A health risk prediction and early warning method for youth physical fitness data combined with deep learning

Abstract

The application discloses a kind of juvenile physical measurement data health risk prediction and early warning method combined with deep learning, it is related to medical health prediction technical field. Including have: S1: model construction: by the historical data of target individual, general deep time sequence meta-model is adjusted training, and personalized evaluation model is obtained;S2: threshold setting: according to the time sequence physical measurement sequence of the target individual, set fluctuation time sequence data, and according to the fluctuation time sequence data, determine personalized dynamic threshold value;S3: risk assessment: according to the personalized evaluation model and the latest physical measurement data of target individual, determine the latest time point prediction result of target individual, and according to the comparison result between the latest time point prediction result of the target individual and personalized dynamic threshold value, determine corresponding early warning level.The present application can accurately reflect the health status of target individual, improve the accuracy and pertinence of risk assessment.

Description

Technical Field

[0001] This invention relates to the field of medical and health prediction technology, specifically to a method for predicting and warning of health risks in adolescents based on physical fitness test data using deep learning. Background Technology

[0002] In recent years, with the rapid development of my country's economy and society and significant changes in lifestyles, the physical health of adolescents has received increasing attention from all sectors of society. However, traditional methods for assessing adolescent health often rely on static threshold judgments (such as BMI classification standards) or simple statistical analysis, making it difficult to comprehensively depict the dynamic trends and potential health risks of individual physical development. On the one hand, adolescents are in a critical stage of growth and development, and their physiological indicators exhibit high individual variability and non-linear evolution characteristics; on the other hand, a single indicator often fails to accurately reflect overall health status. For example, obesity may simultaneously affect cardiopulmonary function, exercise capacity, and mental health, and these correlations are often overlooked in traditional methods.

[0003] Although existing research has attempted to introduce machine learning models (such as support vector machines and random forests) for classification or regression analysis of physical fitness test data, these methods still face problems such as insufficient generalization ability and reliance on human experience in feature engineering when dealing with high-dimensional, time-series-dependent physical fitness test data with missing values. In addition, existing systems generally lack early identification and dynamic warning mechanisms for health risks, making it difficult to achieve the health management goal of "early detection and early intervention".

[0004] Chinese invention patent application CN118155851A discloses a data-driven employee health early warning method and system, comprising: collecting employee health data and processing and analyzing the data using big data and machine learning technologies; constructing a predictive model to predict employee health status and issue early warnings based on historical and real-time monitoring data; quantifying the degree of health risk by combining the changing trends of health indicators with historical data to determine the employee's health status level; sending personalized early warning information to employees and administrators, and recording intervention results for subsequent analysis and improvement. The advantages of this invention are: early detection of health problems, provision of personalized health management services, improvement of employee productivity and job satisfaction, and reduction of enterprise health costs.

[0005] However, the above and similar technical solutions still have the following shortcomings: the current method mainly predicts adolescent health risks by comparing individual single or static time point measurement data with uniform static group norms. However, it not only ignores the differences between individuals at the starting point of growth and development, but also fails to capture the dynamic development trend or growth trajectory of individuals. Therefore, it cannot achieve personalized dynamic risk assessment and early warning based on the longitudinal development trend of individuals, which will lead to inaccurate risk assessment and missed early warning opportunities. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting and warning of health risks in adolescents based on physical fitness test data using deep learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting and warning of health risks from adolescent physical fitness test data using deep learning, comprising:

[0008] S1: Model Construction: By adjusting and training the constructed general deep time-series meta-model using historical data of the target individual, a corresponding personalized evaluation model is obtained, including:

[0009] S1.1: Constructing a general model: By acquiring the relevant longitudinal data, the set deep temporal meta-model is trained, and the model parameters are adjusted through the backpropagation algorithm to obtain a general deep temporal meta-model;

[0010] S1.2: Data processing: Based on the complete physical test records of the target individual, obtain the time-series physical test sequence of the target individual, and combine the time-series physical test sequence with the static feature vector of the target individual to obtain the target individual data;

[0011] S1.3: Model Setup: The target individual data is divided into multiple time series data, and the time series data is used as training data for a general deep time series meta-model. By adjusting the decision layer of the general deep time series meta-model in real time, a personalized evaluation model is constructed.

[0012] S2: Threshold setting: Based on the time series physical test sequence of the target individual, set the volatility time series data, and determine the personalized dynamic threshold based on the volatility time series data;

[0013] S3: Risk Assessment: Based on the personalized assessment model and the latest physical test data of the target individual, determine the latest time point prediction result of the target individual, and determine the corresponding warning level based on the comparison result between the latest time point prediction result of the target individual and the personalized dynamic threshold.

[0014] Furthermore, data on the target individual is obtained, including:

[0015] S1.2.1: Time-series physical test sequence: Based on the historical physical test records of the target individual, core element data is obtained, including but not limited to the individual's unique identifier, timestamp, and complete indicator set. At the same time, the historical physical test records are sorted in chronological order to obtain the longitudinal data sequence of the target individual.

[0016] S1.2.2: Static Feature Vector: By setting a feature list, the gender and baseline health status of the target individual are encoded, and all data encodings are concatenated in sequence according to a predefined concatenation order to obtain the static features of the target individual.

[0017] Furthermore, the complete set of indicators includes, but is not limited to, height data, weight data, vital capacity data, sit-and-reach data, 50-meter sprint data, standing long jump data, and BMI data.

[0018] Furthermore, the static characteristics of the target individual include gender, baseline health status, and genetic background indicators.

[0019] Furthermore, based on the network hierarchy of the general deep temporal meta-model, during the training and adjustment process of the general deep temporal meta-model, the parameters in the feature extraction layer of the general deep temporal meta-model remain unchanged, while the parameters in the decision layer of the general deep temporal meta-model are adjusted in real time.

[0020] Furthermore, based on the structural framework of the general deep temporal meta-model and the target individual data, a personalized evaluation model is constructed, including:

[0021] W1: Constructing the meta-model: The input of the general deep temporal meta-model is set as a dual input interface with a dynamic data input port and a static data input port, and the dual input interface transmits data with the general deep temporal meta-model through a feedforward neural network unit to construct the corresponding general meta-model;

[0022] W2: Personalized assessment: Based on the target individual data, the time series physical test sequence and the static feature vector of the target individual are determined, and the target individual data is used as input through the dual input interface of the general meta-model to construct a personalized assessment model.

[0023] Furthermore, the feedforward neural network unit includes multiple fully connected layers, and each fully connected layer is provided with an activation function. At the same time, the feedforward neural network unit is used to perform weighted summation or weighted average calculation on the fused feature vector.

[0024] Furthermore, the dynamic data in the target individual data is converted into dynamic feature vectors through an LSTM engine, and the static data in the target individual data is converted into static feature vectors through nonlinear transformation. At the same time, the dynamic feature vectors and static feature vectors are concatenated to form a fused feature vector.

[0025] Furthermore, personalized dynamic thresholds are determined, including:

[0026] S2.1: Data Creation: The time series physical measurement sequence of the target individual is used as the input of the personalized evaluation model, and the corresponding time point prediction results are output. The time point prediction results of the target individual at each time point are compared with the actual measurement values ​​to obtain the growth residuals of the target individual at each time point. At the same time, the time series data within the set sliding time window are obtained, and all time series data are sorted and spliced ​​in chronological order to obtain the volatility time series data.

[0027] S2.2: Model prediction: The deep learning model is trained using the volatility time series data to obtain the corresponding time series prediction model. The latest physical measurement data of the target individual is used as the input of the time series prediction model, and the corresponding volatility prediction value is obtained as the output.

[0028] S2.3: Threshold Determination: The latest physical test data of the target individual is used as the input of the personalized assessment model, and the corresponding predicted value of the physical test index is obtained as the output. At the same time, the predicted value of the physical test index, the predicted value of the fluctuation, and the set sensitivity coefficient are combined to determine the personalized dynamic threshold.

[0029] Furthermore, the corresponding warning levels are determined, including:

[0030] S3.1: Data prediction: The latest physical test data of the target individual is concatenated with the target individual data in chronological order to obtain the latest target individual data. At the same time, the latest target individual data is used as the input of the personalized assessment model, and the corresponding latest physical test index prediction value is output.

[0031] S3.2: Threshold Comparison: The latest predicted value of the physical fitness index is compared with the personalized dynamic threshold range, and the corresponding alarm signal is determined based on the comparison result, specifically:

[0032] When the predicted value of the latest physical test indicator is less than the lower limit of the personalized dynamic threshold range, there is a risk of insufficiency; when the predicted value of the latest physical test indicator is within the personalized dynamic threshold range, it is considered normal; when the predicted value of the latest physical test indicator is greater than the upper limit of the personalized dynamic threshold range, there is a risk of exceeding the standard.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] Firstly, this invention sets up corresponding personalized assessment models and personalized dynamic threshold ranges based on the target individual's data, thereby more accurately reflecting the target individual's health status and improving the accuracy and pertinence of risk assessment.

[0035] Secondly, by analyzing historical data sequences, this invention can predict the physical test index values ​​at the next time point and compare the obtained physical test index values ​​with the corresponding set safety threshold range, thereby identifying adverse trends before health problems fully manifest and seizing the critical window period for health risk management.

[0036] Thirdly, this invention processes time-series data using a deep learning model, which can not only analyze changes in individual indicators, but also combine multiple indicators such as height, weight, and lung capacity, thereby enabling a more comprehensive assessment of overall health status and the discovery of potential associated risks. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the method for predicting and warning of health risks from physical fitness test data of adolescents in this invention.

[0038] Figure 2 This is a schematic diagram of the process for constructing a personalized evaluation model using time-series data in this invention;

[0039] Figure 3 This is a schematic diagram of the process for obtaining personalized dynamic thresholds in this invention;

[0040] Figure 4 This is a schematic diagram of the process for constructing a personalized evaluation model through a dual-input interface in this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Current methods primarily predict adolescent health risks by comparing individual measurements at single or static time points with uniform, static group norms. However, this approach ignores differences in the starting points of growth and development between individuals and fails to capture dynamic developmental trends or growth trajectories. Consequently, it cannot achieve personalized dynamic risk assessment and early warning based on individual longitudinal developmental trends, leading to inaccurate risk assessments and missed early warning opportunities. The technical solution presented in this application, however, generates a personalized assessment model using a constructed general deep time-series meta-model and target individual data. Simultaneously, it generates dynamically changing early warning thresholds based on the target individual data and compares the predicted values ​​of physical indicators obtained from the personalized assessment model with these dynamically changing early warning thresholds. Based on the comparison results, it determines corresponding warning levels, thereby achieving precise health risk identification with a unique threshold for each individual.

[0043] Example 1

[0044] refer to Figures 1-3 This embodiment provides a method for predicting and warning of health risks from adolescent physical fitness test data by combining deep learning. The method specifically includes the following steps:

[0045] Step S1: Model Construction. This involves adjusting and training the constructed general deep time-series meta-model based on the target individual's historical data to obtain the corresponding personalized evaluation model. Details are as follows:

[0046] Step S1.1: Construct a general model. This involves collecting multiple relevant longitudinal data points through student physical fitness reporting systems, large-scale cohort research projects (such as the "Chinese Student Physical Fitness and Health Survey"), and research projects in collaboration with schools / hospitals. These data points include, but are not limited to, individual anonymous IDs, measurement time points, and physical fitness test values. The physical fitness test values ​​include, but are not limited to, height, weight, vital capacity, sit-and-reach test, 50-meter sprint, standing long jump, and BMI. Simultaneously, the acquired longitudinal data undergoes data cleaning and alignment to obtain preprocessed longitudinal data. It is worth noting that the data cleaning and alignment methods used in this embodiment are conventional data processing techniques, and therefore are not specifically described in this embodiment.

[0047] Furthermore, based on the preprocessed relevant longitudinal data, the established deep temporal meta-model is trained. The preprocessed relevant longitudinal data serves as the input to the deep temporal meta-model, and the output is the corresponding time-point prediction result. Specifically, during the training process of the deep temporal meta-model, the parameters in the deep temporal meta-model are adjusted using the backpropagation algorithm to minimize the difference between the obtained time-point prediction result and the actual time-point result, thereby obtaining the corresponding general deep temporal meta-model.

[0048] Step S1.2: Data Processing. This involves identifying the complete physical examination record of the target individual using their physical examination profile, and then extracting the corresponding target data sample, i.e., the time-series physical examination sequence, from this record. Simultaneously, the obtained time-series physical examination sequence is combined with the static feature vector corresponding to the target individual and converted into a format identical to the preprocessed longitudinal data obtained in Step S1.1, which is the corresponding target individual data. Specifically:

[0049] Step S1.2.1: Time-series physical test sequence. This involves extracting corresponding historical physical test records from the target individual's school physical test database, regional student health platform, or personal electronic health record through a secure interface such as an API or authorized data export function. Simultaneously, from the extracted historical physical test records, the core data elements of each record are obtained, including but not limited to the individual's unique identifier, timestamp, and complete indicator set. This complete indicator set is the physical test indicator, including height data, weight data, vital capacity data, sit-and-reach data, 50-meter sprint data, standing long jump data, and BMI data.

[0050] Furthermore, based on the timestamp corresponding to each physical examination record, they are sorted chronologically to obtain the longitudinal data sequence of the target individual. Simultaneously, outlier processing is performed on the obtained longitudinal data sequence of the target individual to remove abnormal data. It is worth noting that the outlier processing in this embodiment is a conventional data processing method, namely, verifying or smoothing or removing abnormal data in the longitudinal data sequence of the target individual according to preset rules; therefore, it is not specifically described in this embodiment.

[0051] Step S1.2.2: Static Feature Vector. This involves obtaining the static features of the target individual based on their gender and baseline health status. It's worth noting that the obtained static features also include genetic background indicators. Specifically, through a set feature list, the target individual's gender, date of birth, father's height, mother's height, baseline BMI classification, and history of major illnesses are defined, and corresponding data codes are determined. For example, the gender code for female is 0, and the gender code for male is 1.

[0052] Furthermore, based on the data encoding corresponding to each feature data and the predefined concatenation order, all data encodings are concatenated in sequence to obtain the corresponding concatenated feature vector, which is the corresponding static feature vector.

[0053] Step S1.3: Model Setup. Based on the target individual data obtained in Step S1.2 and the general deep temporal meta-model constructed in Step S1.1, the corresponding personalized evaluation model is determined. Specifically, based on the architecture design of the general deep temporal meta-model, the corresponding network hierarchy is determined, including but not limited to layer types and order, the number of layer parameters, and layer function descriptions. Simultaneously, based on the determined network hierarchy, the bottom layer of the general deep temporal meta-model, i.e., the feature extraction layer, is frozen, ensuring that the parameters in the feature extraction layer remain unchanged. At the same time, the top layer of the general deep temporal meta-model, i.e., the decision layer, is unfrozen, meaning that the parameters in the decision layer are adjusted in real time based on the target individual data.

[0054] In this embodiment, the target individual data is divided into multiple time-series data based on the time points of the target individual data. It is worth noting that the number of time-series data divisions in this embodiment can be specifically set according to actual needs, so it is not specifically described in this embodiment. Simultaneously, based on the multiple time-series data divisions, each time-series data is used as training data for a general deep temporal meta-model (i.e., the feature extraction layer is frozen, and the decision layer is unfrozen). Through iterative training of the general deep temporal meta-model, the final model parameters of the general deep temporal meta-model are determined. In other words, based on the general deep temporal meta-model with its final model parameters, the corresponding personalized evaluation model is obtained.

[0055] Step S2: Threshold Setting. This involves setting volatility time series data based on the time series data obtained in step S1.2.1, and determining the corresponding personalized dynamic threshold based on the set volatility time series data. Details are as follows:

[0056] Step S2.1: Data Creation. The time-series physical measurement sequence obtained in Step S1.2.1 is used as input to the personalized assessment model constructed in Step S1.3. The output is the prediction result at the corresponding time point, which is the predicted measurement value of the target individual at the corresponding time. Simultaneously, the predicted measurement value of the target individual at the corresponding time point is compared with the actual measurement value to obtain the measurement difference of the target individual at each time point, which is the corresponding growth residual.

[0057] Furthermore, based on the growth residuals of the target individual at various time points, a sliding time window is used to divide all the growth residuals of the target individual. Then, based on all the growth residuals within the sliding window, the standard deviation of all growth residuals within the sliding window is determined, which is the time series data for the corresponding sliding window. In other words, based on the time series data corresponding to each sliding window, the volatility time series data of the target individual is obtained.

[0058] Step S2.2: Model Prediction. Based on the volatility time series data obtained in Step S2.1, the constructed deep learning model, such as a lightweight time series prediction model, is trained to obtain the corresponding time series prediction model. Simultaneously, based on the latest physical examination data of the target individual, the corresponding latest time series data is determined, and this determined latest time series data is used as the input to the time series prediction model, outputting the corresponding volatility prediction value.

[0059] Step S2.3: Threshold Determination. This involves combining the predicted fluctuation value obtained in step S2.2 with the set sensitivity coefficient to obtain the corresponding predicted fluctuation value. It is worth noting that the sensitivity coefficient in this embodiment can be specifically set according to actual needs; therefore, it is not specifically described in this embodiment.

[0060] Furthermore, the latest physical fitness test data of the target individual is used as input to the personalized assessment model constructed in step S1.3, and the output is the predicted result of the corresponding target individual at the corresponding time point, which is the predicted value of the corresponding physical fitness test index. Simultaneously, the obtained predicted value of the physical fitness test index is combined with the obtained predicted fluctuation value to determine the corresponding personalized dynamic threshold, specifically:

[0061]

[0062] in: For personalized dynamic thresholds, These are predicted values ​​for physical fitness test indicators. This is the sensitivity coefficient. This is the predicted value for fluctuations.

[0063] Step S3: Risk Assessment. This involves obtaining the latest time-point prediction results for the target individual based on their latest physical examination data and the personalized assessment model constructed in Step S1.3. Simultaneously, these latest time-point prediction results are compared with the personalized dynamic thresholds obtained in Step S2.3, and the corresponding warning level is determined based on the comparison results. Details are as follows:

[0064] Step S3.1: Data Prediction. This involves obtaining the latest physical examination data of the target individual through API interfaces, file uploads, or database monitoring. Simultaneously, this latest physical examination data is concatenated with the target individual data obtained in Step 1.2 in chronological order to update the time-series data of the target individual. In other words, by updating the time-series data of the target individual, the corresponding latest target individual data is determined.

[0065] Furthermore, the latest target individual data obtained is used as the input to the personalized assessment model constructed in step S1.3, and the output is the time point prediction result of the corresponding target individual, which is the corresponding latest physical test index prediction value.

[0066] Step S3.2: Threshold Comparison. This involves comparing the latest predicted physical fitness index value obtained in Step S3.1 with the personalized dynamic threshold obtained in Step S2.3, and determining the corresponding alarm signal based on the comparison result. It is worth noting that, according to the formula for obtaining the personalized dynamic threshold in Step S2.3, the personalized dynamic threshold in this embodiment is set as a threshold range, which includes an upper limit threshold and a lower limit threshold of the personalized dynamic threshold range.

[0067] Furthermore, the latest predicted values ​​of physical fitness indicators are compared with personalized dynamic threshold ranges, and based on the comparison results, corresponding alarm signals are determined, specifically as follows:

[0068] When the predicted value of the latest physical test result is lower than the lower limit of the personalized dynamic threshold range, there is a risk of deficiency, such as malnutrition and growth retardation. When the predicted value of the latest physical test result is within the personalized dynamic threshold range, it is considered normal and no warning is needed. When the predicted value of the latest physical test result is higher than the upper limit of the personalized dynamic threshold range, there is a risk of exceeding the standard, such as obesity.

[0069] Example 2

[0070] This embodiment provides a method for predicting and warning of health risks in adolescent physical fitness test data using deep learning. The specific implementation method is the same as in Embodiment 1, except that the general deep time-series meta-model obtained in step S1.1 is designed with dual input channels to construct the corresponding general meta-model. Simultaneously, the target individual data obtained in step S1.2 is combined with the constructed general meta-model to construct the corresponding personalized assessment model. The invention will be illustrated below with specific examples of this embodiment.

[0071] refer to Figure 4 In this embodiment, a corresponding personalized evaluation model is constructed using the general meta-model and the obtained target individual data, as follows:

[0072] Step W1: Construct the meta-model. This involves setting the input ports of the general deep time-series meta-model obtained in step S1.1 to dual-input interfaces, including a dynamic data input port and a static data input port. Simultaneously, dynamic data is transmitted to the general deep time-series meta-model through the dynamic data input port, and static data is transmitted to the general deep time-series meta-model through the static data input port. The dynamic data is then processed by the LSTM engine configured within the general deep time-series meta-model.

[0073] Furthermore, the LSTM engine converts dynamic data into dynamic feature vectors and performs a non-linear transformation on static data to obtain corresponding static feature vectors. Simultaneously, the dynamic and static feature vectors are concatenated to form a fused feature vector.

[0074] Specifically, the general deep temporal meta-model in this embodiment includes a feedforward neural network unit, which comprises multiple fully connected layers, each with an activation function such as the ReLU function. That is, dynamic and static data are input through dynamic and static data input ports, and the merged fusion feature vector is transmitted to the general deep temporal meta-model via the feedforward neural network unit, outputting the corresponding time-point prediction result. It is worth noting that the feedforward neural network unit performs weighted summation or weighted averaging on the fusion feature vector; that is, it adds or averages the dynamic and static feature vectors element-wise.

[0075] Step W2: Personalized Assessment. Based on the target individual data obtained in Step S1.2, the temporal biometric sequence (i.e., dynamic feature data) and static feature vector of the target individual are determined. The obtained temporal biometric sequence is transmitted to the general deep temporal meta-model through the dynamic data input port, and the obtained static feature vector is transmitted to the general deep temporal meta-model through the static data input port. Simultaneously, in the general deep temporal meta-model, the dynamic feature data and static feature vector are concatenated and fused to obtain the corresponding target individual fused feature vector.

[0076] Furthermore, the obtained target individual fused feature vector is transmitted to a general deep temporal meta-model through a feedforward neural network unit, outputting the corresponding time-point prediction result for the target individual. In other words, the general deep temporal meta-model constructed using the target individual fused feature vector is the personalized evaluation model.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for predicting and warning health risks from adolescent physical fitness test data using deep learning, characterized in that, Including: S1: Model Construction: By adjusting and training the constructed general deep time-series meta-model using historical data of the target individual, a corresponding personalized evaluation model is obtained, including: S1.1: Constructing a general model: By acquiring the relevant longitudinal data, the set deep temporal meta-model is trained, and the model parameters are adjusted through the backpropagation algorithm to obtain a general deep temporal meta-model; S1.2: Data Processing: Based on the complete physical examination records of the target individual, obtain the time-series physical examination sequence of the target individual, and combine the time-series physical examination sequence with the static feature vector of the target individual to obtain the target individual data, including: S1.2.1: Time-series physical test sequence: Based on the historical physical test records of the target individual, core element data is obtained, including but not limited to the individual's unique identifier, timestamp, and complete indicator set. At the same time, the historical physical test records are sorted in chronological order to obtain the longitudinal data sequence of the target individual. S1.2.2: Static Feature Vector: By setting the feature list, the gender and baseline health status of the target individual are encoded, and all data encodings are concatenated in sequence according to the predefined concatenation order to obtain the static features of the target individual; S1.3: Model Setup: The target individual data is divided into multiple time-series data, and the time-series data is used as training data for a general deep time-series meta-model. By adjusting the decision layer of the general deep time-series meta-model in real time, a personalized evaluation model is constructed, including: W1: Constructing the meta-model: The input of the general deep temporal meta-model is set as a dual input interface with a dynamic data input port and a static data input port, and the dual input interface transmits data with the general deep temporal meta-model through a feedforward neural network unit to construct the corresponding general meta-model; The feedforward neural network unit includes multiple fully connected layers, and each fully connected layer is provided with an activation function. The feedforward neural network unit is used to perform weighted summation or weighted average calculation on the fused feature vectors. W2: Personalized assessment: Based on the target individual data, determine the target individual's time-series physical test sequence and the target individual's static feature vector, and construct a personalized assessment model by taking the target individual data as input through the dual-input interface of the general meta-model; The dynamic data in the target individual data is converted into dynamic feature vectors through an LSTM engine, and the static data in the target individual data is converted into static feature vectors through nonlinear transformation. At the same time, the dynamic feature vectors and static feature vectors are concatenated to form a fused feature vector. S2: Threshold setting: Based on the time series physical test sequence of the target individual, set the volatility time series data, and determine the personalized dynamic threshold based on the volatility time series data; S3: Risk Assessment: Based on the personalized assessment model and the latest physical test data of the target individual, determine the latest time point prediction result of the target individual, and determine the corresponding warning level based on the comparison result between the latest time point prediction result of the target individual and the personalized dynamic threshold.

2. The method for predicting and warning health risks of adolescents based on physical fitness test data using deep learning, as described in claim 1, is characterized in that... The complete set of indicators includes, but is not limited to, height data, weight data, vital capacity data, sit-and-reach data, 50-meter sprint data, standing long jump data, and BMI data.

3. The method for predicting and warning health risks of adolescents based on physical fitness test data combined with deep learning, as described in claim 1, is characterized in that... The static characteristics of the target individuals include gender, baseline health status, and genetic background indicators.

4. The method for predicting and warning health risks of adolescents based on physical fitness test data combined with deep learning, as described in claim 1, is characterized in that... Based on the network hierarchy of the general deep temporal meta-model, during the training and adjustment process of the general deep temporal meta-model, the parameters in the feature extraction layer of the general deep temporal meta-model remain unchanged, while the parameters in the decision layer of the general deep temporal meta-model are adjusted in real time.

5. The method for predicting and warning health risks of adolescents based on physical fitness test data combined with deep learning, as described in claim 1, is characterized in that... The personalized dynamic thresholds are determined, including: S2.1: Data Creation: The time series physical measurement sequence of the target individual is used as the input of the personalized evaluation model, and the corresponding time point prediction results are output. The time point prediction results of the target individual at each time point are compared with the actual measurement values ​​to obtain the growth residuals of the target individual at each time point. At the same time, the time series data within the set sliding time window are obtained, and all time series data are sorted and spliced ​​in chronological order to obtain the volatility time series data. S2.2: Model prediction: The deep learning model is trained using the volatility time series data to obtain the corresponding time series prediction model. The latest physical measurement data of the target individual is used as the input of the time series prediction model, and the corresponding volatility prediction value is obtained as the output. S2.3: Threshold Determination: The latest physical test data of the target individual is used as the input of the personalized assessment model, and the corresponding predicted value of the physical test index is obtained as the output. At the same time, the predicted value of the physical test index, the predicted value of the fluctuation, and the set sensitivity coefficient are combined to determine the personalized dynamic threshold.

6. The method for predicting and warning health risks of adolescents based on physical fitness test data combined with deep learning, as described in claim 1, is characterized in that... The corresponding warning levels are determined, including: S3.1: Data prediction: The latest physical test data of the target individual is concatenated with the target individual data in chronological order to obtain the latest target individual data. At the same time, the latest target individual data is used as the input of the personalized assessment model, and the corresponding latest physical test index prediction value is output. S3.2: Threshold Comparison: The latest predicted value of the physical fitness index is compared with the personalized dynamic threshold range, and the corresponding alarm signal is determined based on the comparison result, specifically: When the predicted value of the latest physical test indicator is less than the lower limit of the personalized dynamic threshold range, there is a risk of insufficiency; when the predicted value of the latest physical test indicator is within the personalized dynamic threshold range, it is considered normal; when the predicted value of the latest physical test indicator is greater than the upper limit of the personalized dynamic threshold range, there is a risk of exceeding the standard.

Images