Evaluation method for centrifugal force of North European popliteal muscle test based on deep learning

Abstract

The invention discloses a method for evaluating centrifugal force of a North European popliteal muscle test based on deep learning, and relates to the technical field of sports biomechanics evaluation, and the method comprises the following steps: S1, data acquisition: a near-infrared motion capture system and four three-dimensional dynamometers synchronously acquire kinematics and dynamics data and time-align the kinematics and dynamics data; s2, simulation and database construction: constructing an individualized muscle-bone model based on OpenSim, obtaining a knee joint torque true value through inverse kinematics and inverse dynamics analysis, and constructing a standardized three-dimensional time series data set; s3, constructing a model, and realizing knee joint moment prediction by adopting a double-layer BiLSTM network fused with biomechanical physical constraints; s4, training optimization: standardizing the preprocessed data and improving the model performance by adopting a specific strategy; and S5, evaluation calculation: verifying the model through RMSE, Rand MAE, and calculating the centrifugal force of the popliteal muscles in combination with the arm-of-force constants. The deep learning time sequence modeling capability and the biomechanical physical constraint are considered, low-cost popularization is realized, the precision is guaranteed, and the method is suitable for various mechanisms such as sports, rehabilitation and fitness.

Description

Technical Field

[0001] This invention relates to the field of sports biomechanics assessment technology, specifically a deep learning-based method for assessing eccentric force in the Nordic hamstring test. Background Technology

[0002] With the development of computer technology, machine learning evaluation methods based on data such as label-free motion capture, electromyography (EMG) signals, and ground reaction force (GRF) have gradually emerged. For example, Luh, Song, and others used elbow joint angles, angular velocities, and EMG signals obtained from label-free motion capture to establish a feedforward artificial neural network to predict elbow joint torque. However, due to the difficulty of label-free motion capture and EMG signal acquisition, and the high requirements for experimental environments, research using GRF data to replace motion capture systems has gradually emerged in recent years.

[0003] However, due to the difficulties in label-free motion capture and electromyography (EMG) signal acquisition, and the high requirements for experimental environments, research using GRF data to replace motion capture systems has gradually emerged in recent years. Ardestani et al. combined GRF and EMG signals to predict hip and knee joint torques during human walking; LiuYu et al. constructed a GRF-based neural network model to analyze lower limb joint torques during jumping, verifying the application potential of GRF data in muscle force analysis. However, existing GRF-based models mostly use feedforward neural networks, relying on manual extraction of statistical features, making it difficult to capture the temporal dynamic correlations during movement. Furthermore, specialized modeling for eccentric hamstring force is not yet mature, and it is still unable to accurately predict hamstring force, failing to meet the refined assessment needs for sports injury prevention and performance improvement.

[0004] In addition, in the field of sports biomechanics, the traditional isokinetic dynamometer, as the "gold standard" for assessing hamstring strength, can measure the eccentric, concentric and isometric forces of muscles at different angular velocities with high precision. However, it is limited by the high cost of the equipment, the complex operation process and the high technical requirements for the operators, making it difficult to popularize and apply in ordinary sports colleges, amateur sports teams and mass fitness scenarios.

[0005] Therefore, a deep learning-based method for assessing eccentric strength in the Nordic hamstring test is proposed to address the above issues. Summary of the Invention

[0006] 1. The technical problem this solution aims to solve: In view of this, the technical problem to be solved by the present invention is to propose a deep learning-based method for evaluating eccentric force in the Nordic hamstring test, in order to solve the following technical problems in the prior art: (1) Insufficient accuracy: Existing assessment models based on ground reaction force (GRF) mostly use feedforward neural networks and rely on manual extraction of statistical features. They are difficult to capture the temporal dynamic correlation of force signals in hamstring testing and lack specialized modeling for eccentric hamstring force. They cannot accurately predict knee joint torque and corresponding muscle strength, making it difficult to meet the refined assessment needs for sports injury prevention and performance improvement. At the same time, traditional general models are not fully adapted to the body characteristics and movement habits of different subjects. General parameters are prone to error accumulation, which further affects the accuracy of assessment.

[0007] (2) Difficulty in popularization and application: The traditional "gold standard" for hamstring strength assessment, the isokinetic dynamometer, has limitations such as high equipment cost, complex operation procedures, and high requirements for the technical level of operators, making it difficult to promote in ordinary sports colleges, amateur sports teams, and mass fitness scenarios. On the other hand, markless motion capture, electromyography signal acquisition and other technologies face the problems of difficult data collection and strict requirements for experimental environment, which also limit their widespread application, resulting in a lack of convenient and feasible hamstring strength assessment methods in most scenarios.

[0008] 2. Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for evaluating eccentric strength in the Nordic hamstring test, specifically comprising: S1, Data Acquisition: Kinematic and dynamic data of the subject during the Nordic hamstring test are acquired synchronously through a near-infrared motion capture system and four three-dimensional force platforms, and the data time alignment is completed synchronously. S2, Simulation and Database Construction: Based on OpenSim software, an individualized musculoskeletal model is constructed. The true value of knee joint torque is obtained through inverse kinematics and inverse dynamics analysis. A standardized three-dimensional time series dataset is constructed and divided into training set, validation set and test set. S3, Deep learning model construction: A two-layer bidirectional long short-term memory network (BiLSTM) combined with biomechanical physical constraints is used to predict knee joint torque by sequentially passing through the input layer, temporal feature extraction layer, temporal distribution fully connected layer and output layer; S4, Model Training and Optimization: Standardize and preprocess the data, and use specific training strategies and optimization techniques to improve model performance; S5, Model Evaluation and Force Calculation: The model effect is verified by preset evaluation indicators, and the eccentric force of the hamstring muscle is calculated by combining the human lever arm constant with the predicted knee joint torque.

[0009] Preferably, the simulation and database construction described in S2 specifically includes: S2.1, Construction of individualized musculoskeletal model: Based on the Raabe Full-Body Model template of OpenSim software, the subject's weight parameters are input, and an individualized musculoskeletal model adapted to the subject's body shape is generated through the model scaling function; S2.2, Inverse kinematics analysis: Import the motion capture data collected in S1 into OpenSim software to calculate the curves of the change of joint angles over time, including the knee flexion angle θ(t). S2.3, Inverse Dynamics Analysis: Combining the ground reaction force time series collected in S1, the net joint torque M(t) of the knee joint is calculated using OpenSim software and used as the true value for model training; S2.4, Dataset Construction: Integrate the time series data of the x, y, and z components of Fknee(t) and Fankle(t) with the corresponding M(t) to form a three-dimensional time series dataset. The dataset shape is [number of samples, time steps, number of features], where the number of features is 12. S2.5, Dataset Partitioning and Standardization: Divide the dataset into training, validation, and test sets according to a ratio of 80%, 10%, and 10%, respectively, and normalize all data to 100 time steps to ensure coverage of the entire eccentric contraction of the hamstrings.

[0010] As a preferred embodiment, the deep learning model construction described in S3 specifically includes: S3.1, Input layer construction: Receives time series data of the x, y, and z components of Fknee(t) and Fankle(t), performs standardization processing on the data, and converts it into a three-dimensional tensor in the format of [number of samples, time steps, number of features], where the time steps are 100 and the number of features is 12; S3.2, Construction of Temporal Feature Extraction Layer: A two-layer bidirectional LSTM network is built, with 64 memory units in each layer; the first layer of bidirectional LSTM captures the forward and backward dependencies of temporal data through forward and backward neurons and returns the output of each time step; the second layer of bidirectional LSTM performs high-level refinement on the features output by the first layer, and adds the dropout function to achieve regularization and prevent the model from overfitting. S3.3, Construction of Temporal Distributed Fully Connected Layer: A temporal distributed fully connected layer is set after the temporal feature extraction layer. It contains 32 neurons and uses ReLU as the activation function. The fully connected operation is performed independently on the LSTM output at each time step to achieve feature transformation and dimensionality reduction. S3.4, Incorporation of Biomechanical and Physical Constraints: Based on a single-degree-of-freedom model of the knee joint (allowing only rotation within the sagittal plane), a torque equation is constructed. The inertial torque M0 is calculated using the tension F at the subject's ankle, the distance L1 from the knee to the center of mass, the lower leg length L2, the gravity G on the segment above the lower leg, the lower leg gravity Gshank (6% of the subject's total weight), and the angle θ between the body and the horizontal plane. M0 is then incorporated into the model loss function, with the specific formula as follows:

[0011] In the formula, F is the tension force at the ankle of the subject, L1 is the distance from the knee to the center of mass, L2 is the length of the lower leg, G is the gravity of the segment above the lower leg of the subject, Gshank is the gravity at the lower leg of the subject, and is the angle between the subject's body and the horizontal plane. S3.5, Output layer construction: Set the output layer as a time series prediction structure, and output a three-dimensional tensor in the format of [number of samples, 100, 2], directly outputting the knee joint torque time series M(t).

[0012] Preferably, the model training and optimization described in S4 specifically includes: S4.1, standardize the input features using Z-score and normalize the output true value; S4.2 employs the Adam optimizer with an initial learning rate of 0.0001, a weight decay of 1e-4, and a cosine annealing learning rate scheduling strategy. S4.3, the loss function is a weighted combination of mean squared error (MSE) and M0, an early stopping mechanism is introduced, if the loss does not improve for 80 consecutive rounds, training stops, the batch size is set to 8, the maximum number of training rounds is 500, and the maximum gradient norm of gradient clipping is limited to 0.5; S4.4 calculates the weights of each input feature based on mutual information regression, thereby improving the model's sensitivity to key force signals.

[0013] As a preferred method, the model evaluation and force calculation of S5 are evaluated using three indicators on the test set: root mean square error (RMSE), coefficient of determination (R²), and mean absolute error (MAE). The specific formula for the root mean square error (RMSE) is as follows:

[0014] In the formula, n represents the total number of time steps in the sample sequence; y i Represented as the first i True value of knee joint torque at each time step; Represented as the model's predicted first... i Predicted knee joint torque values ​​at each time step.

[0015] The specific formula for the coefficient of determination (R²) is as follows:

[0016] The specific formula for the mean absolute error (MAE) is as follows:

[0017] The calculation of hamstring eccentric force is based on the formula: hamstring force = knee joint torque × lever arm, where the human lever arm is a measurable constant.

[0018] Compared with existing technologies, the present invention provides a deep learning-based method for evaluating eccentric strength in the Nordic hamstring test, which has the following advantages: I. Prediction accuracy has been significantly improved; This approach overcomes the limitations of traditional feedforward neural networks that rely on manual feature extraction. It employs a two-layer bidirectional LSTM network to accurately capture the temporal dynamics and forward-backward dependencies of force signals in the Nordic hamstring test. Furthermore, it incorporates biomechanical constraints, ensuring the prediction process conforms to mechanical principles and effectively reducing errors. Experimental data show that the predicted R² for left knee torque reaches 0.867, and for right knee, it reaches 0.844, significantly outperforming the specialized prediction performance of existing general-purpose models.

[0019] Second, it has a stronger ability to adapt to individual needs; Based on the OpenSim software, an individualized musculoskeletal model is constructed. By scaling the subject's weight parameter to adapt to different body characteristics, and combining mutual information regression to optimize the input feature weights, the model can automatically adapt to the movement habits and physical conditions of different subjects. This overcomes the error accumulation problem caused by general parameters in traditional methods and improves the personalization and reliability of the evaluation results.

[0020] Third, the testing threshold has been significantly lowered; Without relying on expensive isokinetic dynamometers or complex label-free motion capture and electromyography (EMG) signal acquisition equipment, data acquisition can be completed using only four 3D force platforms and simple fixing devices. The equipment is low-cost and easy to operate, requiring no professional technicians to operate it throughout the process, thus solving the pain point that traditional assessment methods are difficult to popularize in sports colleges, amateur sports teams, and public fitness scenarios.

[0021] IV. The model balances interpretability and practicality; By incorporating the inertial torque derived from the knee joint biomechanics equations into the loss function, the model output not only possesses high accuracy but also aligns with biomechanical principles, enhancing the interpretability of the results. Furthermore, through standardized data processing and model training procedures, the eccentric force of the hamstrings can be directly calculated from the predicted knee joint torque combined with the human lever arm constant, meeting the refined assessment needs for sports injury prevention, rehabilitation training, and athletic performance improvement. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the process for evaluating the eccentric force of the Nordic hamstring test based on deep learning, according to the present invention. Figure 2 This is a schematic diagram of the human skeleton collection points of the present invention; Figure 3 This is a schematic diagram illustrating the simulation process performed by the OpenSim software of this invention. Figure 4 This is a schematic diagram of the experimental scenario for the Nordic hamstring test of this invention; Figure 5 This is a schematic diagram of the LSTM-based hamstring strength assessment network structure of the present invention. Figure 6 This is a schematic diagram simulating human skeletal movement according to the present invention; Figure 7 This is a schematic diagram of the feature weight distribution based on mutual information according to the present invention; Figure 8 This is a schematic diagram illustrating the joint torque prediction performance of the model of the present invention; Figure 9 This is a schematic diagram of the simulation prediction of left knee sample collection in this invention; Figure 10 This is a schematic diagram of the simulation prediction of right knee sample collection according to the present invention. Detailed Implementation

[0023] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0025] For an example, please refer to... Figures 1 to 10 As shown: To address the problems mentioned in the technical solutions, this application provides a deep learning-based method for assessing eccentric strength in the Nordic hamstring test. Step 1: Experimental data collection; For subject preparation, 39 reflective markers were affixed to key anatomical landmarks of the lower limbs (anterior superior iliac spine, lateral epicondyle of the femur, lateral malleolus, etc.). These markers were used by the motion capture system to accurately track the real-time movement trajectory of each joint in the lower limbs.

[0026] Equipment configuration and data synchronization are crucial. To ensure accurate data acquisition, the testing environment must maintain a constant indoor temperature of 20-25℃ and relative humidity of 40%-60%, free from strong light interference. The ground must be flat, with a height difference not exceeding 2mm. A 5m x 5m testing area must be reserved to ensure sufficient space for the subject to perform movements. An 8-camera near-infrared motion capture system (such as Optitrack) is used to capture joint motion trajectories at a frequency of 200Hz. Four 3D force platforms (such as Kistler) are deployed simultaneously at a frequency of 1000Hz. Two force platforms are placed below the subject's knee joint to collect the vertical ground reaction force Fknee(t) in real time. The other two force platforms are connected to the subject's ankle via hooks to collect the ground reaction force Fankel(t) corresponding to the ankle tension. After equipment deployment, strict time synchronization calibration is performed on the motion capture system and force platforms to ensure complete consistency between the timestamps of the kinematic trajectory data and the dynamic force signals, guaranteeing data timing matching.

[0027] Action execution and data recording; Subjects knelt in a standardized posture at the center of the force testing platform, keeping their knees flexed at 90° and their bodies upright. After stabilizing their ankles, they performed the Nordic hamstring test at a slow and controlled pace until their bodies leaned forward and their palms touched the ground. Two types of core experimental data were collected simultaneously throughout the entire movement: kinematic parameters such as lower limb joint angles and angular velocities, and dynamic parameters such as ground reaction forces at the knee and ankle joints.

[0028] Step 2: Musculoskeletal simulation analysis and dataset construction based on OpenSim; 1. Construction of individualized musculoskeletal models; Using the Raabe Full-Body Model built into the OpenSim software as the basic template, and taking the subject's weight as the core scaling parameter, we completed the generation and adaptation of the individualized musculoskeletal model of the subject's lower limb.

[0029] 2. Simulation analysis and calculation; (1) Inverse kinematics analysis: The lower limb motion trajectory data collected by the motion capture system is imported into the musculoskeletal model, and the dynamic change curves of each joint angle over time are calculated, such as the knee flexion angle time series curve θ(t).

[0030] (2) Inverse dynamics analysis: Combining the ground reaction force time series data Fknee(t) and Fanckle(t) collected by the force measuring table, the net joint torque time series curve M(t) of the knee joint is obtained by dynamic inversion calculation. This torque value is used as the true label value (true value) for subsequent deep learning model training.

[0031] 3. Construction of time-series datasets; By integrating the Fknee(t), Fangle(t), and the corresponding true value M(t) of the knee joint torque obtained from simulation analysis, a three-dimensional time-series dataset was constructed. The tensor dimension of the dataset was defined as "number of samples - number of time steps - number of features", with 12 features. The dataset was split and standardized: it was divided into a training set (80%), a validation set (10%), and a test set (10%) in an 8:1:1 ratio; all time-series data were uniformly normalized to 100 time steps, fully covering the entire process of hamstring eccentric contraction.

[0032] Step 3: Building a deep learning model that integrates biomechanical and physical constraints This study constructs a two-layer bidirectional long short-term memory network (BiLSTM) to integrate biomechanical physical constraints into the network modeling process, thereby achieving accurate prediction of knee joint torque based on the temporal characteristics of force signals.

[0033] Construction of biomechanical and physical constraints; Based on the mechanical characteristics of the Nordic hamstring test, the knee joint is simplified into a single-degree-of-freedom model, considering only rotational motion in the sagittal plane. A torque equation is established based on the principle of mechanical equilibrium. Here, F represents the tension at the subject's ankle, L1 is the distance from the knee joint to the body's center of mass, L2 is the lower leg length, G is the weight of the subject's body segment above the lower leg, Gshank is the weight of the lower leg, and θ is the angle between the subject's body and the horizontal plane. The inertial torque M0 generated by angular acceleration is calculated by solving the mechanical equation. This inertial torque M0 is incorporated as a core physical constraint into the model's loss function, thereby improving the physical rationality of the torque prediction and the model's prediction accuracy.

[0034] Design of core modules for network architecture; 1. Input layer: Time series data standardization processing The model input features are: time-series data of the x, y, and z-axis components of the vertical reaction force Fknee(t) collected by the force platform below the knee joint and the tensile force Fankle(t) collected by the force platform at the ankle joint. The input data format is a three-dimensional tensor sample number time step feature number. All time-series data are uniformly normalized to 100 time steps, and the number of features is fixed at 12.

[0035] 2. Temporal Feature Extraction Layer: Two-layer bidirectional LSTM network As the core feature extraction unit of the model, a two-layer cascaded bidirectional LSTM structure is used to complete the temporal feature mining of the force signal. The number of memory units in both layers is set to 64. The first layer of bidirectional LSTM fully captures the forward and backward temporal dependencies of force signals through parallel computation of forward and backward neurons, such as the continuous change trend of force signals during the process of leaning forward. The network is configured to return sequence mode, retaining the feature output of each time step, and providing complete temporal information for the feature extraction of the next layer.

[0036] The second layer, a bidirectional LSTM, further refines the high-level temporal features of the force signal based on the features of the first layer, such as the acceleration changes and periodic fluctuation patterns of the force signal; the dropout regularization mechanism is embedded in the network to effectively suppress model overfitting.

[0037] 3. Temporally Distributed Fully Connected Layers: Feature Transformation and Dimensionality Optimization A time-distributed fully connected layer is added after the two-layer bidirectional LSTM layer. The fully connected transformation is performed independently on the LSTM feature vector output at each time step. This layer has 32 neurons and uses ReLU as the activation function to complete the nonlinear mapping and dimensionality optimization of the features, thereby preserving the effective temporal features while reducing the computational complexity.

[0038] 4. Output layer: Knee joint torque timing prediction The feature output of the time-distributed fully connected layer is linearly mapped to finally output the time-series prediction result M(t) of the knee joint torque; the output data format is three-dimensional tensor sample number time step feature number, where the time step is 100 and the feature number is 2.

[0039] Step 4: Model Training and Optimization Strategies 1. Data Preprocessing Z-score standardization is performed on the input feature data of the model to eliminate the impact of dimensional differences on model training; normalization is performed on the true value of the knee joint torque at the output end to improve the model's convergence efficiency and prediction stability.

[0040] 2. Core Strategies for Model Training (1) Optimizer and learning rate scheduling: The Adam optimizer is used for gradient updates, with an initial learning rate of 0.0001 and a weight decay coefficient of 1×10⁻⁶. 4; Combined with cosine annealing learning rate scheduling strategy, the learning rate is adaptively adjusted to balance the model convergence speed and optimization accuracy. (2) Loss function design: The weighted combination loss function of mean squared error (MSE) and moment of inertia M0 is adopted to ensure the numerical fit between the predicted value and the true value, and to ensure the biomechanical rationality of the prediction result through the physical constraint term M0. (3) Overfitting suppression strategy: An early stopping mechanism is introduced. When the validation set loss has not improved for 50 consecutive rounds, the training is terminated immediately. At the same time, a gradient clipping strategy is applied to limit the maximum gradient norm to 0.5 to avoid the gradient explosion problem. (4) Training hyperparameter settings: The batch training size is 8, and the maximum training rounds of the model are 300.

[0041] 3. Input Feature Weight Optimization The mutual information regression algorithm is used to calculate the weight coefficients of each input feature. Based on the weights, the features are weighted and optimized, which effectively improves the model's sensitivity to key force signal features and further optimizes the model's prediction performance.

[0042] Step 5: Model Evaluation and Application 1. Evaluate model performance on the test set using the following metrics: Root Mean Square Error (RMSE)

[0043] In the formula, n represents the total number of time steps in the sample sequence; y i Represented as the first i True value of knee joint torque at each time step; Represented as the model's predicted first... i Predicted knee joint torque values ​​at each time step.

[0044] Coefficient of determination (R²)

[0045] Mean Absolute Error (MAE)

[0046] The calculation of hamstring eccentric force is based on the formula: hamstring force = knee joint torque × lever arm, where the human lever arm is a measurable constant.

[0047] Step 6: Experimental Results and Analysis; 1. Data scale: 30 subjects, 6 replicates per subject, totaling 180 samples. 2. Model performance; like Figure 7 As shown, after calculation based on mutual information, the Vy component of the force in the vertical direction has the largest weight among the four force platforms. like Figure 8 As shown, right knee joint torque prediction: RMSE=13.106N·m, R²=0.844, MAE=9.06N·m like Figure 9 As shown, the predicted torque of the left knee joint is as follows: RMSE=13.646N·m, R²=0.867, MAE=9.39N·m In summary, based on the above tests, this solution has the following advantages over existing technologies: (1) The prediction accuracy is significantly better than existing models, meeting the requirements for refined evaluation: Experimental data show that the determination coefficient R² of this solution for predicting the left knee torque is 0.867, and for the right knee it is 0.844, with root mean square errors (RMSE) of 13.646 N respectively. m and 13.106N m, mean absolute error (MAE) is 9.39N. m and 9.06N Compared to existing GRF models based on feedforward neural networks (which rely on manual feature extraction, struggle to capture temporal correlations, and have poor specialized prediction performance), this solution accurately mines the forward and backward temporal dependencies of force signals through a two-layer bidirectional LSTM network. Furthermore, it incorporates biomechanical physical constraints to ensure predictions conform to mechanical laws, significantly reducing prediction errors. Its R² value is significantly higher than the specialized prediction level of existing general-purpose models, accurately supporting the refined assessment needs for sports injury prevention and performance improvement.

[0048] (2) Achieving high-precision assessment with low-cost equipment, breaking through the bottleneck of widespread application: In the existing technology, although the isokinetic force measuring instrument has high accuracy, it is expensive and complicated to operate, making it difficult to popularize in ordinary sports colleges, amateur sports teams and mass fitness scenarios; while the labelless motion capture and electromyography signal acquisition technology have strict requirements for experimental environment and operation, and data acquisition is difficult. This solution completes data acquisition with only four three-dimensional force measuring platforms and simple fixing devices. The equipment cost is low, the operation process is simple, and no professional technicians are required to operate it throughout the process, but it achieves assessment accuracy similar to traditional high-precision equipment (R² of both left and right knees exceeds 0.84), successfully breaking through the bottleneck of "high precision and high threshold coexistence" of the existing technology, and enabling the assessment of hamstring eccentric strength to be widely used in more scenarios.

[0049] (3) Strong individualized adaptability and more reliable assessment results: Existing general models mostly use fixed parameters, which are prone to error accumulation due to differences in subject body shape and movement habits. This scheme constructs an individualized musculoskeletal model based on OpenSim, adapts to different body shapes by scaling the weight parameter, and optimizes the input feature weights by combining mutual information regression, so that the model can automatically adapt to the individual differences of different subjects. From the experimental results, the goodness of fit between the predicted value and the true value is high for both the left and right knees (R² is over 0.84), and there is no large deviation due to individual differences. Compared with existing general models, the personalization and reliability of the assessment results are significantly improved.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for assessing eccentric strength in the Nordic hamstring test, characterized in that: S1, Data Acquisition: Kinematic and dynamic data of the subject during the Nordic hamstring test are acquired synchronously through a near-infrared motion capture system and four three-dimensional force platforms, and the data time alignment is completed synchronously. S2, Simulation and Database Construction: Based on OpenSim software, an individualized musculoskeletal model is constructed. The true value of knee joint torque is obtained through inverse kinematics and inverse dynamics analysis. A standardized three-dimensional time series dataset is constructed and divided into training set, validation set and test set. S3, Deep learning model construction: A two-layer bidirectional long short-term memory network (BiLSTM) combined with biomechanical physical constraints is used to predict knee joint torque by sequentially passing through the input layer, temporal feature extraction layer, temporal distribution fully connected layer and output layer; S4, Model Training and Optimization: Standardize and preprocess the data, and use specific training strategies and optimization techniques to improve model performance; S5, Model Evaluation and Force Calculation: The model effect is verified by preset evaluation indicators, and the eccentric force of the hamstring muscle is calculated by combining the human lever arm constant with the predicted knee joint torque.

2. The method for evaluating eccentric strength of the Nordic hamstring test based on deep learning according to claim 1, characterized in that, The simulation and database construction described in S2 specifically include: S2.1, Construction of individualized musculoskeletal model: Based on the Raabe Full-Body Model template of OpenSim software, the subject's weight parameters are input, and an individualized lower limb musculoskeletal model adapted to the subject's body shape is generated through the model scaling function; S2.2, Inverse kinematics analysis: Import the motion capture data collected in S1 into OpenSim software to calculate the curves of the change of joint angles over time, including the knee flexion angle θ(t). S2.3, Inverse Dynamics Analysis: Combining the ground reaction force time series collected in S1, the net joint torque M(t) of the knee joint is calculated using OpenSim software and used as the true value for model training; S2.4, Dataset Construction: Integrate the time series data of the x, y, and z components of Fknee(t) and Fankle(t) with the corresponding M(t) to form a three-dimensional time series dataset. The dataset shape is [number of samples, time steps, number of features], where the number of features is 12. S2.5, Dataset Partitioning and Standardization: Divide the dataset into training, validation, and test sets according to a ratio of 80%, 10%, and 10%, respectively, and normalize all data to 100 time steps to ensure coverage of the entire eccentric contraction of the hamstrings.

3. The method for evaluating eccentric strength in the Nordic hamstring test based on deep learning as described in claim 1, characterized in that, The deep learning model construction described in S3 specifically includes: S3.1, Input layer construction: Receives time series data of the x, y, and z components of Fknee(t) and Fankle(t), performs standardization processing on the data, and converts it into a three-dimensional tensor in the format of [number of samples, time steps, number of features], where the time steps are 100 and the number of features is 12; S3.2, Construction of Temporal Feature Extraction Layer: A two-layer bidirectional LSTM network is built, with 64 memory units in each layer; the first layer of bidirectional LSTM captures the forward and backward dependencies of temporal data through forward and backward neurons and returns the output of each time step; the second layer of bidirectional LSTM performs high-level refinement on the features output by the first layer, and adds the dropout function to achieve regularization and prevent the model from overfitting. S3.3, Construction of Temporal Distributed Fully Connected Layer: A temporal distributed fully connected layer is set after the temporal feature extraction layer. It contains 32 neurons and uses ReLU as the activation function. The fully connected operation is performed independently on the LSTM output at each time step to achieve feature transformation and dimensionality reduction. S3.4, Incorporation of Biomechanical and Physical Constraints: Based on a single-degree-of-freedom model of the knee joint (allowing only rotation within the sagittal plane), a torque equation is constructed. The inertial torque M0 is calculated using the tension F at the subject's ankle, the distance L1 from the knee to the center of mass, the lower leg length L2, the gravity G on the segment above the lower leg, the lower leg gravity Gshank (6% of the subject's total weight), and the angle θ between the body and the horizontal plane. M0 is then incorporated into the model loss function, with the specific formula as follows: In the formula, F is the tension force at the ankle of the subject, L1 is the distance from the knee to the center of mass, L2 is the length of the lower leg, G is the gravity of the segment above the lower leg of the subject, Gshank is the gravity at the lower leg of the subject, and is the angle between the subject's body and the horizontal plane. S3.5, Output layer construction: Set the output layer as a time series prediction structure, and output a three-dimensional tensor in the format of [number of samples, 100, 2], directly outputting the knee joint torque time series M(t).

4. The method for evaluating eccentric strength in the Nordic hamstring test based on deep learning as described in claim 1, characterized in that: The model training and optimization described in S4 specifically include... S4.1, standardize the input features using Z-score and normalize the output true value; S4.2 employs the Adam optimizer with an initial learning rate of 0.0001, a weight decay of 1e-4, and a cosine annealing learning rate scheduling strategy. S4.3, the loss function is a weighted combination of mean squared error (MSE) and M0, an early stopping mechanism is introduced, if the loss does not improve for 50 consecutive rounds, training stops, the batch size is set to 8, the maximum number of training rounds is 300, and the maximum gradient norm of gradient clipping is limited to 0.5; S4.4 calculates the weights of each input feature based on mutual information regression, thereby improving the model's sensitivity to key force signals.

5. The method for evaluating eccentric strength in the Nordic hamstring test based on deep learning according to claim 1, characterized in that: The model evaluation and force calculation described in S5 are evaluated on the test set using three indicators: root mean square error (RMSE), coefficient of determination (R²), and mean absolute error (MAE). The specific formula for the root mean square error (RMSE) is as follows: In the formula, n represents the total number of time steps in the sample sequence; y i Represented as the first i True value of knee joint torque at each time step; Let be the model prediction of the th i Predicted knee joint torque values ​​at each time step. The specific formula for the coefficient of determination (R²) is as follows: The specific formula for the mean absolute error (MAE) is as follows: The calculation of hamstring eccentric force is based on the formula: hamstring force = knee joint torque × lever arm, where the human lever arm is a measurable constant.

Images