A stroke patient adaptive anti-fall training method based on multi-dimensional gait parameters

Abstract

The application discloses a kind of based on multi-dimensional gait parameter's stroke patient adaptive anti-fall training method, it is related to the technical field of rehabilitation robot.The method includes: real-time gait parameter acquisition step, through multiple sensors configured on rehabilitation robot, real-time acquisition multi-dimensional gait parameter of patient in training process;Dynamic fall risk assessment step, based on the multi-dimensional gait parameter, through a preset dynamic fall risk assessment model that gait symmetry, gait variability and time sequence attention mechanism are fused, dynamic fall risk index is calculated in real time;Adaptive control instruction generation step, the dynamic fall risk index is input to an adaptive controller, generate torque control instruction for adjusting rehabilitation robot auxiliary torque and feedback control instruction for providing biological feedback;Closed-loop training execution step, rehabilitation robot adjusts auxiliary torque according to torque control instruction, while biological feedback device provides sensory feedback to patient.The application can dynamically and prospectively assess fall risk, realize personalized adaptive assistance and active feedback, significantly improve the safety and effect of rehabilitation training.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation robot technology, and in particular to a gait training method for stroke patients.

[0002] More specifically, the present invention relates to an adaptive fall prevention training method for stroke patients based on multidimensional gait parameters. Background Technology

[0003] Stroke, commonly known as "cerebrovascular accident," is an acute cerebrovascular disease with a high incidence and disability rate. Most survivors suffer from varying degrees of motor dysfunction, with hemiplegic gait being one of the most common sequelae, severely impacting their daily living abilities and quality of life. Rehabilitation training is a key means of promoting the recovery of patients' motor function, and rehabilitation robots, as an effective supplement to traditional physical therapy, can provide high-intensity, repeatable, and quantifiable training, demonstrating enormous application potential in the field of stroke rehabilitation.

[0004] Existing control strategies for lower limb rehabilitation robots mainly include passive training modes and active assisted training modes. In the active assisted training mode, the robot needs to provide appropriate assistance based on the patient's movement intentions to encourage active participation. However, stroke patients, due to weakened muscle strength, poor coordination, and impaired balance, are highly susceptible to falls during walking training, causing secondary injuries. This not only interrupts the rehabilitation process but also causes significant psychological fear for the patients. Therefore, ensuring the safety of the training process, especially how to actively prevent falls, is a core issue that rehabilitation robot technology urgently needs to address.

[0005] For example, Chinese patent application number CN202010355333.1 discloses an "active lower limb rehabilitation robot control method based on healthy and affected side coupling". This method collects multi-sensor signals of the patient's healthy lower limb, identifies the patient's movement intention based on deep learning and other methods, and controls the movement of the affected side in combination with a preset pathological gait correction strategy.

[0006] However, the aforementioned existing technologies have the following shortcomings: their core control logic responds to the patient's "motor intention" rather than actively predicting and intervening in "fall risk." This method relies on recognizing the intention of the healthy side to drive the affected side. When a patient's overall gait is unstable and they are about to fall, their motor intention itself may be disordered or erroneous. In this case, intention-based control may not provide timely and effective fall prevention protection. Furthermore, the feature extraction methods used in this method, such as mean and standard deviation, are relatively simple and struggle to capture the dynamic changes in gait parameters over time. Falls are typically a dynamic process resulting from the accumulation of a series of minor gait abnormalities; therefore, this method has limited ability to prospectively assess fall risk.

[0007] For example, Chinese patent application number CN202411215261.5 discloses an "optimal constraint following control method for a lower limb rehabilitation exoskeleton robot considering human-computer interaction". This method establishes a robust constraint following control framework that includes human-computer interaction, aiming to ensure that the system achieves optimal control while satisfying the constraint conditions.

[0008] However, this technical solution has its shortcomings: its proposed control framework is too theoretical and generalized, focusing primarily on meeting mathematically defined performance constraints, without establishing a specific assessment model for the specific biomechanical characteristics of stroke patients' falls. The so-called "human decision-making" component is rather abstract, lacking a clear mechanism to transform real-time collected biomechanical parameters (such as gait parameters) into concrete, quantifiable fall risk indices. Therefore, this method is difficult to directly apply to clinical rehabilitation scenarios requiring real-time, accurate prediction and intervention for specific events (such as falls).

[0009] Therefore, current technologies generally lack an effective method to dynamically and proactively assess a patient's fall risk from multi-dimensional real-time gait data, and to provide adaptive, closed-loop fall prevention assistance and feedback accordingly. Thus, there is an urgent need to develop a more comprehensive technological solution to address these issues. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive fall prevention training method for stroke patients based on multidimensional gait parameters. This aims to solve the problems that existing rehabilitation training methods are unable to dynamically and prospectively assess fall risk and cannot provide accurate adaptive fall prevention assistance.

[0011] To achieve the above objectives, the present invention provides the following technical solution: an adaptive fall prevention training method for stroke patients based on multidimensional gait parameters, characterized by comprising: a real-time gait parameter acquisition step, wherein multiple sensors configured on a rehabilitation robot are used to collect multidimensional gait parameters of the patient in real time during the training process; a dynamic fall risk assessment step, wherein based on the multidimensional gait parameters, a dynamic fall risk index characterizing the patient's current fall probability is calculated in real time using a preset dynamic fall risk assessment model; the core of the dynamic fall risk assessment model is based on a weighted summation formula that integrates gait symmetry, gait variability, and temporal attention mechanisms; an adaptive control command generation step, wherein the dynamic fall risk index is input to an adaptive controller to generate torque control commands for adjusting the auxiliary torque of the rehabilitation robot, and feedback control commands for providing biofeedback to the patient; and a closed-loop training execution step, wherein the rehabilitation robot adjusts the auxiliary torque on the patient's affected limb according to the torque control commands, and the biofeedback device provides sensory feedback to the patient according to the feedback control commands, thereby realizing adaptive closed-loop control of fall prevention training.

[0012] As a preferred embodiment, the formula for calculating the dynamic fall risk index is as follows: in, For at a certain point in time Dynamic fall risk index; For at a certain point in time The A normalized gait feature parameter; For the first Gait feature parameters at time point Dynamic weights; The total number of feature parameters included in the multidimensional gait parameters; This is the index of the feature parameter.

[0013] As a preferred embodiment, the gait feature parameters At least include the gait symmetry index calculated by the following formula. and gait variability index : in, and At different time points, for the healthy side and the affected side respectively. Step size; This refers to the number of gait cycles within the time window used to calculate variability; For the first The gait feature parameter in the past... The value of each period; For the first Gait feature parameters in the time window The average value within the range.

[0014] As a preferred embodiment, the multidimensional gait parameters include: stride length, cadence, gait speed, stance phase time, swing phase time, hip joint angle, knee joint angle, and plantar pressure distribution on both healthy and affected sides.

[0015] As a preferred option, the dynamic weight The weights are calculated by an attention-based weight generation module. This module takes the current and historical gait feature parameter sequences as input and calculates the relationship between the query vector, key vector, and value vector to generate the weights for each gait feature parameter at the current time point. Assign attention weights.

[0016] As a preferred option, the dynamic weight The specific calculation process includes: processing the multidimensional gait parameter sequence The query matrix is ​​obtained through linear transformation. Key matrix Sum matrix The attention score is calculated using the following formula, and then scaled and normalized to obtain the attention weight matrix: The dynamic weight The attention weight matrix is ​​related to the current time point and the The elements corresponding to each feature parameter; where... Represents matrix multiplication. Scaling factor is the dimension of the key vector.

[0017] As a preferred embodiment, the weight generation module is a pre-trained model based on the Transformer encoder structure. The pre-trained model is trained using a historical dataset of gait parameters containing normal walking, near-fall, and fall events to learn the mapping relationship from gait parameter sequences to the correlation of fall risk.

[0018] As a preferred embodiment, the adaptive controller generates the torque control command according to the following formula: in, For at a certain point in time The generated auxiliary torque value; This is the preset maximum auxiliary torque; This is the gain coefficient, used to adjust the steepness of the torque change; This is the preset fall risk response threshold.

[0019] As a preferred embodiment, the biofeedback device is a visual feedback device or an auditory feedback device; when it is a visual feedback device, the feedback control command is used to control the color or size of a graphical interface on the screen, the color or size being related to the dynamic fall risk index. The value is positively correlated; when it is an auditory feedback device, the feedback control command is used to control the frequency or volume of the emitted prompt sound, and the frequency or volume is positively correlated with the dynamic fall risk index. The values ​​are positively correlated.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. This invention achieves dynamic and proactive fall risk assessment. By constructing an assessment model that integrates multi-dimensional parameters such as gait symmetry and variability, and innovatively introducing a Transformer-based attention mechanism to dynamically allocate weights, it effectively captures subtle changes and complex dependencies in gait parameters over time. Compared to existing technologies that respond solely to movement intentions or employ simple feature extraction, this invention can identify gait instability patterns that foreshadow falls earlier and more accurately, achieving a shift from "passive response" to "active prediction" and significantly improving safety.

[0022] 2. Personalized adaptive assistive control is achieved. This invention maps a quantified dynamic fall risk index to a smoothly varying assistive torque through a non-linear Sigmoid function. This means that the magnitude of the assistive force is precisely matched to the patient's actual fall risk level: minimal intervention is provided when the risk is low, encouraging the patient to actively exert force; when the risk is high, the assistive torque is increased on demand and steplessly, providing just the right amount of stable support. This overcomes the drawbacks of fixed assistive modes or "one-size-fits-all" control based on simple threshold switching in existing technologies, achieving truly personalized and refined assistive control.

[0023] 3. A closed-loop biofeedback mechanism was established, enhancing patient participation. This invention not only provides physical assistance but also transforms abstract fall risks into intuitive signals perceptible to patients through biofeedback channels such as vision or hearing. This allows patients to understand their gait stability in real time and proactively adjust their walking strategies to reduce risk, thus transforming them from passive recipients of training into active participants. This helps promote neural remodeling and motor learning, fundamentally improving rehabilitation efficiency and effectiveness. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating the overall method of an adaptive fall prevention training method for stroke patients based on multidimensional gait parameters, provided in an embodiment of the present invention.

[0026] Figure 2 This is a detailed flowchart of the dynamic fall risk assessment steps in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 The present invention provides an adaptive fall prevention training method for stroke patients based on multidimensional gait parameters, the overall process of which includes the following steps:

[0029] Step S100: Real-time gait parameter acquisition. In this step, the patient wears a lower limb rehabilitation robot for walking training. The rehabilitation robot integrates multiple sensors to collect multi-dimensional gait parameters of the patient in real time and synchronously during movement. Specifically, the sensors may include, but are not limited to: inertial measurement units (IMUs) configured on the linkages of the healthy and patient limbs to measure the posture and angular velocity of the limb segments; angle encoders configured at the hip and knee joints to accurately measure joint angles; and a flexible pressure sensor array or pressure insole configured on the sole of the foot to measure the plantar pressure distribution.

[0030] The sensors described above can acquire multidimensional gait parameters in real time, including at least: spatiotemporal parameters such as stride length, cadence, gait speed, single-leg stance phase time, double-leg stance phase time, and swing phase time for both healthy and affected sides; kinematic parameters such as real-time angular change curves of the hip, knee, and ankle joints in the sagittal plane; and dynamic parameters such as the trajectory and size of the center of pressure (CoP). This raw data is acquired at a high sampling rate (e.g., 100Hz) via a data acquisition card and transmitted to the central processing unit, providing a high-fidelity data foundation for subsequent risk assessment.

[0031] Step S200: Dynamic fall risk assessment step. This step is the core of the method of this invention. Its purpose is to transform the high-dimensional, time-series gait parameter stream collected in step S100 into a single, intuitive dynamic fall risk index in real time. The detailed procedure for this step is as follows: Figure 2 As shown.

[0032] First, the acquired raw multidimensional gait parameters are preprocessed and feature extracted. Preprocessing includes filtering the signal using, for example, a Butterworth low-pass filter to remove noise, and normalizing parameters of different dimensions to ensure they fall within the [0, 1] interval, thus forming normalized gait feature parameters. .

[0033] Next, key gait biomechanical parameters are calculated from the preprocessed data. For example, the gait symmetry index is calculated. This index is used to quantify the symmetry of leg movements, and its calculation formula is as follows: in, and These represent the stride lengths of the healthy and affected sides within the current gait cycle. A value closer to 0 indicates better symmetry, while a value closer to 1 indicates poorer symmetry. Increased asymmetry is a significant precursor to falls.

[0034] Simultaneously, calculate the gait variability index. This index is used to measure gait stability, and its calculation formula is as follows: in, This is the number of gait cycles within the time window used to calculate variability (e.g., the most recent 10 gait cycles). For the first The gait feature parameter in the past... The value of each period, For this parameter in the time window The average value within, This represents the total number of characteristic parameters. A higher variability index indicates poor gait consistency and decreased stability.

[0035] Then, the calculated gait symmetry index, variability index, and other normalized gait characteristic parameters are used... (Such as the proportion of support phase time, joint angle range, etc.) constitute a feature vector sequence. This sequence is input into an attention-based weight generation module to calculate the value of each feature parameter at the current time point. Dynamic weights .

[0036] In one specific embodiment, the weight generation module is a pre-trained model based on a Transformer encoder structure. This model is trained offline using a large historical gait database, which contains a large number of multidimensional gait parameter sequences from stroke patients under different states, including normal walking, experiencing disturbances, near-fall, and actual falls. In this way, the model can learn the dynamic changes in the role of different gait features at different stages predicting a fall.

[0037] During real-time evaluation, the weight generation module receives a sequence of gait feature parameters within a time window. As input, this sequence is passed through a linear transformation layer to generate query matrices. Key matrix Sum matrix Then, the attention score is calculated using a scaled dot product attention mechanism: in, The similarity between the features at each time point in the sequence and the features at all other time points was calculated. It is a scaling factor to prevent the gradient from becoming too small. This is the dimension of the key vector. The softmax function normalizes the scores into weights. The output of this mechanism includes each feature. The importance of dynamic weighting in assessing overall risk at the present moment. For example, in the early stages of gait instability, the model may assign higher weights to gait variability; while in the moments before a fall, the model may focus more on the sharp shortening of the support phase time.

[0038] Finally, the final dynamic fall risk index is calculated using a weighted summation formula. : The index is a continuous value between 0 and 1, with a higher value indicating a greater likelihood of the patient falling at that moment.

[0039] Step S300: Adaptive control command generation step. This step is based on the dynamic fall risk index calculated in step S200. This generates two types of control commands.

[0040] The first type is torque control command. This is used to adjust the assist torque provided by the rehabilitation robot to the patient's affected limb. The instruction is generated by an adaptive controller, the core of which is a sigmoid function: in, This is the preset maximum allowable auxiliary torque to ensure safety; It is a preset risk response threshold (e.g., 0.5). When the risk index exceeds this threshold, the auxiliary torque begins to increase significantly. The gain coefficient determines the steepness of the torque's change with the risk index. This formula ensures that the output of the auxiliary torque is smooth and proportional to the risk level, achieving refined adaptive assistance.

[0041] The second type is feedback control commands, used to drive biofeedback devices. For example, when the biofeedback device is a visual feedback device (such as a screen), the feedback control commands can control the color and length of a progress bar on the screen, with the color gradually changing from green (low risk) to yellow and then to red (high risk), and the length varying accordingly. The value is positively correlated. When used as an auditory feedback device, the feedback control command can control the frequency of a buzzer's alert tone; the higher the risk, the faster the frequency. This intuitive feedback allows patients to perceive their stability status in real time.

[0042] Step S400: Closed-loop training execution step. In this step, the control system of the rehabilitation robot and the biofeedback device simultaneously execute the instructions generated in step S300.

[0043] The rehabilitation robot's underlying controller receives torque control commands. Then, the motors installed on the affected hip and knee joints are driven, with a real-time output size of... The robot provides auxiliary torque. When an increased risk of fall is detected, the robot will actively provide support to help the patient stabilize their body and prevent a fall.

[0044] Simultaneously, the biofeedback device presents visual or auditory feedback to the patient based on feedback control commands. After receiving risk alerts, patients can consciously and actively adjust their gait and posture, such as focusing more on maintaining balance and adjusting stride length, in an attempt to reduce the risk index.

[0045] This process forms an efficient dual closed loop: one is a human-computer physical interaction closed loop consisting of "patient gait - risk assessment - robot assistance - patient gait", which ensures the physical safety of training; the other is a human-computer information interaction closed loop consisting of "patient gait - risk assessment - biofeedback - patient active adjustment - patient gait", which promotes the patient's active learning and neural function remodeling.

[0046] Here's a specific application scenario: A stroke patient is undergoing walking training on a rehabilitation robot. In the initial stage, their gait is relatively stable. Maintaining a low level of around 0.2, the robot provides only a small following torque, and the indicator bar on the screen is green. Midway through training, due to fatigue, the swing phase time of the patient's affected leg begins to shorten irregularly, and the gait variability index increases. The attention mechanism in the dynamic fall risk assessment model captures this key change, assigning a higher weight to the variability index. ,lead to It quickly climbed to 0.6. The adaptive controller immediately calculated a moderate auxiliary torque. The force is applied by a knee joint motor to help the patient complete the swinging motion more stably. Simultaneously, the screen indicator bar turns yellow, accompanied by intermittent beeping sounds. After seeing and hearing the feedback, the patient focuses their attention and consciously controls the lifting and swinging of the affected leg. After a few steps, gait variability decreases. As the value dropped to 0.3, the assist torque and audible / visual cues also weakened. In this way, the method of the present invention successfully prevented a potential instability event and guided the patient to complete self-correction.

[0047] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive fall prevention training method for stroke patients based on multidimensional gait parameters, characterized in that, include: The real-time gait parameter acquisition step involves using multiple sensors configured on the rehabilitation robot to collect multi-dimensional gait parameters of the patient during the training process in real time. The dynamic fall risk assessment step, based on the multidimensional gait parameters, uses a preset dynamic fall risk assessment model to calculate in real time a dynamic fall risk index that represents the patient's current likelihood of falling. The core of the dynamic fall risk assessment model is based on a weighted summation formula that integrates gait symmetry, gait variability and temporal attention mechanism; The adaptive control command generation step involves inputting the dynamic fall risk index into an adaptive controller to generate torque control commands for adjusting the assist torque of the rehabilitation robot and feedback control commands for providing biofeedback to the patient. The closed-loop training execution step involves the rehabilitation robot adjusting the assist torque on the patient's affected limb according to the torque control commands, while the biofeedback device provides sensory feedback to the patient according to the feedback control commands, thus achieving adaptive closed-loop control for fall prevention training.

2. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The formula for calculating the dynamic fall risk index is as follows: in, For at a certain point in time Dynamic fall risk index; For at a certain point in time The A normalized gait feature parameter; For the first Gait feature parameters at time point Dynamic weights; The total number of feature parameters included in the multidimensional gait parameters; This is the index of the feature parameter.

3. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The gait feature parameters At least include the gait symmetry index calculated by the following formula. and gait variability index : in, and At different time points, for the healthy side and the affected side respectively. Step size; This refers to the number of gait cycles within the time window used to calculate variability; For the first The gait feature parameter in the past... The value of each period; For the first Gait feature parameters in the time window The average value within the range.

4. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The dynamic weight The weights are calculated by an attention-based weight generation module. This module takes the current and historical gait feature parameter sequences as input and calculates the relationship between the query vector, key vector, and value vector to generate the weights for each gait feature parameter at the current time point. Assign attention weights.

5. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The dynamic weight The specific calculation process includes: processing the multidimensional gait parameter sequence The query matrix is ​​obtained through linear transformation. Key matrix Sum matrix The attention score is calculated using the following formula, and then scaled and normalized to obtain the attention weight matrix: The dynamic weight The attention weight matrix is ​​related to the current time point and the The elements corresponding to each feature parameter; where... Represents matrix multiplication. Scaling factor is the dimension of the key vector.

6. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The multidimensional gait parameters include: stride length, cadence, gait speed, stance phase time, swing phase time, hip angle, knee angle, and plantar pressure distribution on both healthy and affected sides.

7. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The adaptive controller generates the torque control command according to the following formula: in, For at a certain point in time The generated auxiliary torque value; This is the preset maximum auxiliary torque; This is the gain coefficient, used to adjust the steepness of the torque change; This is the preset fall risk response threshold.

8. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 1, characterized in that, The biofeedback device is either a visual feedback device or an auditory feedback device; when it is a visual feedback device, the feedback control command is used to control the color or size of a graphical interface on the screen, the color or size being related to the dynamic fall risk index. The value is positively correlated; when it is an auditory feedback device, the feedback control command is used to control the frequency or volume of the emitted prompt sound, and the frequency or volume is positively correlated with the dynamic fall risk index. The values ​​are positively correlated.

9. The adaptive fall prevention training method for stroke patients based on multidimensional gait parameters according to claim 4, characterized in that, The weight generation module is a pre-trained model based on the Transformer encoder structure. The pre-trained model is trained using a historical dataset of gait parameters containing normal walking, near-fall, and fall events to learn the mapping relationship from gait parameter sequences to the correlation of fall risk.

Images