An orthopedic patient fall risk monitoring method based on posture recognition

Abstract

The application provides a kind of orthopedic patient fall risk monitoring method based on gesture recognition, comprising: collecting the reference video image and monitoring video image when target patient wears orthopedic brace, identifying joint key points, limb profile, brace components and connection position, form patient-brace coupling observation sequence;Establish brace constraint topology graph, calculate the constraint reliability, constraint degradation speed and constraint importance of constraint edge, determine the key constraint edge;Extract residual micro-motion around target joint, construct joint maintenance micro-strain field, calculate muscle compensation load and muscle compensation synergy;Establish individualized support contribution transfer model using reinforcement learning, predict brace constraint contribution, available muscle compensation capacity and support margin, determine double-layer support failure time and remaining maintenance time;According to risk level, generate risk disposal instruction, and update individualized monitoring parameters using continuous video image after disposal.

Description

Technical Field

[0001] This invention relates to the field of human posture recognition and rehabilitation safety monitoring technology, and in particular to a method for monitoring the risk of falls in orthopedic patients based on posture recognition. Background Technology

[0002] Orthopedic patients typically require the use of orthopedic braces such as knee and ankle braces for support during joint injury repair and lower limb rehabilitation. Because their weight-bearing capacity, muscle control, and proprioception have not fully recovered, they are prone to falls during standing, starting, walking, turning, or stopping due to joint instability, brace loosening, or decreased compensatory ability, leading to secondary injuries and hindering the rehabilitation process.

[0003] Current fall risk monitoring methods primarily rely on human motion parameters such as trunk tilt, joint angles, gait speed, plantar pressure, or acceleration to identify abnormalities. They typically only trigger an alarm when the patient exhibits significant postural deviation, rapid body fall, or when motion parameters exceed limits. These methods mainly focus on macroscopic posture and fail to simultaneously analyze the brace restraint status, including locking component misalignment, hinge abnormalities, strap slippage, and relative displacement of the brace housing. In reality, brace restraint capacity may gradually decrease, while patients can temporarily maintain a normal posture by increasing the compensatory action of muscles around the target joints, making it difficult to identify potential instability risks in a timely manner. Furthermore, existing technologies lack analysis of the correlation between brace constraint degradation and changes in muscle compensation, making it difficult to distinguish between normal movements, brace vibrations, and joint maintenance micro-movements, and also difficult to predict the remaining posture maintenance time when muscle compensation capacity approaches its limit. Therefore, how to simultaneously identify patient posture and brace constraint status, assess changes in muscle compensation, predict the remaining maintenance time before double-layer support failure, and achieve individualized closed-loop early warning are problems that need to be solved. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for monitoring the risk of falls in orthopedic patients based on posture recognition, thereby solving the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for monitoring fall risk in orthopedic patients based on posture recognition includes the following steps: S1. Collect baseline and monitoring video images of the target patient wearing orthopedic braces, perform calibration, correction and time alignment, identify key points of human joints, limb contours, brace components and connection positions, and determine the effective monitoring period through behavioral feature recognition to form a patient-brace coupled observation sequence. S2. Based on the patient and brace coupling observation sequence, establish a brace constraint topology graph, calculate the constraint reliability, constraint degradation rate and constraint importance of each brace constraint edge, determine the key constraint edges, and form a brace constraint state sequence. S3. Determine the micro-motion extraction area around the target joint based on the key constraint edges, deduct camera shake, overall limb movement, baseline behavioral movement and respiratory interference, construct the joint maintenance micro-strain field, calculate muscle compensation load and muscle compensation synergy, and form a muscle compensation state sequence. S4. Align the brace constraint state sequence and muscle compensation state sequence, construct a double-layer support state vector, use reinforcement learning to establish an individualized support contribution transfer model, predict the brace constraint contribution, available muscle compensation capacity, support margin and target joint posture, determine the double-layer support failure time and remaining maintenance time, and form a fall risk evolution sequence. S5. Determine the risk level based on the fall risk evolution sequence, remaining maintenance time, and safety intervention time, generate risk handling instructions, verify the early warning results using continuous video images after handling, and update individualized monitoring parameters.

[0006] S1 specifically includes: acquiring baseline and monitoring video images of the target patient wearing orthopedic braces; performing distortion correction, coordinate transformation, time alignment, camera shake compensation, and brightness normalization on the video images to form a standardized video image sequence; performing joint image recognition of the human body and orthopedic braces on the standardized video image sequence, extracting key joint points, limb contours, clothing texture blocks, brace components, and connection positions, establishing an identity association between the orthopedic brace and the target patient, and forming a multi-target feature sequence of the patient and brace; and performing behavioral feature recognition based on the multi-target feature sequence of the patient and brace, identifying the stages of static standing, starting, continuous walking, turning, and stopping, screening effective monitoring periods where the brace is visible and the target patient is in a state of autonomous weight-bearing, and forming a patient-brace coupled observation sequence containing data source markers.

[0007] S2 specifically includes: establishing human body nodes, brace nodes, human body connection edges, brace structural edges, and brace constraint edges based on the patient and brace coupling observation sequence; recording the type of each brace constraint edge, the corresponding target joint, and the initial geometric relationship to form a brace constraint topology map; extracting the offset of locking components, relative rotation angle of hinges, slippage of straps, relative displacement of brace shell, and abnormal vibration parameters for locking constraint edges, hinge guiding constraint edges, strap fixing constraint edges, and shell following constraint edges, respectively; calculating constraint reliability based on individual reference parameters and allowable deviations; calculating constraint degradation rate based on constraint reliability; determining constraint importance based on the change in the allowable range of motion of the target joint before and after the simulation release of brace constraint edges; screening key constraint edges; and combining constraint reliability, constraint degradation rate, constraint importance, and key constraint edge markings in chronological order to form a brace constraint state sequence.

[0008] S3 specifically includes: determining the target joint and its micro-motion extraction area based on the patient, brace coupling observation sequence and brace constraint state sequence; subtracting camera shake, overall limb movement, baseline behavioral movement and respiratory interference to form a standard micro-motion image sequence; performing subpixel displacement estimation on the standard micro-motion image sequence, calculating the micro-strain components of the residual displacement, extracting the micro-strain amplitude, dominant frequency, phase and duration, screening the joint maintenance micro-motion corresponding to the degradation of key constraint edges, and constructing a joint maintenance micro-strain field; comparing the joint maintenance micro-strain field with the baseline micro-strain field of the same behavioral stage and gait phase, calculating muscle compensation load and muscle compensation synergy, and associating the target joint, key constraint edges and acquisition time to form a muscle compensation state sequence.

[0009] S4 specifically includes: aligning the brace constraint state sequence and the muscle compensation state sequence; constructing a two-layer support state vector containing constraint reliability, constraint degradation rate, constraint importance, muscle compensation load, muscle compensation synergy, and target joint posture; establishing an individualized support contribution transfer model using reinforcement learning; calculating the brace constraint contribution based on the constraint reliability and constraint importance of key constraint edges; determining the brace constraint gap, available muscle compensation capacity, and support margin; updating the predicted values ​​according to the prediction step size; and forming a posture evolution sequence through the posture prediction module; comparing the predicted brace constraint contribution, predicted support margin, predicted muscle compensation state, and predicted target joint posture with the corresponding thresholds to determine the two-layer support failure time and remaining maintenance time, thus forming a fall risk evolution sequence.

[0010] S5 specifically includes: reading the fall risk evolution sequence; determining the combined risk based on the constraint reliability, muscle compensation load, muscle compensation synergy, and predicted support margin corresponding to the same target joint; determining the risk level by combining the remaining maintenance time and safe intervention time; generating risk treatment instructions based on the risk level; sending the risk treatment instructions to the patient alert device, nursing terminal, or safety protection device; and recording the instruction confirmation, device execution, and nursing response results; verifying real early warnings, effective interventions, false alarm candidates, missed alarms, or invalid data events based on continuous video images after treatment; writing effective events into the individualized experience sample library; updating the individualized support contribution transfer model and individualized monitoring parameters; and forming an individualized monitoring parameter set for use in the next monitoring cycle.

[0011] The beneficial effects of this invention are as follows: This invention achieves simultaneous monitoring of patient posture and orthopedic brace status by jointly recognizing key points of human joints, limb contours, brace components, and connection positions, overcoming the limitations of judging fall risk solely based on macroscopic human posture. By establishing a brace constraint topology and calculating constraint reliability, constraint degradation rate, and constraint importance, it can promptly identify locking component misalignment, hinge abnormalities, strap slippage, and relative displacement of the brace shell, improving the accuracy of brace constraint degradation detection.

[0012] This invention constructs a joint-maintaining micro-strain field by subtracting camera shake, overall limb movement, normal behavioral movements, and respiratory interference. This field can distinguish between normal movements, independent vibrations of the brace, and joint-maintaining micro-movements, improving the reliability of muscle compensation state identification. By calculating muscle compensation load and muscle compensation synergy, it can detect increased compensation load and decreased synergy even before the patient's macroscopic posture shows obvious abnormalities, thus identifying instability risks hidden beneath a normal appearance.

[0013] This invention establishes a transfer relationship between brace constraint contribution and muscle compensation contribution through reinforcement learning, predicting support margin, failure time of double-layer support, and remaining maintenance time, thus allowing nurses more time for safe intervention. By generating risk management instructions based on risk levels and using continuous video images after management to verify early warning results and update individualized monitoring parameters, the monitoring thresholds and prediction models are adapted to differences in patient recovery stage, brace type, and nursing response, reducing false alarms and missed alarms. Attached Figure Description

[0014] Figure 1 This is a flowchart of the orthopedic patient fall risk monitoring method based on posture recognition according to the present invention. Detailed Implementation

[0015] 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.

[0016] Example: Figure 1 As shown in the figure, this embodiment provides a method for monitoring the fall risk of orthopedic patients based on posture recognition, including the following steps: S1. Collect baseline and monitoring video images of the target patient wearing orthopedic braces, perform calibration, correction and time alignment, identify key points of human joints, limb contours, brace components and connection positions, and determine the effective monitoring period through behavioral feature recognition to form a patient-brace coupled observation sequence. S2. Based on the patient-brace coupling observation sequence, establish a brace constraint topology graph, calculate the constraint reliability, constraint degradation rate and constraint importance of each brace constraint edge, determine the key constraint edges, and form a brace constraint state sequence. S3. Determine the micro-motion extraction area around the target joint based on the key constraint edges, deduct camera shake, overall limb movement, baseline behavioral movement and respiratory interference, construct the joint maintenance micro-strain field, calculate muscle compensation load and muscle compensation synergy, and form a muscle compensation state sequence. S4. Align the brace constraint state sequence and muscle compensation state sequence, construct a double-layer support state vector, use reinforcement learning to establish an individualized support contribution transfer model, predict the brace constraint contribution, available muscle compensation capacity, support margin and target joint posture, determine the double-layer support failure time and remaining maintenance time, and form a fall risk evolution sequence. S5. Determine the risk level based on the fall risk evolution sequence, remaining maintenance time, and safety intervention time, generate risk handling instructions, verify the early warning results using continuous video images after handling, and update individualized monitoring parameters.

[0017] S1 specifically includes the following sub-steps: S110. Collect and standardize baseline and monitoring video images of the target patients. Identify individuals wearing orthopedic braces who require fall risk monitoring as target patients, and define the hinges, locking components, straps, and brace housing used to restrict the movement of the target joints in the orthopedic brace as brace components.

[0018] Continuous video images are acquired using fixed cameras in the ward or rehabilitation training area. One camera is used when a single viewpoint can cover the entire body of the target patient and the orthopedic brace. If the target joint or brace component is easily obscured, a second camera is added on the side where the brace is located. The sampling frequency of the camera is set to 25-60fps, and the image resolution is not less than 1280×720px, ensuring that the height of the standing patient's body occupies 40%-90% of the image height.

[0019] Before data collection, medical staff confirm that the locking components are closed, the straps are fixed, the brace shell fits the limb, and that the target patient can safely complete standing still, starting, walking continuously, turning, and stopping. Under the condition that the nursing staff protects but does not directly support the target patient, at least 3 complete action cycles are collected for each behavior to form a baseline video image. The videos collected during the monitoring period form the monitoring video image.

[0020] At least four non-collinear ground calibration points are set up in the effective monitoring area. The actual coordinates of the calibration points are obtained from on-site measurements. The hinge diameter, the length of the support shell, or the distance between adjacent strap fixing points are obtained from the product parameters of the support or the results of manual measurement. One of these dimensions is selected as the dimensional reference of the support.

[0021] The following steps are performed sequentially: lens distortion correction, ground coordinate transformation, multi-view time alignment, camera shake compensation based on fixed background features, and background brightness normalization. The conversion from image coordinates to actual ground coordinates is as follows:

[0022] Where u and v are the horizontal and vertical pixel coordinates, respectively, in pixels (px); X and Y are the actual horizontal and vertical coordinates in the ground reference plane, respectively, in millimeters (mm); and H is the homography matrix obtained from the ground calibration points. This is the homogeneous proportionality coefficient.

[0023] A standardized video image sequence containing image frame number, acquisition time, camera device number, corrected image, and ground transformation parameters is obtained for use in step S120. For example, two camera devices with a sampling frequency of 50fps can be used, and the measured distance of 120mm between the two strap fixation points of the knee brace can be used as the brace dimensional reference.

[0024] S120. Perform joint image recognition of the human body and orthopedic braces on standardized video image sequences. Using benchmark video images and historical videos of the same model of brace, annotate the human body contour, key joint points, brace contour, and brace components to form a sample set and train a joint image recognition model. Preferably, the joint image recognition model adopts a multi-task deep learning model based on network architecture improvements such as HRNet or YOLOv8-Pose to achieve synchronous output of the spatial positions of human body key points, limb contours, and brace components.

[0025] The human body region and orthopedic brace region are segmented frame by frame. From the human body region, key joint points, limb contours, and clothing texture blocks around the target joints of the head, shoulder, hip, knee, ankle, and foot are extracted. From the orthopedic brace region, hinges, locking components, straps, brace shells, and connection points are extracted. Connection points refer to the image areas where the brace components form a fixed, restricted, or follow-along relationship with the target patient's limb, including the strap connection area, the brace shell connection area, and the joint corresponding area between the brace hinge axis and the rotation center of the target joint.

[0026] Based on the outline of the brace shell as the main positioning basis, the hinges, locking components and straps are searched according to the relative position, outline and texture of each brace component in the reference video image; based on the spatial adjacency relationship between the human body area and the orthopedic brace area, the synchronous movement direction and the corresponding limb position, the identity association between the orthopedic brace and the target patient is established, and fixed target numbers are assigned to the target patient, the target joint and each brace component.

[0027] When a joint key point or brace component is temporarily occluded, it is completed based on the length of adjacent limb segments and the position and direction of movement of the preceding and following frames; if the continuous missing value exceeds 0.4 s, or if the normalized position deviation between the completed target joint and the center of the brace hinge exceeds the joint's corresponding error threshold, the corresponding image frame is marked as an unusable frame.

[0028] The normalized positional bias is:

[0029] in, The joint correspondence deviation at time t is the acquisition time of the current image frame; and These are the pixel positions of the target joint key points and the center of the support hinge, respectively, both in pixels (px). The pixel length of the support scale reference within this frame, in pixels (px). This is a dimensionless value. The joint correspondence error threshold is the sum of the 95th percentile of the joint correspondence deviation in the reference video image and the normalized position deviation converted from the 1px positioning error.

[0030] A patient-brace multi-target feature sequence is formed, including at least the image frame number, acquisition time, target number, joint key point coordinates and recognition confidence, limb contour, clothing texture patch, brace component position and recognition confidence, connection position contour, and data availability marker, for behavioral feature recognition in step S130. For example, if the locking component is obscured by the trouser leg for 4 image frames, it can be completed based on its position relative to the hinge and the displacement of the brace shell; if the obscuration exceeds 0.4s, the corresponding time period is not included in the subsequent analysis.

[0031] S130. Perform behavioral feature recognition and determine the effective monitoring period based on the patient-brace multi-target feature sequence. Using a sliding time window of 0.5-2s, extract trunk direction, pelvic movement speed, target joint angular velocity, left and right foot contact status, and brace follow-up status.

[0032] The following stages are identified as follows: a stationary standing stage, where both feet remain in contact and both pelvic movement speed and target joint angular velocity are below the individual threshold; a starting stage, where one foot transitions from contact to lift off the ground and the pelvis moves continuously along the direction of travel; a continuous walking stage, where the left and right feet alternate in their support states for at least two gait cycles; a turning stage, where the trunk and pelvic directions change continuously and the movement directions of both feet change synchronously; and a stopping stage, where both feet return to contact after continuous walking and the pelvic movement speed drops below the individual threshold. The individual threshold is the sum of the 95th percentile of the corresponding feature in the baseline video image and the image measurement tolerance.

[0033] The brace follow-up state refers to the state in which the brace shell moves in the same direction as the corresponding limb segment, and the normalized positional deviation between the center of the brace hinge and the key point of the target joint does not exceed the corresponding error threshold of the joint; the autonomous weight-bearing state refers to the state in which at least one foot forms support with the ground, and the caregiver does not continuously support the target patient's trunk, pelvis, or axilla; the effective monitoring period refers to the continuous time interval in which the target patient and the orthopedic brace are continuously identified, the key point of the target joint and the brace components are both available, the target patient is in the static standing, starting, continuous walking, turning or stopping stage, the orthopedic brace should restrict the movement of the target joint and the target patient is in the autonomous weight-bearing state.

[0034] When the target patient's hands come into contact with the walking aid, it is not directly considered as a non-voluntary weight-bearing state; the corresponding time period is excluded only when the lower limbs do not provide effective support or the main body weight is borne by the upper limbs.

[0035] From the effective monitoring period, key joint points, target joint angles and angular velocities, limb contours, clothing texture blocks, brace components, connection positions, target behavior stages, foot contact states, brace follow-up states, and data availability markers are extracted and associated according to a unified collection time to form a patient-brace coupled observation sequence.

[0036] The patient-brace coupling observation sequence also includes data source markers to indicate whether the corresponding data originates from a reference video image or a monitoring video image. Based on these markers, reference observation subsequences and monitoring observation subsequences are formed, respectively, for use in step S210 to establish the brace constraint topology map and in step S310 to determine the micro-motion extraction region. For example, if the target patient's right foot leaves the ground, the left foot supports the body, and the pelvis moves forward when starting, and the nursing staff does not directly support the patient while the knee brace components are continuously visible, then this period is considered a valid monitoring period; if the nursing staff continuously supports the armpit, it is excluded.

[0037] S2 specifically includes the following sub-steps: S210. Establish a human body-brace constraint topology diagram. Divide the patient-brace coupling observation sequence output in step S130 into a baseline observation subsequence and a monitoring observation subsequence according to the data source label. The baseline observation subsequence comes from the baseline video images collected in step S110 and confirmed as safe by medical staff, while the monitoring observation subsequence comes from the actual monitoring video images.

[0038] Based on the fixed target number, the key points of the target joint and the center of the adjacent limb segment are set as human body nodes, and the center of the brace hinge, the center of the locking component, the center of the strap fixing area and the center of the brace shell are set as brace nodes; the physiological connection relationship between human body nodes is set as human body connection edge, the connection relationship between brace components that is only used to maintain the brace structure is set as brace structure edge, and the functional relationship that can restrict the relative movement between brace components or between orthopedic brace and human limb segment is set as brace constraint edge.

[0039] The brace constraint edges include the locking constraint edge between the locking component and the locking position, the hinge guiding constraint edge between the hinge and the upper and lower brace housings, the strap fixing constraint edge between the strap fixing area and the corresponding human limb segment, and the housing follow-up constraint edge between the brace housing and the corresponding human limb segment.

[0040] Based on the position of each node, the contour of the connection position, and the follow-up state of the brace in the benchmark observation subsequence, the starting node, ending node, constraint type, corresponding target joint, initial geometric relationship, and data availability marker of each brace constraint edge are recorded to form a brace constraint topology map. The initial geometric relationship includes at least the initial distance between the center of the locking component and the center of the locking position, the initial angle between the longitudinal centerlines of the upper and lower brace shells, the initial distance of the strap edge relative to the limb texture marker, and the initial positional relationship of the brace shell relative to the corresponding limb segment.

[0041] For example, for a knee brace with locking function, the center of the knee, the center of the thigh segment, and the center of the lower leg segment are set as human body nodes, and the hinge, locking component, upper and lower straps, and upper and lower brace housings are set as brace nodes. Locking constraint edge, hinge guiding constraint edge, strap fixing constraint edge, and housing follow constraint edge are established respectively.

[0042] The constraint topology diagram is used in step S220 to calculate the visual state parameters of each constraint edge, and in step S230 to evaluate the constraint importance of each constraint edge to the stability of the target joint.

[0043] S220. Calculate the visual state parameters and constraint reliability of each support constraint edge. For the locking constraint edge, calculate the difference between the current distance between the center of the locking component and the initial distance relative to the center of the locking position to obtain the offset of the locking component; for the hinge guide constraint edge, calculate the difference between the current angle formed by the longitudinal center lines of the upper and lower support housings at the hinge center and the initial angle to obtain the relative rotation angle of the hinge.

[0044] For the fixed constraint edge of the strap, select continuously identifiable clothing texture points or skin marker points in the vicinity of the strap edge in the reference observation subsequence, calculate the difference between the current strap edge relative to the marker point along the longitudinal direction of the human limb segment and the initial distance, and obtain the strap slippage amount; for the follow constraint edge of the shell, subtract the overall translation and rotation of the corresponding human limb segment from the movement of the support shell to obtain the relative displacement of the support shell.

[0045] Then, the residual shaking of the camera device, the overall movement of the target patient, and the movement of the corresponding limbs are deducted from the trajectory of the brace components. The main frequency and vibration direction of the remaining high-frequency displacement components within the 0.5-2s sliding time window are calculated using Fast Fourier Transform (FFT) to obtain the main frequency and abnormal vibration direction of the abnormal vibration.

[0046] The median of each visual state parameter under safe load-bearing conditions in the baseline observation subsequence was determined as the individual baseline parameter. The permissible deviation of each parameter was preferentially determined using the technical parameters provided by the orthopedic brace manufacturer; if the manufacturer did not provide these parameters, calibration test data of the same model brace on an artificial limb model was used; if these data were still unavailable, the sum of the 95th percentile of the absolute deviation of the corresponding parameter in the baseline observation subsequence and the image recognition measurement error was used. The image recognition measurement error was obtained by continuously acquiring no fewer than 100 image frames from a stationary brace and calculating the standard deviation of the position recognition results for the same brace component.

[0047] The constraint reliability of the i-th support edge at time t is:

[0048] in, Let be the constraint reliability of the i-th support constraint edge, with a value ranging from 0 to 1; The number of visual state parameters used for the constraint edge of the support; This represents the current value of the m-th visual state parameter; For the corresponding individual baseline parameters; This corresponds to the allowable deviation; These are the parameter weights, and the sum of the weights of all parameters is 1.

[0049] Parameter weights are determined based on the sensitivity of each parameter to constraint failure identification in calibration tests of the same type of brace. When calibration data is unavailable, equal weights are used. For example, in a preferred embodiment, for a knee brace with locking function, the parameter weights are set as follows: locking component offset weight 0.4, hinge relative rotation angle weight 0.3, abnormal vibration direction weight 0.2, and abnormal vibration dominant frequency weight 0.1. Constraint reliability is a brace constraint state index estimated from video images and does not represent the actual mechanical constraint force. When a parameter becomes unavailable due to occlusion, the weights of the remaining valid parameters are renormalized; if no valid parameter exists for more than 0.4 seconds, the corresponding data availability flag is set to unavailable.

[0050] This generates support constraint edge state data that includes constraint reliability, abnormal vibration dominant frequency, abnormal vibration direction, and data availability markers, which is then used in step S230.

[0051] S230. Determine the constraint degradation rate, constraint importance, and critical constraint edges. Calculate the constraint degradation rate for each support constraint edge using a 0.5-2s sliding time window:

[0052] in, Let be the constraint degradation rate of the i-th support constraint edge at time t; The constraint reliability at the start of the sliding time window; The constraint reliability at the current moment; This represents the length of the sliding time window, measured in seconds (s). This indicates a decrease in the reliability of the constraint.

[0053] If the constraint degradation rate is higher than the degradation rate threshold for at least three consecutive sliding time windows, and the current constraint reliability is lower than the constraint reliability at the start of the first sliding time window, the corresponding support constraint edge is determined to be in a state of continuous degradation. The degradation rate threshold is the sum of the 95th percentile of the absolute value of the constraint reliability fluctuation rate in the benchmark observation subsequence and the measurement error tolerance.

[0054] Obtain the allowable range of motion of the target joint when the constraint edges of each brace remain effective and when the simulation is released from the structural parameters of the brace manufacturer, calibration tests of the artificial limb model of the same model, or a pre-established brace type configuration library.

[0055] The constraint importance of a brace constraint edge is determined by comparing the increase in the allowable range of motion of the target joint after simulating the removal of a brace constraint edge with the safe range of motion of the target joint as determined by the doctor's orders or rehabilitation plan at the current rehabilitation stage of the target patient. Available brace constraint edges whose constraint importance reaches the importance threshold, and whose constraint reliability is below the constraint reliability threshold or is in a state of continuous degradation, are identified as critical constraint edges.

[0056] The importance threshold is taken as the 75th percentile of the constraint importance of each constraint edge in the configuration library of the same type of brace; the constraint reliability threshold is taken as the boundary value between the calibration samples of the artificial limb model in normal and loose states. For example, if the knee joint's allowable range of motion is 5° and 25° when the locking constraint edge is effective and when it is simulated to be released, respectively, and the patient's current safe range of motion is 30°, then the constraint importance is 0.67; if its constraint reliability decreases from 0.72 to 0.54 within 1.5 seconds, and the constraint degradation rate exceeds the degradation rate threshold, then it is identified as a critical constraint edge.

[0057] The data acquisition time, target patient number, target joint number, brace constraint edge number and type, constraint reliability, constraint degradation rate, constraint importance, key constraint edge marker, abnormal vibration dominant frequency, abnormal vibration direction, target behavior stage corresponding to step S130, and data availability marker are combined in chronological order to form a brace constraint state sequence. This sequence is used by step S310 to determine the micro-motion extraction area and by step S410 to perform time alignment with the muscle compensation state sequence.

[0058] S3 specifically includes the following sub-steps: S310. Determine the micro-motion extraction region corresponding to the target joint and form a standard micro-motion image sequence. According to the target patient number, target joint number, key constraint edge number, and acquisition time, align the patient-brace coupling observation sequence output in step S130 with the brace constraint state sequence output in step S230 in time to obtain the target joint, target behavior stage, joint key points, limb contour, clothing texture patch, brace edge, constraint reliability, and constraint degradation rate corresponding to the key constraint edge.

[0059] The peri-joint muscle region refers to the area delineated on the body surface image based on the key points of the target joint and the contours of adjacent limb segments. It reflects minor deformations of the soft tissue or covering clothing around the target joint and does not represent direct observation of subcutaneous muscles. When the target joint is the knee joint, the peri-joint muscle region is defined as the area at the distal end of the thigh and the proximal end of the lower leg, each accounting for 20%-35% of the corresponding limb segment length. A clothing texture region is included within this region, a brace edge region is defined at the edge of the brace shell, and a reference stability region is defined for the corresponding limb segment that is far from the target joint and has stable movement.

[0060] For the static standing and stationary phases, image frames with target joint angular velocities below the joint static threshold and joint angle fluctuations less than the attitude stability threshold are selected. The joint static threshold and attitude stability threshold are respectively the sum of the 95th percentile of the absolute value of the joint angular velocity and the joint angle fluctuation corresponding to the same behavioral phase in the benchmark observation subsequence and the image measurement error tolerance.

[0061] For the starting, continuous walking and turning phases, the gait phase is determined based on the time of foot lift-off, foot contact with the ground and pelvic movement, and reference image frames with the same behavioral phase and the same gait phase are selected from the reference observation subsequence.

[0062] Residual camera shake is subtracted using fixed background features, overall limb translation and rotation are subtracted using a reference stable region, macroscopic motion corresponding to normal behavior is subtracted using a reference image frame, respiratory components are eliminated using periodic motion of the chest and abdomen region, and illumination fluctuations are corrected based on changes in background brightness. Image frames with motion blur, overexposure, occlusion exceeding 40% of the area, or fewer than 20 effective texture points are excluded to form a standard micro-motion image sequence.

[0063] A standard micro-motion image sequence is a video image sequence that retains residual motion in the area surrounding the target joint after compensation for camera shake, overall limb movement, baseline behavioral movement, respiratory interference, and lighting changes. It includes at least the acquisition time, target joint number, key constraint edge number, target behavioral stage, gait phase, micro-motion extraction area, macro-motion compensation parameters, and data availability markers, for use in step S320. For example, if the patient is in the right lower limb support phase, and the current knee joint angle is 18°, while the baseline phase angle is 17°, then the normal knee flexion movement in the same phase is first subtracted before analyzing the remaining micro-motion.

[0064] S320. Construct a joint-maintaining micro-strain field from a standard micro-motion image sequence. Calculate the sub-pixel displacements of clothing texture points, body surface contour points, and brace edge points in adjacent image frames using the dense optical flow method. Subtract the camera device jitter, overall limb motion, reference macroscopic motion, breathing component, and illumination interference obtained in step S310 from the original displacements to obtain the residual displacements.

[0065] Establish a local coordinate system for the limb segment with the target joint as the origin. The micro-strain components in the image plane are:

[0066] in, This represents the component of the residual displacement in the transverse direction of the local coordinate system of the limb segment; y represents the component of the residual displacement in the longitudinal direction of the local coordinate system of the limb; lowercase letters x and y represent the transverse and longitudinal coordinates in the local coordinate system of the limb, respectively. These are transverse micro-strain components; For longitudinal micro-strain components; This represents the shear micro-strain component.

[0067] Microstrain represents the relative deformation between adjacent image points, not the actual mechanical strain of human tissue. Short-time Fourier transform (STFT) or continuous wavelet transform (CWT) is used to perform time-frequency analysis on each microstrain component in the range of 0.5-8 Hz to obtain the microstrain amplitude, dominant frequency, phase, and duration.

[0068] Joint maintenance micro-motion is determined only when the macroscopic motion of the target joint has been subtracted, there are residual micro-strains in adjacent regions with opposite directions or mutual restraint, the duration is more than 0.3s, the effective region is continuous, and its change corresponds in time to the decrease in the constraint reliability of the key constraint edge; residual displacement components that are in the same frequency as breathing, appear only at a single texture point, or are consistent with the abnormal vibration main frequency and abnormal vibration direction recorded in step S220 are excluded.

[0069] The spatial position, direction, amplitude, dominant frequency, phase, duration, and effective area ratio of the joint's micro-motion are arranged to form a joint-maintaining micro-strain field for use in step S330. For example, if the knee joint angle changes by less than 1° within 1 second when standing still, and micro-strains with opposite directions and lasting for 0.8 seconds appear on the anterior side of the distal thigh and the posterior side of the proximal calf, while the constraint reliability of the locking constraint edge continues to decrease, then this is written into the joint-maintaining micro-strain field; single-frame jumps of a single texture point are excluded as tracking errors.

[0070] S330. Based on the joint maintenance micro-strain field, identify the characteristics of muscle compensation behavior and form a muscle compensation state sequence. From the benchmark observation subsequence, establish a benchmark micro-strain field according to the target joint, target behavior stage, and gait phase, and record the micro-strain amplitude, duration, effective area ratio, dominant frequency, and median, 5th percentile, and 95th percentile of the phase relationship.

[0071] Muscle compensatory load is a visual state indicator determined based on the amplitude increment, duration, and effective area proportion of the current joint-maintaining microstrain relative to the reference microstrain field. It does not represent actual muscle strength or electromyographic intensity. Its calculation is as follows:

[0072] in, Let be the muscle compensatory load at time t, with a value ranging from 0 to 1; The normalized microstrain amplitude increment is obtained by dividing the difference between the current microstrain amplitude and the reference median by the difference between the 95th percentile of the reference and the reference median. The normalized duration is obtained by the ratio of the duration to the preset longest statistical time, which is the 95th percentile of the duration of the same behavior phase in the baseline observation subsequence. It is the ratio of the effective area of ​​micro-strain to the area of ​​the muscle area around the joint; and These are the corresponding weights, and the sum of the three weights is 1.

[0073] Normalization results are assigned a value of 0 when less than 0 and 1 when greater than 1. Weights are determined by the sensitivity of each feature in historically labeled videos to instability identification; equal weights are used when no calibration data is available. In a specific application scenario, the incremental weight of the micro-strain amplitude is... The preferred value is 0.4, with duration weighting. The preferred value is 0.4, which represents the weight of the effective area ratio of micro-strain. The preferred value is 0.2.

[0074] Muscle compensation synergy is a visual state index determined based on the consistency of dominant frequency, phase stability, and spatial continuity of muscle regions around different joints. The normalized dispersion of dominant frequency in different regions, the normalized deviation of the current phase relationship relative to the baseline phase relationship, and the proportion of spatially discontinuous regions are extracted. A weighted sum of these three parameters is calculated, and the muscle compensation synergy is obtained by subtracting this weighted sum from 1, with the result limited to the 0-1 range. The weights of the three parameters are determined by the sensitivity of corresponding features in historically labeled videos to the identification of unstable states; equal weights are used when no calibration data is available.

[0075] When the muscle compensation load is above the 95th percentile of the corresponding baseline value for three consecutive sliding time windows, the muscle compensation load is considered to have increased; when the muscle compensation synergy is below the 5th percentile of the corresponding baseline value minus the measurement error tolerance for three consecutive sliding time windows, the muscle compensation synergy is considered to have decreased. The confidence score is calculated as the arithmetic mean of the effective proportion of texture points, region continuity, and 1 minus the normalized tracking error. If the extracted effective region area is below the minimum area threshold, the data availability flag for that frame is set to unavailable, and zero-padding or forced interpolation is prohibited to prevent the introduction of false compensation features.

[0076] The acquisition time, target patient number, target joint number, key constraint edge number, target behavior stage, gait phase, muscle compensation load, muscle compensation synergy, effective area ratio of micro-strain, identification confidence, and data availability markers are combined in chronological order to form a muscle compensation state sequence, which is then used for time alignment between step S410 and the brace constraint state sequence.

[0077] S4 specifically includes the following sub-steps: S410. Construct a two-layer support state vector and establish an individualized support contribution transfer model. Align the patient-brace coupling observation sequence output in step S130, the brace constraint state sequence output in step S230, and the muscle compensation state sequence output in step S330 according to the target patient number, target joint number, key constraint edge number, and acquisition time. If the acquisition times are inconsistent, match the most recent valid data with a time difference not exceeding one video sampling period; otherwise, mark the data availability as unavailable. When the same target joint corresponds to multiple key constraint edges, retain the constraint reliability, degradation rate, and importance of each edge separately.

[0078] The constraint reliability, constraint degradation rate, constraint importance, muscle compensation load, muscle compensation synergy, target joint angle, target joint angular velocity, target behavior stage, gait phase, effective micro-strain region ratio, and recognition confidence are combined to form a two-layer support state vector; among which, the target joint angle and target joint angular velocity are derived from step S130, and the effective micro-strain region ratio and recognition confidence are used to evaluate the quality of the input data.

[0079] Training samples were generated using safety state sequences from baseline videos, sequences of brace loosening, posture recovery, nursing intervention, and sudden joint flexion from historical monitoring videos, and constraint degradation test videos of artificial limb models with the same type of brace. Reinforcement learning is a reinforcement learning process that uses a two-layer support state vector as the state, model parameter adjustments as actions, and generates feedback values ​​based on subsequent posture results to update the decision-making strategy.

[0080] The adjustment amounts of brace constraint contribution coefficient, muscle compensation contribution coefficient, and prediction time step are set as reinforcement learning actions; the first two types of adjustment amounts are selected from -0.05, 0, and 0.05, respectively, and the prediction time step adjustment amounts are selected from -0.01s, 0, and 0.01s, respectively.

[0081] The safe state, attitude recovery state, or unstable state obtained from subsequent image recognition are used as environmental feedback; positive feedback is given when the predicted failure time is close to the actual instability time; false alarms are penalized when false alarms occur; and false alarms are penalized with a larger absolute value when false alarms occur. A deep Q-network is used to select parameters to adjust actions.

[0082] In this embodiment, the deep Q-network includes an input layer, 2 to 4 hidden layers, and an output layer. The number of nodes in the input layer is consistent with the dimension of the two-layer support state vector. The hidden layers use the ReLU activation function, and the output layer outputs the Q-values ​​corresponding to the parameter adjustment actions. The training terminates when the verification error fails to decrease for 10 consecutive epochs or when 100 epochs have been completed.

[0083] The input is then normalized using the baseline observation subsequence to form an individualized support contribution transfer model. The posture prediction module is trained using the current double-layer support state vector from the training samples as input and the target joint angle and angular velocity at the next moment as output, and then incorporated into the individualized support contribution transfer model. Instability states are defined as at least one of the following: target joint angle exceeding limits, angular velocity exceeding thresholds, sudden flexion, rapid pelvic descent, or requiring support from a caregiver. The instability angular velocity threshold and pelvic instability threshold are taken as the boundary values ​​between historical safe adjustment samples and instability samples.

[0084] S420. Generate a posture evolution sequence that shows the transfer of brace constraint contribution to muscle compensation contribution. Calculate the brace constraint contribution of the target joint based on the constraint reliability and importance of the critical constraint edges.

[0085] in, The contribution of the support constraint at time t; N is the number of critical constraint edges; Let be the constraint importance of the j-th critical constraint edge; Let be the constraint reliability of the j-th critical constraint edge, where j is the index of the critical constraint edge.

[0086] Brace constraint contribution is a normalized visual indicator characterizing the ability of orthopedic braces to maintain the posture of a target joint. The difference between the baseline brace constraint contribution and the current brace constraint contribution is defined as the brace constraint gap, and the amount of compensation required to fill the brace constraint gap is defined as the required muscle compensation contribution. The available muscle compensation capacity is determined based on muscle compensation load and muscle compensation synergy. Its initial value is represented by the product of muscle compensation synergy and 1 minus muscle compensation load, and is corrected by an individualized brace contribution transfer model.

[0087] The difference between the available muscle compensation capacity and the required muscle compensation contribution is defined as the support margin; a support margin greater than 0 indicates that the support constraint gap can be covered, and a support margin less than or equal to 0 indicates that the gap cannot be filled.

[0088] The prediction time step is set to 0.05-0.2s and the prediction time domain is set to 3-10s. Within each prediction time step, the predicted constraint reliability of each key constraint edge is updated according to the constraint degradation rate, and the predicted brace constraint contribution and predicted brace constraint gap are calculated. The predicted muscle compensation load, predicted muscle compensation synergy, and predicted available muscle compensation capacity are updated by the individualized support contribution transfer model. The current double-layer support state vector and the above predictions are input into the attitude prediction module to obtain the predicted joint angle and predicted joint angular velocity for the next prediction time step.

[0089] Each normalization index is limited to 0-1, and the joint angle is limited to the range allowed by human structure. The prediction confidence is the arithmetic mean of the available data ratio, 1 minus the normalized state distance, and 1 minus the normalized standard deviation at the time of failure, with a value range of 0-1.

[0090] The prediction time, key constraint edge number, prediction constraint reliability, prediction brace constraint contribution, prediction brace constraint gap, prediction available muscle compensation capacity, prediction support margin, prediction muscle compensation load, prediction muscle compensation synergy, prediction joint angle, prediction joint angular velocity, and prediction confidence are combined in chronological order to form a posture evolution sequence.

[0091] For example, if the baseline brace constraint contribution is 0.85 and the current value is 0.60, then the brace constraint gap is 0.25. When the muscle compensation load is 0.65 and the muscle compensation synergy is 0.75, the initial available muscle compensation capacity is 0.2625 and the support margin is 0.0125, indicating that the posture can still be maintained but the compensation margin is small.

[0092] S430. Determine the failure time and remaining maintenance time of the double-layer support. The constraint reliability threshold is derived from the lowest constraint reliability in the calibration test of the artificial limb model of the same type of orthopedic brace, where a single brace constraint edge can still maintain effective constraint; the brace constraint contribution threshold is derived from the lowest brace constraint contribution in the calibration test where the orthopedic brace can still limit the target joint from entering the safe posture range.

[0093] The individual tolerance threshold is the sum of the 95th percentile of the muscle compensation load under safe conditions in the baseline observation subsequence of the target patient and the measurement error tolerance; the synergy threshold is the 5th percentile of the baseline muscle compensation synergy minus the measurement error tolerance, which is determined by the prior bias fixed value of the image recognition algorithm preset by the system; the safe posture range is the minimum range among the doctor's orders, rehabilitation plan, brace limiting angle and baseline movement range; the confidence threshold is the lowest confidence level in the validation set that meets the preset false negative rate.

[0094] In the posture evolution sequence, each predicted time step is used to determine whether the predicted brace constraint contribution is lower than the brace constraint contribution threshold, whether the predicted support margin is less than or equal to 0, whether the predicted muscle compensation load reaches the individual tolerance threshold or whether the predicted muscle compensation synergy is lower than the synergy threshold, and whether the predicted joint angle exceeds the safe posture range or whether the predicted joint angular velocity reaches the instability angular velocity threshold.

[0095] If two consecutive prediction time steps simultaneously satisfy the above-mentioned conditions of insufficient support constraints, insufficient compensation capacity, and tendency towards attitude instability, the first prediction time that meets the conditions is determined as the failure time of the double-layer support; the time difference between the monitoring time and this failure time is the remaining maintenance time. If no double-layer support failure occurs within the entire prediction time domain, the remaining maintenance time is output as greater than the prediction time domain, and the minimum support margin is recorded; if the input state is continuously unusable for more than 0.5s, or the prediction confidence is lower than the confidence threshold, the determination of the specific remaining maintenance time is stopped and a monitoring data insufficiency flag is output.

[0096] The current monitoring time, each prediction time, target patient number, target joint number, key constraint edge number, current constraint reliability, current constraint degradation rate, current muscle compensation load, current muscle compensation synergy, predicted support margin, predicted joint angle, predicted joint angular velocity, double-layer support failure marker, prediction confidence, insufficient monitoring data marker, and remaining maintenance time are combined to form a fall risk evolution sequence, which is then used for combined risk determination in step S510.

[0097] S5 specifically includes the following sub-steps: S510. Determine the combined risk of brace constraint degradation and muscle compensation exhaustion and determine the risk level. Read the fall risk evolution sequence output in step S430 to obtain the target patient number, target joint number, key constraint edge number, current constraint reliability, current constraint degradation rate, current muscle compensation load, current muscle compensation synergy, predicted support margin, remaining maintenance time, prediction confidence, and insufficient monitoring data marker.

[0098] Combined risk refers to the risk state in which the reliability of the constraint continuously decreases, the muscle compensation load continuously increases, the muscle compensation synergy continuously decreases, and the predicted support margin continuously decreases and is less than or equal to 0 in the prediction time domain for the same target joint and key constraint edge within the overlapping time interval.

[0099] The trend is confirmed using the sliding time windows in steps S230 and S330: if the constraint degradation rate is higher than the degradation rate threshold for at least three consecutive sliding time windows, the constraint reliability is determined to be continuously decreasing; if the muscle compensation load is higher than the 95th percentile of the corresponding benchmark value and the last value is higher than the first value, the muscle compensation load is determined to be continuously increasing; if the muscle compensation synergy is lower than the 5th percentile of the corresponding benchmark value minus the measurement error tolerance and the last value is lower than the first value, the muscle compensation synergy is determined to be continuously decreasing; when the three trends overlap for at least two sliding time windows, the combined risk is confirmed.

[0100] The time required from risk identification to the target patient achieving effective protection is defined as the safe intervention time.

[0101] in, For safe intervention time; For risk assessment and instruction generation time; This refers to the time for instruction transmission and terminal confirmation. For the arrival time of the nursing staff; The time allotted for completing support or ceasing action; All values ​​are in seconds (s) to provide a safety margin. and Sourced from system operation logs and Derived from historical nursing response records, The data is set by medical institutions based on the ward layout, and the historical time item is taken as the 95th percentile.

[0102] If the remaining maintenance time is less than or equal to the safety intervention time, or the current predicted support margin is less than or equal to 0, the risk level is determined to be high. If the remaining maintenance time is greater than the safety intervention time but less than or equal to twice the safety intervention time, the risk level is determined to be medium. If the remaining maintenance time is greater than twice the safety intervention time and the combined risk is established, the risk level is determined to be low. When the constraint reliability of a key constraint edge is lower than the constraint reliability threshold, the risk level is increased to at least medium. When the prediction confidence is lower than the confidence threshold or there is a sign of insufficient monitoring data, the monitoring confidence status is output.

[0103] A risk assessment result is generated, including the target joint, key constraint edges, risk level, remaining maintenance time, safety intervention time, predicted support margin, combined risk basis, and prediction confidence level, for use in step S520. For example, if the safety intervention time is 3s, the remaining maintenance time is 2.4s, and the combined risk is valid, it is determined to be high risk.

[0104] S520. Generate and execute risk management instructions based on the risk assessment results. Risk management instructions are structured data sent to patient alert devices, nursing terminals, or safety protection devices, including target patient number, monitoring area number, risk level, target joint, critical constraint edge type, suggested treatment action, generation time, validity period, and instruction priority.

[0105] In low-risk situations, a prompt to check the connection status of the orthopedic brace is sent to the nursing terminal, and the target patient is advised to reduce their activity speed. In medium-risk situations, the target patient is advised to stop walking, maintain their current supporting posture, and wait for nursing staff, while the nurse call interface is activated. In high-risk situations, a stop action prompt is immediately sent to the target patient, an emergency message is sent to the responsible nurse terminal and the ward central terminal, and a safety control command is sent to the safety protection device connected to the system.

[0106] Safety protection devices include nurse call devices, bed height adjustment devices, cushioning pad deployment devices, or safety support devices for rehabilitation training areas. When a risk management instruction is received, the corresponding device independently performs actions that match its type, such as calling, lowering the bed, deploying the cushioning pad, or moving the safety support components.

[0107] After receiving a risk management instruction, the nursing terminal or safety protection device returns a confirmation message; if the confirmation time limit is exceeded, it will resend once, and if no confirmation is received, it will escalate the message to the ward's central terminal. The safety protection device returns a status indicating successful execution, failed execution, or device unavailable; if the device is unavailable, the nursing call will continue to be maintained.

[0108] Record the risk assessment result, instruction content, sending time, receiving terminal, confirmation time, equipment execution result, nursing staff arrival time, treatment completion time, and key video frame index to form a risk treatment record for verification in step S530. For example, if a high-risk instruction is confirmed within 1 second after being generated, and the nursing staff arrives and assists the target patient 2.2 seconds later, then 2.2 seconds is recorded as the actual nursing response time.

[0109] S530. Verify the risk prediction results and update individualized monitoring parameters. Determine the risk confirmation time, the start time of the intervention, and the verification observation window based on the key video frame index in the risk response record. After the risk response, continue to acquire continuous video images. The larger of the remaining maintenance time and the safety intervention time, plus a 2-5s observation margin, is determined as the verification observation window; if the remaining maintenance time cannot be determined, the verification observation window is set to 5-15s.

[0110] Continue performing image recognition and behavioral feature recognition within the window to determine whether the target joint and key constraint edges have been restored, whether the muscle compensatory load has decreased, whether the muscle compensatory synergy has been restored, and whether there has been sudden joint flexion, rapid pelvic descent, or support from nursing staff.

[0111] Events that cause instability near the predicted failure time are marked as true warnings; events in which nursing staff or safety devices protect the target patient and restore them to a safe posture before the predicted failure time are marked as effective intervention events and are not considered false alarms; events in which no intervention occurs and the patient remains safe throughout the verification observation window, with brace restraint and muscle compensation resuming spontaneously, are marked as false alarm candidate events; events in which instability occurs after the initial warning is not given are marked as missed events; events with severe obstruction or interrupted video recording are marked as invalid data events.

[0112] The actual instability moment is determined by the earliest moment when the target joint angle exceeds the safe posture range for two consecutive video sampling cycles, the target joint angular velocity reaches the instability angular velocity threshold, the pelvic height descent rate reaches the pelvic instability threshold, or the nursing staff applies support because the target patient is unable to maintain their posture independently. The prediction time error is the absolute value of the difference between the time of double-layer support failure and the actual instability moment.

[0113] Only genuine early warnings, effective intervention events, false alarm candidate events verified by manual review, and missed events are written into the individualized experience sample database; invalid data events are not included in the update. Genuine early warnings and effective intervention events receive positive feedback, while false alarm candidate events and missed events receive negative feedback, and the absolute value of the penalty for missed events is greater than the penalty for false alarms.

[0114] When 10 new valid samples are added, or when 24 hours have passed since the last update, a batch update is performed. This update includes the individualized support contribution transfer model, muscle compensation load benchmark, muscle compensation synergy benchmark, individual tolerance threshold, prediction confidence parameters, and safety intervention time. Each parameter update is limited to 1%-5% of the original parameters. Brace manufacturer structural parameters, constraint reliability threshold, brace constraint contribution threshold, safe posture range, and safety protection device hardware safety limits are not automatically updated. After the update, the retained benchmark observation subsequence is used for verification. If the false negative rate increases or the safety status prediction error exceeds the allowable range, the update is revoked.

[0115] Finally, an individualized monitoring parameter set is formed, which includes the target patient number, orthopedic brace number, update parameters, model version, sample size and validation results. These parameters are then called by steps S330, S410, S420, S430 and S510 in the next monitoring cycle to complete the closed-loop monitoring.

[0116] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0117] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0118] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0119] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring fall risk in orthopedic patients based on posture recognition, characterized in that, Includes the following steps: S1. Collect baseline and monitoring video images of the target patient wearing orthopedic braces, perform calibration, correction and time alignment, identify key points of human joints, limb contours, brace components and connection positions, and determine the effective monitoring period through behavioral feature recognition to form a patient-brace coupled observation sequence. S2. Based on the patient and brace coupling observation sequence, establish a brace constraint topology graph, calculate the constraint reliability, constraint degradation rate and constraint importance of each brace constraint edge, determine the key constraint edges, and form a brace constraint state sequence. S3. Determine the micro-motion extraction area around the target joint based on the key constraint edges, deduct camera shake, overall limb movement, baseline behavioral movement and respiratory interference, construct the joint maintenance micro-strain field, calculate muscle compensation load and muscle compensation synergy, and form a muscle compensation state sequence. S4. Align the brace constraint state sequence and muscle compensation state sequence to construct a double-layer support state vector. Use reinforcement learning to establish an individualized support contribution transfer model to predict the brace constraint contribution, available muscle compensation capacity, support margin, and target joint posture. Determine the failure time and remaining maintenance time of the double-layer support to form a fall risk evolution sequence.

2. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 1, characterized in that, Also includes: S5. Determine the risk level based on the fall risk evolution sequence, remaining maintenance time, and safety intervention time, generate risk handling instructions, verify the early warning results using continuous video images after handling, and update individualized monitoring parameters.

3. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 1, characterized in that, S1 specifically includes: Baseline and monitoring video images of the target patient wearing orthopedic braces were collected. Distortion correction, coordinate transformation, time alignment, camera shake compensation, and brightness normalization were performed on the video images to form a standardized video image sequence. Joint image recognition of human body and orthopedic brace is performed on standardized video image sequences to extract key joint points, limb contours, clothing texture blocks, brace components and connection positions, establish identity association between orthopedic brace and target patient, and form multi-target feature sequence of patient and brace.

4. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 3, characterized in that, Also includes: Behavioral characteristics are identified based on the multi-target feature sequences of patients and braces, identifying the stages of standing still, starting, continuous walking, turning, and stopping. Effective monitoring periods in which the brace is visible and the target patient is in a state of autonomous weight-bearing are selected, forming a patient-braces coupled observation sequence that includes data source markers.

5. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 1, characterized in that, S2 specifically includes: Based on the patient and brace coupling observation sequence, establish human body nodes, brace nodes, human body connection edges, brace structure edges, and brace constraint edges, record the type of each brace constraint edge, the corresponding target joint, and the initial geometric relationship, and form a brace constraint topology graph. For the locking constraint edge, hinge guide constraint edge, strap fixing constraint edge and shell follower constraint edge, the offset of locking component, relative rotation angle of hinge, strap slippage, relative displacement of support shell and abnormal vibration parameters are extracted respectively. The constraint reliability is calculated based on individual reference parameters and allowable deviations. The constraint degradation rate is calculated based on the constraint reliability, and the constraint importance is determined based on the change in the allowable range of motion of the target joint before and after the constraint is removed in the simulation.

6. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 5, characterized in that, Also includes: Key constraint edges are selected, and constraint reliability, constraint degradation rate, constraint importance, and key constraint edge markings are combined in chronological order to form a support constraint state sequence.

7. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 1, characterized in that, S3 specifically includes: Based on the patient, brace coupling observation sequence and brace constraint state sequence, the target joint and its micro-motion extraction area are determined. After deducting camera shake, overall limb movement, baseline behavioral movement and respiratory interference, a standard micro-motion image sequence is formed. Subpixel displacement estimation is performed on standard micro-motion image sequences, micro-strain components of residual displacement are calculated, micro-strain amplitude, dominant frequency, phase and duration are extracted, joint maintenance micro-motion corresponding to the degradation of key constraint edges is screened, and joint maintenance micro-strain field is constructed. The joint maintenance microstrain field is compared with the baseline microstrain field of the same behavioral stage and gait phase to calculate muscle compensation load and muscle compensation synergy, and the target joint, key constraint edges and acquisition time are associated to form a muscle compensation state sequence.

8. The method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 1, characterized in that, S4 specifically includes: Align the brace constraint state sequence and muscle compensation state sequence, construct a two-layer support state vector that includes constraint reliability, constraint degradation rate, constraint importance, muscle compensation load, muscle compensation synergy, and target joint posture, and use reinforcement learning to establish an individualized support contribution transfer model; The constraint contribution of the brace is calculated based on the constraint reliability and constraint importance of the key constraint edges. The constraint gap, available muscle compensation capacity and support margin of the brace are determined. The predicted values ​​are updated according to the prediction step size, and the attitude evolution sequence is formed through the attitude prediction module.

9. A method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 8, characterized in that, Also includes: By comparing the predicted brace constraint contribution, predicted support margin, predicted muscle compensation state, and predicted target joint posture with the corresponding thresholds, the failure time of the double-layer support and the remaining maintenance time are determined, forming a fall risk evolution sequence.

10. A method for monitoring fall risk in orthopedic patients based on posture recognition according to claim 2, characterized in that, S5 specifically includes: Read the fall risk evolution sequence, determine the combined risk based on the constraint reliability, muscle compensation load, muscle compensation synergy and predicted support margin corresponding to the same target joint, and determine the risk level by combining the remaining maintenance time and safety intervention time. Based on the risk level, generate risk management instructions, send risk management instructions to patient alert devices, nursing terminals or safety protection devices, and record instruction confirmation, device execution and nursing response results; Based on the continuous video images after processing, the true early warning, effective intervention, false alarm candidates, missed reports or invalid data events are verified. The effective events are written into the individualized experience sample library, the individualized support contribution transfer model and individualized monitoring parameters are updated, and an individualized monitoring parameter set is formed for use in the next monitoring cycle.

Images