Child obesity risk prediction system based on deep learning

Abstract

The invention discloses a children obesity risk prediction system based on deep learning, and belongs to the technical field of data processing, and the system comprises a children sequence data integration module, an intelligent children obesity prediction module, a learnable parameter optimization module and a prediction report generation module. According to the method, a graph structure between children is constructed, local time features are extracted through convolution operation, position coding and a multi-head self-attention mechanism are introduced, global time features are extracted, the local time features and the global time features are fused, a learnable adjacent matrix and a dynamic weighting matrix are combined, and spatio-temporal features are obtained for prediction. According to the method, individual positions are divided by setting a division threshold, mutation positions are generated based on the selected individual positions, the fitness change rate and the motion control coefficient, updating is completed according to fitness comparison and individual position detection, optimal parameters are found, the obesity risk of children can be evaluated more accurately, and a more reliable basis is provided for early intervention.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and specifically refers to a children's obesity risk prediction system based on deep learning. Background Art

[0002] A children's obesity risk prediction system is a system that uses deep learning technology to predict the future obesity risk of children by analyzing multi-dimensional factors such as children's physiological data, diet data, exercise data, and health data. By establishing a neural network model, it mines the potential laws in historical data, identifies the important factors affecting children's obesity, and early warns of the possible obesity risk, providing a scientific basis for parents, schools, communities, and relevant medical institutions, helping to formulate personalized health intervention and prevention measures, thereby reducing the incidence of children's obesity and improving the overall health level of children.

[0003] However, in the existing children's obesity risk prediction systems, there are problems that children's obesity is a long-term and multi-factor process, affected by multiple factors, it is difficult to accurately predict based on short-term data alone, and the existing predictions mostly rely on the independent characteristics of individual children, ignoring the potential similarities between children, resulting in inaccurate prediction results; in the existing children's obesity risk prediction systems, there are problems that children's obesity risk is affected by multiple factors, these factors are often very complex and there are interaction effects, and it is difficult for existing predictions to capture this fine interaction effect, resulting in a decline in the accuracy of children's obesity risk prediction. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the existing technology, the present invention provides a child obesity risk prediction system based on deep learning. In the existing child obesity risk prediction system, child obesity is a long-term and multi-factor process, affected by various factors, and it is difficult to accurately predict relying solely on short-term data. Moreover, the existing predictions mostly rely on the independent characteristics of individual children, ignoring the potential similarities between children, resulting in inaccurate prediction results. This solution constructs a graph structure between children, extracts local time features through convolutional operations, and introduces positional encoding and multi-head self-attention mechanisms to extract global time features. The local time features and global time features are fused to capture multi-scale time dependencies, making the prediction results more accurate and comprehensive. Combining a learnable adjacency matrix and a dynamic weighting matrix, spatio-temporal features are obtained for prediction, which can automatically adjust the mutual influence according to the similarity between children, improving the prediction accuracy of child obesity risk. In the existing child obesity risk prediction system, child obesity risk is affected by various factors, and these factors are often very complex and have interaction effects. It is difficult for existing predictions to capture such fine interaction effects, resulting in a decrease in the prediction accuracy of child obesity risk. This solution uses each individual position to represent a set of learnable parameters to ensure effective exploration at the beginning, thus providing a more reliable starting point for child obesity risk prediction. A division threshold is set to divide individual positions, and mutation positions are generated based on the selected individual positions, fitness change rate, and designed dynamic control coefficient. The optimization process can give priority to focusing on parameter combinations with higher prediction ability, and updates are completed according to fitness comparison and individual position detection to find the optimal parameters, more effectively capturing the complex factor interaction effects in child obesity risk prediction and being able to more accurately evaluate the obesity risk of children.

[0005] The child obesity risk prediction system based on deep learning provided by the present invention includes a child sequence data integration module, an intelligent child obesity prediction module, a learnable parameter optimization module, and a prediction report generation module;

[0006] The child sequence data integration module collects and processes historical child sequence data;

[0007] The intelligent child obesity prediction module constructs a graph structure between children, extracts local time features through convolutional operations, and introduces positional encoding and multi-head self-attention mechanisms to extract global time features. The local time features and global time features are fused, and combined with a learnable adjacency matrix and a dynamic weighting matrix, spatio-temporal features are obtained for prediction;

[0008] The learnable parameter optimization module sets a division threshold to divide individual positions, generates mutation positions based on the selected individual positions, fitness change rate, and dynamic control coefficient, and completes the update according to fitness comparison and individual position detection to find the optimal parameters;

[0009] The prediction report generation module inputs the real-time child sequence data into the child obesity risk prediction model constructed based on the optimal parameters for analysis, obtains the obesity risk level corresponding to the real-time child sequence data, and generates a prediction report.

[0010] Further, the child sequence data integration module collects and processes historical child sequence data;

[0011] The collection of historical child sequence data is to collect physiological data, diet data, exercise data, health data, timestamps, and obesity risk levels;

[0012] The processing of historical child sequence data is to perform data cleaning, data normalization, data encoding, and dataset construction on the historical child sequence data to obtain a training dataset and a test dataset.

[0013] Further, the intelligent child obesity prediction module constructs a child obesity risk prediction model based on a neural network, and is provided with a defined child graph structure unit, a time feature extraction unit, a spatio-temporal feature extraction unit, and a prediction output unit, specifically including the following:

[0014] Define the child graph structure unit; use the training dataset as the input data X, take each child in the input data X as a node, and the node attributes are the corresponding child sequence data to obtain a node set and a node attribute matrix; preset a similarity threshold, if the cosine similarity between two node attributes is less than the similarity threshold, there is an edge between the corresponding two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; obtain an edge set and an adjacency matrix C;

[0015] Time feature extraction unit; extract time features from the input data X based on the P1 layer convolutional layer, where P1 is the number of convolutional layers, including the following:

[0016] Local time feature extraction; use convolutional operations to capture the temporal dependencies in the data, and extract local time features through multiple convolutional layers;

[0017] Global time feature extraction; includes the following:

[0018] Position embedding; preset a period cycle, perform position encoding on each timestamp in the input data X to obtain a position encoding matrix WB, and then embed WB into X to obtain embedded data ; The formula used is as follows:

[0019] ;

[0020] In the formula, WB his the position encoding value corresponding to the timestamp h, WB is the position encoding matrix composed of the position encoding values corresponding to all timestamps h, and h is the timestamp index. is the cosine function;

[0021] Multi-head self-attention mechanism; captures the long-term dependencies between different timestamps in the data through multi-head self-attention processing, weights and aggregates the information using multiple attention heads to obtain the multi-head self-attention output.

[0022] Residual connection and layer normalization processing; combines the multi-head self-attention output in the (p - 1)-th convolutional layer and the global time feature L p-1 extracted by the (p - 1)-th convolutional layer, performs a residual connection, and conducts layer normalization processing to obtain the residual feature.

[0023] Feed-forward network; inputs the residual feature into the feed-forward network for processing.

[0024] Obtains the global time feature; combines the output B p of the feed-forward network in the p-th convolutional layer and the residual feature E p extracted by the p-th convolutional layer, performs a residual connection, and conducts layer normalization processing again to obtain the global time feature.

[0025] Local-global time feature fusion; concatenates the local time feature S p and the global time feature L p performs convolution operations to extract more spatio-temporal dependencies from the concatenated features to obtain the comprehensive feature, and then uses the attention mechanism to weight and fuse the local time feature, global time feature, and comprehensive feature to obtain the fused time feature.

[0026] Spatio-temporal feature extraction unit; analyzes the spatial dependencies between different timestamps, combines the learnable adjacency matrix and the dynamic weighting matrix to obtain the spatio-temporal feature, and the formula used is as follows:

[0027] ;

[0028] In the formula, R p is the spatio-temporal feature extracted by the p-th convolutional layer, is the row normalization matrix, and is the column normalization matrix, and is the learnable parametric adjacency matrix, is the dynamic weighting matrix of the p-th convolutional layer, and C is the adjacency matrix, is the rowsums activation function, is the softmax activation function, T is the transpose operation of the matrix, and H is the number of timestamps. , , and are the four weight matrices of the p-th convolutional layer at timestamp h. , , and are learnable parameters.

[0029] The prediction output unit inputs the spatio-temporal feature R extracted by the last convolutional layer into the fully connected layer for prediction operations and outputs the predicted data labels. P1

[0030] Furthermore, the learnable parameter optimization module optimizes the learnable parameters of the neural network in the child obesity risk prediction model based on the swarm intelligence algorithm, and is equipped with an initialization unit, an individual position division unit, a design dynamic control coefficient unit, a mutation unit, an update unit, and a parameter determination unit, specifically including the following:

[0031] The initialization unit establishes a parameter search space for the learnable parameters of the neural network in the child obesity risk prediction model, randomly initializes N individual positions within the parameter search space, represents each individual position with a set of learnable parameters, and uses the cross-entropy loss value of the child obesity risk prediction model based on the learnable parameters for the test data set as the fitness value corresponding to the individual position.

[0032] The individual position division unit sets a division threshold based on the mean μ and standard deviation δ of the fitness values of all individual positions, as well as the fitness change rate, and regards the individual positions with fitness values greater than the division threshold F as pioneer individual positions, and the individual positions with fitness values less than or equal to the division threshold F as pioneer individual positions. The formula for calculating the fitness change rate is as follows:

[0033] ;

[0034] In the formula, is the fitness change rate at the a-th iteration, n and a are the individual position index and iteration number index respectively, and are the n-th individual positions at the a-th and a - 1-th iterations respectively, and are respectively and 's fitness values;

[0035] Design a dynamic control coefficient unit; preset two initial values V0 and V1, and calculate the dynamic control coefficient for each iteration in segments based on the two initial values; the formula used is as follows:

[0036] ;

[0037] In the formula, V a is the dynamic control coefficient at the a-th iteration, a max is the maximum number of iterations, is the initial value of the n-th individual position, is 's fitness value;

[0038] Mutation unit; randomly select two individual positions from the pioneer individual position and the pioneer individual position respectively, and generate the mutation positions corresponding to all individual positions based on the selected individual positions, the fitness change rate, and the dynamic control coefficient; the formula used is as follows:

[0039] ;

[0040] In the formula, is 's corresponding mutation position, and are two individual positions randomly selected from the pioneer individual position, and are two individual positions randomly selected from the pioneer individual position;

[0041] Update unit; includes the following:

[0042] First fitness comparison; compare the fitness value of the mutation position with the fitness value of the original individual position. If , then keep the original individual position unchanged; otherwise, perform individual position detection;

[0043] Individual position detection; detect whether the original individual position is the pioneer individual position. If is the pioneer individual position, then use to directly replace ; otherwise, perform the second fitness value comparison;

[0044] Second fitness value comparison; compare the fitness value of the mutation position with the fitness value of the worst pioneer individual position. If , then use to replace , and use Replace ; otherwise, use to directly replace ; where is the position of the worst pioneer individual at the a-th iteration;

[0045] The parameter determination unit; preset the fitness threshold to detect whether there is an individual position with a fitness value less than the fitness threshold z th , if so, use the learnable parameters represented by this individual position as the optimal parameters of the neural network, and construct a child obesity risk prediction model based on the optimal parameters; otherwise, if the maximum number of iterations is reached, go to the initialization unit to re-initialize the individual position; otherwise, increment the number of iterations by 1 and go to the individual position division unit to continue the iteration.

[0046] Furthermore, the prediction report generation module collects and processes real-time child sequence data; inputs the processed real-time child sequence data into the child obesity risk prediction model constructed based on the optimal parameters for analysis, and according to the output data label, obtains the obesity risk level corresponding to the real-time child sequence data, and generates a prediction report.

[0047] The beneficial effects achieved by the present invention using the above solution are as follows:

[0048] (1) Aiming at the problems in the existing child obesity risk prediction system that child obesity is a long-term and multi-factor process, affected by multiple factors, it is difficult to accurately predict by relying solely on short-term data, and existing predictions mostly rely on the independent characteristics of individual children, ignoring the potential similarities between children, resulting in inaccurate prediction results. This solution constructs a graph structure between children, extracts local time features through convolution operations, pays attention to the short-term changes of child obesity, and introduces position encoding and multi-head self-attention mechanisms to extract global time features, captures the long-term trend and periodic changes of child obesity, fuses the local time features with the global time features, captures multi-scale time dependencies, synthesizes short-term and long-term information, makes the prediction results more accurate and comprehensive, combines the learnable adjacency matrix and dynamic weighting matrix to obtain spatio-temporal features for prediction, can automatically adjust the mutual influence according to the similarity between children, improves the adaptability to individual differences, and thus more accurately captures the risk factors and development trends of obesity, improving the prediction accuracy of child obesity risk.

[0049] (2)In view of the problem that in the existing children's obesity risk prediction system, the children's obesity risk is affected by multiple factors, which are often very complex and have interactions, and it is difficult for existing predictions to capture such fine interactions, resulting in a decline in the accuracy of children's obesity risk prediction. In this solution, a parameter search space is established for the learnable parameters of the neural network in the children's obesity risk prediction model, and each individual position represents a set of learnable parameters to ensure effective exploration at the beginning, so as to provide a more reliable starting point for children's obesity risk prediction, quickly find the optimal parameters, accurately capture the key factors affecting obesity risk, set a division threshold to divide individual positions, generate mutant positions based on the selected individual positions, fitness change rate and the designed dynamic control coefficient, so that the optimization process can give priority to focusing on parameter combinations with higher prediction ability, improve the efficiency of children's obesity risk prediction, complete the update according to fitness comparison and individual position detection, find the optimal parameters, more effectively capture the complex factor interactions in children's obesity risk prediction, be able to more accurately evaluate children's obesity risk, and provide a more reliable basis for early intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 FIG. is a schematic diagram of a children's obesity risk prediction system based on deep learning provided by the present invention;

[0051] Figure 2 FIG. is a schematic diagram of an intelligent children's obesity prediction module;

[0052] Figure 3 FIG. is a schematic diagram of a learnable parameter optimization module.

[0053] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0055] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the present invention.

[0056] Example 1. Refer to Figure 1 , the child obesity risk prediction system based on deep learning provided by the present invention includes a child sequence data integration module, an intelligent child obesity prediction module, a learnable parameter optimization module, and a prediction report generation module;

[0057] The child sequence data integration module collects and processes historical child sequence data, and sends the data to the intelligent child obesity prediction module;

[0058] The intelligent child obesity prediction module receives the data sent by the child sequence data integration module, constructs a graph structure between children, extracts local time features through convolution operations, and introduces positional encoding and multi-head self-attention mechanism to extract global time features, fuses the local time features with the global time features, combines the learnable adjacency matrix and the dynamic weighting matrix, obtains spatio-temporal features for prediction, and sends the data to the learnable parameter optimization module;

[0059] The learnable parameter optimization module receives the data sent by the intelligent child obesity prediction module, sets a division threshold to divide individual positions, generates mutant positions based on the selected individual positions, fitness change rate, and the designed dynamic control coefficient, completes the update according to fitness comparison and individual position detection, finds the optimal parameters, and sends the data to the prediction report generation module;

[0060] The prediction report generation module receives the data sent by the learnable parameter optimization module, inputs the real-time child sequence data into the child obesity risk prediction model constructed based on the optimal parameters for analysis, obtains the obesity risk level corresponding to the real-time child sequence data, and generates a prediction report.

[0061] Example 2. Refer to Figure 1 , this embodiment is based on the above embodiment. In the child sequence data integration module, historical child sequence data is collected and processed;

[0062] The collection of historical child sequence data is to collect physiological data, diet data, exercise data, health data, time stamps, and obesity risk levels;

[0063] The physiological data includes age, gender, height, weight, body fat percentage, waist circumference, hip circumference, blood pressure, blood sugar, blood lipid, and metabolic rate;

[0064] The dietary data includes daily food intake, food types, eating frequency, sugar intake, and water intake;

[0065] The exercise data includes daily exercise volume, exercise types, exercise frequency, and sedentary time;

[0066] The health data includes past medical history and types of drug use;

[0067] The obesity risk levels include high risk, medium risk, and low risk, and the obesity risk levels are used as data labels;

[0068] The processed historical child sequence data is to perform data cleaning, data normalization, data encoding, and dataset construction processing on the historical child sequence data to obtain a training dataset and a test dataset;

[0069] The data cleaning is to remove error values, missing values, and outliers in the data;

[0070] The data normalization is to unify the data into the same range using the maximum-minimum scaling method;

[0071] The data encoding is to convert the data labels into numerical data using One-Hot encoding;

[0072] The dataset construction is to select 70% of the data from the collected historical child sequence data to construct a training dataset, and the remaining 30% of the data is used as a test dataset.

[0073] Example 3, refer to Figure 1 and Figure 2 , this example is based on the above example. In the intelligent child obesity prediction module, a child obesity risk prediction model is constructed based on a neural network, which is provided with a defined child graph structure unit, a time feature extraction unit, a spatio-temporal feature extraction unit, and a prediction output unit, and specifically includes the following contents:

[0074] Define the child graph structure unit; integrate different child sequence data into a unified framework, capture the potential similarity relationships between children through the graph structure, rather than relying solely on the features of a single child, so as to enhance the accuracy and robustness of child obesity risk prediction; use the training dataset as the input data X, take each child in the input data X as a node, and the node attribute is the corresponding child sequence data to obtain a node set and a node attribute matrix; preset a similarity threshold, if the cosine similarity between two node attributes is less than the similarity threshold, then there is an edge between the corresponding two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; obtain an edge set and an adjacency matrix C;

[0075] Time Feature Extraction Unit; Childhood obesity is a long-term development process that is affected by various factors over a period of time. The Time Feature Extraction Unit can capture these temporal dependencies. Local time features focus on short-term changes, while global time features focus on long-term trends. The combination of the two can more comprehensively describe the obesity development trajectory of children, comprehensively consider the impact of factors at different time scales on childhood obesity, and improve the accuracy of prediction; Based on the P1 convolutional layer, time features are extracted from the input data X. P1 is the number of convolutional layers and includes the following:

[0076] Local Time Feature Extraction; Use convolutional operations to capture temporal dependencies in the data and extract local time features through multiple convolutional layers; The formula used is as follows:

[0077] ;

[0078] In the formula, S p is the local time feature extracted by the p-th convolutional layer, A p-1 is the spatio-temporal feature extracted by the (p - 1)-th convolutional layer. When p = 1, the corresponding A0 is the input data X. p is the convolutional layer index. and are the first weight matrix and the second weight matrix respectively. and are the first bias and the second bias respectively. , , and are learnable parameters. is the convolutional operation. is the Hadamard product. is the tanh activation function. is the sigmoid activation function.

[0079] Global Time Feature Extraction; Includes the following:

[0080] Position Embedding; Considering the impact of time periodicity on childhood obesity, position embedding can effectively capture and model the regular changes within different time periods by integrating time information into the model; Preset the period cycle, perform position encoding on each timestamp in the input data X to obtain the position encoding matrix WB, and then embed WB into X to obtain the embedded data ; The formula used is as follows:

[0081] ;

[0082] In the formula, WB h is the position encoding value corresponding to the timestamp h. WB is the position encoding matrix composed of the position encoding values corresponding to all timestamps h. h is the timestamp index. is the cosine function;

[0083] The multi-head self-attention mechanism; captures the long-term dependencies between different timestamps in the data through multi-head self-attention processing, weights and aggregates the information using multiple attention heads to obtain the multi-head self-attention output; the formula used is as follows:

[0084] ;

[0085] ;

[0086] In the formula, , and are the query weight matrix, key weight matrix, and value weight matrix respectively, is the mapping matrix, , , and are learnable parameters, is the multi-head self-attention output in the (p - 1)-th convolutional layer, L p-1 is the global temporal feature extracted by the (p - 1)-th convolutional layer, and L0 corresponding to p = 1 is the embedded data , i is the attention head index, I is the number of attention heads, head1, head i and head I are the outputs of the 1st, i-th, and I-th attention heads respectively, is the self-attention function, is the concatenation function;

[0087] Residual connection and layer normalization processing; performs a residual connection on the multi-head self-attention output in the (p - 1)-th convolutional layer and the global temporal feature L p-1 extracted by the (p - 1)-th convolutional layer, and performs layer normalization processing to obtain the residual feature; the formula used is as follows:

[0088] ;

[0089] In the formula, E p is the residual feature extracted by the p-th convolutional layer, is the layer normalization function;

[0090] Feed-forward network; inputs the residual feature into the feed-forward network for processing; the formula used is as follows:

[0091] ;

[0092] In the formula, B p is the output of the feed-forward network in the p-th convolutional layer, and are two weight matrices of the feed - forward network, and are two biases of the feed - forward network, , , and are learnable parameters, is the ReLU activation function;

[0093] Obtain the global time feature; The output B of the feed - forward network in the p - th convolutional layer p and the residual feature E extracted by the p - th convolutional layer p are subjected to residual connection and layer normalization again to obtain the global time feature; The formula used is as follows:

[0094] ;

[0095] In the formula, L p is the global time feature extracted by the p - th convolutional layer;

[0096] Local - global time feature fusion; Through the fusion of local and global time features, the model can comprehensively capture the trends of children's obesity development from both short - term and long - term aspects, so that the model has stronger adaptability when processing time - series data, capturing subtle time changes without ignoring long - term trends; The local time feature S p and the global time feature L p are concatenated, and more spatio - temporal dependencies are extracted from the concatenated features through convolutional operations to obtain the comprehensive feature, and then the attention mechanism is used to weight and fuse the local time feature, global time feature and comprehensive feature to obtain the fused time feature; The formula used is as follows:

[0097] ;

[0098] ;

[0099] In the formula, D p and G p are the comprehensive feature and time feature extracted by the p - th convolutional layer respectively, is the convolutional function;

[0100] Spatio-temporal feature extraction unit; The spatio-temporal feature extraction unit analyzes the comprehensive impact of similar children at different timestamps on obesity risk prediction by combining the spatial dependence and temporal features among children. Through the dynamic weighting matrix, the model can adaptively adjust the connection weights between nodes according to the similarity between different children, so as to better capture the mutual influence among individuals. The learnable adjacency matrix can automatically learn the connection relationships between nodes through model training, reflect the real relationships among children individuals, and improve the adaptability of the model; By analyzing the spatial dependence relationships between different timestamps, combining the learnable adjacency matrix and the dynamic weighting matrix, spatio-temporal features are obtained, and the formula used is as follows:

[0101] ;

[0102] In the formula, R p is the spatio-temporal feature extracted by the p-th convolutional layer, is the row normalization matrix, , is the column normalization matrix, , is the learnable parameterized adjacency matrix, is the dynamic weighting matrix of the p-th convolutional layer, , C is the adjacency matrix, is the rowsums activation function, is the softmax activation function, T is the transpose operation of the matrix, H is the number of timestamps, , , and are the four weight matrices of the p-th convolutional layer at timestamp h, , , and are learnable parameters;

[0103] Prediction output unit; The spatio-temporal feature R P1 extracted by the last convolutional layer is input into the fully connected layer for prediction operation, and the predicted data label is output.

[0104] By performing the above operations, in view of the problems in the existing children's obesity risk prediction system that children's obesity is a long-term and multi-factor process, affected by multiple factors, it is difficult to accurately predict based on short-term data alone, and existing predictions mostly rely on the independent characteristics of individual children, ignoring the potential similarities between children, resulting in inaccurate prediction results. This solution constructs a graph structure between children, extracts local time features through convolutional operations, focuses on the short-term changes in children's obesity, and introduces positional encoding and multi-head self-attention mechanisms to extract global time features, capturing the long-term trends and periodic changes in children's obesity. By fusing local time features with global time features, multi-scale time dependencies are captured, integrating short-term and long-term information, making the prediction results more accurate and comprehensive. Combining a learnable adjacency matrix and a dynamic weighting matrix, spatio-temporal features are obtained for prediction, which can automatically adjust the mutual influence according to the similarity between children, improving the adaptability to individual differences, thereby more accurately capturing the risk factors and development trends of obesity and enhancing the prediction accuracy of children's obesity risk.

[0105] Example 4. Refer to Figure 1 and Figure 3 This example is based on the above example. In the learnable parameter optimization module, the learnable parameters of the neural network in the children's obesity risk prediction model are optimized based on the swarm intelligence algorithm. There are an initialization unit, an individual position division unit, a design dynamic control coefficient unit, a mutation unit, an update unit, and a parameter determination unit, which specifically include the following:

[0106] Initialization unit; Children's obesity risk is affected by multiple factors, which are often very complex and interact with each other, affecting the accurate prediction of children's obesity. By randomly initializing multiple parameter combinations, it is ensured that effective exploration can be carried out at the beginning, thus providing a more reliable starting point for children's obesity risk prediction, quickly finding the optimal parameters, and accurately capturing the key factors affecting obesity risk; The learnable parameters of the neural network in the children's obesity risk prediction model , , , , , , , , , , , , , , , and Establish a parameter search space, randomly initialize the positions of N individuals within the parameter search space, use each individual position to represent a set of learnable parameters, and take the cross-entropy loss value of the child obesity risk prediction model established based on the learnable parameters for the test data set as the fitness value corresponding to the individual position;

[0107] Individual position division unit; Based on the mean μ and standard deviation δ of the fitness values of all individual positions, as well as the fitness change rate, set a division threshold , and take the individual positions with fitness values greater than the division threshold F as pioneer individual positions, and the individual positions with fitness values less than or equal to the division threshold F as pioneer individual positions; The formula for calculating the fitness change rate is as follows:

[0108] ;

[0109] In the formula, is the fitness change rate at the a-th iteration, n and a are the individual position index and iteration number index respectively, and are the n-th individual positions at the a-th and a - 1-th iterations respectively, and are respectively and 's fitness values;

[0110] Design a dynamic control coefficient unit; Preset two initial values V0 and V1, and calculate the dynamic control coefficient at each iteration based on the two initial values in segments; The formula used is as follows:

[0111] ;

[0112] In the formula, V a is the dynamic control coefficient at the a-th iteration, a max is the maximum number of iterations, is the initial value of the n-th individual position, is 's fitness value;

[0113] Mutation unit; By dividing the individual positions and the designed dynamic control coefficient, the optimization process can give priority to focusing on parameter combinations with higher prediction ability, improving the efficiency of child obesity risk prediction; Randomly select two individual positions from the pioneer individual positions and the pioneer individual positions respectively, and generate the mutation positions corresponding to all individual positions based on the selected individual positions, fitness change rate and dynamic control coefficient; The formula used is as follows:

[0114] ;

[0115] In the formula, is The corresponding mutation positions, and are two individual positions randomly selected from the positions of the pioneer individuals, and are two individual positions randomly selected from the positions of the pioneer individuals;

[0116] Update unit; includes the following:

[0117] First fitness comparison; compare the fitness value of the mutation position with the fitness value of the original individual position If , then keep the original individual position unchanged; otherwise, perform individual position detection;

[0118] Individual position detection; detect whether the original individual position is the position of a pioneer individual. If it is the position of a pioneer individual, then use to directly replace ; otherwise, perform the second fitness value comparison;

[0119] Second fitness value comparison; compare the fitness value of the mutation position with the fitness value of the worst pioneer individual position If , then use to replace , and use to replace ; otherwise, use to directly replace ; where is the worst pioneer individual position at the a-th iteration. The worst pioneer individual position is the position with the largest fitness value among all pioneer individual positions;

[0120] Parameter determination unit; by dynamically optimizing the learnable parameters, it can find the optimal parameters, analyze the characteristics of children more accurately and effectively, thereby improving the prediction accuracy of children's obesity risk; preset a fitness threshold to detect whether there is a fitness value of an individual position less than the fitness threshold z th , if there is, then use the learnable parameters represented by this individual position as the optimal parameters of the neural network, and build a children's obesity risk prediction model based on the optimal parameters; otherwise, if the maximum number of iterations is reached, go to the initialization unit to re-initialize the individual positions; otherwise, increment the number of iterations by 1 and go to the individual position division unit to continue the iteration.

[0121] By performing the above operations, in view of the problem in the existing children's obesity risk prediction system that the children's obesity risk is affected by multiple factors, which are often very complex and have interactions, and it is difficult for existing predictions to capture such fine interactions, resulting in a decrease in the accuracy of children's obesity risk prediction, this solution establishes a parameter search space for the learnable parameters of the neural network in the children's obesity risk prediction model. Each individual position represents a set of learnable parameters to ensure effective exploration at the beginning, thereby providing a more reliable starting point for children's obesity risk prediction, quickly finding the optimal parameters, accurately capturing the key factors affecting obesity risk, setting a division threshold to divide individual positions, generating mutant positions based on the selected individual positions, the fitness change rate, and the designed dynamic control coefficient, enabling the optimization process to prioritize parameter combinations with higher prediction capabilities, improving the efficiency of children's obesity risk prediction, completing the update according to fitness comparison and individual position detection, finding the optimal parameters, more effectively capturing the complex factor interactions in children's obesity risk prediction, being able to more accurately evaluate children's obesity risk, and providing a more reliable basis for early intervention.

[0122] Example Five. Refer to Figure 1 , this example is based on the above example. In the prediction report generation module, real-time children's sequence data is collected and processed; the real-time children's sequence data collection is to collect physiological data, diet data, exercise data, health data, and timestamps; the processing of real-time children's sequence data is to perform data normalization processing on the real-time children's sequence data; the processed real-time children's sequence data is input into the children's obesity risk prediction model constructed based on the optimal parameters for analysis, and according to the output data label, the obesity risk level corresponding to the real-time children's sequence data is obtained, and a prediction report is generated.

[0123] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0124] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0125] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. In summary, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural manners and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. A child obesity risk prediction system based on deep learning, characterized by: It includes a children's sequence data integration module, an intelligent children's obesity prediction module, a learnable parameter optimization module and a prediction report generation module; The child sequence data integration module collects and processes historical child sequence data; The intelligent child obesity prediction module constructs a graph structure between children, extracts local time features through convolution operations, and introduces position encoding and multi-head self-attention mechanisms to extract global time features, fuses local time features with global time features, and combines learnable adjacency matrices and dynamic weighted matrices to obtain spatiotemporal features for prediction; The learnable parameter optimization module sets a division threshold to divide individual positions, generates mutation positions based on the selected individual positions, fitness change rate and dynamic control coefficient, completes the update according to fitness comparison and individual position detection, and finds the optimal parameters; The prediction report generation module inputs the real-time children's sequence data into the children's obesity risk prediction model constructed based on the optimal parameters for analysis, obtains the obesity risk level corresponding to the real-time children's sequence data, and generates a prediction report; The learnable parameter optimization module is provided with a dynamic control coefficient design unit; Two initial values ​​V0 and V1 are set in advance, and the dynamic control coefficients at each iteration are calculated segmentally based on the two initial values; the formula used is as follows: ； Where V a is the dynamic control coefficient at the ath iteration, a max is the maximum number of iterations, is the position of the nth individual at the ath iteration, yes The fitness value of is the initial value of the nth individual position, yes The fitness value of , n and a are the individual position index and iteration number index respectively.

2. The child obesity risk prediction system based on deep learning according to claim 1, characterized in that: The intelligent child obesity prediction module is based on a neural network to build a child obesity risk prediction model, and is provided with a child graph structure definition unit, a time feature extraction unit, a spatiotemporal feature extraction unit and a prediction output unit, which specifically includes the following contents: Define the child graph structural unit; take the training data set as input data X, take each child in the input data X as a node, and the node attribute is the corresponding child sequence data, and obtain the node set and node attribute matrix; A similarity threshold is set in advance. If the cosine similarity between the attributes of two nodes is less than the similarity threshold, there is an edge between the corresponding two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; the edge set and adjacency matrix C are obtained; Temporal feature extraction unit; extracts temporal features from input data X based on P1 convolutional layers, where P1 is the number of convolutional layers; Spatiotemporal feature extraction unit; By analyzing the spatial dependencies between different timestamps, combining the learnable adjacency matrix and the dynamic weighting matrix, the spatiotemporal features are obtained. The formula used is as follows: ； In the formula, R p is the spatiotemporal feature extracted by the pth convolutional layer, is a row normalized matrix, , is a column-normalized matrix, , is a learnable parameterized adjacency matrix, is the dynamic weight matrix of the p-th convolutional layer, , C is the adjacency matrix, is the rowsums activation function, is the softmax activation function, T is the transpose of the matrix, H is the number of timestamps, , , and are the four weight matrices of the p-th convolutional layer at timestamp h, , , and is a learnable parameter; Prediction output unit; the spatiotemporal features R extracted by the last convolutional layer P1 Input into the fully connected layer, perform prediction operations, and output the predicted data label.

3. The child obesity risk prediction system based on deep learning according to claim 2, characterized in that: The time feature extraction unit includes the following contents: Local temporal feature extraction: Use convolution operations to capture the temporal dependencies in the data and extract local temporal features through multiple convolutional layers; Global temporal feature extraction; Includes the following: Positional embedding; Preset the cycle cycle, perform position encoding on each timestamp in the input data X, obtain the position encoding matrix WB, and then embed WB into X to obtain the embedded data ; The formula used is as follows: ； Where WB h is the position coding value corresponding to the timestamp h, WB is the position coding matrix composed of the position coding values ​​corresponding to all timestamps h, h is the timestamp index, is the cosine function; Multi-head self-attention mechanism: Capture the long-term dependencies between different timestamps in the data through multi-head self-attention processing, and use multiple attention heads to weight and aggregate the information to obtain multi-head self-attention output; Residual connection and layer normalization processing; Output the multi-head self-attention in the p-1th convolutional layer and the global temporal feature L extracted by the p-1th convolutional layer p-1 Perform residual connection and layer normalization to obtain residual features; Feedforward network; input the residual features into the feedforward network for processing; Get the global time characteristics; The output B of the feedforward network in the p-th convolutional layer p And the residual feature E extracted by the p-th convolutional layer p Perform residual connection and perform layer normalization again to obtain global temporal features; Local-global temporal feature fusion; local temporal feature S p and the global temporal feature L p The concatenation is performed, and more spatiotemporal dependencies are extracted from the concatenated features through convolution operations to obtain comprehensive features. Then, the attention mechanism is used to weight and fuse local time features, global time features, and comprehensive features to obtain fused time features.

4. The child obesity risk prediction system based on deep learning according to claim 1, characterized in that: The learnable parameter optimization module optimizes the learnable parameters of the neural network in the child obesity risk prediction model based on the swarm intelligence algorithm, and is provided with an initialization unit, an individual position division unit, a design dynamic control coefficient unit, a mutation unit, an update unit and a parameter determination unit, and specifically includes the following contents: Initialize the unit; A parameter search space is established for the learnable parameters of the neural network in the child obesity risk prediction model. N individual positions are randomly initialized in the parameter search space, and each individual position represents a set of learnable parameters. The cross entropy loss value of the child obesity risk prediction model established based on the learnable parameters for the test data set is used as the fitness value of the corresponding individual position. Individual position division unit; based on the mean μ and standard deviation δ of the fitness values ​​of all individual positions, as well as the fitness change rate, the division threshold is set , the individual position with a fitness value greater than the division threshold F is taken as the pioneer individual position, and the individual position with a fitness value less than or equal to the division threshold F is taken as the pioneer individual position; the formula used to calculate the fitness change rate is as follows: ； In the formula, is the fitness change rate at the ath iteration, n and a are the individual position index and iteration number index, respectively. and are the nth individual positions at the ath and a-1th iterations, respectively. and They are and The fitness value of Design dynamic control coefficient unit; Mutation unit; Update unit; a parameter determination unit; Pre-set the fitness threshold and detect whether there is an individual position whose fitness value is less than the fitness threshold z th , if it exists, the learnable parameters represented by the individual position are used as the optimal parameters of the neural network, and a child obesity risk prediction model is constructed based on the optimal parameters; otherwise, if the maximum number of iterations is reached, it goes to the initialization unit and re-initializes the individual position; Otherwise, increase the number of iterations by 1 and go to the individual position division unit to continue iterating.

5. The child obesity risk prediction system based on deep learning according to claim 4, characterized in that: The mutation unit randomly selects two individual positions from the pioneer individual position and the pioneer individual position, and generates mutation positions corresponding to all individual positions based on the selected individual positions, fitness change rate and dynamic control coefficient; the formula used is as follows: ； In the formula, yes The corresponding mutation position, and are two individual positions randomly selected from the pioneer individual positions, and are two individual positions randomly selected from the pioneer individual positions.

6. The child obesity risk prediction system based on deep learning according to claim 4, characterized in that: The updating unit includes the following contents: The first fitness comparison: the fitness value of the mutation position The fitness value of the original individual position For comparison, if , the original individual position is maintained constant; Otherwise, individual position detection is performed; Individual position detection; detect the original individual position Is it a pioneer individual position? If is the pioneer individual position, then use Directly replace ; Otherwise, perform a second fitness value comparison; The second fitness value comparison: the fitness value of the mutation position The fitness value of the worst pioneer individual position For comparison, if , then use Replace , and use Replace Otherwise, use Directly replace ;in, is the worst pioneer individual position at the ath iteration.

7. The child obesity risk prediction system based on deep learning according to claim 1, characterized in that: The children's sequence data integration module is used to collect and process historical children's sequence data; The collecting of historical children's sequence data is collecting physiological data, dietary data, exercise data, health data, timestamp and obesity risk level; The processing of the historical children's sequence data is to perform data cleaning, data normalization, data encoding and data set construction on the historical children's sequence data to obtain a training data set and a test data set.

8. The child obesity risk prediction system based on deep learning according to claim 1, characterized in that: The prediction report generation module collects and processes real-time child sequence data, inputs the processed real-time child sequence data into a child obesity risk prediction model constructed based on optimal parameters for analysis, obtains the obesity risk level corresponding to the real-time child sequence data based on the output data label, and generates a prediction report.

Images