Musculoskeletal shock wave treatment path adaptive optimization method based on behavior feature recognition

Abstract

The application discloses a kind of based on behavior characteristic identification's musculoskeletal shock wave treatment path self-adapting optimization method, belong to medical technical field.The application builds behavior characteristic data acquisition system and extracts the core behavior characteristics related to musculoskeletal lesion treatment, the path parameters of shock wave treatment are dynamically adjusted and optimized, generate precise target treatment path that conforms to the individual adaptability of treatment object, the multidimensional data of treatment object treatment whole cycle is simultaneously acquired, the treatment path parameters are optimized by multi-objective optimization model, realize the precise screening of core characteristics and the individualized adaptation of path parameters, ensure that treatment path and lesion site, individual state depth match, through treatment initial adaptability evaluation, real-time dynamic adjustment in treatment and data sedimentation reuse after treatment, path deviation is promptly corrected, quickly respond to treatment object state fluctuation, ensure that treatment path continues to adapt to individual real-time situation, solve the path failure problem caused by state change in treatment process.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to an adaptive optimization method for musculoskeletal shockwave therapy pathways based on behavioral feature recognition. Background Technology

[0002] Current musculoskeletal shockwave therapy planning relies heavily on physicians' clinical experience, employing standardized and fixed treatment protocols that lack individualized adaptation to patient differences. Significant differences exist among patients in age, musculoskeletal anatomy, tissue density at the lesion site, muscle strength, and past medical history. Standardized treatment pathways struggle to match individual physiological characteristics, easily leading to uneven energy distribution and insufficient alignment between the pathway and the lesion area, thus affecting treatment efficacy. Furthermore, patients' behavioral states (such as limb movement and pain response) dynamically change during treatment; traditional fixed treatment pathways cannot respond to these changes in real time, potentially posing treatment risks or reducing treatment efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive optimization method for musculoskeletal shock wave therapy pathway based on behavioral feature recognition, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition, comprising the following steps:

[0005] A behavioral feature data collection system was established to acquire multi-dimensional behavioral feature data of subjects undergoing musculoskeletal shockwave therapy at different stages before, during, and after treatment. Core behavioral features related to musculoskeletal lesion treatment were extracted through feature recognition algorithms. Combined with clinical treatment guidelines for musculoskeletal therapy, individual physiological differences, and safe treatment boundary conditions, an adaptive optimization model was constructed to dynamically adjust and optimize the path parameters of shockwave therapy, generating a precise target treatment path that is suitable for the individual patient.

[0006] Furthermore, the process of building the behavioral feature data collection system includes:

[0007] A multimodal sensing network is deployed, which includes a motion capture unit, an electromyography signal acquisition unit, a pressure sensing unit, a joint range of motion detection unit, and a physiological signal monitoring unit.

[0008] The system includes: a motion capture unit for acquiring three-dimensional trajectory, speed, and acceleration data of the patient's limb movements; an electromyography (EMG) signal acquisition unit for acquiring EMG activity potentials, discharge frequencies, and signal amplitude data of the muscle groups associated with the lesion; a pressure sensing unit for detecting pressure distribution, peak pressure, and rate of change of pressure between the patient's limb and the support surface or treatment equipment; a joint range of motion detection unit for measuring the flexion-extension angle, rotation angle, and range of motion of the joints associated with the lesion; and a physiological signal monitoring unit for acquiring the patient's heart rate, respiratory rate, and pain feedback signal data.

[0009] In the sensor acquisition network, each unit synchronously collects data according to a preset acquisition cycle, generating a multi-dimensional behavioral feature dataset.

[0010] Furthermore, the extraction process of the core behavioral features includes:

[0011] A multi-scale feature recognition framework is constructed, which includes a bottom-level feature extraction layer, a middle-level feature fusion layer, and a high-level core feature filtering layer.

[0012] The bottom feature extraction layer uses a convolutional neural network algorithm to extract local features from the simplified standardized behavioral feature dataset to obtain basic features. These basic features include limb movement trajectory segment features, electromyographic signal temporal features, pressure distribution spatial features, joint movement angle change features, and physiological signal fluctuation features.

[0013] The middle-layer feature fusion layer uses an attention mechanism algorithm to assign weights to and fuse basic features, generating a fused feature vector.

[0014] The high-level core feature screening layer calculates the core index corresponding to the fusion feature based on the preset core feature evaluation index. The core index is compared with the preset core index threshold. When the core index is determined to be greater than or equal to the preset core index threshold, the fusion feature corresponding to the core index is taken as the core behavioral feature. The core behavioral features include abnormal electromyographic activity features of the diseased muscle group, joint movement restriction features, uneven pressure distribution features during exercise, pain-induced movement features, and behavioral response features during treatment.

[0015] Furthermore, the individual physiological differences include the patient's age, gender, height, weight, differences in skeletal and muscular anatomy, muscle strength level, skin thickness, tissue density of the lesion site, and past medical history. The individual physiological difference information is obtained through preliminary clinical consultation, physical examination, and medical imaging results to construct an individual physiological difference information database.

[0016] Furthermore, the safe treatment boundary conditions include the energy intensity threshold, treatment frequency threshold, upper limit of single treatment duration, tissue tolerance threshold of the treatment area, and safe area range to avoid damage to blood vessels and nerves. These boundary conditions are determined based on clinical treatment guidelines, evidence-based medicine, and human physiological tissue tolerance experimental data, and are dynamically adjusted according to the individual physiological differences of the treatment subjects to form personalized safe treatment boundaries.

[0017] Furthermore, the construction process of the adaptive optimization model includes:

[0018] Obtain the training dataset for the model;

[0019] The model is trained based on the training dataset to obtain an adaptive optimization model, which outputs the optimal treatment path parameters.

[0020] The treatment path parameters include treatment start coordinates, path node distribution density, path trajectory curve parameters, shock wave energy intensity corresponding to each path node, treatment frequency, shock wave duration, and path dynamic adjustment step size.

[0021] The distribution density of path nodes is determined based on the size, shape, and severity of the lesion area, with higher node distribution density in areas with more severe lesions than in areas with milder lesions. The energy intensity of the shock wave is dynamically allocated based on pain feedback characteristics, tissue density characteristics, and skin thickness characteristics in individual physiological differences among the core behavioral characteristics.

[0022] Furthermore, it also includes the process of validating and iteratively optimizing the target treatment pathway:

[0023] In the early stages of implementing the target treatment pathway, behavioral response data, physiological signal change data, and treatment effect feedback data of the treatment subjects are collected in real time to construct a treatment effect evaluation system. The evaluation indicators of the treatment effect evaluation system include the degree of symptom improvement, pain relief rate, improvement in joint mobility, and incidence of adverse treatment reactions.

[0024] The suitability of the target treatment path is evaluated based on the evaluation indicators. If the suitability evaluation score is lower than the preset qualified threshold, the accuracy of the core behavioral feature extraction, the rationality of the parameter setting of the adaptive optimization model, and the completeness of the individual physiological difference information are analyzed. The optimized treatment path is then adjusted, optimized, and regenerated.

[0025] The process of repeated verification and optimization continues until the suitability assessment score of the treatment pathway reaches or exceeds the preset qualified threshold, thus forming the final precise target treatment pathway.

[0026] Furthermore, during the treatment process, the behavioral characteristics, physiological signals, and treatment response data of the treatment subjects are continuously monitored in real time.

[0027] When the monitored data changes exceed the preset fluctuation range, the real-time adjustment mechanism of the adaptive optimization model is triggered. Based on the latest monitoring data, the core behavioral features are re-extracted and the treatment path parameters are dynamically adjusted online.

[0028] After treatment, complete treatment data, behavioral characteristic data, and treatment effect data are collected and stored in the case database.

[0029] Furthermore, the model is trained based on the training dataset to obtain an adaptive optimization model, including:

[0030] Feature extraction is performed on the training dataset to determine core behavioral features; the training dataset includes multi-dimensional behavioral feature data.

[0031] Based on core behavioral features, a Bayesian multilayer perceptron (B-MLP) is trained to output a set of musculoskeletal tissue mechanical parameters. Based on joint and motion data in the core behavioral data, a parametric modeling method is used to construct a personalized musculoskeletal 3D geometric model. The musculoskeletal tissue mechanical parameter set is assigned to the corresponding tissue regions of the geometric model to form a complete musculoskeletal biomechanical digital twin model. A shock wave propagation mechanical simulation sub-model based on the finite element method is embedded to solve the propagation equation of the shock wave in the musculoskeletal tissue and output energy attenuation coefficient, pressure distribution and tissue vibration response data.

[0032] A multi-dimensional state space is defined, which includes core behavioral features, a set of musculoskeletal tissue biomechanical parameters, simulation results of shock wave propagation mechanics, treatment progress, and real-time safety indicators, and is normalized and mapped to the [0,1] interval. An action space is defined, which includes discrete actions and continuous actions. Discrete actions include a candidate set of treatment start coordinates and path trajectory type. Continuous actions include path node distribution density, shock wave energy intensity, treatment frequency, duration of action, and dynamic adjustment step size. Discrete actions are integrated into an action space vector after one-hot encoding and continuous actions are normalized. Two Q-networks with identical structures and corresponding target networks are constructed. The Q-networks include an input layer, three hidden layers, and an output layer. The target network parameters are updated based on a preset update rate. A multi-objective total reward function is designed. The total reward function includes a core reward item, a safety constraint item, and a treatment progress reward item. The priority of each objective is adjusted by a weight coefficient. An improved dual-Q-SARSA training algorithm is adopted, which is combined with an ε-greedy strategy, temporal difference learning, and a priority experience replay pool for training. The initial treatment path parameters are output to obtain the dual-Q-SARSA reinforcement learning prediction model.

[0033] A path resistance function is constructed based on the set of musculoskeletal tissue mechanical parameters. The flow rate function of the waterdrop algorithm is improved and mechanical resistance is incorporated. The path node distribution is generated by combining the simulated annealing mechanism. The temperature decay coefficient of the simulated annealing mechanism is also included. A model predictive control (MPC) objective function is constructed, and each optimization objective is balanced by weighting coefficients. Path constraints, parameter constraints, and mechanical constraints are set, and the optimal treatment path parameter set is output to construct a mechanically adaptive waterdrop-MPC joint optimization model.

[0034] The training dataset is input into the dual-Q-SARSA reinforcement learning prediction model in batches for training. At the same time, the state-action-reward data required for reinforcement learning training is generated using the complete musculoskeletal biomechanical digital twin model. The initial treatment path parameters output by reinforcement learning are optimized using the mechanical adaptive droplet-MPC joint optimization model. When the training results meet the requirements, the adaptive optimization model is obtained.

[0035] Furthermore, the core indices corresponding to the fusion features are calculated, including:

[0036] Obtain the evaluation indicators corresponding to the fusion features; the evaluation indicators include the correlation coefficient between the features and musculoskeletal lesions, the discriminative power of the features, and the stability index of the features; calculate the feature quality comprehensive index corresponding to the fusion features based on the evaluation indicators;

[0037] ;

[0038] in, The feature quality index corresponding to the i-th fused feature; denoted as the correlation coefficient between the i-th fusion feature and musculoskeletal lesions; Let be the discriminant metric for the i-th fused feature; Let be the stability index of the i-th fusion feature; , , These are the weighting coefficients; , , For dimensional nonlinear adjustment index; The three-dimensional collaborative enhancement coefficient;

[0039] Calculate the core index corresponding to the fused feature based on the comprehensive feature quality index;

[0040] ;

[0041] in, The core index corresponding to the i-th fusion feature; Contributes factors to feature dimensions; ; The principal component number to which the i-th fusion feature belongs after PCA transformation; The maximum principal component dimension of all features; This is the generalization correction coefficient for the i-th fused feature; ; This represents the generalization error of the feature on the validation set. This is the complexity penalty coefficient; This is a metric for feature complexity. ; The dimension of the feature; The maximum dimension of all features; This is the stability enhancement coefficient; for coefficient of variation; To prevent decimals with a denominator of zero.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. This invention establishes a multimodal sensing network to simultaneously acquire multi-dimensional data such as motion, electromyography, pressure, joint activity, and physiological signals throughout the entire treatment cycle of the patient, and performs standardized preprocessing. This breaks through the limitations of single data acquisition, ensuring data comprehensiveness, temporal consistency, and accuracy, providing high-quality data support for subsequent feature extraction, effectively solving the problem of insufficient path adaptability caused by incomplete data, laying a solid foundation for personalized treatment path generation, and improving the initial data quality for treatment path optimization.

[0044] 2. This invention employs a multi-scale feature recognition framework to extract core behavioral features, combines individual physiological differences with personalized safety boundaries, and optimizes treatment path parameters through a multi-objective optimization model and an improved particle swarm optimization algorithm. This achieves precise screening of core features and personalized adaptation of path parameters, ensuring a deep match between the treatment path and the lesion site and individual condition, maximizing treatment efficacy and energy utilization, while strictly adhering to safety constraints, reducing treatment risks, and significantly improving the accuracy, personalization, and clinical safety of musculoskeletal shock wave therapy.

[0045] 3. This invention, through initial suitability assessment, real-time dynamic adjustment during treatment, and post-treatment data accumulation and reuse, promptly corrects path deviations, quickly responds to fluctuations in the patient's state, and ensures that the treatment path continuously adapts to the individual's real-time situation. It solves the problem of path failure caused by state changes during treatment. The accumulation of case data helps the model to continuously upgrade, forming a virtuous cycle, further enhancing the stability, suitability, and practicality of treatment, and providing more reliable technical support for clinical treatment. Attached Figure Description

[0046] Figure 1This is a schematic diagram of the treatment pathway optimization process of the present invention;

[0047] Figure 2 This is a schematic diagram of the treatment pathway verification and iterative optimization process of the present invention. Detailed Implementation

[0048] 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.

[0049] Please see Figure 1-2 The present invention provides the following technical solutions:

[0050] An adaptive optimization method for musculoskeletal shockwave therapy pathway based on behavioral feature recognition includes the following steps:

[0051] A behavioral feature data collection system was established to acquire multi-dimensional behavioral feature data of subjects undergoing musculoskeletal shockwave therapy at different stages before, during, and after treatment. Core behavioral features related to musculoskeletal lesion treatment were extracted through feature recognition algorithms. Combined with clinical treatment guidelines for musculoskeletal therapy, individual physiological differences, and safe treatment boundary conditions, an adaptive optimization model was constructed to dynamically adjust and optimize the path parameters of shockwave therapy, generating a precise target treatment path that is suitable for the individual patient.

[0052] Individual physiological differences include the patient's age, gender, height, weight, differences in musculoskeletal anatomy, muscle strength level, skin thickness, tissue density of the lesion site, and past medical history. This information is obtained through preliminary clinical consultations, physical examinations, and medical imaging results. An individual physiological difference database is constructed to provide individual suitability references for optimizing treatment pathways.

[0053] The safety treatment boundary conditions include the energy intensity threshold, treatment frequency threshold, upper limit of single treatment duration, tissue tolerance threshold of the treatment area, and safe area range to avoid damage to blood vessels and nerves. The boundary conditions are determined based on clinical treatment guidelines, evidence-based medicine, and human physiological tissue tolerance experimental data, and are dynamically adjusted according to the individual physiological differences of the treatment subjects to form personalized safety treatment boundaries.

[0054] The process of building a behavioral characteristic data collection system includes:

[0055] A multimodal sensor acquisition network is deployed, which includes motion capture units, electromyography signal acquisition units, pressure sensing units, joint range of motion detection units, and physiological signal monitoring units.

[0056] The system includes: a motion capture unit for acquiring three-dimensional trajectory, speed, and acceleration data of the patient's limb movements; an electromyography (EMG) signal acquisition unit for acquiring EMG activity potentials, discharge frequencies, and signal amplitude data of the muscle groups associated with the lesion; a pressure sensing unit for detecting pressure distribution, peak pressure, and rate of change of pressure between the patient's limb and the support surface or treatment equipment; a joint range of motion detection unit for measuring the flexion-extension angle, rotation angle, and range of motion of the joints associated with the lesion; and a physiological signal monitoring unit for acquiring the patient's heart rate, respiratory rate, and pain feedback signal data.

[0057] In the sensor acquisition network, each unit synchronously collects data according to a preset acquisition cycle, generating a multi-dimensional behavioral feature dataset.

[0058] In the above embodiments, by building a behavioral feature data acquisition system that includes multimodal sensing units, the comprehensive capture of multidimensional behavioral features of the treatment subject throughout the entire treatment cycle is achieved. The collaborative work of units such as motion capture and electromyography signals breaks the limitations of single data acquisition and can simultaneously acquire key data in multiple aspects such as movement, physiology, and force, ensuring the integrity and correlation of the data. Each unit collects data synchronously according to a preset cycle, providing an accurate and coherent dataset for subsequent feature extraction. The multidimensional and synchronized data acquisition method can comprehensively reflect the behavioral state and lesion correlation features of the treatment subject, providing rich basic data support for treatment path optimization and solving the problem of insufficient path adaptability caused by incomplete data in treatment.

[0059] The process of extracting core behavioral features includes:

[0060] A multi-scale feature recognition framework is constructed, which includes a low-level feature extraction layer, a mid-level feature fusion layer, and a high-level core feature selection layer.

[0061] The bottom feature extraction layer uses a convolutional neural network algorithm to extract local features from the simplified and standardized behavioral feature dataset to obtain basic features. The basic features include limb movement trajectory segment features, electromyographic signal temporal features, pressure distribution spatial features, joint movement angle change features, and physiological signal fluctuation features.

[0062] The middle-layer feature fusion layer uses an attention mechanism algorithm to assign weights and fuse basic features, strengthen the feature weights related to musculoskeletal disease treatment, weaken the interference of irrelevant features, and generate a fused feature vector.

[0063] The high-level core feature screening layer calculates the core index corresponding to the fusion feature based on the preset core feature evaluation index. The core index is compared with the preset core index threshold. When the core index is determined to be greater than or equal to the preset core index threshold, the fusion feature corresponding to the core index is taken as the core behavioral feature. The core behavioral features include abnormal electromyographic activity features of the diseased muscle group, joint movement restriction features, uneven pressure distribution features during exercise, pain-induced movement features, and behavioral response features during treatment.

[0064] In the above embodiments, the core behavioral feature extraction adopts a multi-scale feature recognition framework. Through hierarchical processing, it achieves the goal of accurately selecting key features from massive data. The bottom-level convolutional neural network algorithm can efficiently extract local basic features. The middle-level attention mechanism strengthens core information and weakens interference factors through weight allocation, ensuring the pertinence of feature fusion. The high-level selection process based on multi-dimensional evaluation indicators ensures the relevance, discriminability and stability of core features. The finally extracted core behavioral features closely meet the treatment needs of musculoskeletal diseases and comprehensively cover key dimensions such as muscles, joints and pain response. It effectively solves the problems of redundant information and lack of prominence of core information in feature extraction. The accurate core behavioral features can truly reflect the individual disease state and behavioral response pattern of the treatment object, providing high-quality input parameters for the adaptive optimization model and ensuring the scientific and accurate optimization of the subsequent treatment path.

[0065] The process of constructing an adaptive optimization model includes:

[0066] Obtain the training dataset for the model;

[0067] The model is trained based on the training dataset to obtain an adaptive optimization model, which outputs the optimal treatment path parameters.

[0068] The treatment path parameters include treatment start coordinates, path node distribution density, path trajectory curve parameters, shock wave energy intensity corresponding to each path node, treatment frequency, shock wave duration, and path dynamic adjustment step size.

[0069] The distribution density of path nodes is determined based on the size, shape, and severity of the lesion area, with higher node distribution density in areas with more severe lesions than in areas with milder lesions. The energy intensity of the shock wave is dynamically allocated based on pain feedback characteristics, tissue density characteristics, and skin thickness characteristics in individual physiological differences among the core behavioral characteristics.

[0070] In the above embodiments, the adaptive optimization model takes multi-dimensional key information as input, constructs a multi-objective optimization system, and solves it using an improved particle swarm optimization algorithm. This achieves precise optimization of treatment path parameters, organically integrating core behavioral characteristics, individual physiological differences, and personalized safety boundaries. The objective function considers treatment accuracy, efficacy, and safety, and the constraints comprehensively cover clinical, safety, and equipment performance requirements, ensuring the comprehensiveness and rationality of the optimization direction. The improved particle swarm optimization algorithm can efficiently solve multi-objective optimization problems and quickly lock in the optimal combination of path parameters. The refined setting of treatment path parameters, especially the dynamic adjustment of node distribution density according to the degree of lesion and the allocation of shock wave energy according to individual characteristics, achieves deep adaptation between the treatment path and the individual state of the patient. This precise parameter optimization method not only maximizes the treatment effect but also strictly controls the treatment risk, significantly improving the personalization level and clinical safety of musculoskeletal shock wave therapy.

[0071] The adaptive optimization method for musculoskeletal shock wave therapy pathway based on behavioral feature recognition also includes a process of validating and iteratively optimizing the target treatment pathway.

[0072] In the early stages of implementing the target treatment pathway, real-time data on the patient's behavioral response, physiological signal changes, and treatment effect feedback are collected to construct a treatment effect evaluation system. The evaluation indicators of the treatment effect evaluation system include the degree of symptom improvement, pain relief rate, improvement in joint mobility, and incidence of adverse treatment reactions.

[0073] The suitability of the target treatment path is evaluated based on the evaluation indicators. If the suitability evaluation score is lower than the preset qualified threshold, the accuracy of the core behavioral feature extraction, the rationality of the parameter setting of the adaptive optimization model, and the completeness of the individual physiological difference information are analyzed. The optimized treatment path is then adjusted, optimized, and regenerated.

[0074] The process of verification and optimization is repeated until the suitability assessment score of the treatment pathway reaches or exceeds the preset qualified threshold, thus forming the final precise target treatment pathway.

[0075] During the treatment process, the behavioral characteristics, physiological signals, and treatment response data of the patients are continuously monitored in real time.

[0076] When the monitored data changes exceed the preset fluctuation range, the real-time adjustment mechanism of the adaptive optimization model is triggered. Based on the latest monitoring data, the core behavioral features are re-extracted and the treatment path parameters are dynamically adjusted online to ensure that the treatment path always adapts to the real-time status of the patient.

[0077] After treatment, complete treatment data, behavioral characteristic data, and treatment effect data are collected and stored in the case database to provide data support for optimizing treatment pathways for similar cases in the future.

[0078] The model is trained based on the aforementioned training dataset to obtain an adaptive optimization model, including:

[0079] Feature extraction is performed on the training dataset to determine core behavioral features; the training dataset includes multi-dimensional behavioral feature data.

[0080] Based on core behavioral features, a Bayesian multilayer perceptron (B-MLP) is trained to output a set of musculoskeletal tissue mechanical parameters. Based on joint and motion data in the core behavioral data, a parametric modeling method is used to construct a personalized musculoskeletal 3D geometric model. The musculoskeletal tissue mechanical parameter set is assigned to the corresponding tissue regions of the geometric model to form a complete musculoskeletal biomechanical digital twin model. A shock wave propagation mechanical simulation sub-model based on the finite element method is embedded to solve the propagation equation of the shock wave in the musculoskeletal tissue and output energy attenuation coefficient, pressure distribution and tissue vibration response data.

[0081] A multi-dimensional state space is defined, which includes core behavioral features, a set of musculoskeletal tissue biomechanical parameters, simulation results of shock wave propagation mechanics, treatment progress, and real-time safety indicators, and is normalized and mapped to the [0,1] interval. An action space is defined, which includes discrete actions and continuous actions. Discrete actions include a candidate set of treatment start coordinates and path trajectory type. Continuous actions include path node distribution density, shock wave energy intensity, treatment frequency, duration of action, and dynamic adjustment step size. Discrete actions are integrated into an action space vector after one-hot encoding and continuous actions are normalized. Two Q-networks with identical structures and corresponding target networks are constructed. The Q-networks include an input layer, three hidden layers, and an output layer. The target network parameters are updated based on a preset update rate. A multi-objective total reward function is designed. The total reward function includes a core reward item, a safety constraint item, and a treatment progress reward item. The priority of each objective is adjusted by a weight coefficient. An improved dual-Q-SARSA training algorithm is adopted, which is combined with an ε-greedy strategy, temporal difference learning, and a priority experience replay pool for training. The initial treatment path parameters are output to obtain the dual-Q-SARSA reinforcement learning prediction model.

[0082] A path resistance function is constructed based on the set of musculoskeletal tissue mechanical parameters. The flow rate function of the waterdrop algorithm is improved and mechanical resistance is incorporated. The path node distribution is generated by combining the simulated annealing mechanism. The temperature decay coefficient of the simulated annealing mechanism is also included. A model predictive control (MPC) objective function is constructed, and each optimization objective is balanced by weighting coefficients. Path constraints, parameter constraints, and mechanical constraints are set, and the optimal treatment path parameter set is output to construct a mechanically adaptive waterdrop-MPC joint optimization model.

[0083] The training dataset is input into the dual-Q-SARSA reinforcement learning prediction model in batches for training. At the same time, the state-action-reward data required for reinforcement learning training is generated using the complete musculoskeletal biomechanical digital twin model. The initial treatment path parameters output by reinforcement learning are optimized using the mechanical adaptive droplet-MPC joint optimization model. When the training results meet the requirements, the adaptive optimization model is obtained.

[0084] In this embodiment, a Bayesian multilayer perceptron (B-MLP) is trained based on core behavioral features, outputting a set of musculoskeletal biomechanical parameters. Based on joint and motion data from the core behavioral data, a personalized musculoskeletal three-dimensional geometric model is constructed using a parametric modeling method. The biomechanical parameter set is then assigned to the corresponding tissue regions of the geometric model, forming a complete musculoskeletal biomechanical digital twin model. This includes: using a Bayesian multilayer perceptron (B-MLP) with three hidden layers (256→128→64), introducing Dropout regularization, inputting core behavioral features, and outputting a set of musculoskeletal biomechanical parameters; and combining clinical musculoskeletal biomechanical data. The system employs maximum likelihood estimation to optimize model parameters. Based on joint / motion data from core features, it constructs personalized musculoskeletal geometric models: generating three-dimensional geometric structures of muscles, bones, joints, and tendons through parametric modeling methods; mapping geometric models to mechanical parameters: allocating the mechanical parameters output by B-MLP to the corresponding tissue regions of the geometric models to form complete musculoskeletal biomechanical digital twin models; calibrating the shock wave propagation mechanics simulation sub-model: using measured shock wave data from the training dataset, the finite element model parameters are optimized using the least squares method to ensure that the simulation output error is ≤5% compared to the real data.

[0085] In this embodiment, a shock wave propagation mechanics simulation sub-model based on the finite element method is embedded to solve the propagation equation of the shock wave in musculoskeletal tissue, and output the energy attenuation coefficient, pressure distribution, and tissue vibration response data. This includes: constructing a shock wave propagation mechanics simulation sub-model: based on the finite element method, solving the propagation equation of the shock wave in musculoskeletal tissue, and outputting the energy attenuation coefficient, pressure distribution, and tissue vibration response of the shock wave in each tissue region, which serve as constraints for subsequent path optimization; and a dynamic state generation mechanism: integrating a treatment effect prediction sub-module through a digital twin model, generating treatment progress (0-1, representing the degree of pain relief) and real-time safety indicators (electromyography abnormality rate, excessive pressure probability) in real time based on mechanical parameters and shock wave simulation results, ensuring that the dynamic state is calculable.

[0086] In this embodiment, a multi-dimensional state space is defined, which includes core behavioral features, a set of musculoskeletal tissue biomechanical parameters and simulation results of shock wave propagation mechanics, treatment progress and real-time safety indicators. These are normalized and mapped to the [0,1] interval, including: basic state: core features (128 dimensions) + biomechanical parameters (5 dimensions) + digital twin simulation results (10 dimensions, including energy attenuation, pressure distribution, etc.); dynamic state: treatment progress (0-1, representing the degree of pain relief) + real-time safety indicators (electromyographic abnormality rate, excessive pressure probability), which are generated in real-time through digital twin model simulation; the dimensions of the state space are normalized and mapped to the [0,1] interval, with a total of 145 dimensions.

[0087] In this embodiment, an action space is defined, which includes discrete actions and continuous actions. Discrete actions include a candidate set of treatment starting coordinates and path trajectory type. Continuous actions include path node distribution density, shock wave energy intensity, treatment frequency, duration of action, and dynamic adjustment step size. Discrete actions are integrated into an action space vector after one-hot encoding and continuous actions are normalized. The vector includes: Discrete actions: candidate set of treatment starting coordinates (adaptively generated based on the lesion area of ​​the geometric model, with point density inversely proportional to muscle volume, ensuring bone / nerve distance ≥3mm), path trajectory type (straight line / curve / segmented polyline); Continuous actions: path node distribution density (5-20 nodes / cm), shock wave energy intensity (0.1-0.8mJ / mm²), treatment frequency (1-20Hz), duration of action (10-60s), and dynamic adjustment step size (0.1-1mm); Action space encoding: discrete action one-hot encoding (103 dimensions) + continuous action normalization (5 dimensions), for a total of 108 dimensions.

[0088] In this embodiment, the dual-Q-SARSA network structure is designed as follows: Two independent Q-networks (Q1 and Q2) are constructed, both using the following architecture: input layer (145 dimensions) → hidden layer 1 (512 dimensions, ReLU activation) → hidden layer 2 (256 dimensions, LeakyReLU activation) → hidden layer 3 (128 dimensions, ReLU activation) → output layer (108 dimensions, linear activation). Target networks are set for each Q-network (Q1_target, Q2_target), and the target network parameters are updated every 100 steps (update rate). To reduce training oscillations; Policy update: based on SARSA temporal difference learning, Q-value update formula: ;in, ; The learning rate; Discount factor; The choice is made using the ε-greedy strategy of the current Q-network; The expected cumulative reward (Q value) for performing action a in state s; s is the current state; a is the current action; For instant rewards; Calculate the next state for the target network and the next action Q value; The next state to transition to after performing action a; For the next state The action chosen by a greedy strategy.

[0089] In this embodiment, a multi-objective total reward function is designed, which includes a core reward item, a safety constraint item, and a treatment progress reward item. The priority of each objective is adjusted by weighting coefficients, including:

[0090] Core Rewards Where ΔP is the pain reduction rate (post-treatment - pre-treatment), and ΔA is the joint mobility improvement rate. Normalized treatment time; As the core reward item; For pain reduction rate; , , Weighting coefficients; safety constraint terms ; For safety constraints; , Here, are the weighting coefficients; This represents the probability of abnormal electromyography (EMG). Risk of tissue damage; treatment progress reward ;in, This represents the treatment progress value in a dynamic state. This is the scaling factor; Rewards for treatment progress; Total loss function: ; , , These are the weighting coefficients; This is the total reward function.

[0091] In this embodiment, an improved dual-Q-SARSA training algorithm is used, combined with an ε-greedy strategy, temporal difference learning, and a priority experience replay pool for training. Preliminary treatment path parameters are output, resulting in a dual-Q-SARSA reinforcement learning prediction model, including:

[0092] An ε-greedy strategy is adopted, with ε linearly decreasing from 0.9 to 0.05 (decay step size = 5000). Policy update: based on SARSA temporal difference learning, Q-value update formula: ;in, ; The ε-greedy strategy selection of the current Q network is used to introduce a priority experience replay pool, which samples samples according to the TD error priority to improve training efficiency.

[0093] In this embodiment, a path resistance function is constructed based on the musculoskeletal tissue mechanical parameter set; the flow rate function of the waterdrop algorithm is improved and mechanical resistance is incorporated; a path node distribution is generated by combining a simulated annealing mechanism, wherein the simulated annealing mechanism has a temperature decay coefficient; a model predictive control (MPC) objective function is constructed, and the optimization objectives are balanced by weighting coefficients; path constraints, parameter constraints, and mechanical constraints are set, and the optimal treatment path parameter set is output, thus constructing a mechanically adaptive waterdrop-MPC joint optimization model, including:

[0094] Path drag model construction: Based on the mechanical parameters of the digital twin model, the path drag function is defined as follows: ;in, , , ; For muscle elastic modulus, The viscosity coefficient of the tendon. Bone mineral density; For position Path resistance function at the location; For position The elastic modulus of the muscle at that location; For position The viscosity coefficient of the tendon at the location; For position Bone density at the location; all obtained from the mechanical parameter set output by B-MLP; water droplet flow rate optimization: improved rate function, incorporating mechanical resistance: ;in, The target pressure is the average pressure of the lesion area output by the shock wave propagation simulation sub-model. ; ; ; For denominator protection; For position Shock wave flow velocity at the location; For position Path resistance at the location; For position Muscle density at a given location is obtained from a set of mechanical parameters; path node generation: water droplets propagate along the gradient direction in regions of low resistance, and node distribution density... ; Densification of nodes in the pain area; For position The distribution density of path nodes at the location; For position The degree of pain at the location;

[0095] Optimization of simulated annealing mechanism: temperature decay coefficient ; Let t be the simulated annealing temperature at time t; The initial temperature; For the number of iterations or time steps; construct the model predictive control (MPC) objective function: ;in, The objective function for MPC; Let be the sequence of path parameter vectors from time t to time t+H; This could lead to a loss of therapeutic effect. For safety losses; For efficiency loss; , These are the weighting coefficients; This is a path parameter vector; For prediction in the time domain; effect loss ; To predict the rate of pain reduction; Target pain reduction rate; safety loss ; To predict the probability of tissue damage; efficiency loss ; Let t be the energy intensity action at time t.

[0096] In this embodiment, the path constraints are as follows: the starting coordinates must fall within the projection area of ​​the muscle group associated with the lesion, and the distance between the path trajectory and the bone / nerve is ≥3mm; the parameter constraints are: energy intensity ≤0.8mJ / mm², treatment frequency ≤20Hz, and duration of action ≤60s; the mechanical constraints are: peak pressure of shock wave propagation ≤5MPa, and tissue vibration displacement ≤0.1mm; the interior point method (IPM) is used to solve the quadratic programming problem of MPC, with ≤50 iterations and ≤1e-4 solution accuracy; the output is the optimal treatment path parameter set.

[0097] The working principle and beneficial effects of the above technical solution are as follows: By extracting core behavioral features to train a Bayesian multilayer perceptron, and combining it with parametric modeling methods to construct a personalized musculoskeletal three-dimensional geometric model and assign mechanical parameters, a complete musculoskeletal biomechanical digital twin model tailored to the individual can be formed, laying the foundation for subsequent precise treatment; defining a multi-dimensional state space and action space, covering core behavioral features, mechanical parameters, simulation results, treatment progress and safety indicators, etc., comprehensively considering multiple factors, making treatment planning more comprehensive; adopting an improved dual-Q-SARSA training algorithm, combined with an ε-greedy strategy, temporal difference learning and priority experience replay pool, can output preliminary treatment path parameters and realize intelligent decision-making; constructing a path resistance function based on mechanical parameters, improving the droplet algorithm and incorporating mechanical resistance, and combining simulated annealing mechanism and model predictive control to optimize treatment path parameters and improve the rationality of treatment path; inputting training data in batches, using the digital twin model to generate the data required for training, and combining reinforcement learning prediction model and mechanical adaptive optimization model, the final model is obtained when the requirements are met, improving training efficiency and model performance.

[0098] Calculate the core indices corresponding to the fusion features, including:

[0099] Obtain the evaluation indicators corresponding to the fusion features; the evaluation indicators include the correlation coefficient between the features and musculoskeletal lesions, the discriminative power of the features, and the stability index of the features; calculate the feature quality comprehensive index corresponding to the fusion features based on the evaluation indicators;

[0100] ;

[0101] in, The feature quality index corresponding to the i-th fused feature; denoted as the correlation coefficient between the i-th fusion feature and musculoskeletal lesions; Let be the discriminant metric for the i-th fused feature; Let be the stability index of the i-th fusion feature; , , These are the weighting coefficients; , , For dimensional nonlinear adjustment index; The three-dimensional collaborative enhancement coefficient;

[0102] Calculate the core index corresponding to the fused feature based on the comprehensive feature quality index;

[0103] ;

[0104] in, The core index corresponding to the i-th fusion feature; Contributes factors to feature dimensions; ; The principal component number to which the i-th fusion feature belongs after PCA transformation; The maximum principal component dimension of all features; This is the generalization correction coefficient for the i-th fused feature; ; This represents the generalization error of the feature on the validation set. This is the complexity penalty coefficient; This is a metric for feature complexity. ; The dimension of the feature; The maximum dimension of all features; This is the stability enhancement coefficient; for coefficient of variation; To prevent decimals with a denominator of zero.

[0105] In this embodiment, the maximum dimension of all features is illustrated by the following example: In the optimization of musculoskeletal shockwave therapy pathways, three fusion features are generated based on patient walking behavior data. Fusion feature 1: the average and standard deviation of the knee joint angle during the gait cycle, with a dimension of 2; Fusion feature 2: the peak amplitude of ankle joint acceleration along the x, y, and z axes, with a dimension of 3; Fusion feature 3: the coordinates of the center position of the plantar pressure distribution, with a dimension of 2. Therefore, the set of dimensions for all fusion features is... ,but .

[0106] The working principle and beneficial effects of the above technical solution are as follows: By comprehensively considering multiple evaluation indicators such as the correlation coefficient between features and musculoskeletal lesions, the discriminative power of features, and the stability index of features, the quality of fused features can be evaluated more comprehensively. This helps to screen features that are strongly correlated with musculoskeletal lesions, have high discriminative power, and good stability, thereby improving the accuracy of subsequent analysis and diagnosis. When calculating the comprehensive feature quality index, a dimensional nonlinear adjustment index and a three-dimensional synergistic enhancement coefficient are used, which can better reflect the complex relationships between various aspects of feature indicators, making the evaluation results more accurate and reasonable. When calculating the core index, a feature dimensionality contribution factor and a generalization correction coefficient are introduced, considering the dimensional contribution of features in principal component analysis and their generalization ability on the validation set, avoiding the overfitting problem of high-dimensional features and improving the generalization performance of the model. A complexity penalty coefficient is used to penalize feature complexity to prevent the model from becoming too complex. At the same time, a stability enhancement coefficient is used to enhance the stability of features, making the model more robust and improving the model's performance on different datasets.

[0107] In the above embodiments, the verification and iterative optimization process of the target treatment path effectively ensures the continuous adaptability and clinical effectiveness of the treatment path. The adaptability assessment in the early stage of treatment comprehensively examines the path effect through multi-dimensional evaluation indicators. The targeted adjustment method for cases that fail to meet the assessment standards can quickly correct deviations in feature extraction, model parameters, or information collection, ensuring that the final generated treatment path meets the preset qualification standards. The real-time monitoring and dynamic adjustment mechanism during the treatment process can respond promptly to changes in the state of the treated patient. By triggering real-time model optimization, the treatment path can be updated online, solving the problem of decreased path adaptability due to individual state fluctuations during treatment. The data storage and reuse of post-treatment case data provides valuable reference for subsequent similar cases, helping the model to continuously optimize and upgrade. The entire verification and iterative optimization process significantly improves the stability, adaptability, and clinical applicability of the treatment path.

[0108] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition, characterized in that, Includes the following steps: A behavioral feature data collection system was established to acquire multi-dimensional behavioral feature data of musculoskeletal shockwave therapy subjects at different stages before, during and after treatment. Core behavioral features related to musculoskeletal lesion treatment were extracted through feature recognition algorithms. Combined with clinical diagnosis and treatment guidelines for musculoskeletal therapy, individual physiological differences and safe treatment boundary conditions, an adaptive optimization model was constructed to dynamically adjust and optimize the path parameters of shockwave therapy and generate a precise target treatment path that is suitable for the individual patient. In the early stages of implementing the target treatment pathway, behavioral response data, physiological signal change data, and treatment effect feedback data of the treatment subjects are collected in real time to construct a treatment effect evaluation system. The evaluation indicators of the treatment effect evaluation system include the degree of symptom improvement, pain relief rate, improvement in joint mobility, and incidence of adverse treatment reactions. The suitability of the target treatment path is evaluated based on the evaluation indicators. If the suitability evaluation score is lower than the preset qualified threshold, the accuracy of the core behavioral feature extraction, the rationality of the parameter setting of the adaptive optimization model, and the completeness of the individual physiological difference information are analyzed. The optimized treatment path is then adjusted, optimized, and regenerated. The process of verification and optimization is repeated until the suitability assessment score of the treatment pathway reaches or exceeds the preset qualified threshold, thus forming the final precise target treatment pathway. The extraction process of the core behavioral features includes: A multi-scale feature recognition framework is constructed, which includes a bottom-level feature extraction layer, a middle-level feature fusion layer, and a high-level core feature filtering layer. The bottom feature extraction layer uses a convolutional neural network algorithm to extract local features from the simplified standardized behavioral feature dataset to obtain basic features. These basic features include limb movement trajectory segment features, electromyographic signal temporal features, pressure distribution spatial features, joint movement angle change features, and physiological signal fluctuation features. The middle-layer feature fusion layer uses an attention mechanism algorithm to assign weights to and fuse basic features, generating a fused feature vector. The high-level core feature screening layer calculates the core index corresponding to the fusion feature based on the preset core feature evaluation index. The core index is compared with the preset core index threshold. When the core index is determined to be greater than or equal to the preset core index threshold, the fusion feature corresponding to the core index is taken as the core behavioral feature. The core behavioral features include abnormal electromyographic activity features of the diseased muscle group, joint movement restriction features, uneven pressure distribution features during exercise, pain-induced movement features, and behavioral response features during treatment.

2. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 1, characterized in that, The process of building the behavioral feature data collection system includes: A multimodal sensing network is deployed, which includes a motion capture unit, an electromyography signal acquisition unit, a pressure sensing unit, a joint range of motion detection unit, and a physiological signal monitoring unit. The system includes: a motion capture unit for acquiring three-dimensional trajectory, speed, and acceleration data of the patient's limb movements; an electromyography (EMG) signal acquisition unit for acquiring EMG activity potentials, discharge frequencies, and signal amplitude data of the muscle groups associated with the lesion; a pressure sensing unit for detecting pressure distribution, peak pressure, and rate of change of pressure between the patient's limb and the support surface or treatment equipment; a joint range of motion detection unit for measuring the flexion-extension angle, rotation angle, and range of motion of the joints associated with the lesion; and a physiological signal monitoring unit for acquiring the patient's heart rate, respiratory rate, and pain feedback signal data. In the sensor acquisition network, each unit synchronously collects data according to a preset acquisition cycle, generating a multi-dimensional behavioral feature dataset.

3. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 1, characterized in that, The individual physiological differences include the patient's age, gender, height, weight, differences in skeletal and muscular anatomy, muscle strength level, skin thickness, tissue density of the lesion site, and past medical history. The individual physiological differences information is obtained through preliminary clinical consultation, physical examination, and medical imaging results, and an individual physiological differences information database is constructed.

4. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 1, characterized in that, The safe treatment boundary conditions include the energy intensity threshold, treatment frequency threshold, upper limit of single treatment duration, tissue tolerance threshold of the treatment area, and safe area range to avoid damage to blood vessels and nerves. These boundary conditions are determined based on clinical treatment guidelines, evidence-based medicine, and human physiological tissue tolerance experimental data, and are dynamically adjusted according to the individual physiological differences of the treatment subjects to form personalized safe treatment boundaries.

5. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 1, characterized in that, The process of constructing the adaptive optimization model includes: Obtain the training dataset for the model; The model is trained based on the training dataset to obtain an adaptive optimization model, which outputs the optimal treatment path parameters. The treatment path parameters include treatment start coordinates, path node distribution density, path trajectory curve parameters, shock wave energy intensity corresponding to each path node, treatment frequency, shock wave duration, and path dynamic adjustment step size. The distribution density of path nodes is determined based on the size, shape, and severity of the lesion area, with higher node distribution density in areas with more severe lesions than in areas with milder lesions. The energy intensity of the shock wave is dynamically allocated based on pain feedback characteristics, tissue density characteristics, and skin thickness characteristics in individual physiological differences among the core behavioral characteristics.

6. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 1, characterized in that, During the treatment process, the behavioral characteristics, physiological signals, and treatment response data of the patients are continuously monitored in real time. When the monitored data changes exceed the preset fluctuation range, the real-time adjustment mechanism of the adaptive optimization model is triggered. Based on the latest monitoring data, the core behavioral features are re-extracted and the treatment path parameters are dynamically adjusted online. After treatment, complete treatment data, behavioral characteristic data, and treatment effect data are collected and stored in the case database.

7. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 5, characterized in that, The model is trained based on the training dataset to obtain an adaptive optimization model, including: Feature extraction is performed on the training dataset to determine core behavioral features; the training dataset includes multi-dimensional behavioral feature data. Based on core behavioral features, a Bayesian multilayer perceptron (B-MLP) is trained to output a set of musculoskeletal tissue mechanical parameters. Based on joint and motion data in the core behavioral data, a parametric modeling method is used to construct a personalized musculoskeletal 3D geometric model. The musculoskeletal tissue mechanical parameter set is assigned to the corresponding tissue regions of the geometric model to form a complete musculoskeletal biomechanical digital twin model. A shock wave propagation mechanical simulation sub-model based on the finite element method is embedded to solve the propagation equation of the shock wave in the musculoskeletal tissue and output energy attenuation coefficient, pressure distribution and tissue vibration response data. A multi-dimensional state space is defined, which includes core behavioral features, a set of musculoskeletal tissue biomechanical parameters, simulation results of shock wave propagation mechanics, treatment progress, and real-time safety indicators, and is normalized and mapped to the [0,1] interval. An action space is defined, which includes discrete actions and continuous actions. Discrete actions include a candidate set of treatment start coordinates and path trajectory type. Continuous actions include path node distribution density, shock wave energy intensity, treatment frequency, duration of action, and dynamic adjustment step size. Discrete actions are integrated into an action space vector after one-hot encoding and continuous actions are normalized. Two Q-networks with identical structures and corresponding target networks are constructed. The Q-networks include an input layer, three hidden layers, and an output layer. The target network parameters are updated based on a preset update rate. A multi-objective total reward function is designed. The total reward function includes a core reward item, a safety constraint item, and a treatment progress reward item. The priority of each objective is adjusted by a weight coefficient. An improved dual-Q-SARSA training algorithm is adopted, which is combined with an ε-greedy strategy, temporal difference learning, and a priority experience replay pool for training. The initial treatment path parameters are output to obtain the dual-Q-SARSA reinforcement learning prediction model. A path resistance function is constructed based on the set of musculoskeletal tissue mechanical parameters. The flow rate function of the waterdrop algorithm is improved and mechanical resistance is incorporated. The path node distribution is generated by combining the simulated annealing mechanism. The temperature decay coefficient of the simulated annealing mechanism is also included. A model predictive control (MPC) objective function is constructed, and each optimization objective is balanced by weighting coefficients. Path constraints, parameter constraints, and mechanical constraints are set, and the optimal treatment path parameter set is output to construct a mechanically adaptive waterdrop-MPC joint optimization model. The training dataset is input into the dual-Q-SARSA reinforcement learning prediction model in batches for training. At the same time, the state-action-reward data required for reinforcement learning training is generated using the complete musculoskeletal biomechanical digital twin model. The initial treatment path parameters output by reinforcement learning are optimized using the mechanical adaptive droplet-MPC joint optimization model. When the training results meet the requirements, the adaptive optimization model is obtained.

8. The adaptive optimization method for musculoskeletal shockwave therapy path based on behavioral feature recognition as described in claim 1, characterized in that, Calculate the core indices corresponding to the fusion features, including: Obtain the evaluation indicators corresponding to the fusion features; the evaluation indicators include the correlation coefficient between the features and musculoskeletal lesions, the discriminative power of the features, and the stability index of the features; calculate the feature quality comprehensive index corresponding to the fusion features based on the evaluation indicators; ； in, The feature quality index corresponding to the i-th fused feature; denoted as the correlation coefficient between the i-th fusion feature and musculoskeletal lesions; Let be the discriminant metric for the i-th fused feature; Let be the stability index of the i-th fusion feature; , , These are the weighting coefficients; , , For dimensional nonlinear adjustment index; The three-dimensional collaborative enhancement coefficient; Calculate the core index corresponding to the fused feature based on the comprehensive feature quality index; ； in, The core index corresponding to the i-th fusion feature; Contributes factors to feature dimensions; ; The principal component number to which the i-th fusion feature belongs after PCA transformation; The maximum principal component dimension of all features; This is the generalization correction coefficient for the i-th fused feature; ; This represents the generalization error of the feature on the validation set. This is the complexity penalty coefficient; This is a metric for feature complexity. ; The dimension of the feature; The maximum dimension of all features; This is the stability enhancement coefficient; for coefficient of variation; To prevent decimals with a denominator of zero.

Images