A stroke risk grading early warning method and system based on multi-modal gait analysis

Abstract

The application discloses a stroke risk grading early warning method and system based on multi-modal gait analysis, relates to the technical field of gait data analysis and early warning, generates a healthy gait category through clustering of collected healthy user's two-foot gait time series data, and constructs a bidirectional prediction model library; a similar healthy category model is matched for a new user, an individualized prediction model is generated by combining multi-day data incremental learning, and early warning is realized by continuously monitoring the deviation degree of the gait prediction value and the actual value. The application solves the individual adaptation problem of a general model, shortens the cold start time by adopting a 'transfer learning + fine-tuning' strategy, quantifies gait coordination degradation by using mutual prediction deviation of left and right feet, is more sensitive than traditional static indexes, and can realize non-diagnostic, forward-looking stroke risk grading early warning.

Description

Technical Field

[0001] This invention relates to the field of gait data analysis and early warning technology, and in particular to a method and system for stroke risk classification and early warning based on multimodal gait analysis. Background Technology

[0002] Stroke, a major cardiovascular and cerebrovascular disease with high incidence and disability rates, requires early risk identification for prevention and reduction of the social medical burden. Traditional stroke risk assessment relies primarily on clinical questionnaires, blood biochemical indicators, and imaging examinations. These methods are either subjective or invasive, expensive, and cannot achieve continuous monitoring, making it difficult to capture early, dynamic physiological signal abnormalities caused by subtle changes in nervous system function. In recent years, with the rapid development of flexible electronics, the Internet of Things, and artificial intelligence technologies, gait analysis based on wearable sensors has opened up new avenues for non-invasive, continuous, and objective monitoring of human health. Studies have shown that compensatory changes or minor damage to the central nervous system before a stroke often manifests first in subtle changes in gait coordination, particularly a decrease in the symmetry and predictability of left and right lower limb movements. Therefore, developing a technological solution that can utilize daily gait data for intelligent, graded early warning of potential stroke risk has become a key issue urgently needing to be addressed in the fields of smart healthcare and proactive health.

[0003] Significant progress has been made in gait analysis technology. For example, Chinese invention patent application CN120345891A discloses a "gait analysis method and device for stroke patients," which collects multi-source data through a plantar pressure sensing module and an inertial measurement module, uses a dual-stream neural network for feature extraction and fusion, and then calculates key indicators such as stride length, gait frequency, and gait asymmetry index to assess the rehabilitation status of diagnosed stroke patients. The advantage of this approach lies in the application of multimodal data fusion and deep learning models, which effectively improves the objectivity and accuracy of the assessment. However, its technical essence is the quantification of rehabilitation effects for a specific patient group, and its model training, feature selection, and assessment criteria are all based on known pathological patterns. This approach does not solve, and cannot be directly applied to, the prospective risk warning problem for a broad range of healthy or sub-healthy populations.

[0004] Existing gait monitoring technologies mostly rely on static thresholds or general models to assess user gait, making it difficult to adapt to individual differences. Furthermore, they lack sensitivity and specificity when faced with early, subtle gait abnormalities (such as left-right gait asymmetry caused by stroke precursors). In addition, traditional methods do not effectively utilize the common gait patterns of healthy individuals and lack a dynamic weighting mechanism for the diverse daily movement states of users, leading to susceptibility to noise interference and a high false alarm rate in anomaly detection. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a method and system for stroke risk classification and early warning based on multimodal gait analysis.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, including: A stroke risk grading and early warning method based on multimodal gait analysis includes: Acquire bipedal gait time-series data from multiple healthy users, extract multidimensional time-series feature vectors representing gait patterns, i.e., gait feature vectors, and perform clustering based on the feature vectors to obtain several healthy gait categories; For each healthy gait category, gait prediction models in two directions are trained using bipedal gait time-series data of all healthy users in the category, forming a prediction model group. The prediction model group includes: a first prediction model that uses the left foot gait time-series features as input to predict the right foot gait features at the corresponding time, and a second prediction model that uses the right foot gait time-series features as input to predict the left foot gait features at the corresponding time. The prediction model groups of multiple categories together form a basic prediction model library. For new users awaiting warning, collect their bipedal gait time-series data over multiple days, extract their individual gait feature vectors, and calculate their similarity to the center vectors of each healthy gait category; Based on the similarity, a corresponding prediction model group is selected from the basic prediction model library as the initial prediction model group, and incremental learning is performed in combination with the gait data of the user to be warned to generate a personalized gait prediction model. The personalized gait prediction model is used to continuously predict the daily gait of the user to be warned, and based on the degree of deviation between the prediction results and the actual observation values, the warning information of stroke risk is output.

[0007] Preferably, the first prediction model in all prediction model groups has the same parameter structure, but the parameter values ​​at corresponding structural positions differ; the second prediction model in all prediction model groups has the same parameter structure, but the parameter values ​​at corresponding structural positions differ.

[0008] Preferably, based on the similarity, a corresponding prediction model group is selected from the basic prediction model library as the initial prediction model group, and incremental learning is performed using the new gait data of the user to be warned to generate a personalized gait prediction model, including: When the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, the same structural position parameters of all first prediction models in the basic prediction model library are weighted and fused to form a first personalized gait prediction model; the same structural position parameters of all second prediction models in the basic prediction model library are weighted and fused to form a second personalized gait prediction model. The weight parameters of the weighted fusion personalized gait prediction model in the Lth layer of the neural network The following conditions must be met: ; in, Let s be the weight parameters of the prediction model in the k-th prediction model group at the L-th layer of the neural network. k The similarity between the user to be warned and the k-th healthy gait category is N, where N is the total number of healthy gait categories. An initial prediction model set is constructed based on the fused parameters, and incremental learning is performed using the gait data of the users to be warned.

[0009] Preferably, when the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, the prediction model group corresponding to the healthy gait category with the highest similarity is selected as the initial prediction model group, and fine-tuned using a first learning rate during the incremental learning process; The first learning rate is set to be greater than or equal to 1×10. -3 And it remains constant or gradually decreases according to a preset decay strategy during the first M training cycles of incremental learning, where M≥3.

[0010] Preferably, the bipedal gait time-series data is acquired through an insole integrating a flexible pressure sensor array, and the multidimensional time-series feature vector includes at least: The temporal trajectory of peak pressure values ​​and pressure centers in each region of the left and right feet; Duration of single-step support phase, percentage of double support phase; Left and right foot length ratio, support time asymmetry index T left T right The duration of the single-step support phase for each foot is represented by the duration of the left and right foot respectively. Cross-correlation coefficients or phase differences between the velocity sequences of the left and right foot pressure centers.

[0011] Preferably, the clustering uses the K-means algorithm to divide the gait feature vectors of healthy users into multiple categories; both the first prediction model and the second prediction model use an LSTM or Transformer encoder-decoder structure, and the loss function is the mean squared error.

[0012] Preferably, the personalized gait prediction model is used to continuously predict the daily gait of the user to be warned, including: Based on the bipedal gait time-series data of the users to be warned collected on the same day, the movement state category of each walking segment is identified, and the movement state category includes standard walking state and non-standard movement state; Based on the movement state category of each walking segment collected on that day and the similarity between the predicted gait and the actual gait, a daily gait deviation index D is generated for that day. day ; Methods for determining motion state categories include: If the standard deviation of adjacent step length is less than or equal to 0.15 seconds, the asymmetry index of left and right foot support time is less than or equal to 0.12, and the angle between the main movement directions of the pressure center of the left and right feet is less than or equal to 20°, then it is judged as a standard walking state. Otherwise, it is judged as a non-standard motion state.

[0013] Preferably, the daily gait deviation index D day The following conditions must be met: ; Among them, w i The weight of the motion state category corresponding to the i-th segment of bipedal gait time-series data collected on that day, with w corresponding to the standard walking state. i =1.0, non-standard motion state corresponds to w i =0.4; S i The similarity between the predicted gait and the actual gait corresponding to the i-th bipedal gait time series data collected on the day is denoted as Y, which is the mean of the cosine similarity between the predicted gait and the actual gait of the left and right feet. Y is the number of valid walking segments on the day.

[0014] Preferably, the daily gait deviation index D corresponding to the day of generation is... day Subsequently, the method also includes: Continuously monitor the daily gait deviation index value for H consecutive days. If the following conditions are met: If the linear regression slope p < 0.05, an early warning is triggered and a risk warning message is generated; where δ is the rate of change threshold. This is the daily gait deviation index corresponding to the current date t.

[0015] This invention also provides a stroke risk grading and early warning system based on multimodal gait analysis, comprising: The clustering module is used to acquire bipedal gait time-series data of multiple healthy users, extract multi-dimensional time-series feature vectors that characterize gait patterns, i.e. gait feature vectors, and perform clustering based on the feature vectors to obtain several healthy gait categories. The basic model generation module is used to train gait prediction models in two directions for each healthy gait category using bipedal gait time-series data of all healthy users in the category, forming a prediction model group. The prediction model group includes: a first prediction model that takes the left foot gait time-series features as input and predicts the right foot gait features at the corresponding time, and a second prediction model that takes the right foot gait time-series features as input and predicts the left foot gait features at the corresponding time. The prediction model groups of multiple categories together form the basic prediction model library. The similarity determination module is used to collect multi-day bipedal gait time-series data of new users to be warned, extract personal gait feature vectors, and calculate the similarity between them and the center vectors of each healthy gait category. The personalized model generation module is used to select the corresponding prediction model group from the basic prediction model library as the initial prediction model group based on the similarity, and perform incremental learning by combining the gait data of the user to be warned to generate a personalized gait prediction model. The early warning generation module is used to continuously predict the daily gait of the user to be warned using the personalized gait prediction model, and output early warning information of stroke risk based on the degree of deviation between the prediction results and the actual observation values.

[0016] The advantages of this invention are: (1) This invention effectively solves the technical problem that general models are difficult to adapt to individual differences by constructing a bidirectional gait prediction model library based on healthy population clustering. Traditional methods often use a single, fixed threshold or model to evaluate all users, ignoring the various normal gait patterns that exist within healthy populations (such as gait differences caused by different ages, body types, and habits). This invention first collects a large amount of bipedal gait time-series data from healthy users and divides it into several representative healthy gait categories through clustering. For each category, two dedicated models, "left foot predicts right foot" and "right foot predicts left foot," are trained to form a prediction model group that can accurately capture the gait coupling rules within the category. This technical feature enables the system to match the closest initial model from multiple "health paradigms" when facing new users, rather than forcibly applying an "averaged" general standard. As a result, the accuracy of the model in depicting the user's own "healthy normal state" is significantly improved, laying a high-fidelity benchmark for subsequent detection of minor deviations, thereby greatly enhancing the sensitivity of early, individualized gait abnormalities.

[0017] (2) The personalized incremental learning mechanism of this invention achieves seamless and efficient adaptation from group commonalities to individual characteristics, overcoming the shortcomings of large data volume and long cycle required for training from scratch. For new users (i.e., new users to be warned), this invention does not train the model from scratch, but first calculates the similarity between their personal gait characteristics and each health category, and selects the optimal initial prediction model group from the pre-built basic model library accordingly. On this basis, incremental learning is then carried out using the user's limited multi-day gait data. This "transfer learning + fine-tuning" strategy cleverly integrates group intelligence and individual information. Its core advantage is that even if a new user only provides a small amount of daily walking data, it can quickly generate a personalized prediction model that closely matches their personal gait habits. This not only greatly shortens the system's cold start time, but more importantly, it ensures that the warning benchmark is a dynamically evolving "health baseline" that truly belongs to the user, rather than a static external reference. Therefore, this mechanism can more reliably capture subtle changes in the user's own historical state caused by potential health risks, effectively reducing false alarms or missed alarms caused by individual uniqueness.

[0018] (3) This invention utilizes the mutual prediction deviation of left and right gait patterns as the core criterion to directly and sensitively quantify the degree of damage to gait coordination, providing objective and quantifiable technical indicators for stroke risk early warning. The core monitoring logic of this scheme does not rely on absolute gait parameters (such as stride length and speed), but rather continuously predicts "how the right foot should respond if the left foot takes such a step" through a personalized model, and compares the prediction results with the actual observation data of the right foot, and vice versa. In a healthy state, due to the good coordination of the neuromuscular system, the accuracy of this prediction is extremely high; however, when the nervous system experiences early functional impairment (such as stroke precursors), the motor coordination of the left and right lower limbs will be impaired first, leading to gait asymmetry, which in turn causes the actual gait to deviate significantly from the model's prediction. This degree of deviation from "prediction to reality" directly reflects the level of degradation of the intrinsic gait coordination. Compared with traditional static indicators such as the asymmetry index, this method is a dynamic and functional assessment, which is more sensitive to weak and early changes in neurological function, and can achieve non-diagnostic, prospective stroke risk classification and early warning. Attached Figure Description

[0019] Figure 1 The flowchart illustrates a stroke risk grading and early warning method based on multimodal gait analysis, as provided in an embodiment of the present invention.

[0020] Figure 2 This is a structural block diagram of a stroke risk grading and early warning system based on multimodal gait analysis, provided as an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1

[0023] like Figure 1 As shown, this invention provides a stroke risk grading and early warning method based on multimodal gait analysis, comprising the following steps: S100: Acquire bipedal gait time-series data from multiple healthy users, extract multidimensional time-series feature vectors representing gait patterns, i.e., gait feature vectors, and perform clustering based on gait feature vectors to obtain several healthy gait categories.

[0024] Specifically, bipedal gait time-series data are collected through insoles with integrated flexible pressure sensor arrays. For example, the selected smart insoles have no fewer than 16 flexible piezoresistive sensing units arranged on the bottom of each foot, covering the forefoot, arch, and heel areas. The sampling frequency of the sensors is 100 Hz, and the sequence of pressure values ​​of each unit changing over time is recorded synchronously.

[0025] The raw pressure data is first low-pass filtered (cutoff frequency 10 Hz) to remove high-frequency noise; then, individual gait cycles are segmented using a gait event detection algorithm (such as one based on pressure abrupt change points); based on this, the following four types of quantifiable features are calculated to form a multidimensional temporal feature vector: The time-series trajectory of peak pressure values ​​and center of pressure (COP) in each region of the left and right feet.

[0026] The center of gravity (COP) is the weighted average position of plantar pressure distribution, and its XY coordinate sequence reflects the path of weight transfer. The smoothness and symmetry of the COP trajectory are key indicators for evaluating gait coordination.

[0027] The duration of the single-step support phase and the percentage of the double-step support phase. The support phase refers to the time that the same-side foot is in contact with the ground; the double-step support phase refers to the period when both feet are on the ground simultaneously. In healthy individuals, the percentage of the double-step support phase is typically 10%–20%, and an abnormally high percentage may indicate a decline in balance ability.

[0028] Left and right foot length ratio, support time asymmetry index T left and T right These represent the duration of the support phase for each foot. This index quantifies gait asymmetry over time; a value closer to 0 indicates better symmetry.

[0029] The cross-correlation coefficient or phase difference between the COP movement velocity sequences of the left and right feet. By calculating the cross-correlation function of the COP velocity sequences of the left and right feet, the maximum correlation coefficient and the corresponding time delay (i.e., phase difference) can be obtained, which is used to measure the synchronicity of bilateral movements.

[0030] The aforementioned feature dimensions are composed of various physical quantities of different properties, involving multimodal information. Among them, pressure distribution features (peak pressure, COP trajectory) are pressure modes, time structure features (support phase duration, proportion of double support phases) are time series / event modes, and symmetry / coordination features (step ratio, asymmetry index, COP phase difference) are bipedal relationship modes.

[0031] In this embodiment, bipedal gait timing data is collected by a smart insole integrating a flexible pressure sensor array. Taking the left insole as an example, it typically contains multiple pressure sensing areas (such as the medial forefoot, lateral forefoot, arch, medial heel, and lateral heel), each area outputting a time-varying pressure signal sequence. The system simultaneously records pressure values ​​from at least 10 channels for both feet at a sampling rate of 100 Hz.

[0032] A complete gait cycle is defined as the period from heel strike (HS) of the same foot to the next HS of the same foot. A complete gait cycle can also consist of multiple consecutive single-step cycles. Key events in a complete gait cycle include: HS (Heel Strike): The first contact of the foot with the ground; FF (Foot Flat): Landing on the entire foot; HO (Heel Off): Heel off the ground; TO (Toe Off): Toes lift off the ground, entering the swing phase.

[0033] Using only plantar pressure data, HS and TO events (to segment gait cycles) can be detected using the following rules: HS event determination: When the pressure value of any heel area of ​​a foot suddenly increases from close to 0 (<5 kPa) to exceed the threshold (e.g., 20 kPa) and lasts for ≥ 20 ms, it is marked as an HS event.

[0034] TO Event Determination: When the pressure value of all areas of the foot drops synchronously to near 0 (<5 kPa) and lasts for ≥ 50ms, it is marked as a TO event.

[0035] Assuming that when a user is walking, the pressure on the left heel sensor suddenly jumps from 2 kPa to 35 kPa at t=1.20 s and remains above 20 kPa for more than 20 ms, the system records: Left foot HS @ t=1.20 s.

[0036] Subsequently, at t=1.65 s, the pressure of all sensors on the left foot dropped below 3 kPa and remained below 3 kPa for 60 ms. Then, record: Left foot TO @ t=1.65 s.

[0037] The next left foot HS occurs at t=2.10 s.

[0038] Therefore, the complete left foot phase cycle is [1.20 s, 2.10 s), and the duration is 0.90 seconds.

[0039] By continuously detecting the timestamps {t0,t1,t2,…} of HS events, the original pressure time series data can be automatically divided into a series of non-overlapping single-step data segments: the nth step corresponds to the time interval [tn,tn+1).

[0040] For each segmented single-step data segment (such as the nth step of the left foot), perform the above four types of quantifiable feature calculations to form a multi-dimensional temporal feature vector for that step.

[0041] Clustering uses the K-means algorithm to divide the gait feature vectors of healthy users into multiple categories. Both the first and second prediction models employ LSTM or Transformer encoder-decoder structures, with the mean squared error as the loss function.

[0042] Clustering employs the K-means algorithm to divide the gait feature vectors of healthy users into K categories (e.g., K=5). This number needs to be determined through cross-validation to achieve an optimal balance between model complexity and clustering purity. The number of categories K used in clustering is not fixed in advance but determined through a data-driven approach. Specifically, firstly, the candidate range of K is set to 3 to 8; then, for each candidate K value, clustering is performed and the corresponding prediction model group is trained; furthermore, a comprehensive score is calculated based on the cluster silhouette coefficient and the model's prediction accuracy on the validation set; finally, the K value with the highest comprehensive score is selected as the number of healthy gait categories. For example, cross-validation determines that K=5 results in optimal model performance, therefore healthy users are divided into 5 gait categories.

[0043] Both the first and second prediction models employ either a Long Short-Term Memory (LSTM) network or a Transformer encoder-decoder structure. LSTM excels at capturing temporal dependencies, while Transformer models long-range gait patterns through a self-attention mechanism. The loss function is the Mean Squared Error (MSE), used to minimize the difference between the predicted gait features and the actual observations.

[0044] S200: For each healthy gait category, using the bipedal gait time-series data of all healthy users in the category, gait prediction models in two directions are trained respectively, forming a prediction model group. The prediction model group includes: a first prediction model that uses the left foot's gait time-series features as input to predict the right foot's gait features at the corresponding time; and a second prediction model that uses the right foot's gait time-series features as input to predict the left foot's gait features at the corresponding time. Multiple prediction model groups for different categories together form the basic prediction model library.

[0045] By independently training two models, one for "left-to-right" and the other for "right-to-left," for each health category, the system can learn the inherent coupling patterns of left and right lower limb movements within that group. For example, if a certain group habitually steps with their left foot first and their right foot lands heavier, their "left-to-right" model will learn this specific mapping relationship. Furthermore, different genders, age groups, and body types exhibit differences in their left and right foot usage habits, which also need to be learned separately by the model. This category-specific bidirectional prediction mechanism gives the basic model library a high degree of physiological rationality, providing a precise starting point for subsequent personalized adaptation.

[0046] S300: For new users awaiting warning, collect their bipedal gait time-series data over multiple days, extract their individual gait feature vectors, and calculate their similarity to the center vectors of each healthy gait category.

[0047] For new users (i.e., new users awaiting warning), the system collects their gait data for 3–7 consecutive days. The mean vector of all valid single-step feature vectors for each day is calculated, and then the mean vectors from the multiple days are averaged again to obtain the user's personal gait feature vector, which is used for subsequent similarity calculations with healthy cluster centers. It should be noted that the valid user walking data collected in this step is clear data collected when the user is in a healthy and normal state. The personal gait feature vector is the mean vector of its multiple-day features. Similarity is calculated using cosine similarity.

[0048] S400: Based on similarity, select the corresponding prediction model group from the basic prediction model library as the initial prediction model group, and perform incremental learning by combining the gait data of new users to be warned, to generate a personalized gait prediction model.

[0049] The first prediction model in all prediction model groups has the same parameter structure, but the parameter values ​​differ at corresponding structural locations. Similarly, the second prediction model in all prediction model groups has the same parameter structure, but the parameter values ​​differ at corresponding structural locations.

[0050] Specifically, when the maximum similarity between the personal gait feature vector of the user to be warned and the healthy gait category is greater than a preset threshold (such as 95%), two basic prediction models in the corresponding category prediction model group are directly selected, and incremental learning is performed using the user's own data to generate a personalized prediction model. When the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, S400 includes: S410: When the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, the positional parameters of the same structure of all first prediction models in the basic prediction model library are weighted and fused to form the first personalized gait prediction model.

[0051] S411: Weighted fusion of the same structural position parameters of all second prediction models in the basic prediction model library to form a second personalized gait prediction model.

[0052] The weight parameters W of the weighted fusion personalized gait prediction model in the Lth layer of the neural network L fused The following conditions must be met: ; in, Let s be the weight parameters of the prediction model in the k-th prediction model group at the L-th layer of the neural network. k Let N be the similarity between the user to be warned and the k-th healthy gait category, and N be the total number of healthy gait categories.

[0053] S412: Construct an initial prediction model set based on the fused parameters, and use the gait data of the users to be warned for incremental learning.

[0054] The core function of this solution is to generate a "soft attribution" initial model. When a user's gait falls between multiple health categories (e.g., 70% like category A, 30% like category B), meaning the user's gait characteristics differ from any existing base model for any category of users, the neural network weights of the A and B category models can be proportionally mixed to construct a new initial state that better reflects the user's mixed characteristics. Incremental learning is then performed based on this state to create a predictive model with a high degree of fit to the user. This avoids model bias caused by forcibly classifying a user into a particular category and significantly improves the adaptability to "atypical healthy individuals."

[0055] When the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, S400 can also be implemented in the following ways, specifically including: S420: Select the first and second prediction models from the prediction model group corresponding to the health gait category with the highest similarity as the initial prediction model group, and fine-tune them using the first learning rate during the incremental learning process.

[0056] The first learning rate is set to be greater than or equal to 1×10⁻⁶. -3 And it remains constant or gradually decreases according to a preset decay strategy during the first M training cycles of incremental learning, where M≥3.

[0057] In this approach, the learning direction of the initial model during incremental learning is fine-tuned by adjusting the learning rate. Specifically, the learning rate is set to ≥1×10. -3 (Regular fine-tuning is 10) -4 The model aims to accelerate the "memory overlay" of gait patterns specific to new users by maintaining a high value for the first M ≥ 3 training cycles. When a user is not highly similar to any health class, it indicates that their gait is unique, requiring stronger learning signals to correct the prior assumptions of the base model, thereby converging to the personalized optimal solution faster with limited data.

[0058] S500: Utilizes a personalized gait prediction model to continuously predict the daily gait of users under warning, and outputs early warning information on stroke risk based on the degree of deviation between the prediction results and the actual observed values.

[0059] As another possible embodiment of the present invention, S500 includes: S501: Based on the bipedal gait time-series data of the users to be warned collected on the same day, identify the movement state category of each walking segment. The movement state category includes standard walking state and non-standard movement state.

[0060] Each walking segment refers to a continuous and complete gait sequence automatically segmented from the raw gait time-series data using gait event detection algorithms (such as those based on plantar pressure abrupt changes or acceleration zero-crossing points), typically containing no fewer than three complete gait cycles. The system can obtain multiple such valid walking segments daily.

[0061] The method for determining the category of motion state specifically includes the following three joint criteria: If the standard deviation of the time between adjacent steps is less than or equal to 0.15 seconds.

[0062] Stride length refers to the time interval between two consecutive touches of the same foot on the ground. The standard deviation reflects the stability of stride frequency. When a healthy person walks at a constant speed on flat ground, the stride length fluctuates very little (usually <0.1 s). If the standard deviation is >0.15 s, it indicates an unstable walking rhythm, possibly due to starting, stopping, obstacle avoidance, or fatigue, and is considered non-standard movement.

[0063] And the index of asymmetry in the support time of the left and right feet Less than or equal to 0.12.

[0064] A value ≤0.12 indicates that the support time on both sides is highly symmetrical, which is an important characteristic of standard straight-line walking; significant asymmetry is commonly seen when going up and down stairs, walking on slopes, or in pathological gait.

[0065] If the angle between the main movement directions of the left and right feet (COP) is less than or equal to 20°, it is considered a standard walking state.

[0066] The center of gravity (COP) is the equivalent point of application of plantar pressure distribution, and its trajectory reflects the path of weight transfer. For each segment of walking, principal component analysis (PCA) is performed on the COP trajectories of the left and right feet, and the direction of the first principal component is extracted as the "main direction of movement". If the angle between the two directions is ≤ 20°, it indicates that the feet are moving along an approximately parallel path, which conforms to the characteristics of straight-line walking; if the angle is too large, it indicates that there is a significant turning or lateral movement.

[0067] Otherwise, it is judged as a non-standard motion state.

[0068] All three conditions mentioned above must be met simultaneously for a walking state to be considered a standard walking state; otherwise, it is classified as a non-standard movement state. This multi-dimensional joint discrimination mechanism can effectively distinguish between high-reliability assessment scenarios (such as walking at a constant speed in a corridor) and low-reliability scenarios (such as walking around at home, turning around, and going up and down stairs), providing accurate contextual labels for subsequent weighted assessments. This allows for the addition of corresponding weights to adjust the similarity of gait features in different movement states.

[0069] S502: Based on the motion state category of each walking segment collected on the same day and the similarity between the predicted gait and the actual gait, generate the daily gait deviation index D for that day. day .

[0070] Daily Gait Deviation Index D day The following conditions must be met: ; Among them, w i The weight of the motion state category corresponding to the i-th segment of bipedal gait time-series data collected on that day, and the standard walking state w i =1.0, non-standard motion state corresponds to w i =0.4. S i Y represents the mean cosine similarity between the predicted and actual gait of the i-th bipedal gait segment collected on that day, and Y is the number of valid walking segments on that day.

[0071] The core function of this weighted strategy is to dynamically suppress noise interference and strengthen reliable signals. Non-standard movements (such as turning, climbing stairs, and random walking) can cause a temporary decrease in gait symmetry, but this is a normal physiological behavior and not a neurological abnormality. Treating them the same as standard walking would introduce a large number of false positive deviation signals. By assigning them a lower weight (0.4), the system significantly reduces the influence of such segments when calculating the daily index; while the high-weighted (1.0) standard walking segment dominates the final evaluation result. Thus, D day It can more realistically and stably reflect the user's inherent gait coordination ability under controllable and comparable conditions, significantly improving the specificity and robustness of the warning.

[0072] The daily gait deviation index D corresponding to the day of generation day The method then includes: S503: Continuously monitor the daily gait deviation index value for H consecutive days. If the following conditions are met: If the linear regression slope p < 0.05, an early warning is triggered and a risk warning message is generated. Here, δ is the rate of change threshold. Let H be the daily gait deviation index corresponding to the current date t. H=7, δ=0.1.

[0073] The system continuously records daily D day Value. An alert is triggered when the data for H=7 days meets the following two conditions: (1) Relative decline rate ≥ δ (δ = 0.1): that is, the daily deviation index decline exceeds 10% within 7 days. This threshold needs to be adapted by those skilled in the art according to the actual use scenario in order to effectively distinguish between normal intraday fluctuations and potentially significant continuous degradation.

[0074] (2) The slope of the linear regression is p<0.05: for the most recent 7 days of D day The sequence is linearly fitted to obtain the slope of the trend line. The p-value is a statistical indicator used to test whether the slope is significantly different from zero. p < 0.05 indicates that the downward trend is statistically significant, excluding the possibility of random fluctuations.

[0075] Using only the decline rate may trigger false alarms due to a random low value on a particular day (e.g., a user wearing ill-fitting shoes); using only the p-value may be oversensitive to slow but insignificant small changes. Combining both requires that the magnitude of the change be sufficiently large (≥10%) and that the trend of change be sufficiently stable (p<0.05) to ensure that the warning signal is clinically relevant and reliable.

[0076] Once an alert is triggered, the risk warning information generated by the system is strictly limited to non-diagnostic and advisory information, such as: "Your gait coordination has shown a continuous downward trend recently. It is recommended that you pay attention to your health and consult a professional medical institution if necessary."

[0077] Example 2

[0078] like Figure 2 As shown, this invention provides a stroke risk grading and early warning system based on multimodal gait analysis, comprising: The clustering module is used to acquire bipedal gait time-series data from multiple healthy users, extract multidimensional time-series feature vectors representing gait patterns, and perform clustering based on the feature vectors to obtain several healthy gait categories. The basic model generation module is used to train gait prediction models in two directions for each healthy gait category using bipedal gait time-series data of all healthy users in the category. These models form a prediction model group, which includes: a first prediction model that takes the left foot gait time-series features as input and predicts the right foot gait features at the corresponding time; and a second prediction model that takes the right foot gait time-series features as input and predicts the left foot gait features at the corresponding time. The prediction model groups of multiple categories together form the basic prediction model library. The similarity determination module is used to collect multi-day bipedal gait time-series data of new users to be warned, extract personal gait feature vectors, and calculate the similarity between them and the center vectors of each healthy gait category. The personalized model generation module is used to select the corresponding prediction model group from the basic prediction model library as the initial model based on similarity, and to perform incremental learning by combining the gait data of new users to be warned, so as to generate a personalized gait prediction model. The early warning generation module is used to continuously predict the daily gait of users to be warned using a personalized gait prediction model, and output early warning information on the risk of stroke based on the degree of deviation between the prediction results and the actual observation values.

[0079] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for stroke risk classification and early warning based on multimodal gait analysis, characterized in that, include: Acquire bipedal gait time-series data from multiple healthy users, extract multidimensional time-series feature vectors representing gait patterns, i.e., gait feature vectors, and perform clustering based on the feature vectors to obtain several healthy gait categories; For each healthy gait category, gait prediction models in two directions are trained using bipedal gait time-series data of all healthy users in the category, forming a prediction model group. The prediction model group includes: a first prediction model that uses the left foot gait time-series features as input to predict the right foot gait features at the corresponding time, and a second prediction model that uses the right foot gait time-series features as input to predict the left foot gait features at the corresponding time. The prediction model groups of multiple categories together form a basic prediction model library. For new users awaiting warning, collect their bipedal gait time-series data over multiple days, extract their individual gait feature vectors, and calculate their similarity to the center vectors of each healthy gait category; Based on the similarity, a corresponding prediction model group is selected from the basic prediction model library as the initial prediction model group, and incremental learning is performed in combination with the gait data of the user to be warned to generate a personalized gait prediction model. The personalized gait prediction model is used to continuously predict the daily gait of the user to be warned, and based on the degree of deviation between the prediction results and the actual observation values, the warning information of stroke risk is output.

2. The stroke risk classification and early warning method based on multimodal gait analysis according to claim 1, characterized in that, The first prediction model in all prediction model groups has the same parameter structure, but the parameter values ​​at corresponding structural positions differ; the second prediction model in all prediction model groups has the same parameter structure, but the parameter values ​​at corresponding structural positions differ.

3. The stroke risk classification and early warning method based on multimodal gait analysis according to claim 1, characterized in that, Based on the similarity, a corresponding prediction model group is selected from the basic prediction model library as the initial prediction model group, and incremental learning is performed using the new gait data of the user to be warned to generate a personalized gait prediction model, including: When the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, the same structural position parameters of all first prediction models in the basic prediction model library are weighted and fused to form a first personalized gait prediction model; the same structural position parameters of all second prediction models in the basic prediction model library are weighted and fused to form a second personalized gait prediction model. The weight parameters of the weighted fusion personalized gait prediction model in the Lth layer of the neural network The following conditions must be met: ； in, Let s be the weight parameters of the prediction model in the k-th prediction model group at the L-th layer of the neural network. k The similarity between the user to be warned and the k-th healthy gait category is N, where N is the total number of healthy gait categories. An initial prediction model set is constructed based on the fused parameters, and incremental learning is performed using the gait data of the users to be warned.

4. The stroke risk classification and early warning method based on multimodal gait analysis according to claim 1, characterized in that, When the similarity between the personal gait feature vector of the user to be warned and any healthy gait category is lower than a preset threshold, the prediction model group corresponding to the healthy gait category with the highest similarity is selected as the initial prediction model group, and fine-tuned using the first learning rate during the incremental learning process; The first learning rate is set to be greater than or equal to 1×10. -3 And it remains constant or gradually decreases according to a preset decay strategy during the first M training cycles of incremental learning, where M≥3.

5. The stroke risk classification and early warning method based on multimodal gait analysis according to claim 1, characterized in that, The bipedal gait time-series data were collected through an insole with an integrated flexible pressure sensor array, and the multidimensional time-series feature vector included at least: The temporal trajectory of peak pressure values ​​and pressure centers in each region of the left and right feet; Duration of single-step support phase, percentage of double support phase; Left and right foot length ratio, support time asymmetry index T left T right The duration of the single-step support phase for each foot is represented by the duration of the left and right foot respectively. Cross-correlation coefficients or phase differences between the velocity sequences of the left and right foot pressure centers.

6. The stroke risk classification and early warning method based on multimodal gait analysis according to claim 1, characterized in that, Clustering uses the K-means algorithm to divide the gait feature vectors of healthy users into multiple categories; both the first and second prediction models use LSTM or Transformer encoder-decoder structures, and the loss function is mean squared error.

7. The stroke risk classification and early warning method based on multimodal gait analysis according to claim 1, characterized in that, The personalized gait prediction model is used to continuously predict the daily gait of users under warning, including: Based on the bipedal gait time-series data of the users to be warned collected on the same day, the movement state category of each walking segment is identified, and the movement state category includes standard walking state and non-standard movement state; Based on the movement state category of each walking segment collected on that day and the similarity between the predicted gait and the actual gait, a daily gait deviation index D is generated for that day. day ; Methods for determining motion state categories include: If the standard deviation of adjacent step length is less than or equal to 0.15 seconds, the asymmetry index of left and right foot support time is less than or equal to 0.12, and the angle between the main movement directions of the pressure center of the left and right feet is less than or equal to 20°, then it is judged as a standard walking state. Otherwise, it is judged as a non-standard motion state.

8. A stroke risk classification and early warning method based on multimodal gait analysis according to claim 7, characterized in that, The daily gait deviation index D day The following conditions must be met: ； Among them, w i The weight of the motion state category corresponding to the i-th segment of bipedal gait time-series data collected on that day, with w corresponding to the standard walking state. i =1.0, non-standard motion state corresponds to w i =0.4; S i The similarity between the predicted gait and the actual gait corresponding to the i-th bipedal gait time series data collected on the day is denoted as Y, which is the mean of the cosine similarity between the predicted gait and the actual gait of the left and right feet. Y is the number of valid walking segments on the day.

9. A method for stroke risk classification and early warning based on multimodal gait analysis according to claim 1, characterized in that, The daily gait deviation index D corresponding to the day of generation day Subsequently, the method also includes: Continuously monitor the daily gait deviation index value for H consecutive days. If the following conditions are met: If the linear regression slope p < 0.05, an early warning is triggered and a risk warning message is generated; where δ is the rate of change threshold. This is the daily gait deviation index corresponding to the current date t.

10. A stroke risk grading and early warning system based on multimodal gait analysis, characterized in that, A stroke risk grading and early warning method based on multimodal gait analysis, applicable to any one of claims 1-9, comprises: The clustering module is used to acquire bipedal gait time-series data of multiple healthy users, extract multi-dimensional time-series feature vectors that characterize gait patterns, i.e. gait feature vectors, and perform clustering based on the feature vectors to obtain several healthy gait categories. The basic model generation module is used to train gait prediction models in two directions for each healthy gait category using bipedal gait time-series data of all healthy users in the category, forming a prediction model group. The prediction model group includes: a first prediction model that takes the left foot gait time-series features as input and predicts the right foot gait features at the corresponding time, and a second prediction model that takes the right foot gait time-series features as input and predicts the left foot gait features at the corresponding time. The prediction model groups of multiple categories together form the basic prediction model library. The similarity determination module is used to collect multi-day bipedal gait time-series data of new users to be warned, extract personal gait feature vectors, and calculate the similarity between them and the center vectors of each healthy gait category. The personalized model generation module is used to select the corresponding prediction model group from the basic prediction model library as the initial prediction model group based on the similarity, and perform incremental learning by combining the gait data of the user to be warned to generate a personalized gait prediction model. The early warning generation module is used to continuously predict the daily gait of the user to be warned using the personalized gait prediction model, and output early warning information of stroke risk based on the degree of deviation between the prediction results and the actual observation values.

Images