A monitoring and early warning system for children's scoliosis

Abstract

The present invention relates to the fields of child health monitoring and artificial intelligence technology, and provides a child scoliosis monitoring and early warning system. Existing methods rely on static inspections, and it is difficult to capture abnormal dynamic changes of the spine in real time. The present invention acquires multimodal data through posture sensors, acceleration sensors, and image acquisition devices, combines convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) to perform deep feature extraction, and uses hidden Markov models (HMMs) to identify children's behavioral contexts and dynamically adjust the early warning threshold. The system uses a Q-learning algorithm to generate personalized health intervention strategies, and uses differential privacy and homomorphic encryption technologies to ensure data security. Finally, the system generates a three-level health warning signal to immediately remind children to adjust their postures and notify guardians. The present invention realizes real-time monitoring and personalized intervention of spinal health, provides an efficient and safe health management solution, and promotes the early detection and correction of scoliosis.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent health monitoring systems, and in particular to a monitoring and early warning system for children's scoliosis. Background Art

[0002] With the rapid development of medical technology and information technology, early monitoring and health management of scoliosis in children have gradually become a focus of social attention. Scoliosis is a bone disease characterized by abnormal curvature of the spine. It often occurs in children and adolescents. If not intervened in time, it may lead to serious physical deformities and psychological problems, and even cause irreversible damage to cardiopulmonary function. However, the current scoliosis detection method mainly relies on imaging examinations (such as X-rays) and manual physical examinations in hospital environments. This traditional method has limitations such as low frequency, delayed diagnosis, and insufficient monitoring coverage. In addition, the frequent use of radiological images may pose potential radiation risks to children's health, and the accuracy of manual examinations is highly dependent on the experience level of professional physicians. The physiological characteristics of children and the dynamic changes in their daily behaviors make the monitoring of scoliosis more complicated. Existing monitoring methods are difficult to record children's posture changes and their correlation with spinal health in real time, and cannot achieve early capture and early warning of abnormal postures. At the same time, traditional systems usually use single-modality data acquisition methods, such as static image data or limited posture analysis, and lack the ability to integrate and process multi-dimensional dynamic features, resulting in insufficient monitoring accuracy and coverage. In addition, children's behavioral characteristics have significant individual differences and complexity. The existing system still has great limitations in personalized analysis and situational adaptation capabilities, and is unable to dynamically optimize monitoring and early warning strategies based on the physical characteristics and behavioral situations of different children.

[0003] In recent years, the rapid development of smart wearable devices, artificial intelligence algorithms and multimodal data analysis technologies has provided new possibilities for real-time monitoring and intelligent analysis of scoliosis. These technologies can continuously monitor children's posture angles, dynamic behaviors and spinal image data through posture sensors, accelerometers and image acquisition devices, and use deep learning and multimodal fusion algorithms to achieve comprehensive analysis and prediction of data. However, existing technologies still have shortcomings in multimodal data fusion and personalized learning. On the one hand, the feature differences and dynamic change characteristics between multimodal data make it difficult for traditional fusion algorithms to extract comprehensive features stably and efficiently. On the other hand, the system lacks the ability to analyze situations and cannot adjust the warning threshold according to the activity status of children. In addition, with the popularization of smart monitoring equipment, data privacy and security protection issues have gradually become prominent. Children's health data involves sensitive information. The technical measures of existing systems in data encryption, privacy protection and distributed processing are not perfect, making it difficult to achieve remote monitoring and personalized analysis while ensuring data security.

[0004] In summary, there is an urgent need for a pediatric scoliosis monitoring and early warning system that integrates multimodal data collection, real-time monitoring, personalized analysis, contextual adaptation and privacy protection functions. It can capture posture and dynamic behavior data in real time, accurately analyze spinal health conditions and generate early warning information, thereby achieving early intervention and health management of scoliosis and ensuring children's physical development and health and safety. Summary of the invention

[0005] The purpose of the present invention is to provide a pediatric scoliosis monitoring and early warning system, which aims to utilize multimodal data acquisition and intelligent analysis technology to achieve real-time monitoring, situational analysis and abnormal early warning of children's spinal health status, generate personalized intervention strategies, and solve the problems existing in the prior art, such as lagging monitoring methods, insufficient personalized analysis capabilities, and weak data privacy protection.

[0006] To achieve the above-mentioned purpose, the present invention provides a child scoliosis monitoring and early warning system, comprising:

[0007] The data acquisition module is used to collect the child's multimodal data, including real-time recording of the spinal inclination angle, trunk flexion angle and lateral bending angle through the posture sensor, combining the acceleration sensor to obtain the activity intensity and dynamic characteristics of behavior, and using the image acquisition device to obtain multi-angle image data of the back for generating the spinal curve model. The module improves the data acquisition accuracy by optimizing the sampling frequency and resolution. The sampling frequency of the posture sensor is 50Hz, the acceleration sensor is 10Hz, and the image acquisition device collects 5 frames of image data per second. The acquisition module supports wireless transmission function, and uses low-power Bluetooth or Wi-Fi to transmit multimodal data to the data preprocessing module in real time to ensure the efficiency and real-time performance of the system.

[0008] The data preprocessing module is used to clean, smooth and standardize the multimodal data transmitted by the data acquisition module, including applying Kalman filter to dynamic behavior data for denoising, eliminating high-frequency noise through state transfer matrix and covariance matrix, and smoothing time series data to enhance signal stability; Canny edge detection is used to extract key feature points of the spinal curve from image data, and Hough transform fitting is used to generate curve feature vectors for analysis; in addition, the module normalizes all modal data to the [0,1] interval through standardization operations to eliminate differences in feature units and amplitudes of different modal data, providing consistent input for multimodal data fusion, and applying anomaly detection algorithms to eliminate invalid data caused by improper device wearing or interference, thereby improving the robustness of data processing.

[0009] The multimodal data fusion module is used to integrate posture, dynamic behavior and image features, extract local patterns in image features through convolutional neural networks (CNN), and capture dynamic trends in time series data in combination with long short-term memory networks (LSTM). The module further uses a multi-head attention mechanism to dynamically adjust the weights of modal data, and optimizes the weight allocation strategy through a Bayesian optimization algorithm to adapt to different situations and individual differences, and finally generates a comprehensive health feature vector as input to the situational analysis module and the adaptive learning module to ensure the accuracy and practicality of subsequent analysis.

[0010] The situation analysis module is used to identify children's behavior situations based on the hidden Markov model (HMM), including states such as stillness, standing, walking and sitting. The module optimizes the hidden state parameters through the Baum-Welch algorithm, calculates the probability of the observation sequence using the forward algorithm and the backward algorithm in combination with real-time data, and predicts the hidden state of the next time step. According to the identified situation, the warning threshold of the spinal curvature is dynamically adjusted, so that the system reduces the sensitivity in the motion situation to reduce false alarms, and increases the sensitivity in the static situation to ensure early detection of spinal abnormality risks. The module interacts closely with the data fusion module to improve the system's situational adaptation capabilities.

[0011] The adaptive learning module dynamically optimizes personalized early warning strategies based on the Q-learning algorithm, and continuously updates the early warning thresholds and algorithm parameters through real-time feedback. The module combines long-term data training to generate an individualized monitoring model that adapts to the physiological characteristics and behavioral patterns of different children, and allows users to annotate a small amount of sample data through active learning technology, thereby optimizing system performance when the amount of data is insufficient. The adaptive learning module also interacts with the situational analysis module to dynamically adjust the learning parameters based on the real-time situational recognition results to ensure that the early warning strategy always meets the child’s current health status.

[0012] The privacy protection module is used to protect the privacy of individuals involved in the data processing process. It uses a differential privacy mechanism to achieve data privacy protection by injecting Laplace noise, while ensuring the effectiveness of data analysis. The module combines the Trusted Execution Environment (TEE) technology to protect the integrity and security of data processing, and uses homomorphic encryption technology to encrypt gradient information during data transmission to prevent data leakage or tampering in distributed processing, ensuring the privacy security of the entire data processing chain.

[0013] The early warning module generates real-time health early warning signals based on multimodal fusion data and situational analysis results. The early warning signals are divided into three levels: normal, caution and high-risk. The module sends real-time notifications to guardians through mobile terminals. When high-risk postures or persistent abnormalities are detected, the system triggers the instant vibration feedback function to prompt children to adjust their postures, forming a closed-loop monitoring and intervention. At the same time, the module feeds back user response results to the adaptive learning module for further optimization of the early warning strategy.

[0014] The data acquisition module further includes a variational autoencoder (VAE) for generating a three-dimensional curve model of the child's spine from multi-angle image data acquired by the image acquisition device. The variational autoencoder establishes a mapping relationship between multi-angle images and three-dimensional models, and uses its reconstruction error minimization function to compensate for the integrity of the curve when there is partial missing or occlusion in the acquired data. In order to improve the accuracy of the reconstructed model, the system combines multiple three-dimensional reconstruction loss functions during the training process, including mean square error (MSE) loss and adversarial generation loss, and optimizes the model parameters through back propagation. By introducing VAE, the system can generate three-dimensional spinal curves with high precision, which is suitable for spinal health assessment in complex situations, and significantly improves the accuracy and spatial adaptability of monitoring.

[0015] The data preprocessing module further includes an anomaly detection submodule for identifying and removing abnormal data. The anomaly detection submodule is based on the Isolation Forest algorithm, which isolates abnormal data points by constructing a random split tree in a high-dimensional space, thereby detecting invalid data caused by improper device wearing, interference signals or human operation errors. The Isolation Forest algorithm uses a tree structure to quickly divide data points, and the length of its split path is positively correlated with the degree of abnormality. Abnormal data points have shorter split paths in the model, so they are identified and removed. The anomaly detection submodule also combines the time window mechanism to analyze the continuity of local anomalies in dynamic data to avoid erroneous data removal due to short-term errors, thereby improving the overall robustness and reliability of the data preprocessing module.

[0016] The multimodal data fusion module further combines a feature enhancement method based on a multi-head attention mechanism to dynamically allocate attention weights between different modal features. The attention mechanism parallelizes the global weight allocation tasks of different modal features by constructing multiple independent attention heads to capture more fine-grained modal association features. Specifically, the mechanism prioritizes the weight of dynamic behavior data in time series analysis, and increases the contribution of image data in static posture assessment, thereby optimizing the fused comprehensive health feature vector. In order to improve the computational efficiency of the attention mechanism, the module uses matrix decomposition technology to accelerate the weight calculation process, and suppresses overfitting through regularization strategies, so that the model maintains consistent performance in multiple scenarios.

[0017] The situational analysis module further includes a situational prediction submodule, which is used to predict future situations and their potential impact on spinal health based on real-time data. The prediction submodule combines the hidden Markov model (HMM) and the long short-term memory network (LSTM), uses a forward algorithm to estimate the probability distribution of the current situation, and captures the long-term behavior pattern in the time series through LSTM to generate prediction results for future situations. The situational prediction results will be fed back to the situational analysis module to dynamically adjust the warning threshold and model parameters of the spinal curvature to ensure that the system maintains the accuracy of monitoring and warning when the behavioral situation changes rapidly. The submodule further uses a sliding window mechanism to segment the input data to ensure the real-time nature of the time series analysis while improving the robustness of the situational prediction.

[0018] The adaptive learning module is further combined with a generative adversarial network (GAN) to generate virtual data to expand the training sample space when data is insufficient. The generative adversarial network includes a generator and a discriminator, wherein the generator is responsible for generating data samples that simulate individual spinal curves and behavioral characteristics, and the discriminator is used to evaluate the difference between the generated samples and the real data, and continuously optimize the generation quality through the adversarial training mechanism. In order to improve the generation accuracy of GAN, the module uses progressive growing technology to increase the resolution of data generation, and combines the Wasserstein loss function to stabilize the gradient change of the generator. Through the virtual data generated by GAN, the system can still maintain high monitoring accuracy and model performance under small sample conditions.

[0019] The privacy protection module further combines the joint protection mechanism of differential privacy and homomorphic encryption to enhance data security. The differential privacy mechanism protects individual data privacy by injecting Gaussian noise, and the value of the privacy budget ε is dynamically adjusted to adapt to the accuracy and privacy requirements of different data processing tasks. Homomorphic encryption technology uses encrypted gradients to achieve non-plaintext processing of data in an encrypted state, preventing data from being tampered with or stolen during transmission and distributed learning. In addition, the module uses a privacy budget tracker to monitor the privacy consumption of each data operation to ensure that the system maintains a balance between privacy protection and performance optimization in long-term operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0021] Figure 1 This is a diagram of the architecture of a pediatric scoliosis monitoring and early warning system.

[0022] Figure 2 This is a class diagram for a children's scoliosis monitoring and early warning system

[0023] Figure 3 A data flow diagram of a children's scoliosis monitoring and early warning system

[0024] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0025] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0026] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0027] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0028] Figure 1The system architecture diagram provided for the embodiment of the present invention, the specific implementation process is as follows:

[0029] The present invention provides a pediatric scoliosis monitoring and early warning system, which adopts a modular design, from data acquisition, preprocessing to fusion analysis and early warning generation, covering the complete process from raw data processing to final health intervention. The system architecture consists of a data acquisition module, a data preprocessing module, a multimodal data fusion module, a situational analysis module, an adaptive learning module, a privacy protection module and an early warning generation module, and the modules work together through data flow. Through advanced multimodal data analysis technology and reinforcement learning algorithms, the system realizes real-time monitoring of children's spinal health and the generation of personalized intervention plans, while ensuring the privacy and security of data through differential privacy and homomorphic encryption technology.

[0030] The data acquisition module is the front-end part of the system, which mainly obtains children's spinal health-related data in real time through wearable devices. Specifically, the system includes a posture sensor, an accelerometer and an image acquisition device, which respectively collect children's posture angles, dynamic behavior characteristics and back image data. The posture sensor can record the inclination angle, lateral bending angle and trunk flexion angle of the spine in real time, and the accelerometer is used to capture children's dynamic behavior characteristics, such as step frequency, exercise intensity and behavior duration. The image acquisition device uses a high-definition camera to obtain multi-angle spinal images of the child's back, and combines the embedded image segmentation algorithm to complete the preliminary separation of the spinal area. The module ensures the accuracy of data acquisition by optimizing the sampling frequency and resolution. The sampling frequency of the posture sensor is 50Hz, the accelerometer is 10Hz, and the image capture frequency of the image acquisition device is 5 frames per second. The data is transmitted to the data preprocessing module in real time via low-power Bluetooth or Wi-Fi.

[0031] The data preprocessing module is an important middle layer in the system. Its main task is to clean, reduce noise and extract features from the raw data transmitted by the data acquisition module, so as to provide high-quality input data for subsequent analysis. In terms of processing dynamic behavior data, the system uses a Kalman filter to reduce noise on the raw signal. The Kalman filter predicts the optimal state of the signal based on the state transition matrix, and dynamically adjusts the intensity of the noise in combination with the covariance matrix, thereby eliminating high-frequency interference in the signal and smoothing the changing curve of dynamic data. In terms of image data processing, the system uses the Canny edge detection algorithm to extract the key feature points of the spinal curve, and uses the Hough transform to fit it into a continuous curve, which is input into the subsequent module in the form of a feature vector. In addition, in order to ensure data quality, the preprocessing module detects and removes abnormal data through the isolation forest algorithm. The isolation forest algorithm isolates abnormal data points by constructing a random split tree in a high-dimensional space, and the length of its split path is positively correlated with the degree of abnormality of the data. When the system detects abnormal data, it will be removed or marked to avoid misleading subsequent analysis.

[0032] The multimodal data fusion module is the core module of the system analysis. Its function is to integrate data features from different modalities and generate a comprehensive feature vector that can reflect the health status of the child's spine. In order to achieve efficient fusion of multimodal data, this module adopts a variety of deep learning algorithms. Among them, the convolutional neural network (CNN) is used to extract local features in image data, capture subtle changes in spinal curves through convolution kernels, and use pooling operations to reduce noise interference. The long short-term memory network (LSTM) is used to process time series data, capture long-term dependencies in dynamic behavior features, and generate feature sequences in the time dimension. In addition, the module also introduces a multi-head attention mechanism to dynamically assign weights according to the importance of data modalities, such as enhancing the contribution of image data in static situations and increasing the weight of dynamic behavior data in dynamic situations. The final feature vector of multimodal fusion is further adjusted through the Bayesian optimization algorithm to ensure that the fusion strategy can adapt to individual differences and real-time scene changes.

[0033] The situation analysis module is responsible for real-time identification of behavioral situations based on the fused comprehensive feature vectors, and dynamically adjusts the system's warning parameters based on the identification results. The system uses a hidden Markov model (HMM) to classify children's behavioral situations (such as stillness, walking, running, etc.). By optimizing the model's hidden state transition probability matrix and observation probability matrix using the Baum-Welch algorithm, the system can accurately predict the child's current behavioral situation based on real-time data. At the same time, the forward algorithm is used to calculate the probability distribution of the observation sequence to predict the situation state of the next time step. The situation analysis results directly affect the threshold setting of the warning generation module, such as reducing the warning sensitivity in a motion situation to reduce false alarms, and increasing the sensitivity in a stationary situation to ensure that abnormal postures are captured in time.

[0034] The adaptive learning module is the intelligent part of the system. Its core is to achieve dynamic optimization of personalized early warning strategies through reinforcement learning algorithms (Q-learning). The Q-learning algorithm collects real-time feedback from the system and uses a reward and punishment mechanism to continuously update the warning threshold and strategy parameters, thereby generating personalized health intervention plans that adapt to individual differences. The system also combines active learning technology to allow users to annotate some data to improve the adaptability of the model under conditions of insufficient data. In addition, the adaptive learning module also uses adversarial generative networks (GANs) to generate virtual data samples to expand the training data space, thereby improving the system's learning ability in a small sample environment.

[0035] The privacy protection module protects privacy during data processing through differential privacy and homomorphic encryption technologies. The differential privacy mechanism blurs individual data by injecting Laplace noise, and the dynamic adjustment of the privacy budget ensures privacy protection while maintaining the effectiveness of data analysis. Homomorphic encryption technology allows data to be directly calculated in an encrypted state to avoid data leakage during transmission and processing. In addition, the privacy protection module is also equipped with a privacy budget tracker to monitor the privacy consumption of each data operation in real time to ensure a balance between privacy protection and performance optimization in long-term operation.

[0036] The warning generation module generates three-level health warning signals based on the comprehensive feature vector and situational analysis results, including "normal", "caution" and "high risk". The warning strategy is implemented using the Q-learning algorithm, which processes health status and intervention behavior and optimizes the intervention strategy. t and possible intervention behaviorsa t The system determines the optimal intervention strategy by updating the Q value. The update rule is as follows:

[0037]

[0038] The system is based on real-time health feedbackt Dynamically adjust the Q value and gradually approach the optimal strategy. Through continuous learning, the system can adjust the next intervention strategy after each intervention, so that health fluctuations can be controlled to the greatest extent. During the intervention process, the health data flow flows from the multimodal analysis module to the reinforcement learning module, and the system evaluates the patient's status in real time. After each intervention, the system adjusts the Q value in the Q-learning model based on feedback, and the updated Q value determines the optimal strategy for the next round of intervention.

[0039] When the system detects that the child's spine is abnormal or maintains a bad posture for a long time, it generates a "high-risk" signal and sends a real-time notification to the guardian through the mobile terminal. At the same time, the module has an instant feedback function, which reminds the child to adjust his posture through the vibration of the device, forming a closed-loop monitoring and intervention system.

[0040] The present invention forms a closed loop from data collection to warning generation through the collaborative work of multiple modules. The system can monitor the children's spinal health status in real time and accurately, generate personalized intervention strategies, and at the same time ensure data privacy and security, thereby providing a comprehensive and intelligent solution for children's spinal health management.

[0041] To protect data privacy, the system uses homomorphic encryption technology to encrypt and transmit gradient data. The core principle of the Paillier encryption algorithm is to perform linear operations on ciphertext and implement gradient aggregation without plaintext access on the server side. The gradient data generated by each device is encrypted with a random key, which is dynamically generated and protected by the security module inside the device. During the aggregation process, the server uses a verifiable encryption protocol to verify the integrity of the gradient data to prevent tampering or forgery. To improve the efficiency of encryption calculations, the system relies on the AES-NI instruction set of modern hardware to achieve fast processing of gradient encryption, thereby significantly reducing the computational overhead of encryption operations.

[0042] The differential privacy protection mechanism protects individual data by introducing Laplace noise during the gradient update process. The system defines a privacy budget ∈ to control the intensity of injected noise to ensure privacy protection during the gradient update process. The system equips each device with a privacy budget tracker to monitor the consumption of the privacy budget in real time. The noise injection of differential privacy is supported by a hardware-level random number generator (such as the Intel Secure Key module), which ensures the uniform distribution of noise and the reliability of privacy protection by quickly generating high-quality random numbers.

[0043] After each round of model update, the server calculates the global model weight based on the asynchronous aggregation strategy, and the optimized global model parameters are encrypted and transmitted back to the local device to achieve continuous optimization of the distributed model. At the same time, the system's privacy budget scheduling mechanism dynamically adjusts the privacy budget allocation based on the frequency and scale of each data operation to ensure a balance between model accuracy and data protection.

[0044] The recognition model is optimized through a genetic algorithm. The parameters of the parent model are reorganized using a crossover operation with a crossover probability of 0.7. The model parameters are then perturbed through a mutation operation with a mutation rate of 0.01. The fitness function is called to evaluate the performance of the model in the recognition task, and models with higher fitness are retained for subsequent optimization to improve the accuracy and robustness of the recognition model. The system generates risk warnings and adjusts the next round of intervention strategies based on real-time data and personalized intervention strategies, forming a closed loop in the decision-making process.

[0045] Figure 2 A class diagram provided for an embodiment of the present invention;

[0046] like Figure 2 As shown, the present invention provides a class diagram, including:

[0047] The system includes the following key classes:

[0048] Collection methods:

[0049] Collect data(): Returns the collected raw data, which may be of type String (serialized data) or custom DataObject

[0050] TransferData(): Returns a Boolean value indicating whether the transfer was successful

[0051] Preprocessing class methods:

[0052] Denoise(): Returns the denoised data (ProcessedData)

[0053] EdgeDetect(): Returns the detected edge features (FeatureData)

[0054] Feature Extraction(): Returns the extracted feature vector (FeatureVector)

[0055] AnomalyDetect(): Returns a Boolean value indicating whether an anomaly is detected.

[0056] Fusion class methods:

[0057] Feature fusion(): Returns the fused comprehensive feature vector (HealthFeatureVector)

[0058] WeightAdjust(): Returns the updated modality weights (WeightMatrix).

[0059] Situational analysis methods:

[0060] SituationIdentify(): Returns the identified situation status (String)

[0061] Warning threshold adjustment(): Returns the adjusted warning threshold (ThresholdValue)

[0062] Learning methods:

[0063] StrategyOptimize(): Returns the optimized personalized strategy (PersonalizedStrategy)

[0064] ParameterUpdate(): Returns the update status (Boolean), indicating whether the parameters are updated successfully.

[0065] Privacy protection methods:

[0066] Differential Privacy Protection(): Returns the processed data (ProtectedData)

[0067] EncryptTransfer(): Returns a Boolean value indicating whether the encryption is successful.

[0068] Warning methods:

[0069] GenerateAlert(): Returns the generated alert level (String)

[0070] SendNotification(): Returns a Boolean value indicating whether the notification was sent successfully.

[0071] Data Collection Class (DataCollection): The Data Collection Class is the entry module of the system. Its main function is to collect multimodal data from children's daily activities in real time, including posture data, dynamic behavior data and back image data. The main attributes of this class include sensor data and image data. The sensor data covers the spinal inclination angle, scoliosis angle and dynamic behavior characteristics, while the image data is the multi-angle image of the child's back obtained by high-definition camera equipment.

[0072] In terms of method design, the core method is to acquire raw data in real time by calling the sampling interface of the hardware device; the method of transmitting data () uses the Bluetooth or Wi-Fi module to upload the collected data to the data preprocessing module. In addition, this class monitors the device status through anomaly detection mechanisms, such as data transmission interruption or sensor error. During the implementation process, the acquisition frequency and resolution can be adjusted dynamically. For example, the sampling frequency of the posture data is 50Hz, and the image acquisition device captures 5 frames of image data per second to meet the requirements of high real-time performance and accuracy.

[0073] Data Preprocessing: Data Preprocessing is responsible for cleaning, denoising and feature extraction of the collected raw data to ensure data quality. The main attributes of this class include raw data and processed data, which are used to store input data and preprocessing results. In terms of method design, the denoising () method uses the Kalman filter to smooth dynamic data, predicts the next data point through the state transition matrix, and adjusts the noise suppression strength in combination with the covariance matrix to eliminate high-frequency interference signals.

[0074] In terms of image data processing, the edge detection method uses the Canny edge detection algorithm to extract the spine contour and adjusts the Gaussian filter parameters to enhance the recognition of weak edges. Subsequently, the feature extraction method uses the Hough transform to fit the spine curve and converts it into a mathematical feature vector for use by subsequent modules. In addition, in order to further improve the robustness of the data, the anomaly detection method is based on the isolation forest algorithm, which identifies and eliminates abnormal data by constructing a high-dimensional random splitting tree, such as invalid data caused by improper wearing of the device.

[0075] Multimodal Fusion: The Multimodal Fusion class is the core module of the system analysis. Its function is to integrate feature data from different modalities into a comprehensive health feature vector, thereby comprehensively reflecting the health status of the child's spine. The main attributes of this class include modal weights and health feature vectors. Through the feature fusion() method, this class calls the convolutional neural network (CNN) to extract local features from the image data, and uses the long short-term memory network (LSTM) to capture the dynamic changes of time series data.

[0076] In addition, the weight adjustment() method introduces a multi-head attention mechanism to dynamically assign weights to modal data according to the context. For example, in static contexts, the system will enhance the weight of image data, while in dynamic behavior contexts, it will highlight the contribution of time series data. To ensure the efficiency and stability of the fusion strategy, the Bayesian optimization algorithm is used to dynamically adjust the model parameters during the weight adjustment process.

[0077] Context Analysis: The context analysis class supports the system's warning parameter setting through behavioral context recognition. Its main attributes include behavioral context (such as stillness, walking, running, etc.). This class uses the context recognition() method to call the hidden Markov model (HMM) and uses the Baum-Welch algorithm to optimize the hidden state transition probability and observation probability, thereby classifying the child's current behavioral state.

[0078] On this basis, the warning threshold adjustment() method dynamically adjusts the warning sensitivity of the system according to the identified situation. For example, in a sports situation, the system will appropriately reduce the warning threshold sensitivity to avoid false alarms; while in a static situation, the system will increase the sensitivity to ensure timely detection of abnormal postures.

[0079] Adaptive Learning: Adaptive Learning is the intelligent module of the system. Its main task is to continuously optimize the personalized early warning strategy in combination with real-time feedback. The core attributes of this class include personalized strategies and real-time feedback data, and dynamic optimization is achieved through reinforcement learning algorithms (Q-learning).

[0080] In terms of method design, the strategy optimization () method updates personalized strategy parameters by analyzing historical data and real-time feedback; the parameter update () method combines the reward mechanism and the penalty mechanism to dynamically adjust the warning threshold. In order to maintain the robustness of the model under data scarcity conditions, this class introduces the adversarial generative network (GAN) to generate virtual sample data, thereby expanding the training data space. In addition, this class also interacts with the situation analysis class to adjust the learning strategy according to the real-time situation, making the system more adaptable in dynamic scenarios.

[0081] Privacy Protection Class: Privacy Protection Class plays an important role in data security. Its main properties include privacy budget and encryption gradient. The differential privacy protection () method blurs the data by injecting Laplace noise to protect individual privacy; the encrypted transmission () method uses homomorphic encryption technology to ensure that the data is encrypted during transmission and processing to prevent sensitive information leakage. In addition, this class dynamically monitors the privacy consumption of each data operation through the privacy budget tracker to ensure a balance between privacy protection and performance optimization in long-term operation.

[0082] Warning generation class (WarningSystem): The warning generation class generates three levels of health warning signals (normal, caution, high risk) based on the analysis results. Its main attributes include warning level and notification information. The generate warning() method combines comprehensive health characteristics and situational analysis results to determine the child's spinal health risk in real time and send a notification to the guardian through the Bluetooth or Wi-Fi module. In addition, the send notification() method also has an instant feedback function, such as reminding children to adjust bad postures through device vibration, thus forming a closed-loop monitoring and intervention mechanism.

[0083] Interaction between classes is the key to the implementation of the system. Data flow is achieved through close collaboration through class method calls. The data collection class passes the collected raw data to the data preprocessing class, which stores the processed data in the database or directly passes it to the multimodal data fusion class. After the fusion class integrates the features, it inputs the results into the situation analysis class for situation classification. The results of the situation analysis interact with the adaptive learning class to optimize the strategy and generate personalized intervention plans. In addition, the privacy protection class provides security for data processing and storage, and the warning generation class generates health warning signals based on the final analysis results.

[0084] Figure 3 The data flow process of the child scoliosis monitoring and early warning system in actual use scenarios is demonstrated. Taking the post-exercise scenario as an example, the system obtains the child's spinal status and dynamic behavior data through sensors and image acquisition devices, and generates health warning signals after multi-level mathematical model analysis. The following describes the specific implementation process and key technical details of the data flow in the system.

[0085] In this scenario, the specific implementation process of data flow and logical relationship is as follows:

[0086] First, in the data collection stage, the data collection module acquires the child's multimodal data in real time, including the tilt angle, lateral bending angle, dynamic behavior characteristics of the posture sensor, and back image data generated by the image acquisition device. For each type of data, the system represents it in vectorized form. In order to enhance the sampling accuracy, the system dynamically adjusts the sampling frequency, for example, the posture sensor sampling frequency is set to 50Hz, and the acceleration sensor sampling frequency is 10Hz. The collected data is transmitted to the data preprocessing module in a streaming manner through the low-power Bluetooth module.

[0087] In the data preprocessing stage, the system first performs noise reduction on the dynamic behavior data. Kalman Filter is a recursive algorithm, and its basic process can be divided into two stages: prediction and update. The system uses the state space model of Kalman Filter:

[0088] x t+1 =Ax t +w t, z t =Hx t +v t ,

[0089] where x t is the state variable, z t is the observed variable, A is the state transfer matrix, H is the observation matrix, w t and v t They are process noise and measurement noise respectively. Through the calculation of Kalman gain, the system performs weighted fusion of the measured value and the predicted value to obtain the optimal estimate and effectively eliminate high-frequency noise.

[0090] In the multimodal data fusion stage, the system integrates dynamic behavior characteristics and spinal curve characteristics to generate a comprehensive health feature vector that reflects the overall spinal health status of the child.

[0091] After the feature map is processed by the activation function (such as ReLU) and the pooling layer, an image feature vector is generated. For time series data, the system uses a long short-term memory network (LSTM) to capture long-term dependencies in the time dimension through a recursive structure. The state update formula of the LSTM unit is:

[0092] i t =σ(W i [x t ,h t-1 ]+b i ),

[0093] f t =σ(W f [x t ,h t-1 ]+b f ),

[0094] o t =σ(W o [x t ,h t-1 ]+b o ),

[0095] c t =f t ☉c t-1 +i t ☉tanh(W c [x t ,h t-1 ]+b c ),

[0096] h t =o t ☉tanh(ct),

[0097] where i t ,ft ,o t They are input gate, forget gate and output gate respectively, c t is the memory state. The fused features are dynamically weighted through a multi-head attention mechanism to generate a comprehensive health feature vector.

[0098] The fusion module extracts local features from image data, such as subtle curvature changes in the spine, through a convolutional neural network (CNN). At the same time, the long short-term memory network (LSTM) processes time series data to capture the dynamic changes in behavioral features over time. In addition, the system introduces a multi-head attention mechanism to dynamically adjust the weight distribution according to the importance of different data modalities. For example, in the case of intense exercise, the system enhances the weight of dynamic behavioral data, while in the case of static behavior, it pays more attention to the contribution of image data. The fused comprehensive feature vector is input into the situational analysis module.

[0099] The situational analysis module identifies the child's current behavioral situation (such as still, standing, walking or running) through a hidden Markov model (HMM). Taking this implementation scenario as an example, the system identifies the child's violent behavioral situation after exercise. The module calculates the probability distribution of the observation sequence through forward and backward algorithms, and predicts the situational state of the next time step. In the case of intense exercise, the system dynamically adjusts the warning threshold of the spinal curvature. For example, the system reduces the warning sensitivity in this situation to avoid false alarms caused by temporary abnormal posture caused by exercise. At the same time, the output results of the situational analysis module are provided as input to the adaptive learning module to further optimize the personalized health strategy.

[0100] The adaptive learning module realizes dynamic optimization of personalized strategies based on the Q-learning algorithm. Through the analysis of real-time feedback, the system continuously updates personalized early warning strategies. For example, for children with slow recovery from abnormal posture, the system will dynamically adjust the early warning threshold and generate targeted health intervention plans. At the same time, the module combines active learning technology to allow users (such as guardians or medical staff) to annotate a small amount of sample data, thereby maintaining a high model adaptability when data is insufficient. In this implementation scenario, the optimization strategy of the adaptive learning module is stored in the model and strategy storage unit and is called in subsequent analysis and early warning.

[0101] After situational analysis and personalized strategy optimization, the system's warning generation module generates three-level health warning signals based on comprehensive health characteristics and current situations. For this implementation scenario, the system identifies the risk of abnormal posture of children after exercise, generates a warning signal at the "attention" level, and sends real-time notifications to guardians via Bluetooth connection to mobile terminal devices. At the same time, the system's instant feedback mechanism is activated, and the vibration device is used to remind children to correct bad postures, thus forming a closed-loop health intervention system.

[0102] During the data flow, the system always ensures the security of the data through the privacy protection module. Differential privacy technology protects children's sensitive data by injecting Laplace noise to ensure that individual privacy is not leaked. Homomorphic encryption technology further ensures the security of data during transmission and processing, allowing data to be directly calculated in an encrypted state. In this scenario, all processed data is encrypted and stored, and model training and real-time analysis are completed under the coverage of the privacy protection mechanism.

[0103] In summary, the pediatric scoliosis monitoring and early warning system provided by the present invention realizes the automation and intelligence of the entire process from data collection to early warning generation through the collaborative work of various modules. The system can analyze the spinal health status of children in real time in complex situations such as after exercise, generate personalized intervention strategies, and help children improve bad postures through an instant feedback mechanism, ultimately achieving the goal of early monitoring and intervention of scoliosis.

Claims

1. A child scoliosis monitoring and early warning system, characterized in that: The system comprises: A data acquisition module, which collects multimodal data of children, including using a posture sensor to record the spinal inclination angle, trunk flexion angle and lateral bending angle in real time, combining an acceleration sensor to obtain activity intensity and behavioral dynamic characteristics, and using an image acquisition device to obtain multi-angle image data of the back to generate a spinal curve model; the module uses optimized sampling frequency and resolution to improve data acquisition accuracy, and the acquisition module supports wireless transmission function, using low-power Bluetooth or Wi-Fi to transmit multimodal data to the data preprocessing module in real time; A data preprocessing module, which cleans, smoothes and standardizes the multimodal data transmitted by the data acquisition module, including applying a Kalman filter to the dynamic behavior data for noise reduction, using the state transfer matrix and the covariance matrix to eliminate high-frequency noise, and smoothing the time series data to enhance signal stability; using Canny edge detection to extract key feature points of the spinal curve from the image data, and combining Hough transform fitting to generate a curve feature vector for analysis; the module uses a standardization operation to normalize all modal data to the [0,1] interval, eliminating differences in feature units and amplitudes of different modal data, and applying an anomaly detection algorithm to eliminate invalid data caused by improper device wearing or interference; A multimodal data fusion module, which integrates posture, dynamic behavior and image features, uses a convolutional neural network (CNN) to extract local patterns in image features, and combines a long short-term memory network (LSTM) to capture the dynamic change trend of time series data; the module further uses a multi-head attention mechanism to dynamically adjust the weight of modal data, uses a Bayesian optimization algorithm to optimize the weight allocation strategy to adapt to different situations and individual differences, and generates a comprehensive health feature vector as the input of the situational analysis module and the adaptive learning module; A situation analysis module, which uses a hidden Markov model (HMM) to identify children's behavior situations, including static, standing, walking and sitting states; the module uses the Baum-Welch algorithm to optimize hidden state parameters, and uses forward and backward algorithms in combination with real-time data to calculate the probability of the observation sequence and predict the hidden state of the next time step; the warning threshold of the spinal curvature is dynamically adjusted according to the identified situation; the module interacts with the data fusion module; The adaptive learning module dynamically optimizes the personalized early warning strategy based on the Q-learning algorithm and uses real-time feedback to continuously update the early warning threshold and algorithm parameters. The module combines long-term data training to generate an individualized monitoring model that adapts to the physiological characteristics and behavior patterns of different children, and uses active learning technology to allow users to annotate a small amount of sample data. The adaptive learning module also interacts with the situational analysis module to dynamically adjust the learning parameters. The privacy protection module protects the privacy of individuals involved in the data processing process, adopts a differential privacy mechanism to inject Laplace noise to achieve data privacy protection, and ensures the effectiveness of data analysis; the module combines the trusted execution environment (TEE) technology to protect the integrity and security of data processing, and uses homomorphic encryption technology to encrypt gradient information during data transmission. The early warning module generates real-time health early warning signals based on multimodal fusion data and situational analysis results. The early warning signals are divided into three levels: normal, caution and high-risk. The module uses a mobile terminal to send real-time prompts to guardians, forming a closed-loop monitoring and intervention. At the same time, the module feeds back user response results to the adaptive learning module to further optimize the early warning strategy.

2. The child scoliosis monitoring and early warning system according to claim 1, characterized in that: The data acquisition module further includes a variational autoencoder (VAE), which uses multi-angle image data obtained from an image acquisition device to generate a three-dimensional curve model of the child's spine; the variational autoencoder uses a mapping relationship between multi-angle images and three-dimensional models, and when there is partial missing or occlusion in the acquired data, its reconstruction error minimization function is used to compensate for the integrity of the curve; to improve the accuracy of the reconstructed model, the system combines multiple three-dimensional reconstruction loss functions during the training process, including mean square error (MSE) loss and adversarial generation loss, and uses back propagation to optimize model parameters; VAE is introduced.

3. The child scoliosis monitoring and early warning system according to claim 1, characterized in that: The data preprocessing module further includes an anomaly detection submodule, which is used to identify and remove abnormal data; The anomaly detection submodule is based on the Isolation Forest algorithm, which uses a random splitting tree constructed in a high-dimensional space to isolate abnormal data points, thereby detecting invalid data caused by improper device wearing, interference signals or human operation errors; the Isolation Forest algorithm uses a tree structure to quickly divide data points, and the length of its splitting path is positively correlated with the degree of abnormality; abnormal data points have shorter splitting paths in the model; the anomaly detection submodule also combines the time window mechanism to analyze the continuity of local anomalies in dynamic data.

4. The child scoliosis monitoring and early warning system according to claim 1, characterized in that: The situational analysis module further includes a situational prediction submodule, which uses real-time data to predict future situations and their potential impact on spinal health; the prediction submodule combines a hidden Markov model (HMM) and a long short-term memory network (LSTM), uses a forward algorithm to estimate the probability distribution of the current situation, and uses LSTM to capture long-term behavioral patterns in time series to generate prediction results for future situations; the situational prediction results will be fed back to the situational analysis module to dynamically adjust the warning threshold and model parameters of the spinal curvature to ensure that the system maintains monitoring and warning accuracy when the behavioral situation changes rapidly; the submodule further uses a sliding window mechanism to segment the input data.

5. The child scoliosis monitoring and early warning system according to claim 1, characterized in that: The adaptive learning module is further combined with a generative adversarial network (GAN) to generate virtual data when data is insufficient to expand the training sample space; the generative adversarial network includes a generator and a discriminator, wherein the generator is responsible for generating data samples that simulate individual spinal curves and behavioral characteristics, and the discriminator is used to evaluate the difference between the generated samples and the real data, and the generation quality is continuously optimized using an adversarial training mechanism; to improve the generation accuracy of GAN, the module uses progressive growing technology to increase the resolution of data generation, and at the same time combines the Wasserstein loss function to stabilize the gradient change of the generator; using virtual data generated by GAN, the system still maintains monitoring accuracy and model performance under small sample conditions.

6. The child scoliosis monitoring and early warning system according to claim 1, characterized in that: The privacy protection module further combines the joint protection mechanism of differential privacy and homomorphic encryption to enhance data security; the differential privacy mechanism injects Gaussian noise to protect individual data privacy, and the value of the privacy budget ε is dynamically adjusted to adapt to the accuracy and privacy requirements of different data processing tasks; homomorphic encryption technology uses encrypted gradients to achieve plaintext-free processing of data in an encrypted state to prevent data from being tampered with or stolen during transmission and distributed learning; the module uses a privacy budget tracker to monitor the privacy consumption of each data operation.

Images