably, attempts to enhance the termination criteria by relaxing the RMSE threshold to less than 20 or by increasing the number of epochs yielded insignificant improvements in performance. This underscores the importance of achieving a balance between accuracy and computational efficiency in the training process.

[0194] To further study the attention-based recurrent neural network and to simulate the scenario of applying the pre-trained model to new healthy individuals, a sequential training approach was employed across the four healthy adolescents from the previous experiment, with a training interval of three strides. The procedure was performed as follows:1. Initial training: The GRU-AM model was trained using all trials from the four healthy participants in the open-source dataset.2. Usage by Subject No. 1 : The GRU-AM model was further trained with one trial of subjectspecific data involving the same number of strides as used in the open-source dataset participants’ trials (10 strides). Subsequently, the remaining data from each individual was used for knee angle estimation, with sequential training conducted every 3 strides. The evaluation of this process included monitoring the CC and RMSE for each stride.3. The same procedure as Step 2, but applied to Subject No. 2 after the initial training in Step 1.4. The same procedure as Step 2, but applied to Subject No. 3 after the initial training in Step 1.5. The same procedure as Step 2, but applied to Subject No. 4 after the initial training in Step 1.

[0195] As described above, the study or experiments conducted also focused on the progressive adaptation in a participant with CP. The goal of this experiment was to evaluate the ability of the GRU-AM model to estimate knee joint angle during walking in a child with CP across a series of consecutive visits, which occurred approximately once every two weeks. The GRU-AM model was pre-trained using data from healthy participants from the open source and other datasets, including four participants from each dataset. The model was initially trained using the open-source dataset, then progressively trained using each participant from the other dataset, and finally was tested in each of three sequential visits from the participant with CP.

[0196] A sequential learning policy was employed to continually update the GRU-AM model after every 3 strides, employing the sEMG data from the participant with CP. This experiment was designed to mimic real-world rehabilitation sessions, where the trained estimator is utilized topredict the knee angles of a participant with CP wearing a lower-limb exoskeleton or powered gait assistance system. The data used for testing and sequentially training the GRU-AM model originated from the participant’s first, second, and third visits for an ongoing study.

[0197] To further the GRU-AM sensitivity to the input muscle set, each sEMG was sequentially removed and the RMSE and CC computed. The model’s accuracy demonstrated a distinct pattern as muscles were incrementally removed. The four muscles that resulted in the greatest increase in RMSE when removed were the RF, MG, TA and VL, indicating they were primary contributors to knee joint angle estimation from sEMG. A key observation from this is that the removal of single sEMG channels affects both the shape of individual stride estimation and the degree of noise present, particularly manifesting as high-frequency distortion in the output. Additionally, the removal of RF, TA, MG, and VL significantly impacts the mid and late-stance phase estimate accuracy.

[0198] The transfer learning experiment was conducted to investigate the impact of transfer learning when moving from healthy individuals to individuals with CP. Initially, data from six sEMG channels, collected from four healthy participants in an open source dataset, were employed to train the GRU-AM estimator for knee angle prediction. The primary aim was to leverage a pretrained GRU-AM model from four healthy participants in the open-source dataset. The GRU-AM was personalized through initial training with 10 strides from each individual. Additionally, the experiment aimed to establish a Cyclic Update Period, a crucial consideration given the computational complexity in real-time applications. It was determined that a three-stride interval, equivalent to approximately 3 megabytes of memory resources, effectively balanced computational efficiency with the training data’s quantity. This strategic selection harmonized the update period with processing time and data volume, a critical factor for real-time learning. The estimated knee angle and associated performance metrics, namely RMSE and CC, were computed for each successive stride. This evaluation procedure is applied to four participants, encompassing two trials for each participant where each trial consists of 30 strides, encompassing 10 learning cycles.

[0199] The pre-trained GRU-AM exhibited a notable pattern of gradual improvement in estimation performance across trials in all four participants. The term “average convergence rate” encapsulates this enhancement rate, representing the progressive enhancement in estimation performance with each cycle. The average convergence rate is the slope of the line that connectsthe first stride’s RMSE and the last stride’s RMSE. For all participants within the dataset, the initial trial showcased an average convergence rate of -0.5 RMSE / stride. Notably, the third trial demonstrates a slightly slower rate of estimation enhancement (-0.2 RMSE / stride) compared to the second trial. This finding indicates that intra-participant performance improves through sequential learning, substantiating the efficacy of the approach. These outcomes underscore the potential of sequential learning to enhance the model’s estimation performance over consecutive cycles.

[0200] In the experiment for progressive adaptation in participants with CP, the GRU-AM model was initially trained with data from four participants from the open-source dataset and four healthy participants from the dataset. Then, the GRU-AM was evaluated with data from the participant with CP in a sequential learning experiment whereby the model was updated every three strides. The progressive performance of the knee angle estimation was done by recording the RMSE at the end of the three strides. The first visit of the CP participant comprised 15 strides, and the second and third visits comprised 30 strides. Using sequential learning to update the GRU-AM model reduced the RMSE by 20%, 16% and 15% within the first, second and third visits, respectively. This was true even as the gait speed increased progressively across visits. The sEMG signals recorded from the participant with CP exhibited higher noise levels compared to healthy participants. This noise, coupled with the recurrent nature of GRUs and the presence of the attention mechanism, tended to impact the quality of the estimator’s output resulting in greater RMSE in the first visit. However, the application of the sequential training policy contributes to noise reduction. This phenomenon leads to a discernible improvement in estimator performance over the course of the three visits. The reduction in RMSE values across successive cycles highlights the estimator’s competence in estimating knee angle from sEMG despite the presence of a powered gait assistance system.

[0201] The results of the experiments in using sEMG data with a trained GRU-AM demonstrate the ability to estimate knee angle during overground walking in healthy adolescents through the use of an RNN using GRUs and an attention mechanism. On average, the estimation improved with the attention mechanism compared to the same GRU without it, indicating its ability to focus on the most relevant parts of the inputs within each time window. In this case, this likely results in more heavily weighting muscles that are active during knee excursion in overground walking. A sensitivity analysis corroborated this finding as the GRU-AM prediction is mostsensitive to the knee extensor and flexor muscles. The results of these experiments and studies also underscore the model’s potential to generalize knowledge across diverse datasets via transfer learning, as a model that was trained on open source biomechanics data from healthy adults was successfully deployed to estimate knee angle during walking in healthy adolescents. Further, the sequential learning — in which the model was successively trained over three stride periods — showed continued improvement in knee angle estimation, demonstrating the attention based estimator model’s adaptability and robustness in knee angle estimation over time, supporting its eventual use in real-world scenarios. The experiments also demonstrated successful transfer and sequential learning to estimate knee joint angle during overground walking in a child with cerebral palsy (CP) while a powered gait assistance system was worn / used. Importantly, this participant was assessed clinically with mild-to-moderate spasticity of the knee extensor and flexor muscles and functioning at GMFCS Level III. The GRU-AM approach was capable of estimating knee angle with a reasonable accuracy, despite these alterations in neuromuscular functioning.

[0202] Additional descriptions (e.g., descriptions of the second ML - Al model) is described in Abdelhady, M.; Damiano, D.L.; Bulea, T.C. Knee Angle Estimation from Surface EMG during Walking Using Attention-Based Deep Recurrent Neural Networks: Feasibility and Initial Demonstration in Cerebral Palsy. Sensors 2024, 24, 4217. https: / / doi.org / 10.3390 / s24134217, which is incorporated by reference herein in its entirety.

[0203] It will be appreciated that the figures of the present application and their corresponding descriptions are merely exemplary, and that the application is not limited to these exemplary situations.

[0204] It will further be appreciated by those of skill in the art that the execution of the various machine-implemented processes and steps described herein may occur via the computerized execution of processor-executable instructions stored on a non-transitory computer-readable medium, e.g., random access memory (RAM), read-only memory (ROM), programmable readonly memory (PROM), volatile, nonvolatile, or other electronic memory mechanism. Thus, for example, the operations described herein as being performed by computing devices and / or components thereof may be carried out by according to processor-executable instructions and / or installed applications corresponding to software, firmware, and / or computer hardware.

[0205] The use of the term “at least one” followed by a list of one or more items (for example, “at least one of A and B”) is to be construed to mean one item selected from the listed items (A orB) or any combination of two or more of the listed items (A and B), unless otherwise indicated herein or clearly contradicted by context. The terms “comprising,” “having,” “including,” and “containing” are to be construed as open-ended terms (i.e., meaning “including, but not limited to,”) unless otherwise noted. Recitation of ranges of values herein are merely intended to serve as a shorthand method of referring individually to each separate value falling within the range, unless otherwise indicated herein, and each separate value is incorporated into the specification as if it were individually recited herein. All methods described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The use of any and all examples, or exemplary language (e.g., “such as”) provided herein, is intended merely to better illuminate the application and does not pose a limitation on the scope of the application unless otherwise claimed. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of the application.

[0206] It will be appreciated that the examples of the application described herein are merely exemplary. Variations of these examples may become apparent to those of ordinary skill in the art upon reading the foregoing description. The inventors expect skilled artisans to employ such variations as appropriate, and the inventors intend for the application to be practiced otherwise than as specifically described herein. Accordingly, this application includes all modifications and equivalents of the subject matter recited in the claims appended hereto as permitted by applicable law. Moreover, any combination of the above-described elements in all possible variations thereof is encompassed by the application unless otherwise indicated herein or otherwise clearly contradicted by context.

Claims

CLAIMS:

1. A method, comprising: obtaining a training dataset comprising training sensor data associated with measurements of a plurality of first individuals using one or more first powered gait assistance systems; training a multilayer perceptron (MLP) using the training sensor data; obtaining, from one or more sensors of a powered gait assistance system, sensor data associated with a second individual using the powered gait assistance system; processing the sensor data associated with the second individual using the trained MLP to generate predicted torque values for a second powered gait assistance system; and applying the predicted torque values to a motor of the second powered gait assistance system to cause the motor to apply assistive torque, that corresponds to the predicted torque values, between first and second arms of the second powered gait assistance system to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

2. The method of claim 1, wherein processing the sensor data comprises: inputting the sensor data into the trained MLP to determine a predicted gait phase from a plurality of gait phases of the gait cycle; and determining the predicted torque values for the powered gait assistance system based on the predicted gait phase.

3. The method of claim 2, wherein the one or more sensors comprises a torque sensor, an angle sensor, a foot sensor, and / or an inertial based measurement (IMU) sensor, wherein the measurements from the training sensor data and the sensor data includes kinematic measurements and dynamic measurements that are obtained by the one or more sensors of the first and the second powered gait assistance systems.

4. The method of claim 2, wherein a transition between gait phases of the gait cycle is determined by the MLP using the sensor data or transition thresholds for the gait phases aredetermined by the MLP using the sensor data, and wherein each gait phase of the gait phases includes a certain assistive torque applied by the motor.

5. The method of claim 1, wherein obtaining the sensor data associated with the second individual comprises: obtaining actual sensor data from the one or more sensors of the powered gait assistance system; obtaining surface electromyography (sEMG) data from one or more sEMG sensors of the powered gait assistance system; and processing the sEMG data using an attention-based recurrent neural network (RNN) to generate predicted joint angles, wherein processing the sensor data associated with the second individual comprises processing, using the trained MLP, the actual sensor data and the predicted joint angles that are generated using the attention-based RNN to generate the predicted torque values.

6. The method of claim 1, further comprising processing the sensor data associated with the second individual using the trained MLP to generate a binary value that indicates whether to provide extension assistance to the second individual.

7. The method of claim 5, further comprising training the attention-based RNN by: receiving sEMG training data from a data source, wherein the attention-based RNN comprises an encoder, a decoder, and an attention mechanism; training, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input to the decoder; and subsequent to training the encoder and the decoder, using a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided to an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

8. The method of claim 5, wherein the one or more sEMG sensors are located in an area that corresponds to a foot print of the powered gait assistance system.

9. The method of claim 7, wherein the actual sensor data from the one or more sensors is used as ground truth data, wherein training the attention mechanism includes using the ground truth data.

10. A computing environment, comprising: a training computing system configured to: obtain a training dataset comprising training sensor data associated with measurements of a plurality of first individuals using one or more first powered gait assistance systems; train a multilayer perceptron (MLP) using the training sensor data; and provide the trained MLP to the powered gait assistance device; the powered gait assistance system, comprising: a first arm and a second arm coupled together to allow relative pivoting between the first and second arms, wherein the first arm is configured to be coupled to an upper leg portion of a second individual and the second arm is configured to be coupled to a lower leg portion of the second individual such that relative pivoting between the first and second arms is about an axis proximate the second individual’s native knee pivot axis; a motor coupled to the first and second arms and operable to apply torque between the first and second arms to urge relative pivoting between the first and second arms; one or more sensors that are configured to provide sensor data; and a controller programmed to: obtain the sensor data associated with the second individual, the sensor data obtained by the one or more sensors as the second individual uses the powered gait assistance system; process the sensor data associated with the second individual using the trained MLP to generate predicted torque values for the powered gait assistance system; andapply the predicted torque values to the motor to apply assistive torque, that corresponds to the predicted torque values, between the first and second arms to assist the second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

11. The computing environment of claim 10, wherein processing the sensor data comprises: inputting the sensor data into the trained MLP to determine a predicted gait phase from a plurality of gait phases of the gait cycle; and determining the predicted torque values for the powered gait assistance system based on the gait phase.

12. The computing environment of claim 11, wherein the one or more sensors comprises a torque sensor, an angle sensor, a foot sensor, and / or an inertial based measurement (IMU) sensor, wherein the measurements from the training sensor data and the sensor data includes the kinematic measurements and dynamic measurements that are obtained by the one or more sensors of the powered gait assistance system.

13. The computing environment of claim 11, wherein a transition between gait phases of the gait cycle is determined by the MLP using the sensor data or transition thresholds for the gait phases are determined by the MLP using the sensor data, and wherein each gait phase of the gait phases includes a certain assistive torque applied by the motor.

14. The computing environment of claim 10, wherein obtaining the sensor data associated with the second individual comprises: obtaining actual sensor data from the one or more sensors; obtaining surface electromyography (sEMG) data from one or more sEMG sensors of the powered gait assistance system; and processing the sEMG data using an attention-based recurrent neural network (RNN) to generate predicted joint angles, wherein processing the sensor data associated with the second individual comprises processing, using the trained MLP, the actual sensor data and the predictedjoint angles that are generated using the attention -based RNN to generate the predicted torque values.

15. The computing environment of claim 10, wherein processing the sensor data associated with the second individual using the trained MLP includes generating a binary value that indicates whether to provide extension assistance to the second individual.

16. The computing environment of claim 14, wherein the training computing system is further configured to train the attention-based RNN by: receiving sEMG training data from a data source, wherein the attention-based RNN comprises an encoder, a decoder, and an attention mechanism; training, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input to the decoder; and subsequent to training the encoder and the decoder, using a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided to an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

17. A method, comprising: obtaining a training dataset comprising surface electromyography (sEMG) training data from a data source, wherein the sEMG training data is associated with a plurality of first individuals; training an attention-based model using the sEMG training data, the attention-based model comprising an encoder, a decoder, and an attention mechanism; obtaining sEMG data of a second individual using a powered gait assistance system; processing the sEMG data associated with the second individual using the trained attentionbased model to generate predicted joint angles for the second individual; and based on the predicted joint angles, instructing a motor of the powered gait assistance system to apply assistive torque between a first arm and second arm of the powered gait assistance system, the assistive torque corresponding to a stage of the second individual’s gait cycle, to assistthe second individual’s volitional knee muscle output during stages of the second individual’s gait cycle.

18. The method of claim 17, wherein training the attention-based model comprises: training, using a first portion of the sEMG training data, the encoder and the decoder based on providing an output of the encoder to an input to the decoder; and subsequent to training the encoder and the decoder, using a second portion of the sEMG training data to train the attention mechanism, wherein the output of the encoder is provided as an input to the attention mechanism, and an output of the attention mechanism is provided to the decoder.

19. The method of claim 17, further comprising: providing the predicted j oint angles and actual sensor data from one or more sensors of the powered gait assistance system to a trained multilayer perceptron (MLP), the actual sensor data captured by the one or more sensors as the second individual uses the powered gait assistance system; processing the actual sensor data and the predicted joint angles using the trained MLP to generate predicted torque values for the powered gait assistance system; and applying the predicted torque values to the motor of the powered gait assistance system to cause the motor to apply the assistive torque that corresponds to the predicted torque values.

20. The method of claim 19, wherein processing the actual sensor data comprises: inputting the actual sensor data into the trained MLP to determine a gait phase from a plurality of gait phases of the gait cycle; and determining the predicted torque values for the powered gait assistance system based on the gait phase.

Images