A sports health monitoring and early warning system based on a smart wearable device

Abstract

This invention discloses a sports health monitoring and early warning system based on a smart wearable device. The system includes: a multimodal sensor array that simultaneously acquires electromyography (EMG) signals, photoplethysmography (PPG) signals, and inertial measurement unit (IMU) data; a muscle-vascular coupling analysis module that calculates the muscle-vascular phase synchronization index to distinguish between central and peripheral fatigue; a hemodynamic causal inference module that constructs a multimodal temporal causal graph through conditional mutual information to achieve causal-driven early warning; a dynamic adversarial domain adaptive module that optimizes personalized models within a federated learning framework; an energy-information flow collaborative optimization module that dynamically schedules sensor sampling strategies; and a multimodal data fusion center that outputs a multidimensional health risk assessment vector using a graph neural network. This invention, through causal-driven early warning, adversarial domain adaptation, and energy-information flow collaborative optimization, can significantly reduce false alarm rates, improve personalized monitoring accuracy, and extend device battery life.

Description

Technical Field

[0001] This invention relates to the field of smart wearable device technology, and more specifically, to a sports health monitoring and early warning system based on smart wearable devices, particularly a sports health monitoring and early warning system that integrates multimodal physiological signals, biomechanical parameters, causal inference frameworks, and federated learning architecture. Background Technology

[0002] With increasing health awareness and the rapid development of wearable technology, sports and health monitoring devices have evolved from simple pedometers into intelligent wearable devices integrating multiple physiological sensors. Currently, sports and health monitoring systems mainly employ the following technical solutions: The first type of approach relies on threshold monitoring of a single physiological parameter, such as issuing an alert when the heart rate exceeds a preset threshold. While simple to implement, this approach ignores the coupling relationships between multiple parameters, resulting in a high false alarm rate and an inability to distinguish between fatigue types and injury risk levels.

[0003] The second type of approach employs multimodal sensor data fusion, such as simultaneously acquiring electromyography (EMG), inertial measurement unit (IMU), and photoplethysmography (PPG) signals, and then using machine learning models for comprehensive evaluation. For example, Chinese invention patent CN109589496A discloses a wearable bionic rehabilitation system for the entire human movement process, including functions such as joint torque calculation, muscle force analysis, and multi-signal fusion. Chinese invention patent CN113229831A discloses a method for monitoring and managing movement function based on EMG and muscle oxygenation signals, constructing a muscle-vascular state monitoring model.

[0004] However, existing technologies still have the following technical shortcomings: First, although existing technologies mention "muscle-blood vessel coupling," they are limited to qualitative descriptions and do not provide quantitative methods for calculating coupling strength. Furthermore, they do not establish a correlation between coupling strength and fatigue type (central fatigue vs. peripheral fatigue), resulting in a lack of physiological basis for fatigue diagnosis.

[0005] Second, existing injury warning technologies are based on correlation analysis, which means that a warning is issued when multiple parameters are abnormal at the same time. However, correlation does not equal causation. Such schemes are prone to false alarms due to "spurious correlations." For example, simultaneous muscle fatigue and increased heart rate may be caused by a common cause (increased exercise intensity) rather than having a causal relationship.

[0006] Third, existing federated learning frameworks only use weighted average aggregation, without considering the impact of different user data distributions on the model's generalization ability, resulting in insufficient accuracy in personalized monitoring.

[0007] Fourth, the existing sensor sampling strategy uses a fixed frequency sampling, which does not consider the dynamic balance between energy consumption and information acquisition, resulting in limited device battery life and affecting user experience.

[0008] Therefore, there is an urgent need in this field for a sports health monitoring and early warning system that can quantitatively assess the strength of neurovascular coupling, introduce a causal inference framework to suppress false alarms, improve personalized accuracy through adversarial domain adaptation, and achieve energy-information flow collaborative optimization. Summary of the Invention

[0009] This invention provides a sports health monitoring and early warning system based on smart wearable devices, including a multimodal sensor array, a muscle-blood vessel coupling analysis module, a hemodynamic causal inference module, a dynamic adversarial domain adaptive module, an energy-information flow collaborative optimization module, a dynamic joint stiffness estimation module, a multimodal data fusion center, and an early warning engine.

[0010] The technical solution of the present invention will be described in detail below with reference to specific technical features.

[0011] I. Multimodal sensor array The multimodal sensing array includes the following components: (1) Electromyography sensor: A dry electrode surface electromyography sensor is used, which is deployed on key muscle groups such as the quadriceps femoris, hamstrings, and gastrocnemius muscles of the wearer. The sampling frequency is 1000Hz, which is used to collect electromyography signals EMG(t).

[0012] (2) Photoplethysmography sensor: A reflective PPG sensor is used and deployed on the wrist or earlobe. The sampling frequency is 250Hz. It is used to collect blood volume pulse signal PPG(t) and extract parameters such as heart rate HR(t) and blood oxygen saturation SpO2(t) from it.

[0013] (3) Inertial Measurement Unit: A nine-axis inertial measurement unit (three-axis accelerometer, three-axis gyroscope, three-axis magnetometer) is used, deployed at the thoracic vertebra (T12) and sacrum (S1) with a sampling frequency of 200Hz, to collect acceleration a(t), angular velocity ω(t) and attitude angle θ(t) data.

[0014] (4) Sweat electrolyte sensor: A flexible electrochemical sensor is integrated into the inside of the wearable device's strap to detect the sodium ion concentration C in sweat. Na and chloride ion concentration C Cl .

[0015] Each sensor connects to the edge computing node via Bluetooth Low Energy or Near Field Communication, supporting synchronous triggering and data alignment.

[0016] II. Muscle-Blood Vessel Coupling Analysis Module The muscle-vascular coupling analysis module is one of the key innovations of this invention, and its core lies in calculating the muscle-vascular phase synchronization index. This enables a quantitative assessment of the neurovascular coupling strength and distinguishes between central and peripheral fatigue based on the index.

[0017] (a) Signal preprocessing Electromyography signal envelope extraction: extraction of raw electromyography signals Full-wave rectification is performed, i.e., the absolute value is taken, and then the envelope is extracted by passing it through a 2Hz low-pass filter (Butterworth filter, order 4). : (1); Photoplethysmography (PPG) signal preprocessing: The PPG signal is bandpass filtered (0.5Hz-5Hz) to remove baseline drift and high-frequency noise, yielding the PPG signal. filtered(t) .

[0018] (II) Hilbert Transform The Hilbert transform is defined as: (2); Here, PV represents the Cauchy principal value integral. The Hilbert transform converts a real signal into an analytic signal, thereby extracting the instantaneous phase: (3); (III) Calculation of Phase Synchronization Index The muscle-vascular phase synchronization index is defined as the instantaneous phase θ of the electromyographic envelope. EMG(t) instantaneous phase θ with PPG signal PPG(t) The sliding window average of the absolute values ​​of the cosine differences is given by formula (4):

[0019] Where N is the number of sampling points in the sliding window, the window length is 10 seconds, corresponding to 2500 sampling points at a sampling frequency of 250Hz, and the window step size is 1 second, to achieve real-time phase synchronization monitoring.

[0020] Since the signal after Hilbert transform is an analytic signal, the real part of its complex product corresponds to the product of the amplitude of the two signals and the cosine of the phase difference. Therefore, the above expression is equivalent to the following formula (5):

[0021] This form directly reflects the degree of phase synchronization, with a value range of [0,1], which facilitates threshold comparison.

[0022] (iv) Fatigue type determination Based on the phase synchronization index distribution of 20 healthy subjects in an incremental exercise experiment (Bruce protocol), the following thresholds were determined: First threshold Φ high =0.75 (75th percentile): when Φ sync A value >0.75 indicates a state of central fatigue. In this state, muscle-blood vessels are highly synchronized, indicating enhanced central nervous system regulation of the cardiovascular system and coordinated motor center and autonomic nervous system. This state is commonly seen in the middle and late stages of high-intensity continuous exercise.

[0023] Second threshold Φ low =0.25 (25th percentile): when Φ sync A value <0.25 indicates a peripheral fatigue state. In this state, muscle-vascular phase desynchronization occurs, indicating that the accumulation of peripheral muscle metabolites leads to local vascular regulation dysfunction and decoupling of the phase relationship between electromyography and blood flow. This is commonly seen in scenarios with excessive local muscle load.

[0024] The aforementioned thresholds can be adaptively adjusted based on individual historical data or population distribution, for example, by optimizing through ROC curves to obtain optimal sensitivity and specificity.

[0025] (v) Recommendations for Adjusting Exercise Intensity Differentiatedly Generate differentiated adjustment suggestions based on fatigue type: Central fatigue state: The system issues a voice prompt through bone conduction headphones: "Central fatigue, it is recommended to reduce exercise intensity by 30%, pay attention to hydration and rest."

[0026] Peripheral fatigue state: The system issues a prompt: "Peripheral fatigue, it is recommended to maintain the current intensity and perform stretching and relaxation of the target muscle group."

[0027] III. Hemodynamic Causal Inference Module The hemodynamic causal inference module is another core innovation of this invention. By constructing a multimodal temporal causal graph and calculating conditional mutual information, it achieves a leap from "correlation warning" to "causal warning", significantly reducing the false alarm rate.

[0028] (a) Conditional Mutual Information and Causal Discovery Conditional Mutual Information (CMI) measures the dependency between variables X and Y given a condition variable Z. (6) CMI = 0 if and only if X and Y are conditionally independent given Z. This system uses CMI as the basis for causal discovery: if... (If ε is 0.05 bits), then a direct dependency is considered to exist between X and Y, and a directed edge is added to the causal graph.

[0029] (ii) Probability density estimation Since the joint probability density function of the continuous variables is unknown, this system uses the k-nearest neighbor estimation method (Kraskov et al., 2004) for nonparametric estimation. The calculation steps for conditional mutual information are as follows: For each sample point, find its k nearest neighbors in the joint space (X,Y,Z); count the number of nearest neighbors in the X, Y, and Z subspaces respectively; calculate the mutual information estimate based on the Digamma function: (7); in For the Digamma function, , , , respectively, represent the number of nearest neighbors in the XZ, YZ, and Z subspaces, and <·> indicates averaging over all samples. The value of k is 10, which is determined based on the empirical rule (k≈√n, where n is the sample size) and can also be optimized through cross-validation.

[0030] (III) Construction of Cause-and-Effect Graph Based on the CMI calculation results, a causal graph G=(V,E) in the form of a directed acyclic graph (DAG) is constructed: Node set V: Contains all physiological and motor variables, including EMG signals from various muscle groups (quadriceps EMG). Q Hamstring EMG H Gastrocnemius muscle EMG G ), IMU measurements of joint angles and angular velocities (knee joint angle θ) knee Ankle angle θ ankle angular velocity ω knee angular acceleration α ankle The PPG-derived heart rate and blood oxygen (HR, SpO2), as well as risk event nodes (ACL injury, myocardial ischemia, dehydration, etc.).

[0031] Edge set E: If If the timestamp is 0.05 bits, a directed edge is added between X and Y, with the direction determined by the time sequence: if the timestamp of X is earlier than that of Y, the edge direction is X→Y.

[0032] (iv) Determining the causal path When a causal path exists in the causal graph from "abnormal electromyographic activation pattern" through "abnormal joint kinematics" to "risk event," the system identifies it as a causally driven risk and generates a warning. Specific judgment rules: Abnormal electromyographic activation pattern: electromyographic amplitude exceeds the individual baseline mean + 3 standard deviations and the duration is >100ms.

[0033] Abnormal joint kinematics: Joint angles or angular velocities exceed the preset safety range (e.g., knee flexion angle > 45° and varus angle > 15°).

[0034] Existence of causal path: There exists at least one directed path from the abnormal electromyography node through the abnormal kinematic node to the risk event node, and the CMI of each edge on the path is greater than 0.05 bits.

[0035] If the aforementioned causal path does not exist, even if multiple parameters are abnormal at the same time, the system will determine it as a false alarm due to correlation and suppress the warning.

[0036] (v) Causal path strength For the risk of anterior cruciate ligament (ACL) injury, the system further calculates the causal path strength. As in formula (8):

[0037] Mutual Information (MI) is defined as follows: (9); Information entropy H(X) is defined as: (10); The ratio of mutual information to information entropy represents the degree to which a parent node interprets its child nodes. Path strength is the product of the interpretability of each edge on the path, and its value ranges from (0,1). > hour( To preset the causal threshold (0.32 in this embodiment), an early warning is triggered and electrical stimulation intervention is initiated. This threshold can be obtained by optimizing the ROC curve of individual historical data to achieve the best balance between sensitivity and specificity.

[0038] (vi) Conditional transition entropy and fatigue source identification To differentiate the source of fatigue (muscular vs. cardiac), the system calculates the Conditional Transfer Entropy: (11); Transfer entropy measures the intensity of the causal flow from X to Y, eliminating the effects of common causes given Z.

[0039] when When γ is a preset ratio threshold (γ is 2 in this embodiment), and the permutation test p < 0.05, it is determined that muscle-related fatigue is dominant (electromyographic changes drive heart rate changes); otherwise, it is determined that cardiac-related fatigue is dominant (heart rate changes drive electromyographic changes). The ratio threshold γ can be adjusted according to the specific application scenario, for example, by determining it through cross-validation.

[0040] IV. Dynamic Adversarial Domain Adaptive Module This system employs a federated learning architecture to achieve personalized model optimization while protecting user privacy. The dynamic adversarial domain adaptation module addresses the model generalization problem caused by differences in data distribution among different users in traditional federated learning.

[0041] (a) Federated Learning Framework The system consists of a cloud server and multiple local devices. The cloud maintains the global model parameters Θ. global Maintain personalized model parameters Θ on the local device side local In each training round, the local device trains on local data and then uploads the gradients, which are then aggregated and updated in the cloud.

[0042] (ii) Adaptive Adversarial Domain Traditional federated learning uses a weighted average aggregation (FedAvg) as shown in the following formula (12):

[0043] However, when the data distributions of different users vary significantly, weighted averaging can dilute personalized features. This system introduces an adversarial domain adaptation mechanism, adding a domain discriminator D and a feature extractor F during local training. Through adversarial training, the local feature distribution is aligned with the global feature distribution while preserving personalized information.

[0044] Domain discriminator loss function: (13) In this process, the feature extractor F attempts to "deceive" the domain discriminator, making it impossible for D to distinguish whether the features originate from the local or global domain; the domain discriminator, on the other hand, attempts to accurately distinguish the domain origin. Adversarial training enables F to learn robust features that are invariant to the domain.

[0045] (III) Global Distribution Acquisition Global domain data distribution The statistics, including the mean vector, of the global feature distribution are obtained by the cloud every 24 hours. The covariance matrix Σ global The local device then generates a Gaussian distribution N(...) based on this. ,Σ global ) serves as the source for global domain sampling.

[0046] When the global feature distribution exhibits multimodal or non-Gaussian features, a Gaussian Mixture Model (GMM) can be used for modeling. The statistics delivered from the cloud include the weights, mean, and covariance matrices of each Gaussian component.

[0047] (iv) Privacy protection gradient perturbation To prevent gradient backpropagation from affecting user privacy, this system adds Gaussian noise before uploading local gradients. The noise variance is adaptively adjusted based on the KL divergence between the local and global distributions. (14); in, Let η be the preset base noise variance (0.01 in this embodiment, corresponding to a Gaussian mechanism with a privacy budget ε=1), and let η be the preset privacy-utility tradeoff coefficient (0.5 in this embodiment, determined through grid search on the validation set). The KL divergence is defined as: (15); When the local distribution differs significantly from the global distribution (due to users having unique physiological characteristics), the noise intensity increases, providing stronger privacy protection; when the distribution difference is small, the noise intensity decreases, retaining more useful information. η can be adjusted according to actual privacy requirements and data characteristics.

[0048] V. Energy-Information Flow Collaborative Optimization Module Battery life is crucial for user experience in wearable devices. This system introduces an energy-information flow collaborative optimization module to dynamically determine the sensor's sampling strategy, minimizing energy consumption while ensuring monitoring accuracy.

[0049] (I) Optimization Problem Modeling Suppose the system contains M sensors, the energy consumption of the m-th sensor per sampling is Em (in joules), the current battery charge is Battery(t) (in joules), and the sensor's active state is s. m(t) ∈{0,1} (1 indicates enabled, 0 indicates disabled). The system objective is to maximize the difference between information gain and energy consumption penalty in each sampling period, as shown in the following formula (16):

[0050] The constraints are as follows, as shown in formula (17):

[0051] in, The maximum allowable power is preset (100mW in this embodiment), and λ is the energy-information trade-off coefficient (0.1 in this embodiment).

[0052] (II) Rolling Time Domain Optimization Strategy This system employs a retceding horizontal optimization strategy: each optimization cycle covers a 1-second sampling window, meaning the optimization problem is resolved every second. At the beginning of each cycle, Info(m,t) for each sensor is calculated based on the current Vobs (acquired data), and the optimization problem for that cycle is solved. After sampling is completed, Vobs is updated, and the next cycle begins. This strategy decomposes the original problem into a series of static optimization subproblems.

[0053] (III) Greedy Algorithm Since the number of sensors is limited (M≤8), this system uses a greedy algorithm to solve the problem: 1. Initialize s m(t) =0 (all sensors are disabled), current power P curr =0; 2. Calculate the utility value of each sensor: Um = Info(m,t)·Battery(t) / Em; 3. Arrange the sensors in descending order of Um; 4. Consider each sensor in turn: If P curr +Em≤ Then the sensor (s) is enabled. m =1), P curr =Em; 5. Output the set of sensors enabled.

[0054] The greedy algorithm can be proven to have an approximate ratio of (1-1 / e) for each static subproblem.

[0055] VI. Dynamic Joint Stiffness Estimation Module Joint stiffness is a key parameter for assessing joint stability and injury risk. This system corrects joint stiffness calculations based on the muscle-vascular phase synchronization index, achieving more accurate joint load assessment.

[0056] (a) Passive stiffness estimation Passive joint stiffness K passive Online estimation using a system identification method: Joint angle θ(t) is monitored using IMU data, and external torque τ is estimated using electromyographic signals and a joint torque model. ext(t) By fitting a second-order linear system using least squares: (18); Where I is the moment of inertia and B is the damping coefficient, obtained through system identification. .

[0057] (ii) Active stiffness estimation Active stiffness The contribution of muscle contraction to joint stiffness is estimated based on electromyographic signals: (19); Wherein, γ is a proportionality coefficient (2.5 Nm / ° in this embodiment). This represents the root mean square value of the electromyographic signal. This represents the electromyographic value at maximum voluntary contraction. The γ value can be determined individually through calibration experiments.

[0058] (III) Phase Synchronization Correction Muscle-vascular coupling affects muscle mechanical properties: enhanced phase synchronization improves muscle-vascular coordination and increases active stiffness contribution; conversely, phase desynchronization indicates muscle metabolic disorder and decreased stiffness. This system employs a linear correction model. (20); Wherein, β is a preset coupling coefficient (0.3 in this embodiment), determined through regression analysis of a joint experiment with 10 subjects (simultaneous measurement of EMG, PPG, and joint stiffness). The β value can be calibrated according to different joints or populations.

[0059] (iv) Safety threshold calibration For each user, the system records the initial usage. Historical distribution. The safety threshold is set based on individual history. The 95th percentile of the distribution. In real time... When the threshold is exceeded, a joint overload warning is generated to remind the user to adjust their posture.

[0060] VII. Multimodal Data Fusion Hub The multimodal data fusion hub adopts a graph neural network (GNN) architecture, embedding a causal graph structure into a deep learning model to achieve efficient fusion of multi-source information.

[0061] (a) Graph structure embedding The causal graph G=(V,E) is used as the input graph structure of the GNN. The node set V includes all sensor signal nodes and risk event nodes, and the edge set E has weights w. uv This corresponds to the CMI value.

[0062] (ii) Initial characteristics of nodes Initial feature vector of each node It is composed of the time-domain and frequency-domain statistical characteristics of the corresponding sensor signals: Time-domain features (6 dimensions): mean μ, standard deviation σ, maximum value max, minimum value min, kurtosis, skewness; Frequency domain features (10-dimensional): The power spectral density was calculated using the Welch method (window function: Hamming window, window length 256, overlap 50%), and the power spectral density integral values ​​for 10 frequency bands were extracted: 0-0.5Hz, 0.5-1Hz, 1-2Hz, 2-4Hz, 4-8Hz, 8-12Hz, 12-20Hz, 20-30Hz, 30-40Hz, and 40-50Hz.

[0063] The initial feature vector dimension is 16.

[0064] (III) Message Passing in Graph Neural Networks The message passing layer of GNN is defined as follows: (twenty one); in: N(v) is the neighborhood of node v (including the in-neighbor and out-neighbor neighborhoods). AGGREGATE is an aggregation function that uses weighted summation based on edge weights, as shown in formula (22):

[0065] Let be the weight matrix of the l-th layer. For bias terms; It is the ReLU activation function; The graph neural network contains multiple message passing layers, and the number of layers is determined by performance optimization on the validation set (in this embodiment, L=3 layers).

[0066] (iv) Output and Early Warning The output layer of the graph neural network is a multidimensional health risk assessment vector R=[r ACL ,r cardiac ,r dehydration [,...], where each component takes values ​​in the range [0,1], representing the probability of the corresponding risk event occurring.

[0067] The early warning engine generates differentiated early warning signals based on risk vectors: r > 0.7: Red alert, triggering strong vibration and voice alarm, automatically recording the event and uploading it to the cloud; 0.4 < r ≤ 0.7: Yellow warning, triggering weak vibration + alert sound; r≤0.4: Normal state, only record data and do not issue warnings.

[0068] Compared with the prior art, the present invention has the following beneficial effects.

[0069] 1. This invention is the first to propose the muscle-vascular phase synchronization index. By extracting the instantaneous phase of the electromyographic envelope and PPG signal using Hilbert transform, the neurovascular coupling strength is quantitatively assessed. Based on this index, central and peripheral fatigue can be distinguished, providing a physiological basis for adjusting exercise intensity and overcoming the limitation of existing technologies that only qualitatively describe coupling.

[0070] 2. This invention introduces multimodal temporal causal graphs and conditional mutual information, elevating the early warning mechanism from "correlation judgment" to "causality judgment." By examining whether abnormal electromyography and abnormal kinematics have causal paths pointing to risk events, it effectively suppresses false alarms caused by common causes.

[0071] 3. This invention introduces a domain discriminator into the federated learning framework for adversarial training, which aligns the local model feature distribution with the global distribution while preserving personalized information.

[0072] 4. This invention defines sensor information gain based on information theory, establishes an optimization model in conjunction with energy consumption constraints, and employs a greedy algorithm to dynamically schedule sensor sampling strategies. Compared to fixed-frequency sampling, this invention extends the device's battery life while maintaining the same monitoring accuracy.

[0073] 5. This invention embeds a causal graph structure into a graph neural network, utilizing edge weighted aggregation to transmit causal information and output a multidimensional health risk assessment vector. This architecture integrates prior causal knowledge with data-driven learning, resulting in better interpretability and generalization ability. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0075] Example 1

[0076] This embodiment uses a 35-year-old amateur marathon runner as an example to illustrate the application of the system of the present invention in long-distance running scenarios.

[0077] I. System Deployment Runners wear smart knee braces and smart wristbands that integrate multimodal sensor arrays: Intelligent knee brace: Built-in 8-channel electromyography sensor (rectus femoris, vastus medialis, vastus lateralis, semitendinosus, biceps femoris, medial head of gastrocnemius, lateral head of gastrocnemius, tibialis anterior) and 9-axis IMU (deployed above the patella), with sampling frequencies of 1000Hz and 200Hz respectively.

[0078] Smart wristband: Built-in reflective PPG sensor (dual LEDs with wavelengths of 660nm and 940nm), sweat electrolyte sensor (selective electrodes for sodium and chloride ions), sampling frequency 250Hz.

[0079] II. Monitoring of the Exercise Process Runners trained for a half marathon at a pace of 5:00 / km (simulated test data). The system collected multimodal data in real time and processed it according to the following workflow: (a) Calculation of phase synchronization index At the 45-minute mark of the run, the system collected the following data: Electromyographic signals (rectus femoris) were subjected to full-wave rectification and a 2Hz low-pass filter to extract the envelope. PPG signals were subjected to a 0.5-5Hz band-pass filter. The instantaneous phase was extracted using Hilbert transform, and the phase synchronization index within a 10-second sliding window was calculated.

[0080] The result is 0.82.

[0081] (II) Fatigue Type Determination =0.82 > the first threshold of 0.75, the system judges it as a state of central fatigue.

[0082] (III) Generation of Adjustment Suggestions The system issued a voice prompt through the bone conduction headphones: "Central fatigue, it is recommended to reduce the pace to 5:30 / km, pay attention to hydration and breathing rhythm."

[0083] Runners adjust their pace according to the prompts, and after 5 minutes The level dropped to 0.68, alleviating fatigue.

[0084] (iv) Energy-information flow optimization In this embodiment, the system dynamically adjusts the sensor sampling strategy during the runner's stable running phase (pace change <5%). When Battery(t) = 85%, =100mW. The information gain of each sensor is as follows: sensor <![CDATA[E m(μJ) ]]> Info(m,t) Um=Info·Battery / Em EMG - Rectus femoris 50 0.32 0.544 EMG - Gastrocnemius muscle 50 0.28 0.476 IMU 30 0.45 1.275 PPG 40 0.38 0.8075 Sweat electrolytes 80 0.12 0.1275 Sort by Um: IMU (1.275) > PPG (0.8075) > EMG-rectus femoris (0.544) > EMG-gastrocnemius (0.476) > sweat electrolytes (0.1275). Activate sensors sequentially until the power is close to... .

[0085] Final sampling strategy: IMU, PPG, and EMG-rectus femoris are enabled, while EMG-gastrocnemius and sweat electrolytes are disabled. Compared to full sensor sampling, energy consumption is reduced by approximately 40%, while monitoring accuracy remains above 95%.

[0086] Example 2

[0087] This embodiment uses a 22-year-old professional basketball player as an example to illustrate the causal early warning and main intervention of the system of the present invention for anterior cruciate ligament (ACL) injury in explosive sports scenarios.

[0088] I. System Deployment Athletes wear smart knee and ankle braces integrated with multimodal sensor arrays: Smart knee brace: Deployed around the knee joint, it includes EMG sensors for the quadriceps and hamstring muscles, and an IMU (integrated unit) for knee joint angle.

[0089] Smart ankle support: Deployed around the ankle joint, including EMG sensors for the gastrocnemius and tibialis anterior muscles, and an ankle angle sensor (IMU).

[0090] II. Monitoring of the Exercise Process Athletes practice changing direction and sudden stop jump shots. The system monitors in real time and constructs a cause-effect graph.

[0091] (a) Abnormal state identification During an emergency stop and lane change maneuver, the system detected the following anomaly: The quadriceps EMG amplitude suddenly increased to 4.2 times the individual baseline mean (>3 standard deviations) for 150 ms, which was identified as abnormal electromyographic activation.

[0092] If the knee joint angle is 15° flexion at the moment of landing (normal range 20-40°) and the valgus angle reaches 12° (safe threshold <8°), it is judged as abnormal joint kinematics.

[0093] (II) Determining the Causal Path The system constructs a cause-effect graph for the current time window (the first 5 seconds) and calculates conditional mutual information: =0.087 bits > 0.05 =0.094 bits > 0.05 Paths exist in the causal graph: → →ACL damage.

[0094] (III) Calculation of Causal Path Strength Calculate path strength:

[0095] The MI was calculated using the k-nearest neighbor estimation method (k=10), and the result was 0.45.

[0096] (iv) Electrical stimulation intervention because =0.45> =0.32, the system determines that there is a causally driven risk of ACL damage and triggers active intervention: ΔT before muscle natural activation stim The timing (calculated as 85ms using a neuromechanical delay compensation model) involves stimulating the hamstrings with electrical stimulation electrodes to promote synergistic contraction and stabilize the knee joint.

[0097] Electrical stimulation parameters: frequency 50Hz, pulse width 200μs, intensity 1.2 times the motor threshold (approximately 35mA). These parameters were determined based on the safe and effective parameter range reported in the literature.

[0098] Simultaneously, it provides tactile feedback (vibration) and a voice prompt: "Pay attention to knee stability, you are about to land."

[0099] (v) Effect verification Subsequent high-speed camera analysis showed that during the landing maneuver, the athlete's knee valgus angle was reduced from the predicted 15° to 7° due to electrical stimulation intervention, successfully avoiding the risk of ACL injury.

[0100] Example 3

[0101] This embodiment uses a user group of 30 users as an example to illustrate the model training and performance verification of the system of the present invention under the federated learning framework.

[0102] I. Experimental Design Number of participants: 30 healthy adult volunteers (15 males and 15 females, aged 22-45 years). Experimental equipment: customized smart knee brace and wristband, sampling frequency as described above; Data annotation: Risk events (ACL injury risk, myocardial ischemia risk, dehydration risk) were independently annotated by 3 sports medicine experts, and the majority consensus results were used as the gold standard; Statistical methods: Paired t-tests were used to compare the false alarm rate and F1-score of the same subject under different methods. p < 0.05 was considered statistically significant. Sample size: A total of 12,000 valid data samples were collected, of which 1,200 were risk event samples (accounting for 10%).

[0103] II. False Alarm Rate Comparison Experiment Thirty participants were selected and performed progressively heavier exercise on a treadmill until exhaustion, while data was collected simultaneously. Three early warning methods were compared: 1. Traditional threshold method: Warning when heart rate > 180 bpm or joint angle > threshold. 2. Multimodal fusion method (D1): Torque comparison early warning based on EMG+IMU fusion 3. Causal Early Warning Method of this Invention Experimental results: method True positive rate False positive rate (false alarm rate) Traditional threshold method 72.3% 28.5% Multimodal fusion method (D1) 85.1% 15.2% This invention is a causal early warning method. 89.4% 8.7% The false alarm rate of this invention is 42.7% lower than that of D1 ((15.2%-8.7%) / 15.2%≈42.8%), and 69.5% lower than that of the traditional threshold method. The difference is statistically significant (paired t-test, p<0.01).

[0104] III. Battery Life Improvement Experiment On the same hardware platform (MTK2503, 400mAh battery), the battery life was compared between fixed frequency sampling (100Hz across all sensors) and the energy-information flow collaborative optimization sampling of this invention: Sampling strategy Battery life (hours) Monitoring accuracy (F1-score) Fixed frequency sampling 8.2 0.94 This invention optimizes sampling 11.1 0.91 This invention improves battery life by 35.4% while reducing monitoring accuracy by only 3.2%, which is within an acceptable range. The difference in battery life improvement is statistically significant (paired t-test, p<0.01).

[0105] IV. Personalized Precision Improvement Experiment Federated learning was trained on 30 users, and the model accuracy of the adversarial domain adaptation method of this invention was compared with that of traditional FedAvg: User type Traditional FedAvg F1-score This invention's F1-score promote Typical users (18 people) 0.87 0.91 +4.6% Special users (12 people) 0.71 0.86 +21.1% average 0.81 0.92 +13.6% The results show that this invention significantly improves the accuracy of personalized monitoring while maintaining overall generalization ability, especially for users with unique physiological characteristics. The average improvement of 13.6% is statistically significant (paired t-test, p<0.01).

[0106] V. Experiment on the accuracy of fatigue type differentiation Twenty subjects were selected and subjected to high-intensity continuous running (inducing central fatigue) and localized weighted squats (inducing peripheral fatigue). The type of fatigue was determined by exercise physiology experts based on gold standards such as lactate concentration and muscle oxygen recovery rate.

[0107] This invention is based on Fatigue type determination results: Actual type Determined as central Determined as peripheral accuracy Central fatigue (10 people) 9 1 90% Peripheral fatigue (10 people) 1 9 90% Actual type Determined as central Determined as peripheral accuracy An overall accuracy rate of 90% proves It can effectively distinguish between two types of fatigue.

[0108] The sports and health monitoring and early warning system based on smart wearable devices provided by this invention can be widely applied in the following fields: 1. Professional Athlete Training Monitoring: Real-time monitoring of athletes' fatigue status and injury risk, providing personalized training adjustment suggestions to reduce the incidence of sports injuries.

[0109] 2. General Fitness and Rehabilitation: Provides scientific guidance for general fitness enthusiasts to prevent overexertion and sports injuries, and is especially suitable for cardiovascular risk warning for middle-aged and elderly people.

[0110] 3. Telemedicine and Health Management: Protecting user privacy through a federated learning architecture while enabling group health data analysis to provide data support for chronic disease management.

[0111] 4. Military and Special Operations: Used for soldier physical fitness monitoring and heatstroke early warning, improving personnel safety in extreme environments.

[0112] The technical solution of this invention has a complete hardware implementation path and software development framework, and can utilize existing smart wearable device hardware platforms (such as Apple Watch, Huawei Watch, Xiaomi Band, etc.) to expand its functions, and has good prospects for industrial application.

[0113] Those skilled in the art should understand that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., 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 sports health monitoring and early warning system based on smart wearable devices, characterized in that, include: A multimodal sensor array is configured to simultaneously acquire the wearer's electromyography (EMG) signals, photoplethysmography (PPG) signals, and inertial measurement unit (IMU) data. The muscle-vascular coupling analysis module is configured to calculate the muscle-vascular phase synchronization index. : ； Where H is the Hilbert transform. The envelope of the electromyographic signal. The signal is a photoplethysmogram. The muscle-blood vessel coupling analysis module is configured such that: when When the value exceeds the first threshold, it is determined to be a state of central fatigue. If the intensity is below the second threshold, it is determined to be a peripheral fatigue state, and differentiated exercise intensity adjustment suggestions are generated based on the fatigue type.

2. The system according to claim 1, characterized in that, Also includes: The hemodynamic causal inference module is configured to construct a multimodal temporal causal graph. ,in: Node set ; The risk events include at least one of the following: anterior cruciate ligament injury, myocardial ischemia, and dehydration; Edge set Causal discovery through conditional mutual information: ； Where X and Y are any two physiological or motor variables, and Z is a set of conditional variables; The hemodynamic causal inference module is configured to generate an early warning signal when a causal path from abnormal electromyographic activation mode through abnormal joint kinematics to a risk event exists in the causal graph; otherwise, false alarms caused by correlation are suppressed.

3. The system according to claim 1, characterized in that, Also includes: The dynamic adversarial domain adaptation module is configured to perform local model personalized training under the federated learning architecture; The dynamic adversarial domain adaptive module includes a domain discriminator D and a feature extractor F, and its training loss function is: ； in, Data distribution for local users, Data distribution generated for global feature distribution statistics distributed from the cloud; By minimizing the difference in feature distribution between the local and global domains through adversarial training, the accuracy of personalized monitoring can be improved while maintaining generalization ability.

4. The system according to claim 1, characterized in that, Also includes: The energy-information flow collaborative optimization module is configured to dynamically determine the sampling strategy of the multimodal sensing array; The energy-information flow collaborative optimization module determines the activation status of each sensor by solving the following optimization problem. : ； Constraints: ； in, Let m be the energy consumption of a single sampling of the m-th sensor. This is the current remaining battery level. For sensor information gain calculated based on causal contribution, This is the energy-information tradeoff coefficient. This is the preset maximum allowable power.

5. The system according to claim 2, characterized in that, The risk events mentioned include the risk of anterior cruciate ligament injury; The hemodynamic causal inference module is configured to calculate the causal path strength. ； in, To establish a causal path from abnormal EMG nodes to ACL injury nodes, mutual information MI is calculated using the k-nearest neighbor estimation method, and information entropy H is calculated based on the same estimation framework. when When the risk exceeds a preset causal threshold, a causally driven risk of injury is identified, and electrical stimulation intervention is triggered.

6. The system according to claim 2, characterized in that, The hemodynamic causal inference module is also configured to calculate conditional transition entropy to distinguish between cardiac fatigue and peripheral fatigue. ； Where X represents electromyographic signal characteristics, Y represents heart rate variability characteristics, and Z represents blood oxygen saturation; when and When the ratio is greater than the preset ratio threshold and the permutation test is statistically significant, it is determined that muscle-related fatigue is dominant; otherwise, it is determined that cardiac-related fatigue is dominant.

7. The system according to claim 3, characterized in that, The dynamic adversarial domain adaptive module also includes: A privacy-preserving gradient perturbation mechanism adds noise before uploading the local gradient. ; The noise variance Adaptive adjustment based on the KL divergence between local and global data distributions: ； in, The pre-defined basis noise variance is... To pre-determine the privacy-utility tradeoff coefficient, the KL divergence is calculated using the k-nearest neighbor estimation method to achieve differentiated privacy protection.

8. The system according to claim 1, characterized in that, Also includes: The dynamic joint stiffness estimation module is configured to correct joint stiffness calculations based on the muscle-blood vessel phase synchronization index. ； in, The passive joint stiffness is estimated using the system identification method based on inertial measurement unit data. For the active stiffness contribution estimated based on electromyography signals, The preset coupling coefficient; when When the joint overload warning is exceeded, a joint overload warning is generated.

9. The system according to claim 4, characterized in that, The sensor information gain Based on causal contribution calculation: ； in, Let m be the current reading of the m-th sensor. For the target set of predictor variables, This is the set of sensor data collected within the current sampling period; Sensors are prioritized when the information gain is greater than a preset information gain threshold, ensuring that sampling resources are focused on causally correlated signals.

10. The system according to any one of claims 1 to 9, characterized in that, Also includes: The multimodal data fusion hub is configured to input electromyography signals, photoplethysmography signals, inertial measurement unit data, and causal graph structures into the graph neural network; The message passing layer of the graph neural network is defined as follows: ； in, Let v be the neighborhood of node v in the causal graph, and let AGGREGATE be the aggregation function based on the weights of the causal edges. The output of the graph neural network is a multidimensional health risk assessment vector, which is used to drive the early warning engine to generate differentiated early warning signals.

Images