Scoliosis correction identification method and system based on sensing communication equipment

Abstract

The invention relates to a scoliosis correction recognition method and system based on sensing communication equipment, in particular to the technical field of rehabilitation engineering, and realizes the crossing of scoliosis rehabilitation evaluation from universality to individual specificity through fusion of transfer learning and small sample adaptive technologies. The system can quickly construct a highly personalized evaluation model according to a very small amount of user data, and accurately capture unique biomechanical characteristics of different patients; an online learning mechanism driven by uncertainty estimation is further introduced, so that the model can be autonomously optimized and adjusted along with the rehabilitation process of the user to dynamically adapt to the capability change of the user; finally, prospective modeling is carried out on the rehabilitation track by means of a space-time diagram neural network, decision support information with an early warning function and personalized guiding significance is generated, and therefore the problems that a traditional assessment method depends on a fixed threshold value and lacks dynamic and predictive core pain points are effectively solved; and the accuracy, adaptability and intelligent level of family rehabilitation training are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rehabilitation engineering, and more particularly, to a scoliosis correction recognition method and system based on a sensing communication device. BACKGROUND

[0002] Scoliosis, as a common three-dimensional spinal deformity, has a continuously rising incidence in the adolescent population and has become a public health problem that seriously affects the physical and mental health of adolescents. In the non-surgical treatment system of scoliosis, rehabilitation correction training occupies a core position, and the core is to enhance the strength of the trunk muscle group, improve the biomechanical balance of the spine, and delay the progression of scoliosis through a series of targeted action training. However, clinical practice faces a very serious challenge: the individual differences of patients are extremely significant. This difference not only reflects in the static structure indicators of scoliosis such as the type of C or S, the bending angle, the degree of rotation, but also more deeply reflects in the dynamic functional level of the flexibility of the spine, the strength distribution of the surrounding muscles, and the fatigue characteristics. Even for the same standardized correction action, such as standing against the wall or lateral flexion, due to the unique pathological structure and physiological state of different patients, the biomechanical characteristics such as the muscle activation pattern, joint movement trajectory, and body center of gravity distribution corresponding to the ideal standard completion posture of the action are also very different. This means that in the process of rehabilitation training, the evaluation standard of action quality must be highly individualized, and a unified and rigid evaluation scale cannot scientifically adapt to the actual needs of all patients, thereby restricting the maximization of rehabilitation effect.

[0003] With the development of the Internet of Things and wearable sensing technology, some spine posture monitoring systems based on inertial measurement units or surface electromyography sensors have emerged to provide home-based rehabilitation guidance for patients. However, these existing technical solutions have obvious bottlenecks in the core evaluation algorithm. Most systems rely on pre-set, fixed threshold values or rule libraries to determine whether the action is standard. For example, the system sets a uniform spine inclination threshold, and when the sensor reading exceeds the threshold, it is determined that the action is incorrect. This one-size-fits-all evaluation mode completely ignores the significant biomechanical differences between individual patients mentioned above, resulting in evaluation results lacking scientificity and pertinence. More problematic is that although some systems allow rehabilitation therapists to manually adjust parameters to adapt to different patients, this process is highly dependent on the experience of therapists, and the parameter setting is tedious and lacks quantitative basis, making it difficult to achieve precise adaptation, let alone automatically adjust as the patient's rehabilitation process changes. The root problem is that traditional data processing algorithms lack the ability to autonomously learn and extract key biomechanical features from limited patient individual data, and cannot dynamically establish and update the normal posture baseline model belonging to the patient himself. Therefore, developing an intelligent evaluation algorithm that can adapt to different patient individual characteristics and has online learning ability has become a key technical problem to be solved to improve the effectiveness of scoliosis rehabilitation training and the level of intelligent management. SUMMARY

[0004] The present application is directed to the technical problems existing in the prior art, and provides a scoliosis correction recognition method and system based on a sensing and communication device to solve the problems raised in the background art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: a scoliosis correction recognition method based on a sensing and communication device, comprising the following steps: S1, by collecting a large-scale heterogeneous sensor data set containing multiple scoliosis types and rehabilitation actions, and using biomechanical feature labels annotated by rehabilitation experts, a hybrid model of a deep convolutional neural network and a long short-term memory network is trained to minimize the pre-training loss function, thereby obtaining a pre-training model capable of mapping original sensor time series data to high-level general biomechanical feature vectors; S2, based on the pre-training model, 5-10 standard training action data of a new user are collected, which have been annotated by a rehabilitation therapist; on the basis of the pre-training model bottom feature extractor parameters, only the parameters of the task-specific layer are unlocked, and a feature space alignment loss function based on distribution distance measurement is introduced, which together with the task loss forms a joint optimization objective, and the pre-training model is fine-tuned to generate a personalized evaluation model for the user; S3, using the personalized evaluation model, continuously collecting sensor data streams and making predictions in the subsequent training of new users, obtaining the corresponding time series evaluation results and key biomechanical features; simultaneously calculating the prediction uncertainty estimate; when the prediction uncertainty estimate is continuously lower than the preset uncertainty threshold at multiple time points, automatically triggering the model update process, using the data corresponding to the prediction uncertainty estimate continuously lower than the preset uncertainty threshold at multiple time points as new training samples, and performing online incremental update on the personalized evaluation model, and retaining the time series evaluation results and key biomechanical features generated in the process.

[0006] S4, based on the time series evaluation results and key biomechanical features generated in step S3, constructing a multi-dimensional rehabilitation time series sequence of the user; using a spatio-temporal graph neural network model to model the multi-dimensional rehabilitation time series sequence, learning its spatio-temporal evolution pattern, and then predicting the user's future rehabilitation progress curve, and generating decision support information based on the prediction results of the user's future rehabilitation progress curve.

[0007] In a preferred embodiment, in step S1, a hybrid model of a deep convolutional neural network and a long short-term memory network is trained by assembling a large-scale heterogeneous sensor dataset containing multiple scoliosis types and rehabilitation actions, and using biomechanical feature labels annotated by rehabilitation experts, with the objective of minimizing a pre-training loss function. The process of minimizing the pre-training loss function is specifically operated as follows: A joint loss function is constructed by weighted summation of three loss terms, including: a main regression task loss term based on the mean square error between the model prediction output value and the true biomechanical feature label; a feature structure constraint term introduced by calculating the variance of the high-level general biomechanical feature vector extracted by the model in each feature dimension, and constructing the variance by comparing it with a preset ideal variance value; and a weight regularization term that penalizes the L2 norm of all trainable parameters of the hybrid model of the deep convolutional neural network and the long short-term memory network.

[0008] In a preferred embodiment, the operation of the feature structure constraint term is specifically as follows: First, the feature structure constraint term calculates the empirical covariance matrix based on the feature vector of the current training batch sample, and obtains the variance of each feature dimension on the diagonal elements of the empirical covariance matrix; Subsequently, for each feature dimension, the feature structure constraint term calculates the absolute value of the difference between its variance and the ideal variance value, and then calculates the ratio of the sum of the absolute value and the variance of the feature dimension plus a small constant; Finally, the output value of the feature structure constraint term is the result of subtracting the above ratio from the calculation of each feature dimension and summing all feature dimensions; during the process of minimizing the joint loss function, the feature structure constraint term affects the update direction of the model parameters through its gradient signal.

[0009] In a preferred embodiment, the process of fine-tuning the model in step S2 is achieved by minimizing a joint optimization objective function; The joint optimization objective function is composed of the following two parts by weighted addition: The first part is the main task loss term, which is obtained by calculating the arithmetic mean of the mean square error between the prediction output of the personalized evaluation model and the true biomechanical feature label corresponding to the standard training action data of the new user; The second part is the feature space alignment loss term, which is obtained by calculating the maximum mean difference between the empirical distribution of the high-level general biomechanical feature vector of the new user data and the empirical distribution of the high-level general biomechanical feature vector of the base data subset sampled from the pre-trained base data.

[0010] In a preferred embodiment, the specific calculation operation of the feature space alignment loss term is: Based on the theory of reproducing kernel Hilbert space, the kernel function mean between new user feature vectors, the kernel function mean between new user feature vectors and base data subset feature vectors, and the kernel function mean between base data subset feature vectors are calculated respectively based on the preset kernel function, and the three means are combined in a preset weight ratio to form the final loss value.

[0011] In a preferred embodiment, in step S3, the specific operation of using the personalized evaluation model for prediction and triggering the model update process is: First, based on the personalized evaluation model obtained from step S2, continuously collect the sensor data stream of the new user during the subsequent training process, and perform prediction operation on the sensor data at each time point to obtain the corresponding biomechanical feature output; Second, the prediction uncertainty estimate value at each time point is calculated synchronously during the prediction process, which quantifies the dispersion degree of the model prediction output through the Monte Carlo sampling method, specifically, the random dropout layer is enabled during model inference and multiple forward propagations are performed, and the prediction uncertainty estimate value is calculated according to the difference between the multiple prediction results; Then, a preset uncertainty threshold is set as a trigger condition judgment benchmark, when the prediction uncertainty estimate values of consecutive multiple time points are all lower than the preset uncertainty threshold, the system automatically triggers the model update process; Finally, after triggering the update, the sensor data and its predicted output at the time point corresponding to the low prediction uncertainty estimate value are taken as new training samples to perform online incremental update of the personalized assessment model, and during the update process, the model bottom feature extractor parameters are fixed, and only the task-specific layer parameters are adjusted to ensure the stability and adaptability of the model.

[0012] In a preferred embodiment, the specific calculation method of the prediction uncertainty estimate value is as follows: First, for the sensor data at each time point, a predetermined number of Monte Carlo sampling predictions are performed using the personalized assessment model, and each sampling produces a different prediction output by enabling the random dropout layer in the model; Secondly, the arithmetic mean of the predetermined number of prediction results is calculated as the final prediction output at the time point; Then, based on the sum of squared deviations between the final prediction output and the arithmetic mean of the predetermined number of times, the dispersion measure value of the prediction output is calculated; Finally, the dispersion measure value is taken as the prediction uncertainty estimate value, wherein the number of Monte Carlo sampling times needs to ensure statistical significance, and the dispersion calculation uses a sampling variance-based measurement method.

[0013] In a preferred embodiment, in step S4, the specific operation of constructing the multi-dimensional rehabilitation time sequence is as follows: First, based on the time sequence assessment results output by the dynamically updated personalized assessment model obtained from step S3, the action quality score, key biomechanical feature vector, and training compliance indicator in each training session are extracted; Secondly, the data of the action quality score, key biomechanical feature vector, and training compliance indicator are integrated in chronological order to form a state vector at each time point; then, the state vectors of multiple consecutive time points are arranged in chronological order to form a complete multi-dimensional rehabilitation time sequence; Finally, the multi-dimensional rehabilitation time sequence is standardized to eliminate the dimensional differences between different dimensions of data and ensure the stability of subsequent modeling.

[0014] In a preferred embodiment, the specific operation of modeling the multi-dimensional rehabilitation time sequence using a spatio-temporal graph neural network model is as follows: First, the multi-dimensional rehabilitation time sequence is converted into a spatio-temporal graph structure, where the state vector at each time point is taken as a node in the graph, and the directed edges between the nodes are established according to the time adjacent relationship; Secondly, the layer structure of the spatio-temporal graph neural network is designed, including the alternative stacking of the spatial graph convolution layer and the time dimension convolution layer, the spatial graph convolution layer is used to aggregate the information between adjacent time point nodes, and the time dimension convolution layer is used to learn the evolution law of the state vector in the time dimension; Then, the attention ability of the spatio-temporal graph neural network model to the key time point state is enhanced through the multi-head self-attention mechanism, and the capturing efficiency of the spatio-temporal graph neural network model to the key biomechanical feature state information is improved; Finally, a gated recurrent unit network is used to model the long-term time dependence in the multi-dimensional rehabilitation time sequence, and the rehabilitation state prediction results at future time points are output.

[0015] A scoliosis correction recognition system based on a sensing communication device includes a general biomechanics pre-training model construction module, a personalized evaluation model rapid adaptation module, a dynamic evaluation baseline update module, and a rehabilitation trajectory prediction and decision support module. The general biomechanics pre-training model construction module collects a large-scale heterogeneous sensor data set containing various scoliosis types and rehabilitation actions, and uses biomechanics feature labels annotated by rehabilitation experts to minimize the pre-training loss function. A hybrid model of a deep convolutional neural network and a long short-term memory network is trained to obtain a pre-training model that can map raw sensor time series data to high-level general biomechanics feature vectors. The personalized evaluation model rapid adaptation module collects a small amount of standard training action data of a new user that has been annotated by a rehabilitation therapist. On the basis of fixing the parameters of the pre-training model bottom feature extractor, only the parameters of the task-specific layer are unlocked, and a feature space alignment loss function based on distribution distance measurement is introduced. The joint optimization objective is formed with the task loss to fine-tune the model and generate the personalized evaluation model of the user. The dynamic evaluation baseline update module continuously collects sensor data streams and makes predictions during the subsequent training of a new user, and simultaneously calculates the prediction uncertainty estimate. When the prediction uncertainty estimate is continuously lower than the preset uncertainty threshold, the model update process is automatically triggered, the corresponding low-uncertainty data is used as new training samples, and the personalized evaluation model is updated online. The rehabilitation trajectory prediction and decision support module: based on the time series evaluation results output by the personalized evaluation model and dynamically updated by the dynamic evaluation baseline update module, and the key biomechanics features, a multi-dimensional rehabilitation time sequence of the user is constructed. The sequence is modeled using a spatio-temporal graph neural network model to learn its spatio-temporal evolution pattern, and then the user's future rehabilitation progress curve is predicted, and decision support information is generated based on the prediction results.

[0016] The beneficial effects of the present application are: the present application realizes the leap from generality to individuality in scoliosis rehabilitation evaluation by fusing migration learning and small sample adaptive technology, the system can quickly build a highly personalized evaluation model according to a small amount of user data, and accurately capture the unique biomechanical characteristics of different patients; further introduce the uncertainty estimation driven online learning mechanism, so that the model can be self-optimized and adjusted along with the user's rehabilitation process, dynamically adapt to the user's ability changes; finally, with the help of the spatio-temporal graph neural network, the rehabilitation trajectory is prospectively modeled, and the decision support information with early warning function and personalized guiding significance is generated, thereby effectively solving the core pain points of traditional evaluation methods relying on fixed threshold, lack of dynamic and predictive, and significantly improving the accuracy, adaptability and intelligent level of home rehabilitation training. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The method flowchart of the present application is shown in the figure; Figure 2 The system structure block diagram of the present application is shown in the figure; DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] In the description of the present application, the terms "first", "second" are used only for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0020] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope consistent with the principles and characteristics disclosed in the present application.

[0021] The embodiment provides a scoliosis correction recognition method based on a sensing communication device as shown in the specification Figure 1 The embodiment provides a scoliosis correction recognition method based on a sensing communication device as shown in the specification S1, a deep convolutional neural network and a long short-term memory network hybrid model is trained by collecting a large-scale heterogeneous sensor data set containing various scoliosis types and rehabilitation actions, and using the biomechanics feature labels labeled by rehabilitation experts, so as to minimize a pre-training loss function, thereby obtaining a pre-training model capable of mapping original sensor time series data to a high-level general biomechanics feature vector, and the effect of the step is to build a general feature extraction basis rich in prior knowledge, so that the model can learn a stable biomechanics representation from massive data, and provide a high starting point and strong generalization ability for subsequent personalized adaptation; S2, based on the pre-training model, a small amount of standard training action data of a new user is collected, the parameters of the task-specific layer are unlocked on the basis of fixing the parameters of the pre-training model bottom feature extractor, and a feature space alignment loss function based on distribution distance measurement is introduced, and the joint optimization target is formed with the task loss, the model is fine-tuned, and the personalized evaluation model of the user is generated, and the effect of the step is to quickly realize the precise adaptation of the model to the unique biomechanics mode of the new user through small sample learning and feature alignment mechanism, avoid overfitting and distribution deviation, and ensure the reliability and stability of personalized evaluation; S3, using the personalized evaluation model, in the subsequent training of the new user, the sensor data stream is continuously collected and predicted, and the prediction uncertainty estimation value is calculated synchronously; when the prediction uncertainty estimation value is continuously lower than the preset uncertainty threshold, the model updating process is automatically triggered, the corresponding low uncertainty data is used as a new training sample, and the personalized evaluation model is updated online, and the effect of the step is that through dynamic uncertainty monitoring and triggered updating, the model can be optimized automatically with the user's rehabilitation progress, the timeliness and accuracy of the evaluation benchmark are continuously maintained, and the long-term cognitive process of the rehabilitation therapist is simulated; S4, based on the time series evaluation results and key biomechanics features output by the personalized evaluation model, a multi-dimensional rehabilitation time series sequence of the user is constructed; a spatio-temporal graph neural network model is used to model the sequence, learn the spatio-temporal evolution mode, predict the future rehabilitation progress curve of the user, and generate decision support information based on the prediction result, and the effect of the step is to provide prospective rehabilitation trend insight and intelligent decision support such as risk warning and personalized training suggestion through spatio-temporal pattern analysis and multi-step prediction, and finally form a closed-loop intelligent system from perception to decision.

[0022] In this embodiment, it is specifically pointed out that in step S1, step S1 is aimed at constructing a general feature extraction model rich in prior knowledge, which is not directly used for individual assessment, but serves as a cornerstone for all subsequent personalized adaptation operations, and the core goal is to learn a robust mapping from low-level, complex raw sensor data to high-level, feature representation with clear biomechanical significance, so as to compress and encode the general knowledge contained in the massive population data into the model parameters.

[0023] In step S1, the pre-training loss function is minimized The specific operation of the process is as follows: A joint loss function is realized by constructing a weighted sum of three loss terms, including: a main regression task loss term based on the mean square error between the model prediction output value and the true biomechanical feature label; a feature structure constraint term introduced, which is constructed by calculating the variance of the high-level general biomechanical feature vector extracted by the model in each feature dimension, and comparing the variance with a preset ideal variance value; and a weight regularization term that penalizes the L2 norm of all trainable parameters of the model.

[0024] In specific implementation, the pre-training process of step S1 is as follows: first, a large-scale heterogeneous sensor data set is collected, which contains time series data from inertial measurement units (IMU) and surface electromyography sensors (sEMG), and has been labeled with biomechanical feature labels by rehabilitation experts , where is the sample index. The high-level general biomechanical feature vector is extracted by a hybrid model of deep convolutional neural network (CNN) and long short-term memory network (LSTM), which takes the raw sensor data as input, the CNN module extracts spatial features from a local time window, and the LSTM module models the time dependence. The mathematical expression of the joint loss function is as follows: ; , where is the total number of samples, representing the size of the pre-training data set, used for normalizing the loss; is the true biomechanical feature label vector of the i-th sample, labeled by experts to ensure the accuracy of supervised learning; is the high-level general biomechanical feature vector of the i-th sample, extracted by the hybrid model as a feature representation; ​​​The regression layer function, composed of a fully connected layer and a ReLU activation function, achieves a nonlinear mapping to capture the complex relationship between feature vectors and biomechanical labels, thereby transforming the feature vectors... Mapped to the label space for output prediction; The mean squared error between the predicted and actual values ​​is calculated using the L2 norm squared and used as the loss term for the main regression task to ensure that the model output is consistent with clinical knowledge. The empirical covariance matrix is ​​based on the feature vectors of all samples in the current training batch. Calculate and describe the correlations between feature dimensions; The first covariance matrix is ​​the first... The diagonal element represents the first diagonal element. The variance of a feature is used for feature structure analysis; For feature vectors The dimension determines the range of the summation; Small constants (e.g.) This prevents the denominator from being zero and ensures numerical stability. These are hyperparameters that control the weights of the feature structure constraint term and the weight regularization term, balancing the loss component. Their values ​​need to be optimized based on the validation set performance, and a typical empirical range is [insert range here]. The initial value can be set to 0.01. When optimizing, it is recommended to use grid search or Bayesian optimization, and adjust it with reference to the distribution of clinical data (such as the characteristic variance of different types of scoliosis). For hybrid models The set of all trainable parameters is optimized and adjusted.

[0025] This formula addresses the representation quality and generalization issues in general feature learning by weighted summing of three loss terms: the main regression task loss term ensures feature alignment with labels; the feature structure constraint term promotes feature decoupling; and the weight regularization term prevents overfitting.

[0026] The specific operations of the feature structure constraint terms are as follows: First, the feature structure constraint term calculates its empirical covariance matrix based on the feature vectors of the current training batch samples, and obtains the variance of the diagonal elements of the matrix, i.e., each feature dimension. Subsequently, the constraint term calculates the absolute value of the difference between the variance of each feature dimension and the ideal variance value, and then the ratio of that variance to the sum of a small constant. Ultimately, the output value of this constraint term is the sum of all feature dimensions after subtracting the aforementioned ratio; in the process of minimizing the joint loss function, this feature structure constraint term affects the update direction of the model parameters through its gradient signal.

[0027] In practice, the operational mechanism of the feature structure constraint term is refined into mathematical operations: First, for the current training batch, the high-level general biomechanical feature vector of all samples is calculated. empirical covariance matrix ,in , For batch size, The mean of the batch feature vectors; Then, extract the diagonal elements. (i.e., the first) The variance of the features is ideally set to 1, which aims to promote decoupling and scale uniformity of the various dimensions of the feature space and avoid a single dimension dominating model training. This setting is consistent with subsequent data standardization (such as Z-score) and is based on machine learning best practices to ensure the robustness of feature representation. Next, for each dimension Calculate the ratio ,in It is a small constant; Then, the output of the feature structure constraint term is In minimizing the joint loss function At that time, the gradient signal of this constraint term is backpropagated through the gradient descent algorithm, driving the model parameters. Update, adjust the variance of each dimension Approaching 1; this mechanism solves the problem of dimensional redundancy or suppression in feature representation, ensuring that the feature space is decoupled and information-rich.

[0028] Step S1 successfully constructed a hybrid model of a general biomechanical knowledge base model by introducing an innovative multi-task pre-training loss function that includes feature structure constraints. The features extracted by this model have good separability, robustness and rich biomechanical prior knowledge, which lays an indispensable foundation for achieving robust personalized adaptation on a very small number of samples.

[0029] In this embodiment, it should be specifically noted that step S2 is the core innovative step of the overall method, and its goal is to obtain the general pre-trained model in step S1. Based on this, and utilizing a very small amount of labeled data from new users, we achieve rapid and robust personalized adaptation. The pre-trained model produced in step S1... While it provides rich prior biomechanical knowledge, its direct application to individual users can lead to evaluation bias due to individual differences. Step S2 addresses overfitting and distribution shift issues in small-sample scenarios by introducing a feature space alignment mechanism, ensuring that the personalized model adapts to the user's unique patterns while maintaining the constraints of general knowledge. The output of this step is a personalized evaluation model, which provides the basis for the dynamic update in step S3.

[0030] In step S2, the process of fine-tuning the model is achieved by minimizing a joint optimization objective function; The joint optimization objective function is composed of two parts weighted and added together: The first part is the main task loss term, which is obtained by calculating the arithmetic mean of the mean square error between the prediction output of the personalized evaluation model and the true biomechanical feature label corresponding to the new user labeled data; The second part is the feature space alignment loss term, which is obtained by calculating the maximum mean difference between the empirical distribution of the high-level general biomechanical feature vector of the new user data and the empirical distribution of the high-level general biomechanical feature vector of the basic data subset sampled from the pre-training basic data.

[0031] In specific implementation, the fine-tuning process of step S2 is as follows: first, for the new user , a small amount of standard training action data annotated by a rehabilitation therapist is collected , which contains sensor time series data and corresponding biomechanical feature labels ; using the pre-trained model obtained in step S1, the high-level general biomechanical feature vector of the new user data is extracted . At the same time, a basic data subset is randomly sampled from the large-scale data set used in the pre-training stage , the number of samples is (typically , m is the number of new user samples), and its feature vector is extracted , where the pre-training model parameters remain fixed, the personalized adaptation mechanism automatically handles the lateral bending type difference through the feature space alignment loss, without manual adjustment, and the model constrains the new user feature distribution to be close to the multi-type basic distribution of the pre-training data during fine-tuning, thereby adapting to different pathological characteristics such as C-type and S-type; the mathematical expression of the joint optimization objective function is as follows: where is the joint optimization objective function, representing the total loss to be minimized in step S2, used to balance personalization and generalization; is a weighted hyperparameter, taking a positive real number, typically ∈[0.1,1.0], used to control the weight of the feature space alignment loss term; is the feature space alignment loss term, based on maximum mean difference (MMD) calculation, used to measure the distribution distance; The main task loss term is calculated using the following formula: New user annotation data The number of samples is a positive integer, usually m ≤ 10. The setting of 10 samples is based on the theory of small sample learning and clinical feasibility. The pre-trained model provides strong prior knowledge. When fine-tuning, only a small amount of high-quality labeled data (5-10 samples) is needed to capture individual specificity. The samples need to be labeled by rehabilitation therapists to ensure the standardization of movements and cover the user's typical rehabilitation movement patterns. The true biomechanical feature label vector for the i-th new user sample is labeled by the rehabilitation therapist to ensure the monitoring signal; To personalize the evaluation model based on input sensor data The predicted output, where θ is the fine-tuning parameter (the underlying parameters are fixed, and only the task layer parameters are updated); The squared L2 norm is used to calculate the mean square error between the predicted and actual values; the purpose of the main task loss term is to ensure the accuracy of the personalized model's predictions for new user data, adapting to their unique biomechanical patterns.

[0032] Joint optimization objective function The mathematical expression addresses overfitting and distribution shift issues in small-sample fine-tuning by weighted summation of two loss terms: the main task loss term drives the model to fit new user data; the feature space alignment loss term constrains the feature distribution to avoid deviating from general knowledge.

[0033] The specific calculation operation of the feature space alignment loss term is as follows: Based on the Hilbert space theory of regeneration kernel, the mean kernel function between new user feature vectors, the mean kernel function between new user feature vectors and feature vectors of the basic data subset, and the mean kernel function between feature vectors of the basic data subset are calculated by pre-setting kernel functions. These three means are then combined with specific weights to form the final loss value.

[0034] In practical implementation, the feature space alignment loss term The calculation is based on the Regenerative Kernel Hilbert Space (RKHS) theory, and its specific mathematical expression is as follows: ; in, The high-level general biomechanical feature vector of the i-th sample of the new user is extracted by a pre-trained model, with the same dimension as in step S1. basic data subset The high-level general biomechanical feature vector of the j-th sample is extracted using a pre-trained model with fixed parameters. The number of new user samples is defined as before; The value is the number of samples in the subset of basic data, and it is a positive integer, usually n=1000, to ensure the stability of the distribution estimation; The default kernel function is usually a Gaussian kernel function. ,in Kernel bandwidth is a hyperparameter used to control the similarity metric scale. Optimization is based on data distribution characteristics and can employ a median heuristic (such as assuming...). (The median of the distance between samples), or maximizing the power of the MMD statistic, specifically, This ensures that the sensitivity of distribution difference measurement matches the degree of variation in clinical characteristics.

[0035] The purpose of the feature space alignment loss term is to calculate the difference between the new user feature distribution and the basic feature distribution through MMD, and use it as a regularization term to constrain the personalized model from deviating from the normal biomechanical range and improve the robustness of few-shot learning.

[0036] The formula consists of three parts corresponding to the kernel function mean: the first part calculates the similarity within the new user's features; the second part calculates the similarity between the new user and the baseline data; and the third part calculates the similarity within the baseline data. Through linear combination, this loss term quantizes the distribution distance in RKHS, addressing the feature distribution shift problem.

[0037] In this embodiment, it should be specifically noted that in step S3, step S3 follows the personalized evaluation model output in step S2. The aim is to achieve continuous adaptive optimization of the model during the deployment phase. The core of this step is to automatically trigger an incremental update mechanism by monitoring the uncertainty of the model's predictions, so that the model can adapt to the user's dynamically changing biomechanical state and ensure the real-time accuracy of the evaluation results.

[0038] In step S3, the specific operations for using the personalized evaluation model to make predictions and trigger the model update process are as follows: First, based on the personalized evaluation model obtained from step S2, sensor data streams of new users are continuously collected during subsequent training, and prediction operations are performed on the sensor data at each time point to obtain the corresponding biomechanical feature output. In specific implementation, the process of step S3 is as follows: First, based on the personalized evaluation model obtained from step S2... , This indicates the model parameters after fine-tuning in step S2, and the sensor data streams of new users are continuously collected during subsequent training. ,in This represents the total number of time points. Indicates a point in time The sensor time series data is used to perform prediction operations on the sensor data at each time point to obtain the corresponding biomechanical feature output; The specific implementation of the forecasting operation is as follows: for each time point Sensor data Using personalized assessment models conduct Sub-Monte Carlo sampling prediction, each sampling generates a different prediction output by enabling a random dropout layer in the model, resulting in a prediction set. ,in Indicates the first The predicted output vector of the next sample.

[0039] Secondly, during the prediction process, the prediction uncertainty estimate at each time point is calculated simultaneously. This prediction uncertainty estimate is quantified by the Monte Carlo sampling method to quantify the dispersion of the model prediction output. Specifically, a random dropout layer is enabled and multiple forward propagations are performed during model inference, and the uncertainty is calculated based on the degree of difference between the multiple prediction results. The specific calculation of uncertainty estimation is as follows: predict the uncertainty estimate value. Calculated using the following formula: ; in, For time points The estimated value of the prediction uncertainty is used to quantify the degree of dispersion of the model prediction; The number of Monte Carlo samples, taking a positive integer value, is usually ( ≥30), ensuring statistical significance; For the first The predicted output vector of the next sample; To predict the mean, This formula quantifies uncertainty by calculating the variance of multiple prediction results, thus solving the problem of evaluating the confidence level of model predictions.

[0040] Then, a preset uncertainty threshold is set as the benchmark for triggering the condition. When the estimated uncertainty of the prediction at multiple consecutive time points is lower than the preset uncertainty threshold, the system automatically triggers the model update process. The specific implementation of the triggering condition is as follows: Set a preset uncertainty threshold. When the condition is satisfied for (N) consecutive time points (e.g., N=5), At that time, the system automatically triggers the model update process, threshold. Quantiles are determined based on the distribution of historical data.

[0041] Finally, after the update is triggered, the sensor data and its prediction output at the time point corresponding to the low prediction uncertainty estimate are used as new training samples to perform online incremental updates on the personalized evaluation model. During the update process, the parameters of the model's underlying feature extractor are fixed, and only the parameters of the task-specific layer are adjusted to ensure the model's stability and adaptability. The specific implementation of online incremental updates is as follows: during the collection trigger period, the following conditions are met. data points and its predicted mean As pseudo-labels, they form a new training sample set. During the update process, the parameters of the underlying feature extractor of the model are fixed, and only the parameters of the task-specific layer are unlocked. This is achieved by minimizing the loss function. Perform incremental updates.

[0042] The specific calculation method for the estimate of the uncertainty in the forecast is as follows: First, for the sensor data at each time point, a predetermined number of Monte Carlo sampling predictions are performed using a personalized evaluation model. Each sampling produces a different prediction output by enabling a random dropout layer in the model. The specific implementation of Monte Carlo sampling prediction is as follows: set the number of sampling times. During model inference, the randomly dropped layer is kept active, and (K) forward propagations are performed to obtain K predicted outputs. .

[0043] Secondly, the arithmetic mean of the prediction results for the predetermined number of predictions is calculated as the final prediction output for that time point; The specific implementation of the predicted mean calculation is as follows: ; in As the final prediction output at this point in time, it is used for subsequent uncertainty calculations and model updates.

[0044] Then, based on the sum of squared deviations between the predicted output and the arithmetic mean of the predetermined number of predictions, a measure of the dispersion of the predicted output is calculated. The specific formula for calculating the dispersion measure is as follows: ; This formula calculates the variance of the predicted output, which serves as a quantitative indicator of dispersion.

[0045] Finally, the dispersion measure is used as an estimate of the prediction uncertainty. The number of Monte Carlo samplings must be statistically significant, and the dispersion is calculated using a measurement method based on sampling variance.

[0046] The final determination of the uncertainty estimate is as follows: Monte Carlo sampling number To ensure statistical significance, typically ( To ensure the reliability of uncertainty estimation, the dispersion calculation adopts a metric method based on sampling variance, which can effectively reflect the stability of model prediction.

[0047] In this embodiment, it is particularly important to note that step S4 is the final step in the overall technical process. Its core objective is to perform advanced analysis on the time-series data generated by the personalized assessment model after dynamic updates in step S3, thereby achieving predictable management of the rehabilitation process. This step follows the continuously updated assessment results output by step S3, and by constructing a spatiotemporal graph structure and utilizing advanced graph neural network technology, it learns the dynamic evolution law of the rehabilitation trajectory, ultimately forming a forward-looking decision support capability.

[0048] In step S4, the specific operation for constructing the multidimensional rehabilitation time sequence is as follows: First, based on the time-series evaluation results output by the dynamically updated personalized evaluation model obtained from step S3, the action quality score, key biomechanical feature vectors, and training compliance indicators are extracted from each training session. In specific implementation, the sequence construction process of step S4 is as follows: First, based on the dynamically updated personalized evaluation model obtained from step S3... The output temporal evaluation results extract the action quality score for each training session. Key biomechanical feature vectors and training compliance indicators Among them, the quality of the movement is scored. From the model The key biomechanical feature vector is calculated from the evaluation output of the standard movement. From the model The intermediate layer extraction maintains the same dimensions as defined in step S1, and the training compliance metric is used. This includes quantitative data such as training duration and frequency.

[0049] The extraction of action quality scores, key biomechanical feature vectors, and training compliance metrics from each training session is specifically implemented as follows: For each training session, the action quality score is obtained from the model output. This score reflects the quality of action completion in a single training session; key biomechanical feature vectors are extracted. This vector contains detailed information such as spinal curvature and muscle activation patterns, with dimensions of [missing information]. The feature output dimension is consistent with that of the pre-trained model in step S1; calculate the training compliance index. Data such as session duration and number of repetitions together constitute the raw data tuple for each point in time.

[0050] Secondly, the data of movement quality scores, key biomechanical feature vectors, and training compliance indicators are integrated in chronological order to form a state vector at each time point; then, the state vectors of multiple consecutive time points are arranged in chronological order to construct a complete multidimensional rehabilitation time sequence. The specific implementation of integration and construction is as follows: integrate the data tuples at each time point into a state vector. Its mathematical representation is: ; in, No. A state vector at each time point is used to represent the overall recovery status at that time point. The quality score for the movement is a scalar, with a value range of [0,1]. A higher value indicates better movement quality. Key biomechanical feature vectors (vectors), dimensions Defined by the pre-trained model in step S1, typically ( ≥50), including abstract biomechanical characteristics; To train compliance metrics (scalars), values ​​are taken as positive real numbers, such as training duration (minutes) or number of repetitions. Subsequently, the state vectors at multiple consecutive time points are arranged in chronological order to construct a multidimensional rehabilitation time series. Its mathematical representation is: ; in The total number of time points is determined by the user's training history, and this sequence covers all state changes from the start of training to the current state.

[0051] Finally, the sequence is standardized to eliminate the dimensional differences between data of different dimensions, ensuring the stability of subsequent modeling.

[0052] The specific implementation of standardization is as follows: for the sequence Each dimension in the score is standardized separately for the action quality rating. and training compliance indicators The key biomechanical feature vectors are mapped to the range [0,1] using minimum-maximum scaling; Z-score standardization is used to make the mean 0 and the variance 1. The standardization formula is as follows: For scalar dimensions: ; For vector dimensions: ; in and These are the mean and standard deviation vectors of the eigenvector sequence, and the standardized sequence. This is used for subsequent modeling to ensure that data from different dimensions are comparable and to avoid model bias caused by differences in units; The calculation range of the statistics required for standardization (such as minimum, maximum, mean, and standard deviation) needs to be clearly defined during the model training phase, usually using one of the following two methods: Global standardization: Statistics are calculated based on the entire training dataset, for example, and These are action quality scores for all users and at all time points on the training set. The global minimum and maximum values. At this point, and For all feature vectors on the training set The global mean and standard deviation vector.

[0053] Individualized sliding window standardization: Statistics are calculated based on a window of historical data from the current user over a recent period; for example... and The rating sequence can be calculated from the user's most recent N training sessions (e.g., N=30). At this point, and These represent the mean and standard deviation of the feature vectors within the corresponding window.

[0054] The standardized sequences s', c', h' are used for subsequent modeling. This is intended to eliminate the dimensional differences between data of different dimensions, ensure their comparability, and thus avoid model bias caused by them.

[0055] Movement quality score reflects the quality of movement completion in a single training session, key biomechanical feature vectors carry specific movement pattern information, and training compliance indexes quantify training intensity and frequency. Together, these three constitute multidimensional time-series data that comprehensively reflects the user's rehabilitation status.

[0056] The specific steps for modeling multidimensional rehabilitation time-series sequences using a spatiotemporal graph neural network model are as follows: First, the multidimensional rehabilitation time sequence is transformed into a spatiotemporal graph structure, where the state vector of each time point is used as a node in the graph, and directed edges are established between nodes based on temporal adjacency. The sensor data acquisition frequency is set to 50Hz, and time point t corresponds to a training action cycle or a fixed window (e.g., 2 seconds). In practical implementation, the modeling process of step S4 is as follows: First, the standardized multidimensional rehabilitation time sequence is... Transform into a spacetime graph structure Each node corresponds to a state vector at a given time point, and edges are established based on temporal adjacency. Subsequently, a spatiotemporal graph neural network (ST-GNN) model is designed, which includes spatial graph convolutional layers and temporal convolutional layers to learn spatiotemporal patterns in sequences.

[0057] The specific implementation of graph structure construction is as follows: Define the graph. , where the node set edge set ( ),in The time window size is usually taken as ( =3) or ( =5), indicating that each node is only related to the previous one. Nodes at different points in time are connected by directed edges, which capture the causal dependencies in time.

[0058] Secondly, the layer structure of the spatiotemporal graph neural network is designed, which includes alternating stacks of spatial graph convolutional layers and temporal convolutional layers. The spatial graph convolutional layers are used to aggregate information between nodes at adjacent time points, and the temporal convolutional layers are used to learn the evolution law of the state vector in the time dimension. The specific implementation of the network design is as follows: The ST-GNN model consists of multiple spatiotemporal blocks stacked together. Each block contains a spatial graph convolutional layer and a temporal convolutional layer. The spatial graph convolutional layer uses a graph attention mechanism (GAT) to aggregate information, as shown in the formula: ; in, For the first Layer nodes Feature representation; For nodes The set of neighboring nodes is composed of the edge set. definition; Attention weights, calculated through an attention mechanism, reflect the node's... For nodes The importance of; For the first The trainable weight matrix of the layer; For example, ReLU is an activation function.

[0059] The temporal convolutional layer uses one-dimensional convolution (Conv1D) to capture the temporal pattern, and the formula is as follows: ; in For all nodes at the 1st The feature matrix of the layer; the alternation of spatial and temporal layers enables the model to learn spatial correlations and temporal evolution simultaneously.

[0060] Then, a multi-head self-attention mechanism is used to enhance the model's ability to focus on key time point states, thereby improving the model's efficiency in capturing important information, and increasing the attention weights. Calculated using a query-key mechanism: ,in , It is a trainable projection matrix.

[0061] The specific implementation of the attention mechanism is as follows: multi-head self-attention is applied in the time dimension, and the formula is: ; in These are the query, key, and value matrices, respectively, obtained by projecting the input features. is the dimension of the key vector. The multi-head mechanism allows the model to focus on information from different subspaces, enhancing its ability to identify key time points (such as points where training performance changes abruptly).

[0062] Finally, a gated recurrent unit network is used to model the long-term time dependency relationship and output the prediction results of the rehabilitation status at multiple future time points.

[0063] The specific implementation of long-term modeling is as follows: Time series data are processed using a gated recurrent unit (GRU) network. The update formula for the GRU is: ; ; ; ; in, For time points The input features (from the output of the previous layer); For time points The hidden state; To update the gate, control the degree of state update; To reset the door and control the degree to which historical information is forgotten; The weight matrix is ​​trainable. It is a sigmoid activation function.

[0064] After the GRU network processes the entire sequence, the output layer generates the future sequence. Predicted state at each time point ( ),in To predict the window size, we typically take ( =5); the prediction results are used to generate decision support information, such as recovery progress curves and risk warnings.

[0065] It should be noted that when constructing the multidimensional rehabilitation time series, the sensor data acquisition frequency can be set to 50Hz, each time point t can be defined as a complete training action cycle, or a fixed time window (e.g., 2 seconds), the training compliance index can be quantified as the duration of a single training session in minutes, with a typical duration between 15 and 45 minutes, the length T of the time series can be set to 30, that is, using data from the most recent 30 training sessions; the time window size k used to construct the spatiotemporal map can be set to 3 or 5, and the prediction window size W can be set to 5 (i.e., predicting the state of the next 5 training sessions). These parameters can be adjusted according to the actual rehabilitation assessment needs.

[0066] Step S4 transforms the traditional time-series prediction problem into a graph structure learning problem through an innovative spatiotemporal graph neural network architecture, making full use of the complex spatiotemporal dependencies between states at different time points during the rehabilitation process.

[0067] Example 2 This embodiment provides, for example Figure 2 The system shown is a scoliosis correction and recognition system based on sensor communication devices, including a general biomechanical pre-training model construction module, a personalized assessment model rapid adaptation module, a dynamic assessment baseline update module, and a rehabilitation trajectory prediction and decision support module. The general biomechanical pre-training model building module gathers a large-scale heterogeneous sensor dataset containing various types of scoliosis and rehabilitation movements, and uses biomechanical feature labels annotated by rehabilitation experts to train a hybrid model of deep convolutional neural network and long short-term memory network with the goal of minimizing the pre-training loss function. This results in a pre-training model that can map the original sensor time-series data into high-level general biomechanical feature vectors. The personalized assessment model rapid adaptation module collects a small amount of standard training action data that has been labeled by rehabilitation therapists for new users; on the basis of fixing the parameters of the underlying feature extractor of the pre-trained model, it unlocks only the parameters of the task-specific layer and introduces a feature space alignment loss function based on the distribution distance metric, which together with the task loss constitutes a joint optimization objective to fine-tune the model and generate a personalized assessment model for the user. The dynamic evaluation baseline update module continuously collects sensor data streams and makes predictions during the subsequent training of new users, and simultaneously calculates the prediction uncertainty estimate. When the prediction uncertainty estimate is continuously lower than the preset uncertainty threshold, the model update process is automatically triggered, and the corresponding low uncertainty data is used as new training samples to perform online incremental updates to the personalized evaluation model. The rehabilitation trajectory prediction and decision support module: Based on the time-series assessment results and key biomechanical characteristics output by the personalized assessment model and dynamically updated by the dynamic assessment baseline update module, a multidimensional rehabilitation time-series sequence for the user is constructed. The spatiotemporal graph neural network model is used to model this sequence, learn its spatiotemporal evolution pattern, and then predict the user's future rehabilitation progress curve. Based on the prediction results, decision support information is generated, including: risk warning (such as prompting intervention when the predicted progress curve deviates), personalized training suggestions (such as adjusting the range of motion or increasing muscle group training), and visualization report (generating a rehabilitation progress curve).

[0068] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0069] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0070] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0071] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0072] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for scoliosis correction and identification based on sensor communication devices, characterized in that, Includes the following steps: S1. By aggregating a large-scale heterogeneous sensor dataset containing various types of scoliosis and rehabilitation movements, and using biomechanical feature labels annotated by rehabilitation experts, a hybrid model of deep convolutional neural network and long short-term memory network is trained with the goal of minimizing the pre-training loss function, thereby obtaining a pre-trained model that can map the original sensor time-series data into high-level general biomechanical feature vectors. S2. Based on the pre-trained model, collect 5-10 standard training action data that have been labeled by rehabilitation therapists for new users; on the basis of the parameters of the bottom feature extractor of the pre-trained model, only unlock the parameters of the task-specific layer, and introduce a feature space alignment loss function based on the distribution distance metric, which together with the task loss constitutes a joint optimization objective to fine-tune the pre-trained model and generate a personalized evaluation model for the user. S3. Using a personalized evaluation model, in the subsequent training of new users, their sensor data streams are continuously collected and predicted to obtain the corresponding time-series evaluation results and key biomechanical characteristics. Simultaneously calculate the prediction uncertainty estimate; when the prediction uncertainty estimate is lower than the preset uncertainty threshold for multiple consecutive time points, the model update process is automatically triggered, and the corresponding data with prediction uncertainty estimates lower than the preset uncertainty threshold for multiple consecutive time points are used as new training samples to perform online incremental updates to the personalized evaluation model, while retaining the time-series evaluation results and key biomechanical features generated in the process. S4. Based on the time-series assessment results and key biomechanical characteristics generated in step S3, construct the user's multidimensional rehabilitation time-series sequence; use a spatiotemporal graph neural network model to model the multidimensional rehabilitation time-series sequence, learn its spatiotemporal evolution pattern, and then predict the user's future rehabilitation progress curve, and generate decision support information based on the prediction results of the user's future rehabilitation progress curve.

2. The scoliosis correction and identification method based on sensor communication device according to claim 1, characterized in that: In step S1, by aggregating a large-scale heterogeneous sensor dataset containing various types of scoliosis and rehabilitation movements, and utilizing biomechanical feature labels annotated by rehabilitation experts, a hybrid model of a deep convolutional neural network and a long short-term memory network is trained with the goal of minimizing the pre-training loss function. The specific operation of minimizing the pre-training loss function is as follows: This is achieved by constructing a joint loss function consisting of a weighted sum of three loss terms. The function includes: a main regression task loss term based on the mean squared error between the model's predicted output value and the true biomechanical feature labels; an introduced feature structure constraint term, which is constructed by calculating the variance of the high-level general biomechanical feature vector extracted by the model in each feature dimension and comparing this variance with a preset ideal variance value; and a weighted regularization term that penalizes the L2 norm of all trainable parameters of the hybrid model of deep convolutional neural network and long short-term memory network.

3. The scoliosis correction and identification method based on a sensor communication device according to claim 2, characterized in that: The specific operations of the feature structure constraint terms are as follows: First, the feature structure constraint term calculates its empirical covariance matrix based on the feature vectors of the current training batch samples, and obtains the diagonal elements of the empirical covariance matrix, i.e., the variance of each feature dimension. Subsequently For each feature dimension, the feature structure constraint term calculates the absolute value of the difference between its variance and the ideal variance value, and then calculates the ratio of this absolute value to the sum of the variance of that feature dimension plus a small constant. Ultimately, the output value of the feature structure constraint term is the result of calculating a value for each feature dimension minus the aforementioned ratio, and summing the values ​​over all feature dimensions. In the process of minimizing the joint loss function, this feature structure constraint term influences the update direction of the model parameters through its gradient signal.

4. The scoliosis correction and identification method based on a sensor communication device according to claim 3, characterized in that: In step S2, the process of fine-tuning the model is achieved by minimizing a joint optimization objective function; The joint optimization objective function consists of the following two weighted sums: The first part is the main task loss term, which is obtained by calculating the arithmetic mean of the mean square error between the predicted output of the personalized evaluation model and the real biomechanical feature labels corresponding to the standard training action data of the new user. The second part is the feature space alignment loss term, which is obtained by calculating the maximum mean difference between the empirical distribution of the high-level general biomechanical feature vectors of the new user data and the empirical distribution of the high-level general biomechanical feature vectors of the subset of base data sampled from the pre-trained base data.

5. The scoliosis correction and identification method based on a sensor communication device according to claim 4, characterized in that: The specific calculation operation of the feature space alignment loss term is as follows: Based on the Hilbert space theory of regeneration kernel, the mean kernel function between new user feature vectors, the mean kernel function between new user feature vectors and feature vectors of the basic data subset, and the mean kernel function between feature vectors of the basic data subset are calculated by pre-set kernel functions. These three means are then combined according to a pre-set weight ratio to form the final loss value.

6. The scoliosis correction and identification method based on a sensor communication device according to claim 5, characterized in that: In step S3, the specific operations for using the personalized evaluation model to make predictions and trigger the model update process are as follows: First, based on the personalized evaluation model obtained from step S2, sensor data streams of new users are continuously collected during subsequent training, and prediction operations are performed on the sensor data at each time point to obtain the corresponding biomechanical feature output. Secondly, during the prediction process, the prediction uncertainty estimate is calculated synchronously at each time point. This prediction uncertainty estimate is quantified by the Monte Carlo sampling method to quantify the dispersion of the model prediction output. Specifically, a random dropout layer is enabled and multiple forward propagations are performed during model inference. The prediction uncertainty estimate is calculated based on the degree of difference between the multiple prediction results. Then, a preset uncertainty threshold is set as the trigger condition judgment benchmark. When the estimated value of the prediction uncertainty at multiple consecutive time points is lower than the preset uncertainty threshold, the system automatically triggers the model update process. Finally, after the update is triggered, the sensor data and their prediction output at the time point corresponding to the low prediction uncertainty estimate are used as new training samples to perform online incremental updates on the personalized evaluation model. During the update process, the parameters of the model's underlying feature extractor are fixed, and only the parameters of the task-specific layer are adjusted to ensure the model's stability and adaptability.

7. The scoliosis correction and identification method based on a sensor communication device according to claim 6, characterized in that: The specific calculation method for the estimate of the uncertainty in the forecast is as follows: First, for the sensor data at each time point, a predetermined number of Monte Carlo sampling predictions are performed using a personalized evaluation model. Each sampling produces a different prediction output by enabling a random dropout layer in the model. Secondly, the arithmetic mean of the prediction results for the predetermined number of predictions is calculated as the final prediction output for that time point; Then, based on the sum of squared deviations between the final predicted output and the arithmetic mean of the predetermined number of times, the dispersion measure of the predicted output is calculated; Finally, the dispersion measure is used as an estimate of the prediction uncertainty. The number of Monte Carlo samplings must be statistically significant, and the dispersion is calculated using a measurement method based on sampling variance.

8. The scoliosis correction and identification method based on a sensor communication device according to claim 7, characterized in that: In step S4, the specific operation for constructing the multidimensional rehabilitation time sequence is as follows: First, based on the time-series evaluation results output by the dynamically updated personalized evaluation model obtained from step S3, the action quality score, key biomechanical feature vectors, and training compliance indicators are extracted from each training session. Secondly, the data of movement quality scores, key biomechanical feature vectors, and training compliance indicators are integrated in chronological order to form a state vector at each time point; then, the state vectors of multiple consecutive time points are arranged in chronological order to construct a complete multidimensional rehabilitation time sequence. Finally, the multidimensional rehabilitation time series was standardized to eliminate the dimensional differences between different dimensions of data and ensure the stability of subsequent modeling.

9. The scoliosis correction and identification method based on a sensor communication device according to claim 8, characterized in that: The specific steps for modeling multidimensional rehabilitation time-series sequences using a spatiotemporal graph neural network model are as follows: First, the multidimensional rehabilitation time series is transformed into a spatiotemporal graph structure, where the state vector at each time point is used as a node in the graph, and directed edges are established between nodes based on their temporal adjacency. Secondly, the layer structure of the spatiotemporal graph neural network is designed, which includes alternating stacks of spatial graph convolutional layers and temporal convolutional layers. The spatial graph convolutional layers are used to aggregate information between adjacent time point nodes, and the temporal convolutional layers are used to learn the evolution law of the state vector in the time dimension. Then, the spatiotemporal graph neural network model is enhanced to pay attention to the state at key time points by a multi-head self-attention mechanism, thereby improving the efficiency of the spatiotemporal graph neural network model in capturing key biomechanical characteristic state information. Finally, a gated recurrent unit network is used to model the long-term time dependency in the multidimensional rehabilitation time series, and output the rehabilitation status prediction results at multiple future time points.

10. A scoliosis correction and recognition system based on sensor communication devices, characterized in that, The system adopts a scoliosis correction and identification method based on sensor communication device as described in any one of claims 1-9, including a general biomechanical pre-training model construction module, a personalized assessment model rapid adaptation module, a dynamic assessment baseline update module, and a rehabilitation trajectory prediction and decision support module. The general biomechanical pre-training model building module gathers a large-scale heterogeneous sensor dataset containing various types of scoliosis and rehabilitation movements, and uses biomechanical feature labels annotated by rehabilitation experts to train a hybrid model of deep convolutional neural network and long short-term memory network with the goal of minimizing the pre-training loss function. This results in a pre-training model that can map the original sensor time-series data into high-level general biomechanical feature vectors. The personalized assessment model rapid adaptation module collects a small amount of standard training action data that has been labeled by rehabilitation therapists for new users; on the basis of fixing the parameters of the underlying feature extractor of the pre-trained model, it unlocks only the parameters of the task-specific layer and introduces a feature space alignment loss function based on the distribution distance metric, which together with the task loss constitutes a joint optimization objective to fine-tune the model and generate a personalized assessment model for the user. The dynamic evaluation baseline update module continuously collects sensor data streams and makes predictions during the subsequent training of new users, and simultaneously calculates the prediction uncertainty estimate. When the prediction uncertainty estimate is continuously lower than the preset uncertainty threshold, the model update process is automatically triggered, and the corresponding low uncertainty data is used as new training samples to perform online incremental updates to the personalized evaluation model. The rehabilitation trajectory prediction and decision support module: Based on the time-series assessment results and key biomechanical characteristics output by the personalized assessment model and dynamically updated by the dynamic assessment baseline update module, a multidimensional rehabilitation time-series sequence for the user is constructed; the spatiotemporal graph neural network model is used to model the sequence, learn its spatiotemporal evolution pattern, and then predict the user's future rehabilitation progress curve, and generate decision support information based on the prediction results.

Images