Child health monitoring method based on physical sign data

Abstract

The invention provides a child health monitoring method based on physical sign data, and relates to the field of child health monitoring. The child health monitoring process comprises the steps of constructing a child sign data set, mapping a data sequence at each moment in the data set to a high-dimensional space through a lookup table, extracting a feature sequence in combination with linear transformation and a ReLU activation function, and carrying out non-linear adjustment on a generated weight and smooth coefficient matrix to obtain a child health monitoring result. The method comprises the following steps: realizing dynamic weighting and smoothing processing of a feature sequence, obtaining a weighted feature sequence, introducing a growth cycle factor, obtaining a physiological feature sequence, capturing a time sequence feature sequence containing historical health information by using a long-short-term memory network, calculating the difference between the physiological feature sequences at adjacent moments, and calculating feature reconstruction parameters based on the difference value. And finally, the health feature sequence is flattened and then input into the full-connection neural network, a child health monitoring value is output through an activation function and linear transformation, and the accuracy of child health monitoring is improved.

Description

Technical Field

[0001] The present invention belongs to the field of child health monitoring, and in particular relates to a child health monitoring method based on vital sign data. Background Art

[0002] Child health monitoring refers to the real-time or regular collection of multi-dimensional physiological, psychological and other physical data of children, and the use of modern technical means to analyze and evaluate them in order to provide personalized health management recommendations. With the development of social economy and the increasing number of children's health problems, traditional health management methods have been unable to meet the needs of accurate monitoring of children's health. Therefore, the establishment of an accurate and real-time child health monitoring system can help parents and medical staff understand the health status of children in real time, promptly identify potential health risks, and provide an important basis for early intervention and health management of children.

[0003] Existing child health monitoring methods mainly rely on traditional physical examinations and manual analysis, which have problems such as delayed data collection, long monitoring cycles, and insufficient consideration of individual differences. Traditional physical examinations can often only provide children's health information at a certain moment, making it difficult to achieve dynamic tracking and timely response to health status. In addition, the data processing methods in existing technologies mostly rely on simple physiological indicator monitoring, fail to fully utilize multi-dimensional physical sign data, and lack accurate modeling of children's personalized health needs, resulting in insufficient accuracy and timeliness of monitoring results.

[0004] Changes in children's health status are driven by multidimensional vital sign data. These data not only exhibit complex nonlinear fluctuation characteristics in the time dimension, but also present multi-scale and multi-stage dynamic evolution patterns. The use of child health monitoring methods that can integrate multidimensional vital sign data and historical characteristics can systematically depict the long-term trends of children's overall health, thereby achieving accurate monitoring of children's health. Summary of the Invention

[0005] The present invention provides a child health monitoring method based on vital sign data. All vital sign data of the child at each moment are organized into a data sequence, an implicit feature sequence is obtained through lookup table mapping, and a feature sequence is obtained through linear transformation and ReLU activation function. The weight of the current moment is generated by nonlinear adjustment, and a weighted feature sequence is generated in combination with a smoothing coefficient matrix. A growth cycle factor is further introduced to obtain a physiological feature sequence. A time series feature sequence is obtained through a long short-term memory network. The difference between the physiological feature sequence at the current moment and the previous moment is used to generate a feature reconstruction parameter, and adaptive fusion of the physiological and time series feature sequences is achieved to form a health feature sequence. The health feature sequence is flattened and input into a fully connected neural network and an activation function. The final child health monitoring value is output through linear transformation, thereby achieving comprehensive and accurate monitoring of child health.

[0006] The technical method adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps: S1. Collect data on physical signs that affect children's health, preprocess them using the Z-score method, and construct a children's physical sign dataset; S2. All data at each moment in the child vital sign dataset are combined into a data sequence, and the data sequence is mapped to a high-dimensional space using a lookup table to obtain an implicit feature sequence reflecting the child's health status. The feature sequence is obtained through linear transformation and ReLU activation function; S3. Construct an adaptive time series weight module, generate the weight of the current moment based on the feature sequence and the weight of the previous moment using nonlinear adjustment, and calculate the smoothing coefficient matrix at the same time, dynamically weight and smooth the feature sequence to obtain a weighted feature sequence; S4. Constructing a growth cycle feature module, introducing a growth cycle factor, and multiplying the growth cycle factor element by element with the weighted feature sequence to obtain a physiological feature sequence. The physiological feature sequence constitutes time series data, and a long short-term memory network is used to obtain a temporal feature sequence. S5. Construct a health status fluctuation perception module, calculate the difference between the physiological feature sequence at the current moment and the previous moment, obtain the health status fluctuation value, introduce a feature reconstruction parameter, and use the parameter to weightedly fuse the temporal feature sequence and the physiological feature sequence to form a health feature sequence; S6. Flatten the health feature sequence into a feature vector, input the vector into a fully connected neural network and an activation function in sequence, and then obtain the final child health monitoring value through linear transformation.

[0007] Preferably, in S1, vital sign data affecting children's health are collected, including weight, height, body temperature, heart rate, blood oxygen saturation, ambient temperature, number of steps, and sleep duration. For each type of data, the Z-score method is used to detect outliers. Any data value with a Z-score value greater than 3 or less than -3 is judged to be an outlier. For the detected outliers, the mean replacement method is used for processing, that is, the outlier is replaced with the mean of the data of the three days before and after it, so as to ensure the accuracy and consistency of the data and construct a children's vital sign data set.

[0008] Preferably, children's vital signs data are often a simple stack of multiple types of numerical values, which makes it difficult to reveal the complex relationships between different data and cannot fully reflect the potential changes in health status. The present invention automatically maps all children's vital signs data to a high-dimensional space through a lookup table, thereby enriching the semantic expression and internal connection of each data, enabling the model to better understand the deep-level correlation between the data. On this basis, linear transformation and nonlinear activation mechanisms are further introduced to reorganize and enhance these high-dimensional implicit features, thereby solving the problem that the original children's vital signs data has weak expression ability and health information is difficult to capture.

[0009] Preferably, in said S2, all data at each moment in said child vital signs data set constitutes a data sequence for model input. A series of vital signs data that constantly influences children's health ,Will Mapping to a richer feature sequence includes the following steps: S21, the vital signs data sequence By mapping the lookup table to a high-dimensional space, a high-dimensional implicit feature sequence is obtained, which enables the model to automatically extract potential health information. The specific mathematical model is: ; Where, For the The implicit feature sequence that affects children's health data at all times, is the lookup table function; S22, using linear transformation to transform implicit feature sequence Mapping to the physiological feature space enables the output features to better fit the actual physiological patterns of children's health, and enhances the model's ability to express nonlinear features through the ReLU activation function. The specific mathematical model is: ; Where, After transformation The characteristic sequence that affects children's health data at all times, is a learnable weight matrix that maps implicit feature sequences to physiological feature space, is a learnable bias term.

[0010] Preferably, the present invention can effectively mine the potential health information in the data by mapping the children's vital sign data sequence into high-dimensional implicit features. The lookup table mapping method introduces a more discriminative feature expression capability into the model, which is conducive to revealing the complex intrinsic relationship between the original children's vital sign data. It uses linear transformation and ReLU activation function to automatically extract and reorganize high-dimensional implicit features, thereby improving the model's ability to represent nonlinear health information.

[0011] Preferably, the children's vital sign data sequence is highly dynamic and noisy. Traditional methods usually use fixed weights or simple sliding averages to process the feature sequence, which makes it difficult to take into account both the historical trend of health status and sudden anomalies. The present invention designs an adaptive time series weight module, which combines the weight of the previous moment and the information of the feature sequence at the current moment to generate the weight of the current moment, so that the weight can not only retain historical information, but also be adjusted in time according to the current data mutation. Secondly, the smoothing coefficient matrix is ​​calculated to suppress high-frequency fluctuations in real time and reduce the interference of short-term noise on the overall health trend judgment. Finally, the feature sequence is weighted and corrected under the joint action of dynamic weights and smoothing coefficient matrices, which not only fully retains important time series dynamic information, but also prevents false alarms and misjudgments, so that the model can more accurately portray the true health status of children.

[0012] Preferably, in S3, constructing an adaptive timing weight module specifically includes the following steps: S31, Comprehensive The weight of the moment and the The information of the moment feature sequence is combined with nonlinear regulation to control the rate and amplitude of weight change. The specific mathematical model is: ; Where, For the The weight of the moment feature sequence, For the The weight of the moment feature sequence, is a learnable step size parameter for adjusting the weight update amplitude, is a learnable weight suppression parameter that balances the influence of the current feature and the previous weight, It is a learnable nonlinear modulation parameter that adjusts the sensitivity of the weight to feature changes; S32. Calculate the smoothing coefficient matrix and automatically adjust the size of the smoothing coefficient matrix according to the change amplitude of the characteristic sequence at adjacent moments to correct the rate of change of the child health data characteristics and suppress the interference of noise on the fluctuation of the characteristic sequence. The specific mathematical model is: ; Where, For the The smoothing coefficient matrix at time , For the The characteristic sequence that affects children's health data at all times, is a learnable feature change sensitive parameter, To take the absolute value operation; S33, the Moment smoothing coefficient matrix , weight With the Characteristic sequence of children's health data at all times Multiply them together to get the weighted feature sequence. The specific mathematical model is: ; Where, For the The weighted feature sequence at each moment.

[0013] Preferably, the present invention introduces dynamic weights and smoothing coefficient matrices to perceive and adjust changes in vital sign data in real time. The weight update mathematical model integrates historical weights and current feature sequences, and introduces nonlinear modulation to enhance the model's adaptability to mutations, anomalies, and trend changes. The adaptive adjustment of the smoothing coefficient matrix further suppresses the interference of noise on feature fluctuations, preventing false alarms and misjudgments. After dynamic weighting and smoothing, the feature sequence can more accurately reflect the changing trend of the child's true health status.

[0014] Preferably, traditional linear feature processing methods are difficult to fully reflect the potential physiological laws and periodic disturbances in children's growth and development, resulting in limited model recognition capabilities for changes in children's health status and periodic risk signals during their growth stages. To address this problem, the present invention generates growth cycle factors by combining tanh and sine functions to simulate the complex and changeable health evolution trends in children's growth process. The growth cycle modulation factors are combined element by element with the original feature sequence, so that the model can dynamically weight different features according to actual growth laws, highlight signals with strong periodic correlation and significant changes, and significantly suppress the influence of occasional noise. The long short-term memory network is further used to integrate health information at the current moment and multiple historical steps, effectively capturing long-term dependencies, and improving the model's sensitivity and recognition accuracy to dynamic changes in children's health status.

[0015] Preferably, in S4, constructing a growth cycle feature module includes the following steps: S41, generate the first Growth cycle factors at each moment , which is used to simulate the nonlinear physiological changes and periodic disturbances of children's growth process. The specific mathematical model is: ; Where, For the The growth cycle factor of the moment, is the learnable weight matrix, is a learnable bias term, is the learnable periodic frequency, is the learnable phase, To expand the data dimension through the broadcast mechanism; S42, the growth cycle modulation factor and characteristic sequences of child health data The combination is done in an element-by-element multiplication manner to highlight the characteristics of strong periodic correlation and significant changes. The specific mathematical model is: ; Where, is the physiological characteristic sequence after periodic modulation, is element-wise multiplication; S43, will Time to The physiological feature sequence of the moment is combined to form a length of Time series data , , using long short-term memory networks to capture long-term dependencies, the specific mathematical model is: ; Where, To utilize the time series data output by the long short-term memory network, ,in For the The time series feature sequence containing historical information output by the long short-term memory network, Long short-term memory network.

[0016] Preferably, the joint modeling of tanh and sine function modulation can simultaneously express the nonlinear physiological changes and periodic changes in the growth process, which helps to characterize the inherent evolution law in children's physical sign data, the growth cycle modulation factor and the original feature sequence By combining each element, the feature expression weight can be dynamically adjusted to highlight important signals with periodic correlations and suppress noise interference, so that the model can obtain dynamic feature inputs that are more in line with physiological reality. The time series features are further modeled through the long short-term memory network, which not only captures the current state information, but also retains the historical time series dependency, thereby enhancing the timeliness and accuracy of health status judgment.

[0017] Preferably, children's health status is often affected by both environmental and physiological changes, resulting in sudden fluctuations or abnormal changes in health characteristics. Traditional feature processing methods often ignore these fluctuation information, which easily causes the model to be slow to respond to abnormal signals and insensitive to short-term drastic changes. The present invention measures the difference in physiological characteristic sequences between two consecutive moments to timely capture the dynamic fluctuation degree of health status, and further uses nonlinear transformation to generate feature reconstruction parameters, and adaptively adjusts the model's attention weights on current physiological characteristic information and historical time series characteristic information according to the fluctuation intensity. While automatically sensing and adjusting health characteristics, it effectively improves the response sensitivity and discrimination accuracy to abnormal fluctuations and sudden health risks.

[0018] Preferably, in S5, building a health status fluctuation perception module specifically includes the following steps: S51, calculate the The difference between the time series feature sequence at a certain moment and the time series feature sequence at the previous moment captures the dynamic fluctuation value of the health status. The specific mathematical model is: ; Where, For the The fluctuation value of health status at all times, is the L1 norm, For the The physiological sequence of moments; S52. Generate feature reconstruction parameters using nonlinear transformation to adjust the attention weights of the physiological feature sequence and the temporal feature sequence containing historical information. The specific mathematical model is: ; Where, For the The feature reconstruction parameters at time , are the learnable weights, is a learnable bias term; S53, the Time series feature sequence and physiological characteristic sequences Weighted fusion is performed according to feature reconstruction parameters. The specific mathematical model is: ; Where, is the health feature sequence after weighted fusion.

[0019] Preferably, the L1 norm of the time series feature sequence difference is used to accurately quantify the fluctuation amplitude of the current health status, providing sensitivity support for abnormal changes or trend inflection points. By introducing feature reconstruction parameters adaptively generated based on the fluctuation amplitude, the model's attention weights on the time series and physiological feature sequences can be dynamically adjusted to achieve adaptive balance for sudden abnormalities or stable stages. Finally, the time series features and physiological feature sequences are weightedly fused according to the feature reconstruction parameters to improve the robustness and sensitivity of the model under changes in health status.

[0020] Preferably, in the child health monitoring method, the ultimate goal is to accurately and quantitatively evaluate the child's health status. To this end, the present invention flattens the health feature sequence into a feature vector, obtains the output vector through a combination of a fully connected network and a ReLU activation function, enhances the model's ability to express complex health patterns, and then further compresses the output vector and maps it into a single health monitoring value through another linear transformation, thereby realizing the quantification of the child's health monitoring value.

[0021] Preferably, in said S6, the health feature sequence Concatenate by the last dimension and flatten into a feature vector , the feature vector is sequentially input into the fully connected layer and ReLU activation to obtain the output vector, and then the health monitoring value is obtained through linear transformation. The specific mathematical model is: ; ; Where, is the output vector, 、 is the learnable weight matrix, 、 is a learnable bias term, The final output value of child health monitoring.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention organizes all data at each moment in the child's vital signs data into a data sequence, utilizes lookup table mapping, linear transformation, and activation function to realize efficient extraction of high-dimensional feature sequences, combines adaptive weights and smoothing coefficient matrices, dynamically weights and smoothes feature sequences, improves the model's adaptability to mutations, anomalies, and trend changes, introduces growth cycle factors to obtain physiological feature sequences, further utilizes long-term and short-term memory networks to fully explore the historical evolution laws of child's vital signs data, obtains time series feature sequences, and obtains health feature sequences by weighted fusion of time series features and physiological feature sequences through feature reconstruction parameters. Finally, the health feature sequences are flattened and input into the fully connected neural network, activation function, and linear transformation to obtain more comprehensive and accurate child health monitoring values. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a step-by-step diagram of a child health monitoring method based on vital sign data.

[0024] Figure 2 This is the module diagram of adaptive timing weight.

[0025] Figure 3 It is the growth cycle characteristic module diagram.

[0026] Figure 4 This is the module diagram for health status fluctuation perception.

[0027] Figure 5 A trend chart showing the changes in training loss values ​​for child health monitoring methods.

[0028] Figure 6 This is the effect diagram of children's health monitoring. DETAILED DESCRIPTION

[0029] The present invention proposes a child health monitoring method based on vital sign data, which composes all data at each moment in the child's vital sign data into a data sequence, uses a lookup table mapping to convert the data sequence into a high-dimensional implicit feature sequence, further extracts the feature sequence through linear transformation and ReLU activation function, adopts adaptive time series weight and smoothing adjustment mechanism, dynamically adjusts the feature sequence to obtain a weighted feature sequence, introduces a growth cycle factor, and multiplies it element by element with the weighted feature sequence to obtain a physiological feature sequence, combines the long short-term memory network to capture the time series features, obtains the time series feature sequence, calculates the difference between the physiological feature sequence at the current moment and the previous moment, introduces feature reconstruction parameters, adaptively weights and fuses the time series features with the physiological features to form a health feature sequence, finally, flattens the health feature sequence and inputs it into a fully connected neural network and activation function, and outputs the child health monitoring value through linear transformation, thereby realizing comprehensive and accurate monitoring of child health. The technical solution in the embodiment of the present invention is described in detail and completely below, specifically including the following steps: Figure 1 shown.

[0030] S1. Collect data on vital signs that affect children's health, preprocess them using the Z-score method, and construct a children's vital sign dataset.

[0031] Furthermore, in S1, data affecting children's health are collected for 60 consecutive days, including weight, height, body temperature, heart rate, blood oxygen saturation, ambient temperature, number of steps, and sleep duration. For each type of data, the Z-score method is used to detect outliers. Any data value with a Z-score value greater than 3 or less than -3 is judged as an outlier. For the detected outliers, the mean replacement method is used for processing, that is, the outliers are replaced with the mean of the data of the three days before and after them, and a children's physical sign data set is constructed. The data set is divided into a training set and a validation set in a ratio of 7:3.

[0032] S2. All data at each moment in the child vital sign dataset are organized into a data sequence, and a lookup table is used to map the data sequence to a high-dimensional space to obtain an implicit feature sequence reflecting the child's health status. The feature sequence is obtained through linear transformation and ReLU activation function.

[0033] Furthermore, in said S2, all the data at each moment in said child vital signs data set constitutes the data sequence of the model input. A series of vital signs data that constantly influences children's health , ,in For the The collection value of children's physical sign data, The number of data types for children's physical signs, , the vital sign data sequence Mapping to a richer feature sequence includes the following steps: S21, the vital signs data sequence By mapping the lookup table to a high-dimensional space, a high-dimensional implicit feature sequence is obtained, which enables the model to automatically extract potential health information. The specific mathematical model is: ; Where, For the The implicit feature sequence that affects children's health data at all times, The dimensions obtained by mapping the lookup table, is the lookup table function; S22, using linear transformation to transform implicit feature sequence Mapped to the physiological feature space, and the ReLU activation function is used to enhance the model's ability to express nonlinear features. The specific mathematical model is: ; Where, After transformation The characteristic sequence that affects children's health data at all times, is a learnable weight matrix that maps implicit feature sequences to physiological feature space, is a learnable bias term.

[0034] S3. Construct an adaptive time series weight module, generate the weight of the current moment based on the feature sequence and the weight of the previous moment using nonlinear adjustment, and calculate the smoothing coefficient matrix at the same time, dynamically weight and smooth the feature sequence to obtain a weighted feature sequence.

[0035] Furthermore, in said S3, an adaptive timing weight module is constructed, such as Figure 2 As shown, the specific steps include: S31, Comprehensive The weight of the moment and the The information of the moment feature sequence is combined with nonlinear regulation to control the rate and amplitude of weight change. The specific mathematical model is: ; Where, For the The weight of the moment feature sequence, For the The weight of the feature sequence at the moment, the weight at the initial moment is initialized to a full 1 matrix, is a learnable step size parameter for adjusting the weight update amplitude. The initial value is set to 0.05 during the implementation. is a learnable weight suppression parameter, and its initial value is set to 0.05 during implementation. is a learnable nonlinear modulation parameter, and its initial value is set to 1 during implementation; S32. Introduce a smoothing coefficient matrix and automatically adjust the size of the smoothing coefficient matrix according to the variation of the characteristic sequence at adjacent moments. The specific mathematical model is: ; Where, For the The smoothing coefficient matrix at time , For the The characteristic sequence that affects children's health data at all times, It is a learnable feature change sensitive parameter, and its initial value is set to 0.1 during implementation. To take the absolute value operation; S33, the Moment smoothing coefficient matrix , weight With the Characteristic sequence of children's health data at all times Multiply them together to get the weighted feature sequence. The specific mathematical model is: ; Where, For the The weighted feature sequence at each moment.

[0036] S4. Construct a growth cycle feature module, introduce a growth cycle factor, and multiply it element by element with the weighted feature sequence to obtain a physiological feature sequence. The physiological feature sequence constitutes time series data, and a long short-term memory network is used to obtain a temporal feature sequence.

[0037] Furthermore, in said S4, a growth cycle feature module is constructed, such as Figure 3 As shown, the following steps are included: S41, generate the first Growth cycle factors at each moment , the specific mathematical model is: ; Where, For the The growth cycle factor of the moment, is the learnable weight matrix, is a learnable bias term, is the learnable cycle frequency, and the initial value is set to 0.4 during the implementation process. is a learnable phase, and its initial value is set to 0.4 during implementation. In order to expand the data dimension through the broadcast mechanism, the data dimension in the function is expanded to dimension; S42, the growth cycle modulation factor and characteristic sequences of child health data Combined in an element-by-element multiplication manner, the specific mathematical model is: ; Where, is the physiological characteristic sequence after periodic modulation, is element-wise multiplication; S43, will Time to The combination of physiological characteristics sequence at each moment, , forming a length of Time series data , , using long short-term memory networks to capture long-term dependencies, the specific mathematical model is: ; Where, To utilize the time series data output by the long short-term memory network, ,in For the The time series feature sequence containing historical information output by the long short-term memory network, Long short-term memory network.

[0038] S5. Construct a health status fluctuation perception module, obtain the health status fluctuation value by calculating the difference between the physiological feature sequence at the current moment and the previous moment, introduce a feature reconstruction parameter, and weightedly fuse the time series feature sequence and the physiological feature sequence based on the parameter to form a health feature sequence.

[0039] Furthermore, in said S5, a health status fluctuation perception module is constructed, such as Figure 4 As shown, the specific steps include: S51, calculate the The difference between the temporal feature sequence at a certain moment and the physiological feature sequence at the previous moment, the specific mathematical model is: ; Where, For the The fluctuation value of health status at all times, is the L1 norm, For the The physiological sequence of moments; S52. Generate feature reconstruction parameters using nonlinear transformation to adjust the attention weights of the physiological feature sequence and the temporal feature sequence containing historical information. The specific mathematical model is: ; Where, For the The feature reconstruction parameters at time , are the learnable weights, is a learnable bias term; S53, the Time series feature sequence and physiological characteristic sequences Weighted fusion is performed according to feature reconstruction parameters. The specific mathematical model is: ; Where, is the health feature sequence after weighted fusion.

[0040] S6. Flatten the health feature sequence into a feature vector, input the vector into a fully connected neural network and an activation function in sequence, and then obtain the final child health monitoring value through linear transformation.

[0041] Furthermore, in said S6, the health feature sequence Concatenate by the last dimension and flatten into a feature vector , the feature vector is sequentially input into the fully connected layer and ReLU activation to obtain the output vector, and then the health monitoring value is obtained through linear transformation. The specific mathematical model is: ; ; Where, is the output vector, 、 is the learnable weight matrix, 、 is a learnable bias term, The final output value of child health monitoring.

[0042] Furthermore, the model is implemented using Python 3.8 programming language and PyTorch framework, running in CUDA11.3 environment, trained using NVIDIA 3090 24GB GPU, loss function using mean square error loss, and optimizer using , the initial learning rate is set to , the batch size is , the training round is .

[0043] Furthermore, the loss value of the training process of this method changes with the iteration round as follows: Figure 5 As shown, the vertical axis is the training loss value and the horizontal axis is the training round. Figure 5It can be seen that with the increase of training rounds, the training loss value shows a continuous downward trend, indicating that the model is constantly optimizing parameters to minimize the loss function. The effect of this method on child health monitoring is as follows: Figure 6 As shown, the vertical axis is the child health monitoring value (percentage), the horizontal axis is time (days), the gray dotted line in the figure represents the actual child health assessment value, the black solid line is the child health monitoring value generated by this method, and the gray shaded area is the difference between the two. Figure 6 It can be seen that the trends of the monitored values ​​are basically consistent with the actual evaluation values, and the error margin is small, which proves the excellent performance of the proposed method in the child health monitoring task and verifies the accuracy and reliability of the method.

[0044] The above are only preferred embodiments of the present invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A child health monitoring method based on vital sign data, characterized in that: The following steps are involved: S1. Collect data on physical signs that affect children's health, preprocess them using the Z-score method, and construct a children's physical sign dataset; S2. All data at each moment in the child vital sign dataset are combined into a data sequence, and the data sequence is mapped to a high-dimensional space using a lookup table to obtain an implicit feature sequence reflecting the child's health status. The feature sequence is obtained through linear transformation and ReLU activation function; S3. Construct an adaptive time series weight module, generate the weight of the current moment based on the feature sequence and the weight of the previous moment by nonlinear adjustment, and introduce a smoothing coefficient matrix to dynamically weight and smooth the feature sequence to obtain a weighted feature sequence; S4. Constructing a growth cycle feature module, introducing a growth cycle factor, and multiplying the growth cycle factor element by element with the weighted feature sequence to obtain a physiological feature sequence. The physiological feature sequence constitutes time series data, and a long short-term memory network is used to obtain a temporal feature sequence. S5. Construct a health status fluctuation perception module, calculate the difference between the physiological feature sequence at the current moment and the previous moment, obtain the health status fluctuation value, introduce a feature reconstruction parameter, and use the parameter to weightedly fuse the temporal feature sequence and the physiological feature sequence to form a health feature sequence; S6. Flatten the health feature sequence into a feature vector, input the vector into a fully connected neural network and an activation function in sequence, and then obtain the final child health monitoring value through linear transformation.

2. A child health monitoring method based on vital sign data according to claim 1, characterized in that: Vital sign data affecting children's health were collected, including weight, height, body temperature, heart rate, blood oxygen saturation, ambient temperature, number of steps, and sleep duration. For each type of data, the Z-score method was used to detect outliers. Any data value with a Z-score value greater than 3 or less than -3 was judged as an outlier. For the detected outliers, the mean replacement method was used, that is, the outlier was replaced with the mean of the data three days before and after it to construct a children's vital sign dataset.

3. A child health monitoring method based on vital sign data according to claim 2, characterized in that: In the children's vital signs dataset, all data at each moment constitute the data sequence of the model input. A series of vital signs data that constantly influences children's health ,Will Mapping to a richer feature sequence includes the following steps: S21, the vital signs data sequence Mapping to high-dimensional space through lookup table to obtain high-dimensional implicit feature sequence ; S22, using linear transformation to transform implicit feature sequence Mapped to the physiological feature space, and processed nonlinearly through the ReLU activation function, the first Feature sequences that constantly impact children's health data .

4. The method for monitoring child health based on vital sign data according to claim 3, characterized in that: Constructing an adaptive timing weight module includes the following steps: S31, Comprehensive The weight of the moment and the The information of the moment characteristic sequence is combined with the changing rate and amplitude of the nonlinear adjustment control weight. The specific mathematical model is: ； Where, For the The weight of the moment feature sequence, For the The weight of the moment feature sequence, is a learnable step size parameter for adjusting the weight update amplitude, is a learnable weight suppression parameter, is a learnable nonlinear modulation parameter; S32. Introduce a smoothing coefficient matrix and automatically adjust the size of the smoothing coefficient matrix according to the variation of the characteristic sequence at adjacent moments. The specific mathematical model is: ； Where, For the The smoothing coefficient matrix at time , For the The characteristic sequence that affects children's health data at all times, is a learnable feature change sensitive parameter, To take the absolute value operation; S33, the Moment smoothing coefficient matrix With the Characteristic sequence of children's health data at all times Multiply and add weight Multiply to get the weighted feature sequence .

5. The method for monitoring child health based on vital sign data according to claim 4, characterized in that: Constructing a growth cycle feature module includes the following steps: S41, generate the first Growth cycle factors at each moment , the specific mathematical model is: ； Where, For the The growth cycle factor of the moment, is the learnable weight matrix, is a learnable bias term, is the learnable periodic frequency, is the learnable phase, To expand the data dimension through the broadcast mechanism; S42, the growth cycle modulation factor and characteristic sequences of child health data Combined by element-by-element multiplication, the physiological characteristic sequence after periodic modulation is obtained ; S43, will Time to The physiological feature sequence of the moment is combined to form a length of Time series data , , using long short-term memory networks to capture long-term dependencies, we get Time series feature sequence containing historical information output by long short-term memory network .

6. The method for monitoring child health based on vital sign data according to claim 5, characterized in that: Constructing a health status fluctuation perception module includes the following steps: S51, calculate the The L1 norm of the difference between the temporal feature sequence at the moment and the physiological feature sequence at the previous moment is used to obtain the health status fluctuation value ; S52. Generate feature reconstruction parameters using nonlinear transformation. The specific mathematical model is: ； Where, For the The feature reconstruction parameters at time , are the learnable weights, is a learnable bias term; S53, the Time series feature sequence and physiological characteristic sequences Weighted fusion is performed according to feature reconstruction parameters. The specific mathematical model is: ； Where, is the health feature sequence after weighted fusion.

7. The method for monitoring child health based on vital sign data according to claim 6, characterized in that: Health feature sequence Concatenate by the last dimension and flatten into a feature vector , the feature vector is input into the fully connected layer and ReLU activation in turn to obtain the output vector, and then the child health monitoring value is obtained through linear transformation .

Images