A health correlation analysis method based on foot pressure and gait dynamic monitoring

Abstract

The application relates to a health correlation analysis method based on foot pressure and gait dynamic monitoring, which collects original plantar pressure data; adopts double-reference adaptive filtering according to treadmill vibration and trunk acceleration to perform vibration suppression processing on the original plantar pressure data to obtain expected plantar pressure data; divides a gait cycle for the expected plantar pressure data; extracts gait features from the data after gait cycle division; constructs and trains a causal inference model; the causal inference model adopts a three-layer directed graph model, including an observation layer formed by gait feature nodes, a latent variable layer formed by health correlation feature nodes, and an output layer formed by health risk item nodes; causal edges are set between nodes at various levels according to causal constraints; during reasoning, the trained causal inference model is used to establish the output of gait features to health correlation features and health correlation features to risk items; and health risk analysis is realized according to gait features.

Description

Technical Field

[0001] This invention relates to the field of health risk monitoring technology, and in particular to a health correlation analysis method based on foot pressure and gait dynamic monitoring. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] When exercising on a treadmill, plantar pressure reflects the stress on the foot and the characteristics of the movement. As the most popular form of aerobic exercise, running carries a significant risk of injury. Most runners experience at least one running-related injury during their careers, including knee pain, plantar fasciitis, and tibial stress syndrome. Numerous studies have shown that abnormal gait patterns, such as excessive pronation, asymmetrical ground contact time, and unbalanced propulsive force distribution, are key contributing factors to these injuries. The spatiotemporal distribution characteristics of plantar pressure are one of the most direct and sensitive physiological signals reflecting gait abnormalities.

[0004] Traditional treadmills generally lack the ability to objectively quantify the quality of a user's gait, providing only macroscopic parameters such as speed, incline, and heart rate. They cannot detect microscopic changes in the biomechanical behavior of the user's lower limbs, making it difficult to achieve personalized injury warnings. In recent years, some treadmills have begun to integrate pressure sensor pads, but their functions are limited to static balance testing or simple pressure heat map display, and they suffer from the following technical bottlenecks: Force measurement using pressure sensor pads placed on the running belt results in sensor points that are not fixed relative to the foot position, leading to poor ability to perceive the force applied to the feet; moreover, most neglect the temporal evolution characteristics, spatial dynamic distribution, and coupling relationship with kinematic parameters of pressure signals, resulting in weak data feature representation capabilities; and they lack an effective personalized modeling mechanism based on mapping data features to risk, leading to a high false alarm rate. Summary of the Invention

[0005] To solve the above-mentioned technical problems, or at least partially solve them, the present invention provides a health correlation analysis method based on foot pressure and gait dynamic monitoring.

[0006] This invention provides a health correlation analysis method based on foot pressure and gait dynamic monitoring, comprising: Raw plantar pressure data of the human body during treadmill exercise is collected using a flexible thin-film pressure sensor array installed on the heel, toes, and sides of the midfoot of the insole. The raw plantar pressure data is subjected to vibration suppression processing using dual-reference adaptive filtering based on treadmill vibration and trunk acceleration to obtain the desired plantar pressure data. The acquired desired plantar pressure data is then divided into gait cycles. Gait features are extracted from the data after gait cycle segmentation. The gait features include: temporal amplitude features, spatiotemporal distribution features, dynamic mechanical features, symmetry and variability features, frequency domain features, and associated health features. A causal inference model is constructed and trained. The causal inference model adopts a three-layer directed graph model. The first layer uses gait features as observation layer nodes, the second layer uses health-related features as latent variable layer nodes, and the third layer uses health risk items as output layer nodes. The causal edges between nodes at each level are constructed based on biomechanical and clinical knowledge, and the strength weights of the causal edges are learned according to Bayesian linear regression. During inference, the trained causal inference model is used to establish the output from gait features to health-related features, and from health-related features to risk item probabilities. Adjust your exercise method based on the probability of risk factors.

[0007] Furthermore, based on treadmill vibration and torso acceleration, adaptive filtering is used to perform vibration suppression processing on the raw plantar pressure data to obtain the desired plantar pressure data, including: At time n, the original plantar pressure data are represented as follows: ; in, The desired plantar pressure data is related to gait. The noise is caused by the vibration of the treadmill, is related to the treadmill motor, and is periodic or quasi-periodic. White noise; adaptive filtering is used to estimate the noise caused by treadmill vibration. The expected plantar pressure data estimate is obtained. ; Vibration sensors are used to collect treadmill vibrations as the first reference signal. The first reference signal and the noise caused by the treadmill vibration Strong correlation; The vertical acceleration component of the user's torso is collected using an inertial measurement unit as a second reference signal. Second reference signal The spectrum contains harmonic components with the same frequency as the treadmill vibration, and there is a phase coupling relationship with the foot vibration interference; First reference signal Second reference signal Combined to form a reference vector ; Noise caused by treadmill vibration is estimated using an FIR-type multi-channel transverse filter based on reference vector estimation. ,as follows: ; Where L is the filter order covering the dominant vibration frequency period; Let be the filter weight vector of the k-th tap; The error is obtained by subtracting the output of the FIR multi-channel lateral filter from the raw plantar pressure data: , in, This refers to residual noise and white noise caused by treadmill vibration; To minimize mean square error To achieve this, LMS gradient descent is used to update the weights of the FIR-type multichannel transverse filter according to the error and the reference vector: ; in, Step size; The noise caused by treadmill vibration is estimated based on the reference vector using an FIR-type multichannel transverse filter that satisfies the condition of minimizing the mean square error. Original plantar pressure data minus Obtain the desired plantar pressure data.

[0008] Furthermore, the gait periodization of the acquired expected plantar pressure data includes: Ground contact is determined when the weighted result of the pressure gradient in the heel area and the velocity norm of the center of pressure on the plantar surface is greater than a set adaptive threshold. The system detects whether the abrupt change in toe pressure gradient from the midpoint to the end of the gait exceeds a dynamic threshold to determine whether push-off has occurred. The gait phase sequence segmented according to the above judgment criteria is smoothed and corrected by a pre-trained gait segmentation hidden Markov model.

[0009] Furthermore, the gait segmentation Hidden Markov Model divides the gait phase into a swing phase, a ground contact phase, and a support phase, which transition in a predetermined order. A state transition probability matrix is ​​initialized, where each element represents the probability of transitioning from one state to another. Due to the sequential nature of the gait phases, the transition probability between some states is 0. An observation space is defined, including total pressure, heel pressure gradient, second derivative of pressure at the toes, velocity at the center of plantar pressure, and position of the center of plantar pressure. An observation probability matrix is ​​initialized according to a Gaussian distribution, representing the probability of observing a certain observation value in a given state. An initial state probability vector is initialized, representing the probability of being in each state at the start of the gait. The ground contact initiation is determined when the weighted result of the pressure gradient in the heel area and the velocity norm of the center of plantar pressure is greater than the set adaptive threshold; whether the toe pressure gradient change in the interval from the midpoint to the end of the gait exceeds the dynamic threshold is detected to determine whether to push off; after initially dividing the gait cycle into the swing phase, ground contact phase and support phase, the observation vector of each time point is extracted. The observation vector sequence is input into the gait segmentation hidden Markov model, and the optimal gait phase sequence is decoded using the Viterbi algorithm. Based on the optimal gait phase sequence, the ground contact time and push-off time are redefined: the transition point from the swing phase to the ground contact phase is the ground contact time, and the transition point from the support phase to the swing phase is the push-off time.

[0010] Furthermore, the parameters of the gait segmentation hidden Markov model are learned through training data, including: Collect expected plantar pressure data for multiple gait cycles and label the gait phase corresponding to each time point in the expected plantar pressure data, including the swing phase, ground contact phase, and stance phase; For each time point, observation vectors are extracted from the training data, including: total pressure, heel pressure gradient, second derivative of pressure at the toes, velocity at the center of plantar pressure, and location of the center of plantar pressure. The probability of the initial state is estimated by statistically analyzing the probability of each gait phase in the first step of the training data. The frequency of transitions between arbitrary gait phases is obtained from the statistical training data, and then normalized to obtain the estimated state transition probability. For each gait phase, collect all observation vectors belonging to that gait phase, then calculate the mean and covariance matrix of the observation vectors, and determine the Gaussian distribution that determines the observation probability matrix.

[0011] Furthermore, the temporal amplitude characteristics include: peak pressure in each zone of the heel, medial midfoot, lateral midfoot, and toe area; average pressure and standard deviation of the entire foot during the stance phase; the first peak pressure generated by the heel at the initial contact stage and its rise time; the second peak pressure at the toes during the mid-to-late stance phase and its fall time, as well as the ratio of the two peak pressure values; and the propulsive impulse at the toes within a set time range before push-off. Spatiotemporal distribution characteristics include: The total trajectory length and curvature of the plantar pressure center trajectory in a gait cycle, the lateral offset of the plantar pressure center relative to the midline of the foot in each phase, the longitudinal migration rate of the plantar pressure center, and the pressure distribution entropy of each region of the foot in each phase. Dynamic mechanical characteristics include: estimated vertical ground reaction force, impact loading rate, and ankle joint moment; Symmetry and variability characteristics include: left and right foot difference index, interstep variability, and phase synchronization; Frequency domain characteristics include: pressure fast Fourier transform dominant frequency, harmonic energy ratio, and wavelet packet energy entropy; Related health characteristics include: Arch index: the ratio of midfoot pressure integral to total foot pressure integral; Pronation index: the percentage of the area of ​​the plantar pressure center trajectory that is offset towards the arch side to the entire plantar pressure center trajectory envelope area; Achilles tendon load index: peak heel pressure multiplied by ground contact time; Metatarsal stress index: peak pressure in the midfoot area divided by the square root of the contact area. All features are Z-score normalized based on the user's individual historical mean to eliminate dimensional differences and individual baseline drift.

[0012] Furthermore, the training process for a causal reasoning model includes: The expected plantar pressure dataset is categorized by individual users. A portion of the expected plantar pressure data is collected from each user. Experts provide latent variable layer health-related feature node values ​​based on gait characteristics of the observation layer. A Bayesian linear regression model from the observation layer to the latent variable layer is trained using the labeled data. The pre-trained Bayesian linear regression model from the observation layer to the latent variable layer is used to infer labels for the remaining data. Data with latent variable layer label confidence levels higher than a set threshold is selected from the remaining data to train the model. This process is iterated until the parameters of the Bayesian linear regression model from the observation layer to the latent variable layer are obtained. Using a portion of the data with output labels, where the output labels are binary injury events or risk scores given by experts, train a Bayesian linear regression model from the latent variable layer to the output layer; use the trained Bayesian linear regression model from the latent variable layer to the output layer to infer the output layer label for the remaining data; select data from the remaining data where the confidence level of the latent variable layer label is higher than a set threshold to train the model, iterating until the parameters of the Bayesian linear regression model from the latent variable layer to the output layer are obtained; In this process, Bayesian linear regression learns the weights of all causal edges. For any node, its parent node is determined by the causal edges associated with it, and the linear combination of its parent nodes plus noise is used as the value of that node.

[0013] Furthermore, a hierarchical Bayesian framework is adopted. First, at the group level, it is assumed that each user's data is independent and shares the same group weight, with the goal of maximizing the log-likelihood of the group data to estimate the group weight. A group weight is learned, and then at the individual level, an individual weight offset is learned to fine-tune the group weight.

[0014] Secondly, the present invention provides a health correlation analysis device based on foot pressure and gait dynamic monitoring, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit storing a computer program, and the processing unit implementing the health correlation analysis method based on foot pressure and gait dynamic monitoring by running the computer program stored in the storage unit.

[0015] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed, it implements the health correlation analysis method based on foot pressure and gait dynamic monitoring.

[0016] The technical solutions provided in the embodiments of the present invention have the following advantages compared with the prior art: This application utilizes insoles equipped with flexible thin-film pressure sensor arrays to collect raw plantar pressure data. Based on treadmill vibration and trunk acceleration, a dual-reference adaptive filter is employed to perform vibration suppression processing on the raw plantar pressure data to obtain the desired plantar pressure data. This significantly improves the signal-to-noise ratio, solves the long-standing problem of equipment vibration interfering with plantar signals in treadmill scenarios, and enhances the reliability of subsequent analysis.

[0017] This application performs initial gait cycle segmentation on the expected data based on the pressure gradient in the heel area and the velocity at the center of plantar pressure, as well as the abrupt change in the pressure gradient in the toe area from the midpoint to the end of the gait. Then, a pre-trained gait segmentation Hidden Markov Model is used to smooth and correct the initial gait cycle segmentation result sequence, thereby improving the robustness and accuracy of gait cycle segmentation. This ensures that key phases can be stably identified even when fatigued or experiencing gait abnormalities, guaranteeing the effectiveness of subsequent gait feature extraction.

[0018] This application extracts rich gait features with health risk significance from each gait cycle. Based on a constructed three-layer causal inference model, it realizes the analysis from gait features to health risks: the observation layer of the causal inference model is gait features; the latent variable layer is health-related features, which is a semantic abstraction of health risk from the gait features of the observation layer; and the output layer is the specific health risk item. The connections between the layers are not arbitrary, but strictly follow biomechanical and clinical knowledge. The causal inference model uses Bayesian linear regression to learn the weights of each causal edge, and through a semi-supervised iterative training strategy, it efficiently expands its applicability with a small amount of expert annotation. The final output is not only "risk level", but also includes a clear attribution chain, providing support for adjusting exercise patterns. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and constitute this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1A flowchart illustrating a health correlation analysis method based on foot pressure and gait dynamic monitoring, provided as an embodiment of the present invention; Figure 2 A schematic diagram showing the distribution of the insole and the flexible thin-film pressure sensor group thereon provided in an embodiment of the present invention; Figure 3 A flowchart for obtaining desired plantar pressure data by using dual-reference adaptive filtering to perform vibration suppression processing on raw plantar pressure data based on treadmill vibration and trunk acceleration, as provided in an embodiment of the present invention; Figure 4 An architectural diagram of the causal reasoning model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a health correlation analysis device based on foot pressure and gait dynamic monitoring provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 embodiments of the present invention, but 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.

[0023] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0024] Example 1 like Figure 1 As shown, a health correlation analysis method based on foot pressure and gait dynamic monitoring according to the present invention includes: S100 utilizes a flexible thin-film pressure sensor array positioned at a designated location on the insole to collect raw plantar pressure data during treadmill exercise; in specific implementation processes, such as... Figure 2 As shown, the flexible thin-film pressure sensors are installed at the heel, toes, and sides of the midfoot of the insole to collect raw plantar pressure data.

[0025] The S200 uses adaptive filtering to process the raw plantar pressure data based on treadmill vibration and torso acceleration to obtain the desired plantar pressure data.

[0026] Treadmill vibration is measured by an accelerometer installed on the treadmill, while trunk acceleration is measured by an inertial measurement unit worn on the human torso. During plantar pressure data acquisition, plantar pressure noise caused by treadmill vibration is collected, and it is necessary to suppress this pressure noise. The desired plantar pressure data is the result of suppressing treadmill vibration. This invention uses a dual-reference adaptive filter to process the raw plantar pressure data based on treadmill vibration and trunk acceleration to obtain the desired plantar pressure data.

[0027] In the specific implementation process, such as Figure 3 As shown, step S200 includes: S201, the problem of processing the original plantar pressure data with dual-reference adaptive filtering to obtain the desired plantar pressure data is modeled as follows: At time n, the original plantar pressure data are represented as follows: ; in, The desired plantar pressure data is related to gait. The noise is caused by the vibration of the treadmill, is related to the treadmill motor, and is periodic or quasi-periodic. White noise; adaptive filtering is used to estimate the noise caused by treadmill vibration. The expected plantar pressure data estimate is obtained. .

[0028] To solve the above problems, S202, this application uses a vibration sensor to collect the vibration of the treadmill as a first reference signal. The first reference signal and the noise caused by the treadmill vibration The strong correlation means the first reference signal is affected by transmission path attenuation and phase shift, and cannot fully characterize the vibration response at the flexible thin-film pressure sensor; therefore, the vertical acceleration component of the user's torso is collected using an inertial measurement unit worn on the human torso as the second reference signal. When a user runs stably on a treadmill, the vertical acceleration of the torso is excited by the vibration of the treadmill and modulated by the human biomechanics system, the second reference signal. The spectrum contains harmonic components with the same frequency as the treadmill vibration and has a phase coupling relationship with the foot vibration interference, which supports compensation for the transfer function mismatch caused by the coupling between the human body and the treadmill.

[0029] S203, First Reference Signal Second reference signal Combined to form a reference vector .

[0030] S204 uses an FIR-type multi-channel transverse filter to estimate the noise caused by treadmill vibration based on a reference vector. ,as follows: ; Where L is the filter order covering the dominant vibration frequency period; Let be the filter weight vector of the k-th tap.

[0031] S205, the error is obtained by subtracting the output of the FIR multi-channel transverse filter from the raw plantar pressure data: , in, This refers to residual noise and white noise caused by treadmill vibration; S206, to minimize mean square error To achieve this, LMS gradient descent is used to update the weights of the FIR-type multichannel transverse filter according to the error and the reference vector: ; in, The step size.

[0032] S207 uses an FIR-type multichannel transverse filter that satisfies the condition of minimizing the mean square error to estimate the noise caused by treadmill vibration based on the reference vector. Original plantar pressure data minus Obtain the desired plantar pressure data.

[0033] This invention significantly improves the signal-to-noise ratio of plantar pressure signals in treadmill scenarios by using dual physical reference signals and multi-channel adaptive filtering, thereby enhancing the reliability of subsequent analysis.

[0034] S300 involves segmenting the acquired desired plantar pressure data into gait cycles. The process includes initial cycle segmentation and smoothing the initial gait cycle segmentation result sequence using a pre-trained gait segmentation Hidden Markov Model.

[0035] The initial cycle division includes determining the ground contact initiation time based on the combined heel pressure gradient and plantar pressure center velocity; specifically, the ground contact initiation time is determined when the weighted result of the norm of the heel pressure gradient and the plantar pressure center velocity is greater than a set adaptive threshold. ; in, For pressure on the heel area, by The center velocity of plantar pressure, weight , The adaptive threshold is set to 3 times the standard deviation of the local signal.

[0036] in, The center of plantar pressure per unit time The displacement difference, the center of pressure on the sole of the foot is the normalized first-order moment of the horizontal and vertical position coordinates of the four flexible film pressure sensors on the insole by the combined force of the four flexible film pressure sensors.

[0037] When the foot contacts the treadmill during exercise, the heel makes contact first, resulting in two characteristics: first, a sudden increase in heel pressure, with a positive peak appearing in the heel pressure gradient; and second, a rapid shift in the center of plantar pressure: the center of plantar pressure quickly moves from nothing to the heel. Therefore, the moment of contact can be preliminarily determined based on these conditions.

[0038] The initial cycle division includes determining the push-off moment based on the abrupt change in the pressure gradient in the toe region from the midpoint to the end of the gait; specifically, it involves detecting whether the abrupt change in the toe pressure gradient from the midpoint to the end of the gait exceeds a dynamic threshold to determine whether push-off has occurred. ; in, Pressure on the toes, The moment corresponding to the midpoint of the gait, from the center of plantar pressure. The inflection point of the trajectory has been determined. The end time of the gait is determined by the start time of the next ground contact. The dynamic threshold is the historical mean ± 1.5 standard deviation of the pressure gradient change in the toe area. At the moment of push-off, the forefoot quickly leaves the treadmill, and the pressure at the toes drops sharply, which is represented as an inflection point on the pressure-time curve, i.e., the extreme point of the second derivative; therefore, the push-off moment can be preliminarily determined by the above conditions.

[0039] A pre-trained gait segmentation Hidden Markov Model (HMM) is used to smooth and correct the gait phase sequence divided according to the above-mentioned judgment criteria. The gait segmentation HMM divides the gait phase into the swing phase, ground contact phase, and support phase. The swing phase, ground contact phase, and support phase transition in a set order, that is, the swing phase sequentially transitions to itself or the ground contact phase, the ground contact phase sequentially transitions to the support phase, and the support phase sequentially transitions to itself or the swing phase. The state transition probability matrix is ​​initialized, and the elements of the state transition probability matrix represent the probability of transitioning from one state (gait phase) to another. Due to the sequential nature of the gait phases, the transition probability between some states is 0. An observation space is defined, which includes total pressure, heel pressure gradient, second derivative of pressure at the toes, velocity of the plantar pressure center, and position of the plantar pressure center. The observation probability matrix is ​​initialized according to a Gaussian distribution, and the observation probability matrix represents the probability of observing a certain observation value in a certain state. The initial state probability vector is initialized, and the initial state probability vector represents the probability of being in each state at the beginning of the gait.

[0040] The parameters of the gait segmentation hidden Markov model are learned using training data, including: Collect expected plantar pressure data for multiple gait cycles and label the gait phase corresponding to each time point in the expected plantar pressure data, including the swing phase, ground contact phase, and stance phase; For each time point, observation vectors are extracted from the training data, including: total pressure, heel pressure gradient, second derivative of pressure at the toes, velocity at the center of plantar pressure, and location of the center of plantar pressure. The probability of the initial state is estimated by statistically analyzing the probability of each gait phase in the first step of the training data. The frequency of transitions between arbitrary gait phases is obtained from the statistical training data, and then normalized to obtain the estimated state transition probability. For each gait phase, collect all observation vectors belonging to that gait phase, then calculate the mean and covariance matrix of the observation vectors, and determine the Gaussian distribution that determines the observation probability matrix.

[0041] The ground contact initiation is determined when the weighted result of the pressure gradient in the heel area and the velocity norm of the plantar pressure center is greater than a set adaptive threshold; whether the toe pressure gradient change in the interval from the midpoint to the end of the gait exceeds a dynamic threshold is detected to determine whether push-off has occurred; after initially dividing the gait cycle into the swing phase, ground contact phase and support phase according to the above judgment conditions, the observation vector of each time point is extracted.

[0042] The observation vector sequence is input into the gait segmentation hidden Markov model, and the optimal gait phase sequence is decoded using the Viterbi algorithm. Based on the optimal gait phase sequence, the ground contact time and push-off time are redefined: the transition point from the swing phase to the ground contact phase is the ground contact time, and the transition point from the support phase to the swing phase is the push-off time.

[0043] S400, extract gait features from the data after gait period segmentation. The gait features include: temporal amplitude features, spatiotemporal distribution features, dynamic mechanical features, symmetry and variability features, frequency domain features, and associated health features.

[0044] Specifically, the time-domain amplitude characteristics include: Peak pressure in different areas of the heel, medial midfoot, lateral midfoot, and toe area reflects the maximum load borne by different areas of the foot during the gait cycle; The support phase has a full range of average pressure and pressure standard deviation. The average pressure reflects the overall load level, while the standard deviation reflects the stability of the pressure. The initial pressure peak generated at the heel during the initial contact phase and the rise time are as follows: the initial pressure peak comes from the impact of body weight and the treadmill on the foot, and the rise time reflects the urgency of the impact; the second pressure peak at the toes during the mid-to-late support phase and the descent time are as follows: the second pressure peak comes from the force of the active push-off, and the descent time reflects the speed of pressure release; and the ratio of the two pressure peaks reflects the ratio of impact force to active propulsion force, which is an important indicator for assessing running economy and injury risk. The propulsive impulse at the toes within a set time range before push-off.

[0045] Spatiotemporal distribution characteristics include: The total trajectory length and curvature of the plantar pressure center trajectory during a gait cycle reflect the stability of the gait. The lateral shift of the plantar pressure center relative to the midline of the foot in each phase of a gait cycle reflects the degree of inversion or eversion of the foot. The longitudinal migration rate of the plantar pressure center during a gait cycle, the average speed at which the plantar pressure center moves from the heel to the toes, reflects the speed of weight transfer. Pressure distribution entropy of the sole regions at each phase of a gait cycle: , For the normalized pressure of the i-th region, a high entropy value indicates a dispersed pressure distribution, while a low entropy value indicates a concentrated pressure. Dynamic mechanical characteristics include: Estimated vertical ground reaction force: ; in Vertical acceleration measured by an inertial measurement unit on the torso; Impact loading rate, i.e. the slope of the first peak pressure rise, a high loading rate is associated with damage risk; Ankle joint torque: , The distance from the center of pressure on the sole of the foot to the projection of the ankle joint is estimated based on the foot length ratio. The ankle joint torque reflects the load on the ankle joint.

[0046] Symmetry and variability characteristics include: Left-right foot difference index: ; The characteristics of the left and right feet are as follows: peak pressure, ground contact time, and lateral shift of the plantar pressure center. Interstep variability, i.e., the standard deviation / mean of a feature over 10 consecutive steps; Phase synchronization, which is the phase difference between the trajectories of the left and right plantar pressure centers calculated by Hilbert variation, is used to assess the degree of synchronization between the left and right foot movements.

[0047] Frequency domain characteristics include: Pressure Fast Fourier Transform (FFT) Main Frequency: The main frequency component in the range of 0 to 10 Hz after the pressure signal is transformed by FFT, reflecting the gait rhythm.

[0048] Harmonic energy ratio: The ratio of the 2nd to 5th order harmonic energy to the fundamental energy; the harmonic components reflect the nonlinear characteristics of the gait period. Wavelet packet energy entropy: Four-level wavelet packet decomposition is performed using the db6 wavelet to obtain the energy distribution of different frequency bands, and then the energy entropy is calculated. The wavelet packet energy entropy value reflects the complexity of the gait signal.

[0049] Associated health characteristics include: Arch index: The ratio of midfoot pressure integral to total foot pressure integral, reflecting the height and function of the arch. A higher index indicates a flatter foot. Pronation index: The percentage of the area of ​​the center of pressure trajectory of the foot that deviates towards the arch of the foot relative to the entire envelope area of ​​the center of pressure trajectory of the foot. The pronation index quantifies the degree of pronation of the foot.

[0050] Achilles tendon load index: peak heel pressure multiplied by ground contact time, reflecting the load borne by the Achilles tendon during the gait cycle; Metatarsal stress index: The peak pressure in the midfoot region divided by the square root of the contact area, used to assess stress concentration in the metatarsal region.

[0051] All features are Z-score normalized based on the user's individual historical mean to eliminate dimensional differences and individual baseline drift.

[0052] S500, construct and train a causal inference model. The causal inference model adopts a three-layer directed graph model. The first layer uses gait features as observation layer nodes, the second layer uses health-related features as latent variable layer nodes, and the third layer uses health risk items as risk output layer nodes. The causal edges between each layer are constructed based on biomechanical and clinical knowledge, and the strength weights of the causal edges are learned according to Bayesian linear regression.

[0053] The causal reasoning model has three layers: the observation layer, the latent variable layer, and the output layer.

[0054] Suppose that the observation layer has K gait feature nodes, and the gait feature nodes correspond to the gait features extracted above.

[0055] Suppose that the latent variable layer has M health-related feature nodes, such as the impact intensity node reflecting the severity of the ground impact, the propulsion efficiency node reflecting the mechanical efficiency of the push-off phase, the pronation control node reflecting the control ability of foot pronation / external pronation, the dynamic stability node reflecting balance control during movement, the symmetry coordination node reflecting bilateral limb coordination, the fatigue state node reflecting the cumulative fatigue level, the tissue load capacity node reflecting the body tissue's tolerance to load, and the running technique quality node reflecting the rationality of gait.

[0056] The causal edges between nodes at different levels are constructed based on biomechanics and clinical knowledge. For example, gait feature nodes related to heel impact in the observation layer can only be connected to impact intensity nodes in the latent variable layer, and cannot be directly connected to propulsion efficiency nodes in the latent variable layer.

[0057] Suppose the output layer has C health risk item nodes, such as stress bone injury risk node, fascia injury risk node, joint injury risk node, muscle injury risk node, and sprain risk node.

[0058] A first adjacency matrix of dimension M×K is used to represent the edges from the observation layer to the latent variable layer. If a causal edge exists between a gait feature node j in the observation layer and a health-related feature node i in the latent variable layer, the element at position (i,j) in the first adjacency matrix is ​​1; otherwise, it is 0. Similarly, a second adjacency matrix of dimension C×M is used to represent the causal edges from the latent variable layer to the output layer. In one implementation, a third adjacency matrix of dimension M×M can be used to represent the causal edges within the latent variable layer. A complete causal relationship from the observation layer to the output layer is as follows: peak heel impact (observation layer) affects impact intensity (latent variable layer), and impact intensity affects the risk of tibial stress fracture (risk output layer).

[0059] The specific training process includes: The expected plantar pressure dataset is categorized by individual user, with a portion of expected plantar pressure data collected per user. Experts then provide latent variable layer health-related feature node values ​​and output layer labels based on gait characteristics at the observation layer. Health-related feature node values ​​include scores for impact intensity and propulsion efficiency. Output labels are either binary injury events or risk scores provided by experts.

[0060] Train a Bayesian linear regression model from the observation layer to the latent variable layer using labeled data; use the pre-trained Bayesian linear regression model from the observation layer to the latent variable layer to infer the label of the latent variable layer from the remaining data; select data from the remaining data whose latent variable layer label confidence is higher than a set threshold to train the model; iterate until the parameters of the Bayesian linear regression model from the observation layer to the latent variable layer are obtained. Use a portion of the data with output labels to train a Bayesian linear regression model from the latent variable layer to the output layer; use the trained Bayesian linear regression model from the latent variable layer to the output layer to infer the output layer label for the remaining data; select data from the remaining data where the confidence level of the latent variable layer label is higher than a set threshold to train the model; iterate until the parameters of the Bayesian linear regression model from the latent variable layer to the output layer are obtained. In this process, Bayesian linear regression learns the weights of all causal edges. For any node, its parent node is determined by the causal edges associated with it, and the linear combination of its parent nodes plus noise is used as the value of that node.

[0061] In one implementation, group weights are learned using group data from all users. It is assumed that each user's data is independent and shares the same group weights, with the objective of maximizing the log-likelihood of the group data to estimate the group weights. For each user, the group weights are fixed, and then individual weight shifts are learned using that user's data.

[0062] During inference, the trained causal inference model is used to establish the output from gait features to health-related features, and from health-related features to risk item probabilities.

[0063] This application adjusts movement patterns based on the probability of risk items. It extracts rich gait features with health risk significance from each gait cycle. Based on a constructed three-layer causal inference model, it achieves the analysis from gait features to health risks: the observation layer of the causal inference model is gait features; the latent variable layer is health-related features, which is a semantic abstraction of health risk from the gait features of the observation layer; and the output layer is the specific health risk item. The connections between the layers are not arbitrary but strictly follow biomechanical and clinical knowledge. The causal inference model uses Bayesian linear regression to learn the weights of each causal edge and, through a semi-supervised iterative training strategy, efficiently expands its applicability with a small amount of expert annotation. The final output is not only "risk level," but also includes a clear attribution chain, such as "Your risk of tibial stress injury is increased, mainly due to excessive impact intensity; it is recommended to adjust the ground contact pattern or increase cushioning training," thus providing a direct basis for personalized intervention.

[0064] Example 2 This invention provides a health correlation analysis device based on foot pressure and gait dynamic monitoring, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit serving as a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as the software program, computer-executable program, and modules corresponding to the health correlation analysis method based on foot pressure and gait dynamic monitoring in this invention. The processing unit implements the aforementioned health correlation analysis method based on foot pressure and gait dynamic monitoring by running the software program, computer-executable program, and modules stored in the storage unit.

[0065] Of course, the computer program stored in the storage unit of the health association analysis device based on foot pressure and gait dynamic monitoring provided in the embodiments of the present invention is not limited to the method operation described above, but can also execute related operations in the health association analysis method based on foot pressure and gait dynamic monitoring provided in any embodiment of the present invention.

[0066] Example 3 This invention provides a computer-readable storage medium storing a computer program that, when executed, implements the health correlation analysis method based on foot pressure and gait dynamic monitoring.

[0067] In the embodiments provided by this invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, structures, or units, and may be electrical, mechanical, or other forms.

[0068] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0069] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0070] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A health correlation analysis method based on foot pressure and gait dynamic monitoring, characterized in that, include: The system uses a flexible thin-film pressure sensor array located at the heel, toes, and sides of the midfoot of the insole to collect raw plantar pressure data of the human body during exercise on a treadmill. Based on treadmill vibration and torso acceleration, dual-reference adaptive filtering is used to perform vibration suppression processing on the raw plantar pressure data to obtain the desired plantar pressure data; The acquired expected plantar pressure data are divided into gait cycles; Gait features are extracted from the data after gait cycle segmentation. The gait features include: temporal amplitude features, spatiotemporal distribution features, dynamic mechanical features, symmetry and variability features, frequency domain features, and associated health features. A causal inference model is constructed and trained. The causal inference model adopts a three-layer directed graph model. The first layer uses gait features as observation layer nodes, the second layer uses health-related features as latent variable layer nodes, and the third layer uses health risk items as output layer nodes. The causal edges between nodes at each level are constructed based on biomechanical and clinical knowledge, and the strength weights of the causal edges are learned according to Bayesian linear regression. During inference, the trained causal inference model is used to establish the output from gait features to health-related features, and from health-related features to risk item probabilities. Adjust your exercise method based on the probability of risk factors.

2. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 1, characterized in that, Based on treadmill vibration and torso acceleration, adaptive filtering is used to perform vibration suppression processing on the raw plantar pressure data to obtain the desired plantar pressure data, including: At time n, the original plantar pressure data are represented as follows: ; in, The desired plantar pressure data is related to gait. The noise is caused by the vibration of the treadmill, is related to the treadmill motor, and is periodic or quasi-periodic. White noise; adaptive filtering is used to estimate the noise caused by treadmill vibration. The expected plantar pressure data estimate is obtained. ; Vibration sensors are used to collect treadmill vibrations as the first reference signal. The first reference signal and the noise caused by the treadmill vibration Strong correlation; The vertical acceleration component of the user's torso is collected using an inertial measurement unit as a second reference signal. Second reference signal The spectrum contains harmonic components with the same frequency as the treadmill vibration, and there is a phase coupling relationship with the foot vibration interference; First reference signal Second reference signal Combined to form a reference vector ; Noise caused by treadmill vibration is estimated using an FIR-type multi-channel transverse filter based on reference vector estimation. ,as follows: ； Where L is the filter order covering the dominant vibration frequency period; Let be the filter weight vector of the k-th tap; The error is obtained by subtracting the output of the FIR multi-channel lateral filter from the raw plantar pressure data: ， in, This refers to residual noise and white noise caused by treadmill vibration; To minimize mean square error To achieve this, LMS gradient descent is used to update the weights of the FIR-type multichannel transverse filter according to the error and the reference vector: ； in, Step size; The noise caused by treadmill vibration is estimated based on the reference vector using an FIR-type multichannel transverse filter that satisfies the condition of minimizing the mean square error. Original plantar pressure data minus Obtain the desired plantar pressure data.

3. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 1, characterized in that, Gait periodization of the acquired expected plantar pressure data includes: Ground contact is determined when the weighted result of the pressure gradient in the heel area and the velocity norm of the center of pressure on the plantar surface is greater than a set adaptive threshold. The system detects whether the abrupt change in toe pressure gradient from the midpoint to the end of the gait exceeds a dynamic threshold to determine whether push-off has occurred. The gait phase sequence segmented according to the above judgment criteria is smoothed and corrected by a pre-trained gait segmentation hidden Markov model.

4. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 3, characterized in that, The gait segmentation Hidden Markov Model divides the gait phase into swing phase, ground contact phase, and support phase, which transition in a predetermined order. A state transition probability matrix is ​​initialized, where each element represents the probability of transitioning from one state to another. Due to the sequential nature of the gait phases, the transition probability between some states is 0. An observation space is defined, including total pressure, heel pressure gradient, second derivative of pressure at the toes, velocity at the center of plantar pressure, and position of the center of plantar pressure. An observation probability matrix is ​​initialized according to a Gaussian distribution, representing the probability of observing a particular observation in a given state. An initial state probability vector is initialized, representing the probability of being in each state at the start of the gait. The ground contact initiation is determined when the weighted result of the pressure gradient in the heel area and the velocity norm of the center of plantar pressure is greater than the set adaptive threshold; whether the toe pressure gradient change in the interval from the midpoint to the end of the gait exceeds the dynamic threshold is detected to determine whether to push off; after initially dividing the gait cycle into the swing phase, ground contact phase and support phase, the observation vector of each time point is extracted. The observation vector sequence is input into the gait segmentation hidden Markov model, and the optimal gait phase sequence is decoded using the Viterbi algorithm. Based on the optimal gait phase sequence, the ground contact time and push-off time are redefined: the transition point from the swing phase to the ground contact phase is the ground contact time, and the transition point from the support phase to the swing phase is the push-off time.

5. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 4, characterized in that, The parameters of the gait segmentation hidden Markov model are learned using training data, including: Collect expected plantar pressure data for multiple gait cycles and label the gait phase corresponding to each time point in the expected plantar pressure data, including the swing phase, ground contact phase, and stance phase; For each time point, observation vectors are extracted from the training data, including: total pressure, heel pressure gradient, second derivative of pressure at the toes, velocity at the center of plantar pressure, and location of the center of plantar pressure. The probability of the initial state is estimated by statistically analyzing the probability of each gait phase in the first step of the training data. The frequency of transitions between arbitrary gait phases is obtained from the statistical training data, and then normalized to obtain the estimated state transition probability. For each gait phase, collect all observation vectors belonging to that gait phase, then calculate the mean and covariance matrix of the observation vectors, and determine the Gaussian distribution that determines the observation probability matrix.

6. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 1, characterized in that, The temporal amplitude characteristics include: peak pressure in each zone of the heel, medial midfoot, lateral midfoot, and toe area; average pressure and standard deviation of the entire foot during the stance phase; the first peak pressure generated by the heel at the initial contact stage and its rise time; the second peak pressure at the toes during the mid-to-late stance phase and its fall time, as well as the ratio of the two peak pressure values; and the propulsive impulse at the toes within a set time range before push-off. Spatiotemporal distribution characteristics include: The total trajectory length and curvature of the plantar pressure center trajectory in a gait cycle, the lateral offset of the plantar pressure center relative to the midline of the foot in each phase, the longitudinal migration rate of the plantar pressure center, and the pressure distribution entropy of each region of the foot in each phase. Dynamic mechanical characteristics include: estimated vertical ground reaction force, impact loading rate, and ankle joint moment; Symmetry and variability characteristics include: left and right foot difference index, interstep variability, and phase synchronization; Frequency domain characteristics include: pressure fast Fourier transform dominant frequency, harmonic energy ratio, and wavelet packet energy entropy; Related health characteristics include: Arch index: the ratio of midfoot pressure integral to total foot pressure integral; Pronation index: the percentage of the area of ​​the plantar pressure center trajectory that is offset towards the arch side to the entire plantar pressure center trajectory envelope area; Achilles tendon load index: peak heel pressure multiplied by ground contact time; Metatarsal stress index: peak pressure in the midfoot area divided by the square root of the contact area. All features are Z-score normalized based on the user's individual historical mean to eliminate dimensional differences and individual baseline drift.

7. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 1, characterized in that, The training process for a causal reasoning model includes: The expected plantar pressure dataset is categorized by individual users. A portion of the expected plantar pressure data is collected from each user. Experts provide latent variable layer health-related feature node values ​​based on gait characteristics of the observation layer. A Bayesian linear regression model from the observation layer to the latent variable layer is trained using the labeled data. The pre-trained Bayesian linear regression model from the observation layer to the latent variable layer is used to infer labels for the remaining data. Data with latent variable layer label confidence levels higher than a set threshold is selected from the remaining data to train the model. This process is iterated until the parameters of the Bayesian linear regression model from the observation layer to the latent variable layer are obtained. Using a portion of the data with output labels, where the output labels are binary injury events or risk scores given by experts, train a Bayesian linear regression model from the latent variable layer to the output layer; use the trained Bayesian linear regression model from the latent variable layer to the output layer to infer the output layer label for the remaining data; select data from the remaining data where the confidence level of the latent variable layer label is higher than a set threshold to train the model, iterating until the parameters of the Bayesian linear regression model from the latent variable layer to the output layer are obtained; In this process, Bayesian linear regression learns the weights of all causal edges. For any node, its parent node is determined by the causal edges associated with it, and the linear combination of its parent nodes plus noise is used as the value of that node.

8. The health correlation analysis method based on foot pressure and gait dynamic monitoring according to claim 7, characterized in that, Using a hierarchical Bayesian framework, we first assume at the group level that each user's data is independent and shares the same group weight, and estimate the group weight with the goal of maximizing the log-likelihood of the group data. Learn a group weight, and then at the individual level, learn an individual weight offset to fine-tune the group weight.

9. A health correlation analysis device based on foot pressure and gait dynamic monitoring, comprising: At least one processing unit, the processing unit being connected to a storage unit via a bus unit, characterized in that the storage unit stores a computer program, and the processing unit implements the health correlation analysis method based on foot pressure and gait dynamic monitoring as described in any one of claims 1-8 by running the computer program stored in the storage unit.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it implements the health correlation analysis method based on foot pressure and gait dynamic monitoring as described in any one of claims 1-8.

Images