A gait analysis method and system based on hand-knee-foot cooperative motion coupling

Abstract

The application discloses a gait analysis method based on hand-knee-foot cooperative motion coupling, and the method comprises the following steps: step 1, collecting data by using a flexible insole and an inertial knee ring; step 2, unifying the time axis, posture and position of three nodes of hands, feet and knees, establishing a state vector and an observation equation, updating the state vector, calculating the hand-foot phase difference and the neural coordination score, merging the insole features, the knee ring features and the hand ring features to obtain a multi-modal cooperative feature vector and performing verification; and step 3, based on the calculated attention fusion features after verification, inputting the features into a bidirectional LSTM-CNN network to output a Parkinson's disease gait risk probability, a stroke sequelae gait abnormality risk probability and a fall risk probability, and calculating SHAP values of the contribution degrees of the three modalities of hands, knees and feet and performing credibility judgment. The application realizes the balance between high-precision medical-level analysis and low-cost portability application.

Description

Technical Field

[0001] This invention relates to the field of portable multimodal data acquisition and analysis systems for gait function assessment in neurological diseases, and in particular to a gait analysis method and system based on hand-knee-foot coordinated movement coupling. Background Technology

[0002] Gait analysis refers to the quantitative assessment of neuromuscular and skeletal function by measuring the spatiotemporal, kinematic, and dynamic parameters of the human body during walking. Its data accounts for approximately 50% to 70% of the assessment criteria for rehabilitation of neurological diseases (Parkinson's disease, stroke), and has a direct impact on fall risk warning, personalized rehabilitation program development, and follow-up of treatment effectiveness. As rehabilitation medicine shifts towards a model centered on functional recovery and integrating home and institutional care, low-cost, portable, and visualized gait analysis methods are receiving increasing attention.

[0003] Current gait analysis systems have the following problems and shortcomings: 1. Lack of simultaneous joint analysis of multimodal data. Some domestic gait analysis systems only use inertial sensors and lack the collaborative analysis of plantar pressure information and upper limb motion parameters. This results in the inability to simultaneously acquire dynamic (ground reaction force, plantar pressure mapping) and whole-body kinematic parameters, leading to a misjudgment rate of about 12%–18% when identifying Parkinson's frozen gait and hemiplegic gait. 2. High system cost and complex deployment. Although foreign technologies use insole-type pressure + IMU, they require external base station synchronization and on-site time calibration. 30 minutes is a burden that grassroots organizations cannot afford; 3. The data loop between doctors and patients is not fully established. Domestic and international approaches largely focus on the "offline data collection - laboratory analysis" model, lacking interfaces for real-time visual reports and remote training guidance for rehabilitation physicians, therapists, and patients. Clinical surveys show that 80% of primary care rehabilitation therapists report that existing systems cannot directly export gait assessment reports or push personalized home training prescriptions; 4. The conflict between miniaturization and efficiency. Existing portable gait analysis devices often combine multiple sensors in pursuit of accuracy, leading to increased power consumption. 1.2W, continuous working time In terms of data acquisition, visual acquisition is easily affected by lighting and background interference, which affects recognition accuracy. Furthermore, deviations in the layout and calibration of inertial sensors can lead to distortion of gait parameters. In terms of algorithms, the complex algorithms required for high-precision analysis conflict with the limited computing power of portable devices, sacrificing real-time performance. In addition, a single sensor is difficult to capture all features, and some patents lack sufficient multimodal data fusion, resulting in a low abnormal gait recognition rate. At the same time, there is a lack of foldable and easy-to-store structural design, and the patient's home compliance rate after discharge is less than 30%.

[0004] The causes and challenges in solving the above problems lie in the following aspects: Sensor heterogeneity: Foot pressure arrays and IMUs differ significantly in sampling frequency, clock reference, and data format. Achieving millisecond-level hard synchronization analysis of upper and lower limb data requires redesigning the clock allocation and data fusion algorithms. Cost bottleneck: The IMU system relies on imported high-frame-rate industrial cameras and calibration frames; the batch consistency of flexible film pressure insoles is poor, resulting in high cost per pair. 800 yuan. Software fragmentation: The existing algorithm library is incompatible with the HIS / EMR system interface and lacks automatic reporting templates for Chinese rehabilitation terminology, resulting in long secondary development cycles and high costs. Limited clinical validation samples: Frozen gait and hemiplegic gait data need to be collected across centers and diseases, making ethical approval and follow-up difficult and hindering model generalization. Summary of the Invention

[0005] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a gait analysis method based on the coupling of hand, knee, and foot coordinated movements.

[0006] Another objective of this invention is to provide a gait analysis system based on the coupling of hand, knee, and foot coordinated movements.

[0007] The technical solution adopted to achieve the purpose of this invention is: A gait analysis method based on hand-knee-foot coordinated motion coupling includes the following steps: Step 1: Collect gait cycles using flexible pressure insoles on the soles of the feet. Step frequency Step length Three-dimensional position of the foot Foot and ankle joint posture quaternions and shear forces in the anteroposterior and medial directions of the foot Inertial knee rings collect acceleration data at the knee joint. angular velocity Knee phase Knee joint posture quaternion Three-dimensional position of the knee and joint angle The inertial wristband collects the acceleration at the wrist. angular velocity wrist posture quaternion Three-dimensional position of the wrist Arm swing amplitude within one gait cycle and peak angular velocity ; Step 2: Unify the time axes of the three nodes (hand, foot, and knee), and unify the pose and position of the three nodes into the human coordinate system to establish a state vector. and observation equations, through observation equations to Update and calculate hand-foot time difference. and neuro-coordination score Insole features are constructed by combining the data collected in step 1. Knee ring characteristics and wristband features ,Will , and The merging yields a multimodal collaborative feature vector. And verify; Step 3, based on the verified Computational attention fusion features ,Will Input a bidirectional LSTM-CNN network and output the probability of gait risk for Parkinson's disease. Risk of gait abnormalities as a sequela of stroke and the probability of falling Based on , and Calculate the SHAP values ​​of the contributions of the three modalities of hand, knee and foot respectively, and judge their reliability.

[0008] In the above technical solution, in step 1, The calculation formula is: ; In the formula, The moment of liftoff. The moment of contact with the ground; Total vertical force on the sole of the foot The moment when the trigger threshold is first exceeded Within the same area When the time is below the cutoff threshold, The calculation formula is: ; In the formula, for The reaction force is always perpendicular to the ground. Foot pressure values ​​collected in real time by a piezoresistive array Calculations show that The calculation formula is: ; In the formula, This represents the area of ​​a single piezoresistive point. The plantar pressure value is denoted as , where This indicates the longitudinal position of the piezoresistive array. This indicates the lateral position of the piezoresistive array. Sampling time; The calculation formula is: ; The step size Through the center trajectory of the two steps before and after It was estimated; in, Represented as: ; In the formula, Let be the lateral coordinate of the center of pressure in the insole coordinate system. Let be the longitudinal coordinate of the center of pressure in the insole coordinate system; and The calculation formula is: ; In the formula, The coordinates of the pressure resistance point are shown in the coordinate system of the flexible pressure insole on the sole of the foot.

[0009] In the above technical solution, in step 2, the inertial wristband is used as the master node, and the flexible pressure insole and inertial knee ring are used as slave nodes, unifying the time axes of the three nodes (hand, foot, and knee) to the same master clock axis. By calculating the arm swing amplitude-step length coupling coefficient and torso rotation angle Quantifying the coordination of upper and lower limb movements; Among them, the master clock axis The calculation formula is: ; In the formula, The most recent master timestamp received by the flexible pressure insole and inertial knee ring. For the flexible pressure insoles and inertial knee rings, at the local sampling time, For flexible pressure insoles and inertial knee rings, the local time of the master timestamp is received. Clock drift factor for each slave node; Arm swing amplitude-step coupling coefficient The calculation formula is: ; In the formula, This is a reference ratio for healthy individuals; Torso rotation angle The calculation formula is: ; In the formula, This refers to the yaw angle weighting coefficient of the inertial wristband. The inertial knee roll angle weighting coefficient is used. This represents the change in yaw angle of the inertial wristband. This represents the change in inertial knee ring roll angle; The observation equation is expressed as: ; In the formula, For the observation vectors of the inertial wristband, inertial kneeband, and flexible pressure insole, The observation matrix is ​​for the inertial wristband, inertial knee ring, and flexible pressure insole. The noise levels observed were for inertial wristbands, inertial kneebands, and flexible pressure insoles on the soles of the feet.

[0010] In the above technical solution, in step 2, the update equation for the state vector is: ; In the formula, for The state vector at time t, for The state vector at time t, Here is the Kalman gain matrix. Observation vector, The observation matrix; The hand and foot time phase difference The calculation formula is: ; In the formula, This is the peak moment of the inertial wristband's angular velocity. The moment when the ipsilateral plantar pressure drops to 10%; The neuro-coordination score The calculation formula is: ; In the formula, , These represent the mean and standard deviation of hand-foot distance differences in healthy individuals. This represents the coordinated phase difference between the hand joint and the knee joint. This represents the coordinated phase difference between the knee joint and the ankle joint. The multimodal collaborative feature vector Represented as: ; in: ; ; ; In the formula, The reaction force is perpendicular to the ground. The trajectory centered on.

[0011] In the above technical solution, in step 2, The verification includes: When the inertial wristband detects peak torso acceleration satisfy At the same time Changes satisfy hour, If the change in vertical ground reaction force is detected, the data from the flexible pressure insole is deemed abnormal. In this case, the system automatically switches to the IMU-dominated mode of the inertial wristband and estimates the stride frequency using the arm swing frequency. ,accomplish Verification; When the plantar pressure cycle stabilizes after 5 consecutive steps, but the inertial wristband attitude integral displacement error... At 10cm, the zero bias of the inertial wristband IMU is corrected by using the foot step length in the reverse direction. ,accomplish Verification; When step size and step frequency are consistent At that time, calculate the correction step size. Joint angles with soft constraints within the physiological range Initiate bidirectional correction to achieve and Verification.

[0012] In the above technical solution, the zero bias of the inertial wristband IMU The calculation formula is: ; In the formula, For the displacement increment based on inertial wristband measurement data, The displacement increment is based on plantar pressure data. This is an empirical scaling factor; The step size-step frequency consistency is expressed as: ; In the formula, For speed based on insole pressure, The speed is based on the wristband's acceleration. in, The calculation formula is: ; The calculation formula is: ; The correction step size The calculation formula is: ; In the formula, For adaptive fusion weights; in, The calculation formula is: ; In the formula, Confidence level for flexible pressure insoles on the sole of the foot. For the confidence level of the inertial wristband; The joint angles under soft constraint within the physiological range The calculation formula is: ; In the formula, The original joint angle, This represents the physiological limit angle of the joint. It is a sigmoid function. This is the scaling factor.

[0013] In the above technical solution, in step 3, the attention fusion feature The calculation formula is: ; In the formula, For attention mechanism queries, For the key of attention mechanism The value matrix for the attention mechanism, For the normalized attention weight matrix, Let be the scaling factor, where The feature dimension of the key vector; Among them, the query of attention mechanism Key to attention mechanisms Value matrix of attention mechanism The calculation formula is: ; In the formula, The projection matrix; The probability of gait risk in Parkinson's disease Risk of gait abnormalities as a sequela of stroke and the probability of falling The calculation formula is: ; In the formula, , Representing different risk prediction tasks, The deep feature vector output by the bidirectional LSTM-CNN network is... This is the weight matrix of the fully connected layer for the corresponding task. This refers to the bias term for the corresponding task. The function is used to map the network output to The probability value of the interval; The SHAP values ​​of the contributions of the hand, knee, and foot modalities The calculation formula is: ; In the formula, For modal types, , The total number of modes, The risk score output by the network. This represents any combination of modes other than the current mode. The number of modes contained in the combination. Indicates that only a subset is input. The risk score predicted by the model at that time This represents the output after adding the current modal feature to the same combination. These are the combined weighting coefficients.

[0014] In the above technical solution, in step 3, the bidirectional LSTM-CNN network uses multi-task joint loss. To make the three modal features complementary and then calculate the gait risk probability of Parkinson's disease. Risk of gait abnormalities as a sequela of stroke and the probability of falling ; Among them, multi-task joint loss The calculation formula is: ; In the formula, Cross-entropy loss predicted for Parkinson's disease. Cross-entropy loss for stroke prediction For the cross-entropy loss of fall prediction, For feature fusion regularization term, For Parkinson's task weights, For stroke task weights, For the task weight of falling, To incorporate regularization weights.

[0015] In the above technical solution, in step 3, The larger the value, the more critical the current modality feature is in the risk assessment of the current sample; if it is negative, the current modality feature plays a dominant role in reducing risk prediction. The credibility is judged as follows: if the high-risk conclusion of a suspected patient is mainly driven by a modality consistent with clinical symptoms, then the model is considered to be reasonable and the conclusion is credible. If the system predicts a high risk, but If the dominant modality does not match clinical observation, the results should be questioned and the sensor status or feature calculations of that modality should be traced back to avoid misjudgment. For the same test, comparisons were made between the hand, knee, and foot modalities. The absolute value and sign of the data are used to confirm whether the risk outcome is raised or lowered by a single modality, thereby determining whether additional data collection or manual review is required.

[0016] In another aspect of the present invention, a system for implementing the gait analysis method based on hand-knee-foot coordinated motion coupling includes: a sensor group, a data processing module, and a user terminal; The sensor group is used to collect multimodal gait detection data of the hand, knee and foot. The sensor group includes a piezoresistive array built into the flexible pressure insole of the foot, a shear force film strip set at the front end and heel of the piezoresistive array, an IMU integrated into the inertial knee ring and an IMU built into the inertial wrist ring. The data processing module includes a data collaboration layer, a feature collaboration layer, and a decision collaboration layer; The data collaboration layer receives multimodal gait detection data collected by the sensor group, unifies the time axes of the hands, knees, and feet to the main clock axis, unifies the posture and position to the human coordinate system, establishes state vectors and observation equations, updates the state vectors, dynamically adjusts the fusion weights according to the sensor confidence, and performs bidirectional complementary correction between the inertial wristband and the flexible pressure insole data. The feature collaboration layer receives the updated state vector, extracts multimodal collaborative features based on the human gait biomechanical coupling mechanism, calculates the stride length-step frequency consistency and compares it with a threshold. When the threshold is exceeded, bidirectional correction is initiated, and soft constraint verification of joint angles is performed based on physiological limit angles to form a verified multimodal collaborative feature vector. The decision collaboration layer receives the validated multimodal collaborative feature vector, calculates the weights of each modality using the built-in attention mechanism fusion module, and generates attention fusion features. The bidirectional LSTM-CNN network of the decision collaboration layer receives the attention fusion features and outputs the gait risk probabilities of Parkinson's disease, post-stroke sequelae, and falls, as well as the contribution of each modality. The data is transmitted to the user terminal to generate a visual gait assessment report. Based on the risk level, personalized rehabilitation guidance plans and real-time warnings are pushed. The user terminal is wirelessly connected to the flexible pressure insole and sends control commands or predetermined thresholds to the flexible pressure insole in real time. The flexible pressure insole simultaneously sends "collection trigger" commands to the inertial knee ring and inertial wrist ring.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention solves the spatiotemporal registration and drift problems of heterogeneous data from multiple sensors, achieving high-precision collaborative sensing. At the data collaboration layer, this invention introduces an Extended Kalman Filter (EKF) algorithm based on human kinematic chain constraints. It utilizes skeletal length constraints to suppress integral drift of inertial sensors and proposes a hand-foot complementary anomaly detection mechanism (i.e., wristband supplementing foot sole, foot sole calibrating wristband), effectively solving the problem of single-modal failure under complex working conditions. Through this collaborative algorithm, the system achieves millisecond-level multimodal clock synchronization (error... 3ms) and high-precision spatial registration (6DoF error) (4mm), significantly improving the robustness and accuracy of the data; 2. This invention overcomes the limitations of traditional gait analysis, which lacks quantification of upper and lower limb coordination, thus improving the physiological interpretability of features. Based on a biomechanical coupling mechanism, this invention constructs a neuro-coordination score and an arm swing amplitude-step length coupling coefficient at the feature coordination layer, filling the gap in existing technologies that neglect upper and lower limb coordination impairments in patients with Parkinson's disease and stroke. Simultaneously, it introduces step length-step frequency consistency verification and soft constraint algorithms for joint angles, using physiological limit thresholds to bidirectionally correct the calculation results. This avoids erroneous results from purely data-driven models that violate common sense about human anatomy, ensuring the clinical reliability of the evaluation indicators. 3. This invention addresses the problem of the "black box" nature of deep learning models, which makes them difficult to accept in clinical practice, and achieves visualized attribution in risk assessment. The invention designs a bidirectional LSTM-CNN network with a built-in attention mechanism at the decision collaboration layer and introduces a SHAP (Shapley Additive Explanations) attribution analysis module. The system can not only output the risk probability of Parkinson's disease, stroke, or falls, but also quantify the contribution of each modality (hand, knee, and foot) to the risk, providing doctors with intuitive "pathological attribution evidence" (such as clearly indicating whether high risk is dominated by hand tremors or foot dragging), greatly promoting trust and acceptance of AI diagnostic results by both doctors and patients. 4. A balance is achieved between high-precision medical-grade analysis and low-cost, portable applications, creating a closed loop of "assessment-intervention." While ensuring the operation of the aforementioned high-precision collaborative algorithms, this system optimizes the overall hardware cost through low-cost sensor networking and edge computing, keeping it below 10,000 yuan (far lower than the 150,000+ yuan of traditional optical capture systems). The device is lightweight, portable, and easy to operate, making it suitable for primary care and home settings. Furthermore, the system constructs a complete closed loop from real-time risk warning to personalized rehabilitation guidance. Clinical validation data shows that this instant feedback mechanism effectively improves patients' adherence to home rehabilitation (by approximately 37%), significantly improves rehabilitation efficiency, and reduces the risk of falls. Attached Figure Description

[0018] Figure 1 The diagram shows the gait analysis method based on hand-knee-foot coordinated motion coupling of the present invention.

[0019] Figure 2 The diagram shown is a structural schematic of the flexible pressure insole for the sole of the foot according to the present invention.

[0020] Figure 3 The diagram shown is a schematic representation of the inertial knee ring of the present invention.

[0021] Figure 4 The diagram shown is a structural schematic of the inertial wristband of the present invention.

[0022] Figure 5 The image shown is a visualization example of the SHAP value of this invention.

[0023] Among them, 1-flexible pressure insole, 2-inertial knee ring, 21-knee ring shell, 22-knee ring fixing layer, 23-knee ring elastic bandage, 3-inertial wristband, 31-wristband shell, 32-wristband fixing layer. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] Example 1 A gait analysis method based on hand-knee-foot coordinated motion coupling, referring to Figure 1 This includes the following steps: Step 1: Data collection from flexible foot pressure insoles, inertial knee rings, and inertial wristbands: The flexible foot pressure insoles collect gait cycles. Step frequency Step length Three-dimensional position of the foot Foot and ankle joint posture quaternions and shear forces in the anterior-posterior and medial-lateral directions of the foot; inertial knee rings collect acceleration at the knee joint. angular velocity Knee phase Knee joint posture quaternion Three-dimensional position of the knee and joint angle The inertial wristband collects the acceleration at the wrist. angular velocity wrist posture quaternion Three-dimensional position of the wrist Arm swing amplitude within one gait cycle and peak angular velocity ; The gait cycle The calculation formula is: ; In the formula, The moment of liftoff. The moment of contact with the ground; The moment of contact with the ground (recorded as) The total vertical force on the sole of the foot is determined to be... The moment the trigger threshold is first exceeded (30kg for adults, 10kg for children), the MCU pulls high the general purpose input / output interface (GPIO) and outputs the "STEP_START" signal; the moment of lift-off... (recorded as) ) Determined to be in the same area When the pressure drops below the cutoff threshold (ground clearance threshold) by 5 kg again (or falls below 10% of the peak plantar pressure), the cutoff threshold is set to 5 kg instead of 0 kg to filter out background noise from the pressure sensor and ensure that ground clearance is only marked when the pressure is actually released. The cutoff threshold, together with the trigger threshold, forms a complete gait cycle determination logic.

[0026] Among them, the total vertical force of the foot The calculation formula is: ; In the formula, for The vertical ground reaction force at the sampling moment; The vertical ground reaction force is calculated by summing the pressure values ​​at all pressure resistance points. The The calculation formula is: ; In the formula, This represents the area of ​​a single piezoresistive point. Here is the plantar pressure matrix, where, , indicating the longitudinal position of the piezoresistive array. , indicating the lateral position of the piezoresistive array. The sampling time.

[0027] Step frequency is calculated based on the ground contact times of two consecutive steps. The calculation formula is: ; In the formula, The step frequency is the time interval between two consecutive steps touching the ground, which is one gait cycle. This indicates the number of steps completed per unit of time (unit: steps / min or Hz). The normal range for healthy adults is approximately 100–120 steps / min.

[0028] Step length Stride length represents the distance between the centers of the heels (in meters or centimeters) when the same-side foot strikes the ground twice consecutively. The normal range for healthy adults is approximately 0.6–0.8 meters. With step frequency Together they form the basis of gait spatiotemporal parameters, the Through the pressure center trajectory of the two steps The trajectory was estimated. Among them, the trajectory of the pressure center The representation is: ; In the formula, Let be the lateral coordinate of the center of pressure in the insole coordinate system. Let be the longitudinal coordinate of the center of pressure in the insole coordinate system.

[0029] in, and The calculation formula is: ; In the formula, The coordinates of the pressure resistance point in the insole coordinate system are given.

[0030] The inertial knee ring is used to assess lower limb motor function and identify abnormal gait patterns, only during the gait cycle. Internal acceleration and angular velocity acquisition is used to avoid accidental triggering during sitting or lying positions. These events are used to segment gait phases (support phase, swing phase), which form the basis for subsequent calculations of parameters such as stride frequency and stride length.

[0031] The acceleration at the knee joint It is obtained by measuring the acceleration components along three orthogonal axes at the knee joint. The coordinate system is defined as follows: the X-axis points proximally along the femur, the Y-axis points laterally, and the Z-axis points vertically upward.

[0032] The angular velocity It is obtained by measuring the angular velocity components of the knee joint around the X, Y, and Z axes.

[0033] By fusing accelerometer and gyroscope data using complementary filtering or extended Kalman filtering, the three-dimensional rotational attitude of the knee joint in a global coordinate system (such as a geomagnetic north-gravity east coordinate system) is calculated to obtain the knee joint attitude quaternion. Knee joint posture quaternion Used for subsequent calculations of knee flexion-extension angles, adduction-abduction angles, and internal / external rotation angles.

[0034] Displacement is obtained by double integration of accelerometer data, and zero-bias compensation and drift suppression are performed using an extended Kalman filter to obtain the three-dimensional spatial coordinates of the inertial knee loop in the global coordinate system (each component is in meters or millimeters), thus obtaining the three-dimensional position of the knee. .

[0035] acceleration at the wrist It is obtained by measuring the acceleration components at the wrist along the X, Y, and Z axes of a coordinate system. The coordinate system is defined as follows: the X-axis points along the back of the hand towards the fingers, the Y-axis points towards the ulnar side, and the Z-axis is perpendicular to the back of the hand and upwards.

[0036] The angular velocity at the wrist It is obtained by measuring the angular velocity components of the wrist around the X, Y, and Z axes.

[0037] The wrist posture quaternion The wrist's three-dimensional rotational attitude (unit: dimensionless) in the global coordinate system is calculated by fusing accelerometer and gyroscope data using complementary filtering or extended Kalman filtering. Wrist attitude quaternion. It is used to calculate arm swing amplitude and swing phase, and is the basic data for assessing hand-foot coordination.

[0038] The three-dimensional position of the wrist Displacement is obtained by double integration of accelerometer data, and zero-bias compensation and drift suppression are performed using an extended Kalman filter to obtain the three-dimensional spatial coordinates of the wristband in the global coordinate system (each component is in m or mm). The coordinate system is centered at the wrist joint rotation center, with the X-axis pointing along the back of the hand to the fingers, the Y-axis pointing to the ulnar side, and the Z-axis perpendicular to the back of the hand and upward.

[0039] The arm swing It is obtained by calculating the range of change of the magnitude of the angular velocity vector within one gait period. It represents the maximum swing amplitude of the arm within one gait cycle (unit: ° or rad), reflecting the range of motion of the upper limb. The normal range for healthy adults walking is about 30°–50°.

[0040] The peak angular velocity It is obtained by calculating the maximum value of the magnitude of the angular velocity vector within one gait period. It represents the maximum angular velocity of the arm during one gait cycle (unit: ° / s or rad / s), reflecting the swing speed of the upper limb. It is of great significance for recognizing muscle rigidity (a typical symptom of Parkinson's disease) and assessing upper and lower limb coordination.

[0041] Step 2: Unify the time axes of the three nodes (hand, foot, and knee), unify the posture and position to the human coordinate system, and establish a state vector. and observation equations, through observation equations to Update and calculate hand-foot time difference. and Insole features are constructed by combining the data collected in step 1. Knee ring characteristics and wristband features ,Will , and Merging multimodal collaborative feature vectors And verify; Although the sensor nodes in the flexible pressure insole, the inertial knee loop, and the inertial wristband in Step 1 have the same sampling rate, they each have their own independent time reference and local coordinate system. Specifically, each node has a local clock. And local coordinate systems (insole coordinate system, knee joint coordinate system, wrist coordinate system). If these raw data are not uniformly preprocessed and fused, the following problems will occur: Time synchronization error: Because each node has an independent clock reference, even if the sampling frequency is the same (100Hz), directly splicing the data will produce a time jitter of 5~15ms. This will cause a hand-foot time difference. The calculation error can reach over 20%, making the neural coordination score... It is severely distorted and cannot accurately assess typical gait abnormalities such as "frozen gait" in Parkinson's disease.

[0042] Spatial registration error: Due to the inconsistent coordinate systems of each node, directly calculating joint angles will result in an error of 10–15° due to coordinate system misalignment. This will cause key parameters such as gait symmetry and single-support phase ratio to deviate from the normal range, directly affecting the accuracy of disease classification results.

[0043] This invention employs a collaborative time synchronization mechanism to unify the time axes of the inertial wristband, inertial knee ring, and flexible pressure insole onto the same master clock, specifically as follows: Employing a master-slave star topology, the inertial wristband acts as a Bluetooth Low Energy (BLE) master node, broadcasting a 64-bit UTC timestamp every 100ms. Flexible pressure insoles and inertial knee rings serve as nodes for local sampling. Simultaneously record the most recent main timestamp received. And the local time that received the timestamp. Simultaneously, the node is sampled locally at the specified time. The raw sensor data collected (including plantar pressure, acceleration, or angular velocity) is labeled as and send back data packets, the format of which is ( , , ); When the flexible pressure insole first detects a vertical load ≥30kg (adult threshold, 10kg can be set for children), a global 0ms clock reference is triggered. The user terminal broadcasts a 64-bit timestamp via BLE, and the insole, knee band, and wristband synchronously transmit data. If any node experiences packet loss... 5%, the user end automatically downsamples to 50Hz and retransmits the synchronization packet to ensure data integrity.

[0044] The user end uses joint calibration and linear interpolation to unify all data to the master clock axis. The clock drift factor is estimated for each slave node. (The clock drift rate of each slave node is estimated in real time using Kalman filtering), and the slave node time is mapped to the master clock axis. In this embodiment, an inertial wristband is used as the master node, and a flexible pressure insole on the sole of the foot and an inertial knee ring are used as slave nodes, unifying the time axes of the three nodes (hand, foot, and knee) to the same master clock axis. If time synchronization is inaccurate, there will be a time difference between the hands and feet. Phase difference can be contaminated by systematic errors, which may result in a lower reading even if the patient's gait is normal. This can lead to the risk of false positives. Therefore, by calculating the arm swing amplitude-step length coupling coefficient... and torso rotation angle To quantify the coordination of upper and lower limb movements; Among them, the master clock axis The calculation formula is: ; In the formula, For the most recent master timestamp received by the insole and knee ring, For insoles and knee rings, local sampling time, For insoles and knee rings, receive the local time of the master timestamp. Clock drift factor for each slave node.

[0045] The units of measurement are seconds (s) or milliseconds (ms). and The units of measurement are both seconds (s) or milliseconds (ms), therefore, their difference The units of measurement are seconds (s) or milliseconds (ms). Dimensionless, therefore Its units are seconds (s) or milliseconds (ms).

[0046] The arm swing amplitude-step coupling coefficient The calculation formula is: ; In the formula, This is a reference ratio for healthy individuals; The arm swing amplitude during one gait cycle, measured in meters (m) or millimeters (mm), stride length. The units of measurement are meters (m) or millimeters (mm), and the ratio between the two is... Dimensionless Dimensionless, therefore, Dimensionless.

[0047] Torso rotation angle The calculation formula is: ; In the formula, This refers to the yaw angle weighting coefficient of the inertial wristband. The inertial knee roll angle weighting coefficient is used. This represents the change in yaw angle of the inertial wristband. This represents the change in the inertial knee ring roll angle.

[0048] The observation equation is expressed as: ; In the formula, The observation vectors for the inertial wristband, inertial knee ring, and flexible plantar pressure insole contain raw measurements such as acceleration / angular velocity or plantar pressure, respectively. The observation matrix for the inertial wristband, inertial knee ring, and flexible pressure insole is responsible for converting the state vector. Projected onto their respective observation spaces, The observation noise of the inertial wristband, inertial knee ring, and flexible pressure insole is assumed to be zero-mean Gaussian noise, and its covariance is determined by the accuracy of each sensor.

[0049] Using human kinematics chain constraints on the state vector Constraints are applied to reduce the estimated degrees of freedom, thereby suppressing noise. The constraint conditions are as follows: ; ; In the formula, For individualized lower leg segment length, This indicates the equivalent torso length from the knee ring to the wrist ring, both derived from human measurements or calibrations, used to suppress spatial drift and maintain biomechanical consistency.

[0050] The state vector is recursively derived using the Extended Kalman Filter (EKF), and the update equation for the state vector is: ; In the formula, for The state vector at time t, Here is the Kalman gain matrix. For the observation vector, This is the observation matrix.

[0051] When a healthy person walks, the arm swing is antiphase coupled with the contralateral lower limb (the right arm swings forward corresponding to the left leg stepping forward), and momentum is transferred through trunk rotation; this invention introduces the hand-foot phase difference. To quantify the temporal coordination of these upper and lower limb movements. Parkinson's patients often experience upper limb rigidity, which can lead to... Abnormal fluctuations may occur, and in stroke patients with hemiplegia, the reduced swing amplitude of the affected arm can lead to... The regularity is disrupted.

[0052] The hand and foot time phase difference The calculation formula is: ; In the formula, This is the peak moment of the inertial wristband's angular velocity. The moment when the ipsilateral plantar pressure drops to 10%; The neuro-coordination score The calculation formula is: ; In the formula, , These represent the mean and standard deviation of hand-foot distance differences in healthy individuals. , , This represents the coordinated phase difference between the wrist and knee joints. This represents the coordinated phase difference between the knee joint and the ankle joint. The units of measurement are seconds (s) or milliseconds (ms). The unit of measurement is seconds (s) or milliseconds (ms), therefore, Dimensionless Dimensionless Dimensionless, therefore, Dimensionless; and The calculation formula is: ; ; In the formula, For the phase of the arm's angular velocity, Knee flexion-extension phase, For the soles of the feet Phase, healthy people are satisfied (Anti-coupling).

[0053] The multimodal collaborative feature vector Represented as: ; in: ; ; ; In the formula, The reaction force is perpendicular to the ground. The trajectory centered on.

[0054] The verification includes: When the inertial wristband detects peak torso acceleration satisfy At the same time Changes satisfy hour, If the change in vertical ground reaction force is detected, the data from the flexible pressure insole is deemed abnormal. In this case, the system automatically switches to the IMU-dominated mode of the inertial wristband and estimates the stride frequency using the arm swing frequency. ,accomplish Verification; When the plantar pressure cycle stabilizes after 5 consecutive steps, but the inertial wristband attitude integral displacement error... At 10cm, the zero bias of the inertial wristband IMU is corrected by using the foot step length in the reverse direction. ,accomplish Verification; When step size and step frequency are consistent At that time, calculate the correction step size. Joint angles with soft constraints within the physiological range Initiate bidirectional correction to achieve and Verification.

[0055] The zero bias of the wristband IMU The calculation formula is: ; In the formula, This represents the displacement increment based on inertial measurement unit (IMU) data. This represents the displacement increment based on plantar pressure data. This is an empirical scaling factor used to convert displacement differences into an estimation bias for the wristband; The units of measurement are meters (m) or millimeters (mm). The units of measurement are meters (m) or millimeters (mm). Dimensionless, therefore, The units of measurement are meters (m) or millimeters (mm).

[0056] The step size-step frequency consistency is expressed as: ; In the formula, For speed based on insole pressure, The speed is based on the wristband's acceleration. The step size-step frequency consistency is the ratio of speed, therefore it is dimensionless.

[0057] in, The calculation formula is: ; The calculation formula is: ; The correction step size The calculation formula is: ; In the formula, For adaptive fusion weights; Dimensionless The units are meters (m) or millimeters (mm). The unit of measurement is meter (m) or millimeter (mm), therefore, Dimensions and Consistent, in meters (m) or millimeters (mm).

[0058] in, The calculation formula is: ; In the formula, Confidence level for flexible pressure insoles on the sole of the foot. For the confidence level of the inertial wristband; The joint angles under soft constraint within the physiological range The calculation formula is: ; In the formula, The original joint angle, This represents the physiological limit angle of the joint. It is a sigmoid function. This is the scaling factor; and All are joint angles, and the physical units are degrees (°) or radians (rad). It is a sigmoid function, dimensionless. The scaling factor is the unit of measurement. Consistent, therefore, The units are degrees (°) or radians (rad).

[0059] Step 3, based on the verified Calculating attention fusion features using an attention mechanism ,Will Input a bidirectional LSTM-CNN network and output the probability of gait risk for Parkinson's disease. Risk of gait abnormalities as a sequela of stroke and the probability of falling Based on , and The credibility was determined after calculating the SHAP values ​​of the contributions of the three modalities of hand, knee and foot respectively.

[0060] The attention fusion feature The calculation formula is: ; In the formula For attention mechanism queries, For the key of attention mechanism The value matrix for the attention mechanism, For the normalized attention weight matrix, Let be the scaling factor, where The feature dimension of the key vector. Used to suppress gradient instability caused by excessively large inner product scale; , Dimensionless, therefore, Dimensionless.

[0061] Among them, the query of attention mechanism Key to attention mechanisms Value matrix of attention mechanism The calculation formula is: ; In the formula, This refers to the modal splicing features of a single gait cycle, i.e., a multimodal collaborative feature vector. The projection matrix maps the input features to the attention space dimension.

[0062] The probability of gait risk in Parkinson's disease Risk of gait abnormalities as a sequela of stroke and the probability of falling The calculation formula is: ; In the formula, , Representing different risk prediction tasks, The deep feature vector output by the bidirectional LSTM-CNN network is... This is the weight matrix of the fully connected layer for the corresponding task. This refers to the bias term for the corresponding task. The function is used to map the network output to The probability value of the interval.

[0063] The SHAP value of the modal contribution of the hand, knee, and foot. The calculation formula is: ; In the formula, For modal types, , The total number of modes, The risk score output by the network. This represents any combination of modes other than the current mode. The number of modes contained in the combination. Indicates that only a subset is input. The risk score predicted by the model at that time This represents the output after adding the current modal feature to the same combination, with coefficients... The average of all modal permutations is taken so that SHAP measures the average marginal contribution of the current mode across all “cooperation sequences”. Dimensionless, therefore, Dimensionless, coefficient Dimensionless, therefore, Dimensionless; The larger the value, the more critical the current modality is in the risk assessment of the current sample; a negative value means that the current feature plays a dominant role in reducing risk prediction. (Hand, foot, and knee modalities) Together, we can provide doctors with interpretable evidence on "which modalities drive their output conclusions".

[0064] Bidirectional LSTM-CNN networks employ multi-task joint loss The multi-task joint loss makes the three modal features complementary. The calculation formula is: ; In the formula, Cross-entropy loss predicted for Parkinson's disease. Cross-entropy loss for stroke prediction For the cross-entropy loss of fall prediction, For feature fusion regularization term, For Parkinson's task weights, For stroke task weights, For the task weight of falling, To incorporate regularization weights; , , and Dimensionless , , and Dimensionless, therefore, Dimensionless.

[0065] The specific method for determining credibility is as follows: If the high-risk conclusion for a suspected patient is primarily driven by modalities consistent with clinical symptoms (such as wristbands used to indicate high risk in Parkinson's patients) (If the value is positive and significantly greater than other modalities), doctors can consider the model to be based on reasonable grounds and the conclusions to be credible. If the system predicts a high risk, but The dominant modality is inconsistent with clinical observation (e.g., insoles). (The actual plantar data quality is poor, which may indicate that the data is extremely high. Doctors can use this to question the results and trace back the sensor status or feature calculations of that modality to avoid misjudgment.) For the same test, doctors can compare the three modalities. The absolute value and sign of the data are used to confirm whether the risk result is "increased / decreased" by a single modality, thereby determining whether additional data collection or manual review is required.

[0066] Based on the above The interpretability module is validated so that when the assessment conclusion is confirmed to be credible and the risk level is ≥ "Medium", the system automatically initiates a closed-loop intervention process. First, it automatically generates rehabilitation assessment reports in PDF and HL7 FHIR standard formats, and pushes personalized home protection guidelines and warnings through the software on both the doctor and patient sides; the user side has a built-in AI intelligent consultant module, configured to answer patients' common questions about fall risk and training intensity, and for complex questions, it will transfer to the doctor's side for professional answers through a dialogue window; Subsequently, the doctor pushes a personalized rehabilitation prescription based on the patient's specific condition; the system prompts the patient's training through center of gravity shift animation and flexion and extension rhythm, and uses wristband data to estimate fatigue in real time to automatically adjust the training intensity, realizing remote home rehabilitation; After training, the device collects data a second time, recalculates rehabilitation indicators and compares them with rehabilitation thresholds, and feeds the data back to the physician to dynamically update the next cycle plan. All the above records are uploaded to the cloud for backup, supporting historical curve playback and remote consultation.

[0067] Combination Figure 5 Using a fall risk assessment example (baseline risk 0.15, predicted risk 0.75) to illustrate: flexible pressure insoles on the soles of the feet +0.45 (the largest percentage in the chart) objectively indicates that abnormal plantar control is the primary risk factor; inertia wristband A value of +0.25 reflects the positive risk of increased upper limb compensatory decline; inertial knee ring A value of -0.10 (the smallest percentage in the graph) indicates that the knee condition is good and has a negative compensatory effect on risk. Through Figure 5 each mode and These symbols allow doctors to quickly identify the dominant characteristic sequence that triggers risk. This quantitative evidence breaks the "black box" nature of AI early warning systems, providing precise data support for subsequent targeted rehabilitation guidance programs and achieving a closed loop of "evidence-based intervention."

[0068] Example 2 A gait analysis system based on hand-knee-foot coordinated motion coupling includes: a sensor group, a data processing module, and a user terminal; Reference Figure 2 , Figure 3 , Figure 4 The sensor group includes a 256-point piezoresistive array embedded in the flexible pressure insole 1, shear force film strips disposed at the front and heel of the piezoresistive array, a nine-axis IMU integrated in the inertial knee ring 2, and a nine-axis IMU embedded in the inertial wristband 3; the sensor group is equipped with a wireless communication module for collecting multimodal gait detection data of the hand, knee, and foot; the data processing module includes a data collaboration layer, a feature collaboration layer, and a decision collaboration layer. The data collaboration layer is configured to receive multimodal gait detection data collected by the sensor group, unify the time axes of the hands, knees, and feet onto the main clock axis, and unify the posture and position into the human coordinate system.

[0069] The feature collaboration layer is configured to receive the unified spatiotemporal state data, extract multimodal collaborative features based on the human gait biomechanical coupling mechanism, calculate stride length-step frequency consistency and compare it with a threshold. When the threshold is exceeded, bidirectional correction is initiated, and soft constraint verification of joint angles is performed based on physiological limit angles. The fusion weights are dynamically adjusted according to sensor confidence, and bidirectional complementary correction of data from the inertial wristband 3 and the flexible pressure insole 1 is performed to form verified multimodal collaborative features. The decision collaboration layer is configured to receive the verified multimodal collaborative feature vectors, calculate the weights of each modality using the built-in attention mechanism fusion module, and generate attention fusion features. The bidirectional LSTM-CNN network of the decision collaboration layer receives the attention fusion features, outputs the gait risk probabilities of Parkinson's disease, post-stroke sequelae, and falls, and calculates the contribution of each modality. It is wirelessly transmitted to the user's device to generate a visual gait assessment report, and personalized rehabilitation guidance plans and real-time warnings are pushed according to the risk level.

[0070] The flexible pressure insole 1 is composed of three flexible layers: the upper surface is a 0.2mm anti-slip PU film to ensure that the foot surface does not slip; the middle layer is a 256-point piezoresistive array (8×32 grid), with a single-point range of 0–800kPa and a density of 4–5 points / cm²; a shear force film strip (range ±200N, thickness 0.3mm) is added to the front end and the heel of the piezoresistive array, sharing a flexible FPC substrate with the piezoresistive array to form a three-dimensional force field of "vertical + front-back + inside-outside"; the lower surface is a 0.5mm TPE buffer layer, with a 5mm pre-cut edge at the front end to fit shoes of sizes 36–46.

[0071] The flexible pressure insole 1 has a built-in nRF52832 microcontroller and a 12-bit ADC that scans all 256 points at 100Hz. It also reads the differential signals of two shear force bars in parallel. The arch area is equipped with a 400mAh lithium thin-film battery with a thickness of ≤1mm. It can be magnetically charged to 80% in 30 minutes and can work continuously for ≥6 hours. The overall weight is ≤25g. It can be fixed by elastic bandages or silicone anti-slip strips on the insole and can be used as soon as you put on your shoes.

[0072] The local coordinates of the flexible pressure insole 1 are: X-axis pointing to the toe, Y-axis pointing to the outside of the foot, and Z-axis vertically upward; all force values ​​( , , It can be directly mapped to this coordinate system, which facilitates subsequent alignment with the gait data of the inertial knee ring 2 and the inertial wrist ring 3.

[0073] Regarding triggering and period determination, the microcontroller (MCU) continuously calculates the total vertical force on the sole of the foot. For adult mode, the total vertical force threshold for the sole of the foot is set to 30kg. When the MCU pulls high on the GPIO (General Purpose Input / Output Interface), it outputs a "STEP_START" signal; for child mode, the threshold is set to 10kg. This threshold can be dynamically modified via a BLE instruction packet (4 bytes) from the user terminal (mobile phone). In cycle calculation, when... Record the time when the threshold is first exceeded. (Foot touches the ground); record the time when the pressure in the same area drops below 5 kg again. (Foot off the ground). The system defines the complete gait cycle. Subsequently, the inertial knee ring 2 and inertial wrist ring 3 only collect acceleration and angular velocity within this cycle to avoid accidental triggering in sitting or lying positions.

[0074] In terms of wireless communication and data transmission, the flexible pressure insole 1 establishes a bidirectional connection with the user terminal via a wireless communication module (Bluetooth BLE). On one hand, the user terminal can send control commands: when sending the "THRESH_SET" command (carrying a threshold and unit kg), the MCU immediately writes the parameters into the EEPROM (Electrically Erasable Programmable Read-Only Memory), without needing to re-pair; when sending the "DETECT_OFF" command (e.g., in fitness or dance scenarios), the MCU disables posture acquisition and only retains pressure polling; when sending the "DETECT_ON" command, the system immediately restarts. On the other hand, the wireless communication module packages the collected gait detection data (including pressure data and gait cycle signals) in real time and transmits it to the data processing module via Bluetooth for processing.

[0075] Reference Figure 3 The inertial knee ring 2 includes a knee ring outer shell 21, a knee ring sensing layer, and a knee ring fixing layer 22. The knee ring sensing layer is encapsulated inside the knee ring outer shell 21, and the knee ring outer shell 21 is fixedly disposed outside the knee ring fixing layer 22. The knee ring outer shell 21 is a 3D-printed PETG flexible shell with an adjustable inner diameter of 28-45mm, a thickness of 8mm, and a weight of ≤25g. The sensing layer is a "C"-shaped flexible circuit board, which includes a built-in nine-axis IMU (model BMI160). The nine-axis IMU is located at the center of the upper edge of the patella, with a range of ±16g (g=9.8m / s²) and ±2000° / s (angular velocity), and a sampling frequency of 100Hz. The outer side of the knee ring fixing layer 22 is a 50mm wide knee ring elastic bandage 23, and the inner side is a silicone anti-slip strip + Velcro. The wearing pressure is controlled by the bandage. 2.5kPa, no slippage after 20 minutes of running. The knee ring shell 21 has a built-in 400mAh lithium thin-film battery at the rear edge, with a thickness of 7mm. It can be magnetically charged to 80% in 30 minutes via 2-pin charging and can work continuously for ≥6 hours. The overall encapsulation is IP54, and it can work normally in sweat and light rain. It weighs 25g. The elastic bandage 23 of the knee ring is machine washable, and the shell can be wiped with alcohol for disinfection.

[0076] The local coordinates of the inertial knee ring 2 are centered at the center of knee joint rotation, with the X-axis pointing proximally along the femur, the Y-axis pointing laterally, and the Z-axis pointing vertically upwards. The microcontroller calculates and outputs the three-dimensional coordinates of the hip, knee, and ankle, which are then broadcast via BLE 100Hz. The timestamps are aligned with the flexible pressure insole 1 and the inertial wrist ring 3, with an error margin of [missing information]. Data synchronization is achieved in 3ms.

[0077] The inertial knee ring 2 is configured to work in conjunction with the user terminal and the flexible pressure insole 1 on the sole of the foot, specifically including: In terms of collaborative data acquisition, upon receiving the "STEP_START" signal from the flexible pressure insole 1, the MCU of the inertial knee ring 2 immediately starts continuous IMU sampling and synchronizes the sampling period with the gait period of the flexible pressure insole 1. It maintains synchronization, outputs valid frames only during this cycle, and enters a sleep state with a power consumption of only 1mA for the rest of the time.

[0078] In terms of command interaction, the inertial knee ring 2 is configured to receive control commands from the user terminal: according to the “CAM_RES” command sent by the user terminal, it switches the frame rate between 480p and 240p to balance power saving and accuracy; and according to the “IMU_RANGE” command, it dynamically switches the acceleration range to adapt to different sports scenarios such as walking, jogging and going up and down stairs.

[0079] Reference Figure 4 The inertial wristband 3 consists of a double-layer structure of "flexible strap + rigid core board", which takes into account both wearing comfort and device protection. It includes a wristband shell 31, a wristband sensor layer, and a wristband fixing layer 32. The wristband sensor layer is encapsulated inside the wristband shell 31, and the wristband shell 31 is fixedly connected to the wristband fixing layer 32. The outer shell 31 of the wristband is a 3D-printed TPU soft rubber watchband shell, with a total length of 230mm, width of 20mm, thickness of 10mm, and weight ≤20g. The wristband's sensing layer is designed with a nine-axis IMU (BMI160) located on the dorsal side of the wrist bone, with a range of ±8g (acceleration) and ±1000° / s (angular velocity), and a sampling frequency of 100Hz. The outer layer of the wristband fixing layer 32 is an elastic bandage watchband made of medical-grade TPU 55° soft rubber, with a length of 230mm, width of 20mm, and thickness of 2mm. The inner side has an integrally molded 0.5mm thick silicone anti-slip strip to ensure that it will not slip after running for 20 minutes. It is fixed with Velcro, which has a 50mm×20mm hook and loop combination, and the circumference can be continuously adjusted from 140 to 200mm to fit the wrists of children to adults. It is also equipped with a POM plastic buckle for easy wearing and removal.

[0080] The local coordinates of the inertial wristband 3 are based on the wrist joint rotation center as the origin: the X-axis points along the back of the hand towards the fingers, the Y-axis points towards the ulnar side, and the Z-axis is perpendicular to the back of the hand and upwards. After obtaining the data, the IMU outputs 16-bit raw values. The data shares BLE timestamps with the insole and knee band, and the error... 3ms.

[0081] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A gait analysis method based on hand-knee-foot coordinated motion coupling, characterized in that, Includes the following steps: Step 1: Collect gait cycles using flexible pressure insoles on the soles of the feet. Step frequency Step length Three-dimensional position of the foot Foot and ankle joint posture quaternions and shear forces in the anteroposterior and medial directions of the foot ; Inertial knee ring collects acceleration at the knee joint angular velocity Knee phase Knee joint posture quaternion Three-dimensional position of the knee and joint angle The inertial wristband collects the acceleration at the wrist. angular velocity wrist posture quaternion Three-dimensional position of the wrist Arm swing amplitude within one gait cycle and peak angular velocity ; Step 2: Unify the time axes of the three nodes (hand, foot, and knee), and unify the pose and position of the three nodes into the human coordinate system to establish a state vector. and observation equations, through observation equations to Update and calculate hand-foot time difference. and neuro-coordination score Insole features are constructed by combining the data collected in step 1. Knee ring characteristics and wristband features ,Will , and The merging yields a multimodal collaborative feature vector. And verify; Step 3, based on the verified Computational attention fusion features ,Will Input a bidirectional LSTM-CNN network and output the probability of gait risk for Parkinson's disease. Risk of gait abnormalities as a sequela of stroke and the probability of falling Based on , and Calculate the SHAP values ​​of the contributions of the three modalities of hand, knee and foot respectively, and judge their reliability.

2. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 1, The calculation formula is: ； In the formula, The moment of liftoff. The moment of contact with the ground; Total vertical force on the sole of the foot The moment when the trigger threshold is first exceeded Within the same area When the time is below the cutoff threshold, The calculation formula is: ； In the formula, for The reaction force is always perpendicular to the ground. Foot pressure values ​​collected in real time by a piezoresistive array Calculations show that The calculation formula is: ； In the formula, This represents the area of ​​a single piezoresistive point. The plantar pressure value is denoted as , where This indicates the longitudinal position of the piezoresistive array. This indicates the lateral position of the piezoresistive array. Sampling time; The calculation formula is: ； The step size Through the center trajectory of the two steps before and after It was estimated; in, Represented as: ； In the formula, Let be the lateral coordinate of the center of pressure in the insole coordinate system. Let be the longitudinal coordinate of the center of pressure in the insole coordinate system; and The calculation formula is: ； In the formula, The coordinates of the pressure resistance point are shown in the coordinate system of the flexible pressure insole on the sole of the foot.

3. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 2, the inertial wristband is used as the master node, and the flexible pressure insole and inertial knee ring are used as slave nodes, unifying the time axes of the three nodes (hand, foot, and knee) to the same master clock axis. By calculating the arm swing amplitude-step length coupling coefficient and torso rotation angle Quantifying the coordination of upper and lower limb movements; Among them, the master clock axis The calculation formula is: ； In the formula, The most recent master timestamp received by the flexible pressure insole and inertial knee ring. For the flexible pressure insoles and inertial knee rings, at the local sampling time, For flexible pressure insoles and inertial knee rings, the local time of the master timestamp is received. Clock drift factor for each slave node; Arm swing amplitude-step coupling coefficient The calculation formula is: ； In the formula, This is a reference ratio for healthy individuals; Torso rotation angle The calculation formula is: ； In the formula, This refers to the yaw angle weighting coefficient of the inertial wristband. The inertial knee roll angle weighting coefficient is used. This represents the change in yaw angle of the inertial wristband. This represents the change in inertial knee ring roll angle; The observation equation is expressed as: ； In the formula, For the observation vectors of the inertial wristband, inertial kneeband, and flexible pressure insole, The observation matrix is ​​for the inertial wristband, inertial knee ring, and flexible pressure insole. The noise levels observed were for inertial wristbands, inertial kneebands, and flexible pressure insoles on the soles of the feet.

4. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 2, the update equation for the state vector is: ； In the formula, for The state vector at time t, for The state vector at time t, Here is the Kalman gain matrix. Observation vector, The observation matrix; The hand and foot time phase difference The calculation formula is: ； In the formula, This is the peak moment of the inertial wristband's angular velocity. The moment when the ipsilateral plantar pressure drops to 10%; The neuro-coordination score The calculation formula is: ； In the formula, , These represent the mean and standard deviation of hand-foot distance differences in healthy individuals. This represents the coordinated phase difference between the hand joint and the knee joint. This represents the coordinated phase difference between the knee joint and the ankle joint. The multimodal collaborative feature vector Represented as: ； in: ； ； ； In the formula, The reaction force is perpendicular to the ground. The trajectory centered on.

5. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 2, The verification includes: When the inertial wristband detects peak torso acceleration satisfy At the same time Changes satisfy hour, If the change in vertical ground reaction force is detected, the data from the flexible pressure insole is deemed abnormal. In this case, the system automatically switches to the IMU-dominated mode of the inertial wristband and estimates the stride frequency using the arm swing frequency. ,accomplish Verification; When the plantar pressure cycle stabilizes after 5 consecutive steps, but the inertial wristband attitude integral displacement error... At 10cm, the zero bias of the inertial wristband IMU is corrected by using the foot step length in the reverse direction. ,accomplish Verification; When step size and step frequency are consistent At that time, calculate the correction step size. Joint angles with soft constraints within the physiological range Initiate bidirectional correction to achieve and Verification.

6. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 5, characterized in that, The zero bias of the inertial wristband IMU The calculation formula is: ； In the formula, For the displacement increment based on inertial wristband measurement data, The displacement increment is based on plantar pressure data. This is an empirical scaling factor; The step size-step frequency consistency is expressed as: ； In the formula, For speed based on insole pressure, The speed is based on the wristband's acceleration. in, The calculation formula is: ； The calculation formula is: ； The correction step size The calculation formula is: ； In the formula, For adaptive fusion weights; in, The calculation formula is: ； In the formula, Confidence level for flexible pressure insoles on the sole of the foot. For the confidence level of the inertial wristband; The joint angles under soft constraint within the physiological range The calculation formula is: ； In the formula, The original joint angle, This represents the physiological limit angle of the joint. It is a sigmoid function. This is the scaling factor.

7. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 3, the attention fusion feature The calculation formula is: ； In the formula, For attention mechanism queries, For the key of attention mechanism The value matrix for the attention mechanism, To normalize the attention weight matrix, Let be the scaling factor, where The feature dimension of the key vector; Among them, the query of attention mechanism The key to attention mechanisms The value matrix of the attention mechanism The calculation formula is: ； In the formula, The projection matrix; The probability of gait risk in Parkinson's disease Risk of gait abnormalities as a sequela of stroke and the probability of falling The calculation formula is: ； In the formula, , Representing different risk prediction tasks, The deep feature vector output by the bidirectional LSTM-CNN network is... This is the weight matrix of the fully connected layer for the corresponding task. This refers to the bias term for the corresponding task. The function is used to map the network output to The probability value of the interval; The SHAP values ​​of the contributions of the hand, knee, and foot modalities The calculation formula is: ； In the formula, For modal types, , For the total number of modes, The risk score output by the network. This represents any combination of modes other than the current mode. The number of modes contained in the combination. Indicates that only a subset is input. The risk score predicted by the model at that time This represents the output after adding the current modal feature to the same combination. These are the combined weighting coefficients.

8. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 3, the bidirectional LSTM-CNN network uses a multi-task joint loss. To make the three modal features complementary and then calculate the gait risk probability of Parkinson's disease. Risk of gait abnormalities as a sequela of stroke and the probability of falling ; Among them, multi-task joint loss The calculation formula is: ； In the formula, Cross-entropy loss predicted for Parkinson's disease. Cross-entropy loss for stroke prediction For the cross-entropy loss of fall prediction, For feature fusion regularization term, For Parkinson's task weights, For stroke task weights, For the task weight of falling, To incorporate regularization weights.

9. The gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, In step 3 The larger the value, the more critical the current modality feature is in the risk assessment of the current sample; if it is negative, the current modality feature plays a dominant role in reducing risk prediction. The credibility is judged as follows: if the high-risk conclusion of a suspected patient is mainly driven by a modality consistent with clinical symptoms, then the model is considered to be reasonable and the conclusion is credible. If the system predicts a high risk, but If the dominant modality does not match clinical observation, the results should be questioned and the sensor status or feature calculations of that modality should be traced back to avoid misjudgment. For the same test, comparisons were made between the hand, knee, and foot modalities. The absolute value and sign of the data are used to confirm whether the risk outcome is raised or lowered by a single modality, thereby determining whether additional data collection or manual review is required.

10. A system for implementing the gait analysis method based on hand-knee-foot coordinated motion coupling as described in claim 1, characterized in that, include: Sensor array, data processing module, and user terminal; The sensor group is used to collect multimodal gait detection data of the hand, knee and foot. The sensor group includes a piezoresistive array built into the flexible pressure insole of the foot, a shear force film strip set at the front end and heel of the piezoresistive array, an IMU integrated into the inertial knee ring and an IMU built into the inertial wrist ring. The data processing module includes a data collaboration layer, a feature collaboration layer, and a decision collaboration layer; The data collaboration layer receives multimodal gait detection data collected by the sensor group, unifies the time axes of the hands, knees, and feet to the main clock axis, unifies the posture and position to the human coordinate system, establishes state vectors and observation equations, updates the state vectors, dynamically adjusts the fusion weights according to the sensor confidence, and performs bidirectional complementary correction between the inertial wristband and the flexible pressure insole data. The feature collaboration layer receives the updated state vector, extracts multimodal collaborative features based on the human gait biomechanical coupling mechanism, calculates the stride length-step frequency consistency and compares it with a threshold. When the threshold is exceeded, bidirectional correction is initiated, and soft constraint verification of joint angles is performed based on physiological limit angles to form a verified multimodal collaborative feature vector. The decision collaboration layer receives the validated multimodal collaborative feature vector, calculates the weights of each modality using the built-in attention mechanism fusion module, and generates attention fusion features. The bidirectional LSTM-CNN network of the decision collaboration layer receives the attention fusion features and outputs the gait risk probabilities of Parkinson's disease, post-stroke sequelae, and falls, as well as the contribution of each modality. The data is transmitted to the user terminal to generate a visual gait assessment report. Based on the risk level, personalized rehabilitation guidance plans and real-time warnings are pushed. The user terminal is wirelessly connected to the flexible pressure insole and sends control commands or predetermined thresholds to the flexible pressure insole in real time. The flexible pressure insole simultaneously sends "collection trigger" commands to the inertial knee ring and inertial wrist ring.

Images