Children rehabilitation training-based compliance monitoring method and system

Abstract

The invention provides a compliance monitoring method and system based on child rehabilitation training, relates to the field of child rehabilitation training and medical data monitoring, and solves the technical problems that in the prior art, rehabilitation training compliance evaluation is distorted, the weight adaptability is poor, and the trend cannot be predicted. The method comprises the following steps: collecting original training data of child training, carrying out environment interference discrimination and training intention distinguishing processing, and carrying out deviation quantification correction aiming at subjective quality data to obtain training data; performing standardization and fusion processing on the training data to obtain multi-source training data; based on the multi-source training data, dynamically calculating a training compliance sub-index through a nonlinear fusion function, and dynamically adjusting a sub-index weight in combination with a historical training compliance sub-index and a training regression model to obtain a comprehensive training compliance index; and obtaining a compliance trend prediction result and training effect evaluation through a time sequence prediction model. The method and the device are used in the compliance monitoring process of rehabilitation training of children.

Description

Technical Field

[0001] This application relates to the field of children's rehabilitation training and medical data monitoring, and in particular to a method and system for monitoring children's rehabilitation training compliance. Background Technology

[0002] Current technologies typically assess children's adherence to rehabilitation training through subjective ratings by therapists and identification of single movement features. The process mainly involves collecting video recordings of movements or subjective scoring data during training and then obtaining a adherence index through static weighted summation. However, this approach is only suitable for institutionalized, highly controlled professional rehabilitation settings and cannot be applied to daily rehabilitation training in complex and disruptive environments such as the home. In home-based rehabilitation training, individual differences among children, the randomness of environmental disturbances, and variations in the subjective bias of assessors can lead to distorted adherence assessments, poor weighting adaptation, and an inability to predict trends. Therefore, accurately quantifying children's adherence to rehabilitation training, dynamically adapting to individual differences, and achieving trend prediction and closed-loop optimization of the program in complex and disruptive environments such as the home have become pressing technical challenges. Summary of the Invention

[0003] This application provides a method and system for monitoring children's rehabilitation training compliance, which solves the technical problems of distortion in rehabilitation training compliance assessment, poor weighting adaptability, and inability to predict trends in existing technologies.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for monitoring children's rehabilitation training compliance is provided, comprising: collecting raw training data of children, performing environmental interference analysis and training intention differentiation processing, and performing deviation quantification correction on quasi-subjective quality data to obtain training data; wherein, the raw training data includes quasi-subjective quality data, action and behavioral characteristic data, and acoustic characteristic data; the quasi-subjective quality data is the evaluation data of the children's training process; standardizing and fusing the training data to obtain multi-source training data; based on the multi-source training data, dynamically calculating the training compliance sub-index through a nonlinear fusion function, and dynamically adjusting the sub-index weights by combining historical training compliance sub-indexes and a training regression model to obtain a comprehensive training compliance index; inputting the multi-source training data and the comprehensive training compliance index into a time series prediction model to obtain compliance trend prediction results and training effect evaluation.

[0005] Based on the above technical solutions, the method for monitoring children's rehabilitation training compliance provided in this application constructs a complete technical chain, including raw data collection, interference identification and bias correction, multi-source fusion, dynamic sub-index calculation, and trend prediction and evaluation. This solves the core pain points of traditional assessments in home settings, such as large subjective biases, inability to distinguish interference, and static, rigid weights. It achieves accurate quantification, dynamic adaptation, and forward-looking prediction of children's rehabilitation training compliance, providing a reliable quantitative basis for closed-loop optimization of subsequent training programs and significantly improving the efficiency and effectiveness of home-based rehabilitation training.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves collecting raw training data from children's training and processing it for environmental interference analysis and training intent differentiation. This includes: processing video data of the entire training process collected by a camera in real time using an edge computing device, extracting the coordinates, angles, movement trajectories, and gaze focus features of x core joints of the child to obtain action and behavioral feature data; and processing audio data of the entire training process collected by a microphone in real time using an edge computing device, extracting the fundamental frequency, short-time energy, and Mel-frequency cepstral coefficients as acoustic features to obtain acoustic feature data.

[0007] Based on the field-of-view edge of the video data, the Canny edge detection algorithm is used to identify and record sudden motion events and timestamps of non-child targets in the scene, resulting in a visual interference event sequence. Based on the audio data, the ambient background audio is continuously monitored using short-time energy and spectrum mutation anomaly detection technology to identify and record sudden abnormal noise events and timestamps, resulting in an auditory interference event sequence. The visual interference event sequence and the auditory interference event sequence are aligned and fused according to the timestamps to obtain an environmental interference event stream. The environmental interference event stream includes event type, occurrence timestamp, and intensity information.

[0008] From the acoustic feature data, a subset of features generated by children's vocalizations, used to characterize children's emotional state and participation, is extracted and aligned with the action and behavior feature data by timestamp, and fused to obtain children's multidimensional behavioral response time-series data. Based on the environmental interference event stream and the children's multidimensional behavioral response time-series data, a time-series causal association model is established to calculate the causal association probability between the interference event and the subsequent changes in children's behavioral responses within a preset time window. According to the comparison result of the causal association probability and a preset threshold, the training period corresponding to the interference event is labeled with intent tags to obtain the differentiated interference period or active behavior period.

[0009] By employing a dual-modal visual and auditory interference event detection system and a temporal causal association model, this method accurately distinguishes between action pauses caused by environmental interference and children's voluntary inattentiveness, avoiding the distortion in compliance assessment caused by interference events in traditional methods. The generated labels for interfered / voluntary behavior periods provide a basis for weight adjustment in subsequent sub-index calculations, making compliance assessments more closely reflect children's actual training status and effectively improving assessment accuracy.

[0010] In conjunction with the first aspect mentioned above, one possible implementation involves quantifying and correcting the deviation of quasi-subjective quality data, including: calculating the individual rating deviation coefficient and rating confidence based on historical quasi-subjective quality data and corresponding action and behavioral characteristic data, and weighting and fusing the quasi-subjective quality data to obtain subjective quality data.

[0011] By calculating individual bias coefficients and confidence indices through regression analysis, subjective scores are weighted and calibrated, effectively reducing individual bias and noise in therapist / parent subjective scores and improving the reliability of subjective quality data. Dynamically lowering the weight of confidence scores further reduces the interference of subjective data on the overall assessment results, significantly enhancing the correlation between the features of multi-source data fusion and rehabilitation outcomes.

[0012] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the training compliance sub-index includes: determining the sub-index dimension and extracting corresponding feature values ​​based on the multi-source training data to obtain a sub-index sequence; wherein, the sub-index dimension includes completion, continuity, timeliness, preset difficulty coefficient, quality, and consistency.

[0013] The sub-exponential sequence is dynamically solved using a nonlinear fusion function to obtain the dynamic sub-exponential sequence value; wherein, the calculation formula of the nonlinear fusion function is:

[0014] in, It is a dynamic sub-exponential function. For real-time eigenvalue functions, For historical data sequence functions, For dynamic feature weighting function, For the characteristic dynamic quantization function, For activation function, Dimension-specific activation coefficients, α is the global dynamic adjustment factor, i is the dimension index, t is the time index, k is the feature index, N is the total number of features, α is the weight amplification coefficient, and β is the aggregation gain coefficient.

[0015] The values ​​of the dynamic sub-index sequence are calibrated to the preset interval [a,b] to obtain the training compliance sub-index.

[0016] By introducing a nonlinear fusion function, real-time features, historical data, and dynamic weights are deeply coupled, solving the problem that traditional static weighting cannot characterize the dynamic changes in compliance during training. Through dimension-specific activation coefficients and global dynamic adjustment factors, personalized dynamic calculation of sub-indices is achieved, making the sub-indices more closely match the compliance fluctuation patterns of different children and different training stages, thus improving the accuracy and adaptability of the sub-indices.

[0017] In conjunction with the first aspect mentioned above, one possible implementation involves dynamically adjusting the sub-index weights by combining historical training compliance sub-indices with the training regression model. This includes: obtaining historical training compliance sub-indices and corresponding historical rehabilitation effect labels, and removing outliers using the 3σ principle to obtain a historical training dataset; inputting the historical training dataset into the training regression model, and obtaining the contribution of each dimension of historical training compliance sub-indices to the rehabilitation effect through time series cross-validation fitting; and normalizing the contribution to obtain the dynamic weights of each dimension of training compliance sub-indices.

[0018] The 3σ principle was used to remove outliers from historical data, ensuring the purity of the training dataset and avoiding interference from outliers on model fitting. By training a regression model to fit the contribution of sub-indices to rehabilitation effects, data-driven dynamic adjustment of sub-indice weights was achieved, replacing the traditional method of manually pre-setting weights. This significantly improved the correlation between the comprehensive compliance index and actual rehabilitation effects, and increased the accuracy of weight adaptation.

[0019] In conjunction with the first aspect mentioned above, in one possible implementation, the method for obtaining the contribution of each dimension of historical training compliance index to the rehabilitation effect includes: based on the historical training dataset, training a gradient boosting tree regression model through h-fold extended window time series cross-validation, and optimizing the hyperparameters to obtain a preliminary training model; wherein, the hyperparameters include: learning rate and decision tree depth.

[0020] An ensemble regression model is obtained by integrating the arithmetic mean of similar quantifiable parameters from the initial trained model; wherein, the similar quantifiable parameters are the decision tree splitting threshold, the output value of the leaf nodes of the decision tree, the weight coefficient of the decision tree in each iteration, and the total number of decision trees in gradient boosting; an information gain set is obtained by extracting the information gain generated by each sub-index dimension during the decision tree splitting process from the ensemble regression model; the information gain set is then processed... After normalization, the contribution of each dimension of historical training compliance index to the rehabilitation effect was obtained.

[0021] The adoption of h-fold extended window time series cross-validation avoids the leakage of future information from time series data, ensuring the reliability of model training. Fixed learning rate and decision tree depth ensure the model's repeatability and feasibility. Ensemble regression models are obtained through arithmetic averaging of similar parameters, effectively reducing the risk of overfitting in single models and improving the stability and accuracy of contribution calculation. Contribution quantification based on information gain makes weight allocation more interpretable.

[0022] In conjunction with the first aspect mentioned above, in one possible implementation, the method for obtaining the compliance trend prediction result and the training effect evaluation includes: concatenating the multi-source training data with the comprehensive training compliance index to obtain a fusion feature matrix; and performing smoothing and noise reduction and PCA dimensionality reduction on the fusion feature matrix to obtain a low-dimensional fusion feature matrix.

[0023] An LSTM time series prediction model is constructed based on a historical low-dimensional fused time series feature matrix. It is trained using the Adam optimizer and a weighted joint loss function, and iteratively optimized using an early stopping mechanism to obtain the optimal prediction model. The low-dimensional fused feature matrix is ​​then input into the optimal prediction model to obtain compliance trend prediction results and a quantitative score for training effectiveness. The quantitative score is then used to classify training effectiveness into levels based on a preset threshold. The training effectiveness evaluation includes both a quantitative score and a training effectiveness level.

[0024] Moving average denoising and PCA dimensionality reduction are employed to improve the information density and computational efficiency of the feature matrix. The LSTM dual-output model simultaneously achieves trend prediction and effect evaluation, avoiding redundancy in multi-model deployment. The combination of early stopping mechanism and weighted joint loss function ensures the model's generalization ability and prediction accuracy. The generated training effect levels and trend prediction results provide rehabilitation therapists with intuitive and actionable decision-making support, improving prediction accuracy.

[0025] In conjunction with the first aspect mentioned above, one possible implementation method for acquiring multi-source training data includes: aligning the training data, the environmental interference event stream, and the children's multidimensional behavioral response time-series data with the main time axis through dynamic time warping, performing timestamp synchronization and standardization to obtain multi-dimensional feature vectors; wherein, the training data includes subjective quality data, action and behavioral feature data, and acoustic feature data; based on a preset association logic and data model, the multi-dimensional feature vectors are spliced ​​and encapsulated to integrate and obtain multi-source training data.

[0026] By dynamically warping and aligning with the main timeline, the problem of temporal heterogeneity in multi-source data is solved, enabling deep integration of interfering event streams, children's behavioral responses, and training data. Structured encapsulation based on a pre-defined data model provides a unified and standardized input format for subsequent sub-index calculations and model training, improving the scalability and compatibility of the entire technical solution and supporting rapid adaptation to multiple scenarios.

[0027] In conjunction with the first aspect above, one possible implementation further includes: obtaining a rehabilitation training plan; determining whether the rehabilitation training plan needs to be adjusted based on the compliance trend prediction result and the training effect evaluation; if so, screening candidate tasks through static fitness scoring and generating a task recommendation set by combining the multi-armed gambling machine algorithm and collaborative filtering mechanism; and adjusting the rehabilitation training plan based on the comprehensive training compliance index and the trend prediction result using a PID collaborative feedback control mechanism.

[0028] By combining multi-armed gambling machine algorithms with collaborative filtering mechanisms to generate task recommendation sets, a balance is achieved between exploring new tasks and utilizing effective tasks, thus improving the personalized matching degree of training tasks. The introduction of a PID collaborative feedback control mechanism maps the comprehensive compliance index and trend prediction results to PID parameters, enabling dynamic and precise adjustment of the training program. This forms a complete closed loop of monitoring, evaluation, adjustment, and feedback, significantly improving the intelligence level of rehabilitation training.

[0029] Secondly, this application provides a monitoring system for children's rehabilitation training compliance, comprising: a data acquisition and processing module, a training compliance module, and a training effect module; wherein, the data acquisition and processing module acquires the children's original training data, performs environmental interference analysis and training intention differentiation processing, and performs deviation quantification correction on quasi-subjective quality data to obtain training data; wherein, the original training data includes quasi-subjective quality data, action and behavioral characteristic data, and acoustic characteristic data; the training data is standardized and fused to obtain multi-source training data.

[0030] The training compliance module dynamically calculates the training compliance sub-index based on the multi-source training data using a nonlinear fusion function, and dynamically adjusts the sub-index weights by combining historical training compliance sub-indexes with the training regression model to obtain a comprehensive training compliance index. The training effect module inputs the multi-source training data and the comprehensive training compliance index into a time series prediction model to obtain compliance trend prediction results and training effect evaluation.

[0031] The system adopts a modular architecture design, separating data collection and processing, compliance calculation, and effect prediction into independent modules. Each module has clear responsibilities and can be iterated independently, improving the system's maintainability and scalability. Each module corresponds one-to-one with the methodological steps, ensuring the consistency and feasibility of the technical solution. This provides a complete system-level solution for monitoring children's rehabilitation training compliance, reducing deployment and maintenance costs.

[0032] This application provides a method and system for monitoring children's rehabilitation training adherence. It constructs a multi-source data fusion, dynamic quantitative assessment, trend prediction, and closed-loop optimization solution for monitoring children's rehabilitation training adherence, addressing industry pain points such as numerous interferences, subjective assessments, rigid weights, and lack of prediction in home settings. Through technologies such as bimodal interference identification, bias correction, nonlinear dynamic sub-index calculation, ensemble learning weight optimization, LSTM time-series prediction, and PID closed-loop adjustment, it achieves accurate quantification, dynamic adaptation, and proactive intervention of adherence, significantly improving the efficiency and effectiveness of home-based rehabilitation training and providing intelligent and personalized technical support for children's rehabilitation.

[0033] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0034] Figure 1 A system architecture diagram of a children's rehabilitation training compliance monitoring system provided in this application embodiment; Figure 2 A flowchart illustrating a method for monitoring children's compliance with rehabilitation training, provided as an embodiment of this application; Figure 3 A flowchart illustrating another method for monitoring children's compliance with rehabilitation training provided in this application embodiment; Figure 4 A flowchart illustrating another method for monitoring children's compliance with rehabilitation training provided in this application embodiment; Detailed Implementation

[0035] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0036] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0037] To address the technical problems of distorted rehabilitation training compliance assessment, poor weighting adaptability, and inability to predict trends in existing technologies, this application provides a method for monitoring children's rehabilitation training compliance. The method includes: collecting raw training data from children, analyzing environmental interference and distinguishing training intentions, and performing deviation quantification correction on quasi-subjective quality data to obtain training data; wherein the raw training data includes quasi-subjective quality data, movement and behavioral characteristic data, and acoustic characteristic data; and standardizing and fusing the training data to obtain multi-source training data.

[0038] Based on the multi-source training data, the training compliance sub-index is dynamically calculated through a nonlinear fusion function. The weights of the sub-indexes are dynamically adjusted by combining the historical training compliance sub-indexes with the training regression model to obtain a comprehensive training compliance index. The multi-source training data and the comprehensive training compliance index are input into a time series prediction model to obtain compliance trend prediction results and training effect evaluation.

[0039] like Figure 2 As shown in the embodiments of this application, the method for monitoring children's rehabilitation training compliance includes: S201. Collect the raw training data of children, analyze environmental interference and distinguish training intentions, and perform deviation quantification correction on the subjective quality data to obtain training data.

[0040] The raw training data includes quasi-subjective quality data, action and behavioral characteristic data, and acoustic characteristic data. The environmental interference event stream includes event type, occurrence timestamp, and intensity information. The quasi-subjective quality data serves as evaluation data for the child's training process.

[0041] In some implementations, quasi-subjective quality data is obtained by scoring the child's training process; action and behavior feature data are obtained by processing the video data of the entire training process captured by the camera in real time through edge computing devices, extracting the coordinates, angles, movement trajectories, and gaze focus features of x core joints of the child; and acoustic feature data are obtained by processing the audio data of the entire training process captured by the microphone in real time through edge computing devices, extracting the fundamental frequency, short-time energy, and Mel-frequency cepstral coefficients.

[0042] Based on the field-of-view edge of the video data, the Canny edge detection algorithm is used to identify and record sudden motion events and timestamps of non-child targets in the scene, resulting in a visual interference event sequence. Based on the audio data, the ambient background audio is continuously monitored using short-time energy and spectrum mutation anomaly detection technology to identify and record sudden abnormal noise events and timestamps, resulting in an auditory interference event sequence. The visual interference event sequence and the auditory interference event sequence are aligned and fused according to the timestamps to obtain an environmental interference event stream.

[0043] From the acoustic feature data, a subset of features generated by children's vocalizations, used to characterize children's emotional state and participation, is extracted and aligned with the action and behavior feature data by timestamp, and fused to obtain children's multidimensional behavioral response time-series data. Based on the environmental interference event stream and the children's multidimensional behavioral response time-series data, a time-series causal association model is established to calculate the causal association probability between the interference event and the subsequent changes in children's behavioral responses within a preset time window. According to the comparison result of the causal association probability and a preset threshold, the training period corresponding to the interference event is labeled with intent tags to obtain the differentiated interference period or active behavior period.

[0044] It should be noted that the action and behavior feature data, acoustic feature data, environmental interference event streams, and children's multidimensional behavioral response time series data involved in the above implementation methods have all undergone strict timestamp alignment and standardization processing to ensure that all multi-source data are fused and calculated under the same time reference to support the accurate analysis of the time series causal relationship model.

[0045] It should also be noted that the intent label will be embedded as a weighting adjustment factor in the feature extraction process of the subsequent training compliance sub-index. For example, if the child maintains attention or resumes training independently during the interrupted period, the scoring weight of the "continuity" and "focus" sub-indexes for that period will be increased, thereby avoiding the distortion of compliance assessment caused by environmental interference and making the assessment results more consistent with the child's actual training status.

[0046] For example, if, within 2 seconds of detecting a sudden unusual noise, the child's gaze deviates from the training area and the joint movement speed drops below 0.1 m / s, the causal correlation probability is calculated to be 0.85, which is higher than the preset threshold of 0.7. This period is then marked as a period of interference. If no environmental interference event is detected but the child's attention shifts, it is marked as a period of active behavior. This intent label will subsequently be used to assign values ​​to the dynamic feature weight function in the nonlinear fusion function, further improving the accuracy of the training compliance index.

[0047] In some implementations, based on historical quasi-subjective quality data and corresponding action and behavioral characteristic data, the individual rating deviation coefficient and rating confidence are calculated by regression, and the quasi-subjective quality data are weighted and fused to obtain subjective quality data.

[0048] It's important to note that regression analysis aims to build a personalized correction model for each assessor. Its core function is to identify systematic bias patterns in the assessor's scoring. For example, a therapist might habitually assign a higher score to the "movement fluency" dimension than supported by objective data. The calculated "personal rating bias coefficient" is used to perform linear translation and scaling correction on the original scores; while the "rating confidence score" is a weight between 0 and 1, used to measure the reliability of the assessor's ratings during subsequent multi-source data fusion.

[0049] It's important to note that the temporal causal association model is an attention-based temporal neural network designed to quantify the strength of the causal association between a "disruptive event" and "subsequent changes in children's behavior." The model employs an encoder-decoder architecture and incorporates an attention mechanism. The encoder processes multidimensional temporal data of children's behavioral responses, learning dynamic patterns such as continuous gaze, movement, and acoustic feature sequences. The attention layer is trained to focus on the timing of the disruptive event; if the model learns to consistently allocate high attention weights to subsequent behavioral sequences after the disruptive event, it indicates a strong correlation. The decoder / output layer, based on the encoded environmental disruptive event stream and attention distribution, outputs a scalar between 0 and 1, representing the causal probability P of "behavioral change caused by the disruptive event." Supervised training is performed using historical labeled data, such as expert-labeled periods of "disrupted" / "initiated behavior." The model's learning objective is that, given an input disruptive event and its subsequent behavioral sequence, the output probability P should be as close as possible to the expert-labeled label of 1 or 0.

[0050] The probability P is calculated through forward propagation of a pre-trained temporal causal association model. A single disturbance event is taken as the trigger point, and its temporal data of the child's behavior within a subsequent T-second time window is used as a sample pair. This pair is input into the model, which extracts features from the behavior sequence through an encoder. An attention mechanism calculates the association weight between the features of each time point in the behavior sequence and the features of the disturbance event. Finally, the model maps the aggregated information to a probability value P between 0 and 1 through a fully connected layer and a sigmoid activation function. The formula for P is as follows: Where σ is the Sigmoid function, AttentionContext is the weighted aggregated environmental disturbance event stream vector, and W and b are model parameters.

[0051] It should also be noted that the preset time window refers to the length of time from the occurrence of the interference event onwards that children's behavioral data is extracted for causal analysis. This is set based on the temporal characteristics of children's physiological and cognitive responses. The typical reaction time for a child's attention to shift and behavior to pause due to a sudden external interference is usually within a few seconds. Therefore, the window length can be preset to 2 to 5 seconds. This is a configurable parameter that can be adjusted according to specific training scenarios, such as task type and child's age. When the system detects any interference event, it automatically extracts T seconds of "children's multidimensional behavioral response time-series data" from its timestamp as the input data pair for this causal analysis.

[0052] The preset threshold in the phrase "based on the comparison result of the causal association probability and the preset threshold" is used to convert the continuous probability P into a binary classification label "interferenced" / "active behavior" decision boundary. The model is determined using a performance optimization method based on the validation set. After model training, an independent validation set containing expert-annotated true labels is used. All samples from the validation set are input into the model to obtain a series of predicted probabilities P. PR curves or ROC curves are plotted, and the optimal balance point is found. Typically, the probability value corresponding to maximizing the F1 score or the Youden index is chosen as the preset threshold. For example, the optimal threshold may be approximately 0.7. If P is greater than or equal to the threshold, the time period corresponding to the interference event is determined to be an interference period; if P is less than the threshold, it is determined to be an active behavior period.

[0053] For example, if therapist A's historical scores deviate from the group average by a mean of 1.2 points, and the maximum group score is 10, then the individual deviation correction coefficient = 1 - mean deviation / maximum group score = 1 - 1.2 / 10 = 0.88. The therapist A's original score is 8, and the calibrated score = 8 × 0.88 = 7.04. If parent B's original score is 7, the typical response delay is 6 seconds, and the preset delay threshold is 10 seconds, then the state confidence index = 1 - (response delay / preset delay threshold) = 1 - 6 / 10 = 0.4. The assessor's default weight is 0.3. Since parent B's confidence index of 0.4 is less than 0.5, their dynamic weight = default weight × state confidence index = 0.3 × 0.4 = 0.12, and the calibrated score = 7 × 0.12 = 0.84. The objective feature data mapping score shows an objective action completion rate of 85%, which is mapped to a score of 8.5 with a weight of 0.2. The effective score is 8.5 × 0.2 = 1.7. The final subjective quality data = calibrated score of therapist A × 0.7 + original score of parent B × 0.1 + objective feature data mapping score × 0.2 = 7.04 × 0.7 + 0.84 × 0.1 + 1.7 = 6.712.

[0054] S202 performs standardization and fusion processing on the training data to obtain multi-source training data.

[0055] The training data includes subjective quality data, action and behavioral feature data, and acoustic feature data.

[0056] In some implementations, the training data, the environmental interference event stream, and the children's multidimensional behavioral response time series data are aligned with the main time axis through dynamic time warping, and timestamp synchronization and standardization are performed to obtain multidimensional feature vectors. Based on the preset association logic and data model, the multidimensional feature vectors are spliced ​​and encapsulated to integrate multi-source training data.

[0057] It should be noted that the aforementioned timestamp synchronization and standardization are technical prerequisites for ensuring the effective fusion and calculation of subjective quality data, action and behavioral feature data, acoustic feature data, environmental interference event streams, and children's multidimensional behavioral response time-series data under the same time reference. Its core lies in aligning time-series data with different sampling rates using a dynamic time warping algorithm and normalizing them using a main time axis as a global reference, such as video frame timestamps; thereby generating multidimensional feature vectors with consistent time labels, providing standardized input for subsequent modeling.

[0058] For example, if the training data includes a subjective quality score of 7.33 at the 5th minute of the 2nd training cycle, a joint velocity of 0.5 m / s in the action and behavior feature data, a fundamental frequency of 250 Hz in the acoustic feature data, an environmental interference event stream record of no interference events at that moment, and children's multidimensional behavioral reaction time sequence data showing that children's gaze is focused on the training area, then these data are aligned to the main time axis through dynamic time warping to generate a multidimensional feature vector [7.33, 0.5, 250, 0, 1], and then encapsulated into multi-source training data based on a preset data model for subsequent sub-index calculation.

[0059] S203. Based on the multi-source training data, the training compliance sub-index is dynamically calculated through a nonlinear fusion function. The weights of the sub-indexes are dynamically adjusted by combining the historical training compliance sub-indexes with the training regression model to obtain a comprehensive training compliance index.

[0060] In some implementations, based on the multi-source training data, the sub-index dimensions are determined and corresponding feature values ​​are extracted to obtain the sub-index sequence; wherein, the sub-index dimensions include completeness, continuity, timeliness, preset difficulty coefficient, quality, and consistency; The sub-exponential sequence is dynamically solved using a nonlinear fusion function to obtain the dynamic sub-exponential sequence value; wherein, the calculation formula of the nonlinear fusion function is:

[0061] in, It is a dynamic sub-exponential function. For real-time eigenvalue functions, For historical data sequence functions, For dynamic feature weighting function, For the characteristic dynamic quantization function, For activation function, Dimension-specific activation coefficients, α is the global dynamic adjustment factor, i is the dimension index, t is the time index, k is the feature index, N is the total number of features, α is the weight amplification factor, and β is the aggregation gain factor. The values ​​of the dynamic sub-index sequence are calibrated to the preset interval [a,b] to obtain the training compliance sub-index.

[0062] It should be noted that the feature dynamic quantization function integrates the current feature value with its historical trend, outputting a "dynamic value," the formula of which is: ;in, The purpose is to map the original eigenvalues ​​x to the (0,1) interval and introduce nonlinearity; For historical data sequences The slope obtained by performing a first-order linear fit is used to quantify recent performance trends; positive values ​​indicate progress, and negative values ​​indicate regression. The trend influence factor is a configurable positive real number used to adjust the strength of the influence of historical trends on current value.

[0063] The dynamic feature weight function generates normalized weights based on the real-time task and the child's state. Its calculation formula is: ;in, To reflect the immediate importance of the k-th feature for evaluating the i-th sub-index dimension at time t, its value can be calculated by a small neural network or a rule-based query system based on real-time status such as the current training task attributes and the child's fatigue level. θ is a temperature parameter, a positive real number, which controls the concentration of the weight distribution. The larger θ is, the more concentrated the weights are on the most important features.

[0064] Dimension-specific activation coefficients The intermediate fusion results from different dimensions are then subjected to a final nonlinear mapping, and the expression varies depending on the dimension. When the dimension is completion, continuity, or a preset difficulty coefficient, the dimension-specific activation coefficient is... ,in, When the dimension is quality, the dimension-specific activation coefficient is... Where μ is the preset baseline threshold for this dimension, the Sigmoid function smoothly maps the output to the (0,1) interval, and the change is most significant near the μ point; when the dimension is time-sensitive, the dimension-specific activation coefficient is , where v is the delay between the actual start time and the planned start time; The zero-tolerance delay threshold is defined as follows: if v ≤ τ, it indicates on-time delivery or slight delay, and the function value is 1. The penalty slope coefficient, ζ>0, determines the rate at which the compliance index decreases with increasing latency after exceeding the threshold; when the dimension is consistency, the specific activation coefficient is... Where v is the characteristic value of the current period, The historical average value of this feature is derived from the historical data sequence H; σ is the tolerance parameter. It controls the width of the function curve; the larger σ is, the higher the tolerance for fluctuations; the smaller σ is, the more severe the penalty for deviations from the historical mean. This function is applied at v= The maximum value is 1.

[0065] Global dynamic adjustment factor To achieve macro-level fine-tuning of assessment results through rehabilitation phase strategies, its expression is: ,in, This is the identifier for the recovery stage at time t. The policy offset is a predefined constant related to the stage P(t) and dimension i, typically ranging between ±0.2. The adjustment coefficient has a value range of [0, 1]. It can be used to control whether the adjustment is fully effective (η=1) or partially effective.

[0066] is the activation function, a standard function used to provide smooth, non-linear normalization, and its expression is: Its output is always positive, and its growth is approximately linear when u is large, providing a bounded and robust normalization capability for the denominator and avoiding the numerical instability that may be caused by traditional linear denominators.

[0067] For example, suppose the system needs to dynamically adjust the weights of the six sub-indices.

[0068] The first step is to collect data: historical data from the past 6 cycles, each cycle has a 6-dimensional sub-index vector [0.8, 0.7, 0.9, 0.6, 0.85, 0.75], and the rehabilitation effect index for the corresponding cycle is the rehabilitation progress score [75, 80, 70, 85, 78, 82]; the sub-index vector for the current cycle is [0.82, 0.68, 0.88, 0.62, 0.83, 0.78].

[0069] The second step is to train the regression model: Random forest regression is used, with 6 samples × 6-dimensional features of historical data as the training set and rehabilitation effect score as the target variable; the regression model is trained to quantify the relationship between the sub-index and the rehabilitation effect.

[0070] The third step is to extract feature importance: obtain the importance scores of each sub-index from the trained model [0.15, 0.10, 0.20, 0.05, 0.30, 0.20].

[0071] Step 4, normalize the weights: the sum of importance scores is 0.15 + 0.10 + 0.20 + 0.05 + 0.30 + 0.20

[0072] =1.0; the new weight vector is [0.15,0.10,0.20,0.05,0.30,0.20].

[0073] Step 5, Smoothing Adjustment: Smoothing factor λ = 0.7, previous period's old weights were [0.12, 0.15, 0.18, 0.10, 0.25, 0.20]; final weight = 0.7 × new weight + 0.3 × old weight, completion: 0.7 × 0.15 + 0.3 × 0.12 = 0.141, continuity: 0.7 × 0.10 + 0.3 × 0.15 = 0.115, timeliness: 0.7 × 0.20 + 0.3 × 0.18 = 0.194, difficulty coefficient: 0.7 × 0.05 + 0.3 × 0.10 = 0.065, quality: 0.7 × 0.30 + 0.3 × 0.25 = 0.285, consistency: 0.7 × 0.20 + 0.3 × 0.20 = 0.200; normalization verification: the sum is 1.0, no further adjustment is needed.

[0074] Step 6: Update the weights and calculate the comprehensive index: Update the final weights to the system configuration; the current cycle comprehensive training compliance index = 0.141×0.82+0.115×0.68+0.194×0.88+0.065×0.62+0.285×0.83+0.200×0.78≈0.802, and the final comprehensive index is 0.802.

[0075] S204. Input the multi-source training data and the comprehensive training compliance index into the time series prediction model to obtain the compliance trend prediction results and training effect evaluation.

[0076] The training effectiveness evaluation includes a quantitative score for training effectiveness and a training effectiveness level.

[0077] In some implementations, the multi-source training data is concatenated with the comprehensive training compliance index to obtain a fusion feature matrix; the fusion feature matrix is ​​then smoothed, denoised, and reduced in dimension using PCA to obtain a low-dimensional fusion feature matrix; an LSTM time series prediction model is constructed based on the historical low-dimensional fusion time series feature matrix, trained using the Adam optimizer and a weighted joint loss function, and iteratively optimized using an early stopping mechanism to obtain the optimal prediction model; the low-dimensional fusion feature matrix is ​​input into the optimal prediction model to obtain compliance trend prediction results and a quantitative score for training effect, and the quantitative score for training effect is divided into training effect levels according to a preset threshold.

[0078] It should be noted that the weighted joint loss function is composed of the mean squared error (MSE) and the mean absolute error (MAE) in a preset ratio. This is used to simultaneously optimize the overall accuracy of the predicted values ​​and their robustness to abnormal fluctuations during training. The preset ratio is not determined by a single empirical value, but rather by a multi-objective Bayesian optimization framework based on historical validation set performance. The joint loss function is defined as follows: ,in Let be the mixing coefficients to be optimized. The optimization objective is to simultaneously minimize the root mean square error (RMSE) of the predicted sequence and a smoothness index, such as the variance of the difference between adjacent predicted points, on the validation set. Through iterative sampling guided by a Gaussian process surrogate model, the optimal balance of the Pareto front is finally determined. The value is then fixed to the preset ratio.

[0079] For example, the optimal prediction model trained based on historical data outputs a quantitative training effect score sequence [78, 82, 85, 88, 86] for a child after five consecutive training sessions. The system calculates the moving average and variance of this sequence and maps it to a level based on a preset threshold for the rehabilitation cycle in which it occurs. This cycle threshold, after cluster analysis and expert calibration, is set as follows: ≥90 is excellent, [80, 89] is good, [70, 79] is average, and <70 requires intervention. Based on this, three scores (82, 85, 88) in this sequence fall into the good range, and the overall trend is steadily increasing. Therefore, the compliance trend in this assessment is steadily improving, and the training effect level is judged as good. At the same time, the system stores the quantitative score, feature sequence, and final level of this assessment in the historical database for subsequent periodic updates of the threshold model.

[0080] Based on the above technical solutions, the method for monitoring children's rehabilitation training compliance provided in this application focuses on the core needs of children's home rehabilitation training scenarios. It effectively solves the problems of traditional compliance assessments, such as subjectivity, susceptibility to interference, lack of dynamic adaptation and forward-looking guidance, providing a more scientific and practical monitoring and optimization solution for children's rehabilitation. By distinguishing between environmental interference and children's proactive behavior and correcting subjective biases, the authenticity of the assessment data is ensured, making the results more consistent with the children's actual training status.

[0081] By relying on dynamic quantification and adaptive weight adjustment to replace manual experience-based judgment, the accuracy and objectivity of compliance assessment are improved. Combined with trend prediction and closed-loop optimization of training programs, an upgrade from post-assessment to pre-intervention and dynamic adjustment is achieved, facilitating the implementation of personalized rehabilitation training. Simultaneously, through standardized model training and integration methods, the stability and feasibility of the program are ensured, adapting to complex family scenarios and effectively improving the efficiency and quality of children's rehabilitation training.

[0082] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S203 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Obtain the historical training compliance index and corresponding historical rehabilitation effect labels, and use the 3σ principle to remove outliers to obtain the historical training dataset.

[0083] The historical rehabilitation effect label is a quantitative score based on a rehabilitation assessment scale or a comprehensive evaluation by a clinician, used to characterize the actual rehabilitation effect after each training session.

[0084] In some implementations, obtaining historical training compliance sub-indices and corresponding historical rehabilitation effect labels includes: extracting multi-dimensional training compliance sub-indices and their corresponding rehabilitation effect labels for each training period within the past M training cycles from the historical database; calculating the mean and standard deviation for each dimension's sub-index sequence; and considering data points outside the range of [mean - 3 × standard deviation, mean + 3 × standard deviation] as outliers and removing them to obtain a cleaned historical training dataset.

[0085] It should be noted that the process of removing outliers using the 3σ principle is performed independently for each dimension's sub-index sequence to ensure the statistical independence of the data in each dimension. After removing outliers, if there are missing data, linear interpolation is used to fill in the missing data to ensure data continuity.

[0086] For example, suppose we extract the sub-indices of each dimension and their corresponding rehabilitation effect scores for a child's past 30 training sessions from a historical database, with a maximum score of 100. The mean of the completion sub-index is 75, and the standard deviation is 5, so the normal range is [60, 90]. If the completion sub-index for a training session is 95, it is considered an outlier and removed. Similar operations are performed on all dimensions, resulting in a cleaned historical training dataset containing 28 valid training samples and 2 outlier samples removed.

[0087] S302. Input the historical training dataset into the training regression model, and obtain the contribution of each dimension of historical training compliance index to the rehabilitation effect through time series cross-validation fitting.

[0088] The contribution degree is used to quantify the influence of each dimension's sub-index on the rehabilitation effect. A higher contribution degree indicates that the dimension provides more information when predicting the rehabilitation effect. Similar quantifiable parameters include: decision tree splitting threshold, decision tree leaf node output values, decision tree weight coefficients in each iteration, and the total number of decision trees in gradient boosting.

[0089] In some implementations, the training regression model is a gradient boosting tree regression model, using h-fold expanded window time-series cross-validation for training and validation. Specifically, the historical training dataset is divided into h consecutive folds in chronological order. Each time, the first i folds are used as the training set, and the (i+1)th fold is used as the validation set, progressively expanding the training window for a total of h-1 training and validation iterations. In each training iteration, the learning rate and decision tree depth are optimized using grid search to obtain a preliminary training model. Then, the arithmetic mean of similar quantifiable parameters from each preliminary training model is taken and ensembled into a more robust ensemble regression model.

[0090] It should be noted that using time-series cross-validation instead of random cross-validation is a key technical choice that respects the inherent temporal dependence and concept drift characteristics of rehabilitation effect data, ensuring that the model has higher generalization ability when predicting future cycles. The contribution is ultimately obtained by calculating the sum of information gains of each dimension's sub-indices at all decision tree split nodes of the ensemble regression model, and then normalizing it. Information gain is chosen as the metric because it can effectively capture the non-linear relationship between features and rehabilitation effects, thus reflecting the true impact weights better than a simple linear correlation coefficient.

[0091] For example, suppose the historical training dataset contains 28 samples, each with 6 sub-index dimensions. Using 5-fold extended window time series cross-validation, 5 initial trained models are obtained. After ensemble, a single ensemble regression model containing 100 decision trees is obtained. The sum of information gains for each dimension is calculated by counting all split nodes: completion 0.35, continuity 0.20, timeliness 0.15, preset difficulty coefficient 0.10, quality 0.12, and consistency 0.08. After normalization, the contribution of each dimension is: completion 0.35, continuity 0.20, timeliness 0.15, preset difficulty coefficient 0.10, quality 0.12, and consistency 0.08.

[0092] S303. Normalize the contribution to obtain the dynamic weights of the training compliance index for each dimension.

[0093] The dynamic weights are used to weight and fuse the training compliance sub-indices of each dimension when calculating the comprehensive training compliance index, so as to reflect the differences in the importance of different dimensions in predicting the current rehabilitation effect.

[0094] In some implementations, the contribution normalization process employs the Softmax function to process the contribution set, thereby enhancing the discriminative power of the key dimension weights. The specific steps are as follows: Let the set of contributions of each dimension's historical training compliance index to the rehabilitation effect be... Where D is the total number of dimensions; calculate the exponential weights before normalization. , where β is a temperature coefficient used to control the steepness of the weight distribution; for Normalization is performed to obtain the final dynamic weights. The temperature coefficient β is determined through cross-validation. Its function is to give a relatively higher weight to the dimensions with higher contributions, thereby giving them a more dominant position in the comprehensive index.

[0095] It should be noted that the dynamic weights are time-dependent, not fixed. The system periodically re-executes steps S301 to S303, for example, after accumulating N new training samples; it recalculates the contribution and updates the dynamic weights based on the updated historical training dataset, thereby enabling the weights to adaptively adjust with the rehabilitation stage, individual progress, and data pattern evolution, ensuring that the comprehensive training adherence index always sensitively and accurately reflects the most critical adherence dimension at present.

[0096] For example, suppose the system has already collected historical training data and needs to calculate dynamic weights for six sub-index dimensions (completeness, continuity, timeliness, preset difficulty coefficient, quality, and consistency).

[0097] The first step is to determine the set of contributions: assuming that after analyzing historical data using a regression model, we can obtain the set of contributions of each dimension to the rehabilitation effect. Among them, completion rate contribution =0.18, Continuous contribution =0.12, Timeliness Contribution =0.25, Contribution of preset difficulty coefficient =0.08, Quality Contribution =0.30 and consistency contribution =0.20.

[0098] The second step is to set the temperature coefficient: determine the value of the temperature coefficient β through cross-validation. It is assumed that the validation results show that the model's prediction performance is best when β=2, meaning the weight distribution has good discriminative power while avoiding over-concentration.

[0099] The third step is to calculate the indexed weights: according to the formula... Calculate the exponential weights for each dimension to obtain... , , , , , .

[0100] The fourth step is to normalize to obtain dynamic weights: calculate the sum of the exponential weights. =8.8406; according to the formula Normalization is performed to obtain , , , , , The sum of all weights is verified to be 1.0000; therefore, the dynamic weight vector for the current period is... .

[0101] Step 5: Calculate the comprehensive index using dynamic weights: Assuming the training compliance sub-indices for each dimension in the current period are completion 0.85, continuity 0.78, timeliness 0.92, preset difficulty coefficient 0.65, quality 0.88, and consistency 0.80, then the comprehensive training compliance index is a weighted sum of...

[0102] Step 6, Dynamically Update Weights: Assume the system is set to recalculate weights every N=10 new training samples. After a period of time, the system adds 10 periods of training data, including sub-indices and rehabilitation effect labels. The system will train the regression model and extract a new set of contributions based on the expanded dataset containing the new data. The same Softmax normalization process is used to obtain new dynamic weights. To avoid abrupt changes in weights, a smoothing factor of 0.7 is used to perform a weighted average of the old and new weights, resulting in the updated weights. After substituting the values ​​into the calculation, a new weight vector is obtained, which is used for the calculation of the comprehensive index in subsequent cycles.

[0103] Based on the above technical solution, outliers in historical data are removed using the 3σ principle, ensuring the purity of the training dataset and avoiding interference from outliers on model fitting. By training the regression model to fit the contribution of sub-indices to rehabilitation effects, data-driven dynamic adjustment of sub-indice weights is achieved, replacing the traditional method of manually pre-setting weights. This significantly improves the correlation between the comprehensive compliance index and actual rehabilitation effects, and enhances the accuracy of weight adaptation.

[0104] In one possible implementation, combining Figure 3 ,like Figure 4 As shown, following S302, the method for monitoring children's rehabilitation training compliance provided in this application embodiment further includes the following S401 to S404: S401. Based on the historical training dataset, train the gradient boosting tree regression model using h-fold extended window time series cross-validation, and optimize the hyperparameters to obtain the preliminary training model.

[0105] The hyperparameters include the learning rate and the decision tree depth.

[0106] In some implementations, the specific implementation of the h-fold extended window time series cross-validation includes: dividing the historical training dataset into h consecutive, time-disjoint folds in chronological order, with each fold containing approximately equal numbers of samples; initializing the training window to the data of the first fold; For i from 1 to h-1, perform the following iterations: train a gradient boosting tree regression model using the data in the current training window; validate the trained model on the (i+1)th fold of data and calculate the validation error; include the (i+1)th fold of data into the training window to expand the training data range; during each iteration of training, use Bayesian optimization to simultaneously optimize hyperparameters: within a preset search space, with the goal of minimizing the validation error on the previous i folds of training data, automatically find the optimal combination of learning rate and decision tree depth; finally, collect models with different hyperparameter combinations obtained from h-1 iterations of training to form a preliminary training model set.

[0107] It should be noted that using an expanded window instead of a sliding window for time-series cross-validation maximizes the use of historical data in each training iteration while strictly adhering to the principle that future data cannot be used to train past models, preventing data leakage and ensuring the temporal validity of the model's generalization ability assessment. Gradient boosting tree regression models are chosen because they naturally handle nonlinear relationships, feature interactions, and are robust to outliers. Their hyperparameter, the learning rate, controls the strength of residual correction by each new tree, while the decision tree depth controls the model complexity and the depth of feature interactions; both jointly influence the trade-off between model bias and variance.

[0108] For example, suppose the historical training dataset contains 120 samples arranged chronologically, and h=5. It is then divided into 5 folds, each containing approximately 24 samples. The training process is as follows: The model is first trained using fold 1 with samples 1-24, and validated using fold 2 with samples 25-48; the second time, fold 1-2 are used for training with samples 1-48, and validated using fold 3 with samples 49-72; and so on, until fold 1-4 are used for training with samples 1-96, and validated using fold 5 with samples 97-120. During each training iteration, the Bayesian optimizer performs 20 rounds of evaluation within the defined hyperparameter space. Finally, in the third iteration, an optimal set of hyperparameters is determined: a learning rate of 0.12 and a decision tree depth of 6. After all four iterations, four preliminary trained models trained on different data segments with potentially different hyperparameters are obtained, constituting the preliminary trained model set.

[0109] S402. An integrated regression model is obtained by integrating the arithmetic mean of the similar quantifiable parameters of the preliminary training model.

[0110] Among them, the similar quantifiable parameters are the decision tree splitting threshold, the output value of the decision tree leaf nodes, the decision tree weight coefficient in each iteration, and the total number of decision trees in gradient boosting.

[0111] In some implementations, the specific steps for parameter averaging integration include: for each model in the initial training model set, extracting the following parameters from all decision trees within it: the decision tree splitting threshold is the threshold of the feature value used to split the data at each splitting node; the output value of each leaf node is the predicted value corresponding to each leaf node; the decision tree weight coefficient for each iteration is the weight corresponding to each decision tree in the gradient boosting iteration, i.e., the product of the learning rate and the contribution of the tree in the current round or the equivalent coefficient; the total number of decision trees for gradient boosting is the total number of decision trees used by the model.

[0112] For all the four types of parameters corresponding to the initial training models, their arithmetic mean is calculated. For example, for the split threshold, the thresholds at the same structural positions are averaged; for the total number of decision trees, the values ​​are directly averaged. Based on the arithmetic mean of these four types of parameters, an average decision tree structure is reconstructed, and the averaged leaf node output value and tree weight coefficient are assigned. At the same time, the average number of decision trees is used as the tree number constraint of the new model, and finally the ensemble regression model is assembled.

[0113] It should be noted that using the arithmetic mean of similar quantifiable parameters for ensemble integration is a model-level ensemble method, rather than simply weighting the predictions of multiple models. This method effectively integrates the structured knowledge of different initially trained models, smooths out model fluctuations caused by differences in training data time periods or hyperparameter randomness, thereby improving the robustness and stability of the final ensemble model and reducing the risk of overfitting.

[0114] For example, suppose there are three initially trained models (M1, M2, M3), all used to predict rehabilitation outcomes. For a certain split node in the completion sub-index: M1 has a split threshold of 85, corresponding to a leaf output of 10, and a tree weight of 0.1; M2 has a split threshold of 83, corresponding to a leaf output of 12, and a tree weight of 0.09; M3 has a split threshold of 87, corresponding to a leaf output of 9, and a tree weight of 0.11. Then, the average split threshold for this node in the ensemble model is (85+83+87) / 3=85, the average leaf output is (10+12+9) / 3≈10.33, and the average tree weight is (0.1+0.09+0.11) / 3=0.1. By performing such averaging on all nodes and all trees, and taking the average total number of decision trees, the final ensemble regression model is constructed.

[0115] S403. By extracting the information gain generated by each sub-index dimension during the decision tree splitting process from the integrated regression model, an information gain set is obtained.

[0116] The information gain is used to measure the reduction in impurity of the category or regression value when a decision tree node is split using a certain sub-index dimension. It is a key indicator for quantifying the importance of the feature to the prediction target.

[0117] In some implementations, the specific process of extracting the information gain set includes: traversing each decision tree in the ensemble regression model and visiting each split node; at each split node, recording the features used for splitting, i.e., the corresponding sub-index dimensions, and obtaining the information gain value calculated for that split; merging and accumulating the information gains according to the sub-index dimensions, and performing the same operation on other dimensions; organizing all dimensions and their corresponding accumulated information gain values ​​into a set to form the information gain set, denoted as . ;in, Let represent the total information gain of the i-th sub-index dimension, and D be the total number of dimensions.

[0118] It should be noted that the information gain extraction here is a static analysis based on a pre-constructed, structurally defined ensemble regression model. It reflects the frequency and effectiveness with which each feature dimension is used for key segmentation throughout the model's decision-making logic. A higher total information gain indicates a greater role for that dimension in distinguishing different rehabilitation outcomes, meaning its potential contribution may be higher. Compared to analysis methods based on model output weights, this method reveals the importance of features within the model's internal decision-making path more directly and in greater detail.

[0119] For example, suppose the ensemble regression model contains 100 decision trees with a total of 5000 non-leaf nodes. After traversal and summation, the following information gains are obtained: "Completion" dimension was used for 1200 splits, with a cumulative information gain of 185.6; "Continuity" dimension was used for 900 splits, with a cumulative information gain of 112.3; "Quality" dimension was used for 800 splits, with a cumulative information gain of 98.7; "Timeliness" dimension was used for 750 splits, with a cumulative information gain of 85.2; "Preset Difficulty Coefficient" dimension was used for 700 splits, with a cumulative information gain of 74.1; and "Consistency" dimension was used for 650 splits, with a cumulative information gain of 63.5. The resulting information gain set G = {185.6, 112.3, 98.7, 85.2, 74.1, 63.5}. This set will serve as the direct input for subsequent calculations of the contribution of each dimension.

[0120] S404, Perform the information gain set... After normalization, the contribution of each dimension of historical training compliance index to the rehabilitation effect was obtained.

[0121] Among them, the Normalization is a method that transforms the set of information gains into a standard probability distribution, such that the contribution of each dimension represents its relative proportion in the total importance of all dimensions, and the sum of the contributions of all dimensions is 1.

[0122] In some implementations, the The specific steps for normalization include: Let the set of information gains extracted from the ensemble regression model be... ;in, Let be the cumulative information gain of the i-th dimension, and D be the total number of sub-index dimensions. Calculate the sum of the information gains of all dimensions. For each dimension i, calculate its contribution. ; obtain the contribution set ;in, .

[0123] It should be noted that, adopting Normalization is used instead of other normalization methods because it directly satisfies the physical meaning of contribution as "relative weight," meaning that the contribution of each dimension is directly equal to the proportion of its information gain to the total information gain. This ensures that the final calculated comprehensive training compliance index has clear weighting and interpretability, and that the relative importance ranking between dimensions will not change due to different normalization methods.

[0124] For example, assume the information gain set obtained from S403 is {185.6, 112.3, 98.7, 85.2, 74.1, 63.5}, corresponding to the six dimensions of "completeness", "continuity", "quality", "timeliness", "preset difficulty coefficient", and "consistency", respectively. Calculate the total information gain S = 185.6 + 112.3 + 98.7 + 85.2 + 74.1 + 63.5 = 619.4. The contribution of each dimension is as follows: Completion: 185.6 / 619.4≈0.300; Continuity: 112.3 / 619.4≈0.181; Quality: 98.7 / 619.4≈0.159; Timeliness: 85.2 / 619.4≈0.138; Preset Difficulty Coefficient: 74.1 / 619.4≈0.120; Consistency: 63.5 / 619.4≈0.103. The verification sum is 0.300+0.181+0.159+0.138+0.120+0.103=1.001≈10.300+0.181+0.159+0.138+0.120+0.103=1.001≈1. This contribution set C will be used for subsequent calculation of dynamic weights.

[0125] Based on the above technical solution, an h-fold extended window time series cross-validation training model is adopted. This strictly adheres to the temporal order of rehabilitation training data, effectively preventing future information leakage and ensuring that the trained model and its subsequent contribution have stronger generalization ability and predictive reliability in real-world time-series scenarios. An ensemble regression model is constructed by merging similar quantifiable parameters from multiple preliminary training models using an arithmetic mean. This method integrates model knowledge from different training stages and hyperparameter configurations, smoothing out potential random biases or overfitting in a single model, making the final model more robust, and thus providing a more stable and reliable model source for basic information gain extraction.

[0126] By directly extracting information gain from the decision tree splitting process of an ensemble regression model, we can clearly and intuitively measure the reduction in impurity or variance of each sub-index dimension when the model makes decisions. This approach provides an intrinsic measure of feature importance directly linked to predictive performance, avoiding the poor interpretability issues of black-box models.

[0127] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor or by instructions in software form. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0128] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for monitoring children's adherence to rehabilitation training, characterized in that, include: The raw training data of children is collected, and environmental interference is analyzed and training intentions are distinguished. The deviation of the quasi-subjective quality data is quantitatively corrected to obtain the training data. The raw training data includes quasi-subjective quality data, action and behavior characteristic data, and acoustic characteristic data. The quasi-subjective quality data is the evaluation data of the children's training process. The training data is standardized and fused to obtain multi-source training data; Based on the multi-source training data, the training compliance sub-index is dynamically calculated through a nonlinear fusion function. The weights of the sub-indexes are dynamically adjusted by combining the historical training compliance sub-indexes with the training regression model to obtain a comprehensive training compliance index. By inputting multi-source training data and a comprehensive training compliance index into a time series prediction model, compliance trend prediction results and training effect evaluation are obtained.

2. The method for monitoring children's compliance with rehabilitation training according to claim 1, characterized in that, The process of collecting raw training data from children for environmental interference analysis and training intention differentiation includes: By using edge computing devices to process video data of children's training process collected by cameras in real time, the coordinates, angles, movement trajectories and gaze focus features of x core joints of children are extracted to obtain action and behavior feature data. The audio data of children's training process is processed in real time by the microphone through edge computing devices, and the fundamental frequency, short-time energy and Mel-frequency cepstral coefficients are extracted to obtain acoustic feature data. Based on the field of view edge of the video data, the Canny edge detection algorithm is used to identify and record sudden motion events and timestamps of non-child targets in the scene to obtain a sequence of visual interference events; Based on the audio data, the ambient background audio is continuously monitored using short-time energy and spectrum mutation anomaly detection technology to identify and record sudden abnormal noise events and timestamps, thereby obtaining a sequence of auditory interference events; The visual interference event sequence and the auditory interference event sequence are aligned and fused according to timestamps to obtain an environmental interference event stream; wherein, the environmental interference event stream includes event type, occurrence timestamp, and intensity information; From the acoustic feature data, a subset of features generated by children's vocalizations and used to characterize children's emotional state and participation is extracted, and aligned and spliced ​​with the action and behavior feature data according to the timestamp to obtain multidimensional behavioral response time series data of children. Based on the environmental disturbance event stream and the time series data of children’s multidimensional behavioral responses, a time series causal relationship model is established to calculate the causal relationship probability between the disturbance event and the changes in children’s behavioral responses within a subsequent preset time window. Based on the comparison result between the causal association probability and the preset threshold, the training period corresponding to the interference event is labeled with intent tags to obtain the differentiated interference period or active behavior period.

3. The method for monitoring children's compliance with rehabilitation training according to claim 1, characterized in that, The deviation quantification correction of the quasi-subjective quality data includes: calculating the individual rating deviation coefficient and rating confidence based on historical quasi-subjective quality data and corresponding action and behavioral characteristic data, and weighting and fusing the quasi-subjective quality data to obtain subjective quality data.

4. The method for monitoring children's compliance with rehabilitation training according to claim 1, characterized in that, The method for obtaining the training compliance index includes: Based on the multi-source training data, the sub-index dimensions are determined and the corresponding feature values ​​are extracted to obtain the sub-index sequence; wherein, the sub-index dimensions include completeness, continuity, timeliness, preset difficulty coefficient, quality, and consistency; The sub-exponential sequence is dynamically solved using a nonlinear fusion function to obtain the dynamic sub-exponential sequence value; wherein, the calculation formula of the nonlinear fusion function is: in, It is a dynamic sub-exponential function. For real-time eigenvalue functions, For historical data sequence functions, For dynamic feature weighting function, For the characteristic dynamic quantization function, For activation function, Dimension-specific activation coefficients, α is the global dynamic adjustment factor, i is the dimension index, t is the time index, k is the feature index, N is the total number of features, α is the weight amplification factor, and β is the aggregation gain factor. The values ​​of the dynamic sub-index sequence are calibrated to the preset interval [a,b] to obtain the training compliance sub-index.

5. The method for monitoring children's compliance with rehabilitation training according to claim 1, characterized in that, The method of dynamically adjusting the sub-index weights by combining historical training compliance sub-index with the training regression model includes: Obtain the historical training compliance index and corresponding historical rehabilitation effect labels, and use the 3σ principle to remove outliers to obtain the historical training dataset; The historical training dataset is input into the training regression model, and the contribution of each dimension of historical training compliance index to the rehabilitation effect is obtained by time series cross-validation fitting. The contribution values ​​are normalized to obtain the dynamic weights of the training compliance index for each dimension.

6. The method for monitoring children's compliance with rehabilitation training according to claim 5, characterized in that, The methods for obtaining the contribution of each dimension of historical training compliance index to rehabilitation effectiveness include: Based on the historical training dataset, a gradient boosting tree regression model is trained using h-fold extended window time series cross-validation, and hyperparameters are optimized to obtain a preliminary trained model; wherein, the hyperparameters include: learning rate and decision tree depth; An ensemble regression model is obtained by integrating the arithmetic mean of similar quantifiable parameters of the initial training model; wherein, the similar quantifiable parameters are the decision tree splitting threshold, the output value of the leaf nodes of the decision tree, the weight coefficient of the decision tree in each iteration, and the total number of decision trees in the gradient boost. By extracting the information gain generated by each sub-index dimension during the decision tree splitting process from the integrated regression model, an information gain set is obtained; Perform the information gain set After normalization, the contribution of each dimension of historical training compliance index to the rehabilitation effect was obtained.

7. The method for monitoring children's compliance with rehabilitation training according to claim 1, characterized in that, The methods for obtaining the compliance trend prediction results and training effect evaluation include: The multi-source training data is concatenated with the comprehensive training compliance index to obtain a fusion feature matrix; The fused feature matrix is ​​then smoothed and denoised, and its dimensionality is reduced using PCA to obtain a low-dimensional fused feature matrix. An LSTM time series prediction model is constructed based on a historical low-dimensional fused time series feature matrix. It is trained using the Adam optimizer and a weighted joint loss function, and iteratively optimized using an early stopping mechanism to obtain the optimal prediction model. The low-dimensional fusion feature matrix is ​​input into the optimal prediction model to obtain the compliance trend prediction result and the quantitative score of training effect. The quantitative score of training effect is then divided into training effect levels according to a preset threshold. The training effect evaluation includes the quantitative score of training effect and the training effect level.

8. The method for monitoring children's compliance with rehabilitation training according to claim 2, characterized in that, The methods for acquiring the multi-source training data include: The training data, the environmental disturbance event stream, and the children's multidimensional behavioral response time series data are aligned with the main time axis through dynamic time warping, and timestamp synchronization and standardization are performed to obtain a multidimensional feature vector; wherein, the training data includes subjective quality data, action and behavioral feature data, and acoustic feature data; Based on the preset association logic and data model, the multi-dimensional feature vectors are spliced ​​and encapsulated to obtain multi-source training data.

9. A method for monitoring children's adherence to rehabilitation training according to any one of claims 1-8, characterized in that, Also includes: Obtain a rehabilitation training plan; Based on the compliance trend prediction results and training effect evaluation, it is determined whether the rehabilitation training program needs to be adjusted. If so, then candidate tasks are selected through static fit scoring, and a task recommendation set is generated by combining the multi-armed gambling machine algorithm and collaborative filtering mechanism; Based on the comprehensive training compliance index and the trend prediction results, the rehabilitation training program is adjusted using a PID-coordinated feedback control mechanism.

10. A monitoring system for children's adherence to rehabilitation training, characterized in that, The system includes: a data acquisition and processing module, a training compliance module, and a training effect module; The data acquisition and processing module collects raw training data from children's training, performs environmental interference analysis and training intention differentiation processing, and performs deviation quantification correction on quasi-subjective quality data to obtain training data; wherein, the raw training data includes quasi-subjective quality data, action and behavioral feature data, and acoustic feature data; the training data is standardized and fused to obtain multi-source training data; The training compliance module, based on the multi-source training data, dynamically calculates the training compliance sub-index through a nonlinear fusion function, and dynamically adjusts the sub-index weights by combining historical training compliance sub-indexes and the training regression model to obtain a comprehensive training compliance index. The training effect module inputs multi-source training data and a comprehensive training compliance index into a time series prediction model to obtain compliance trend prediction results and training effect evaluation.

Images