Nursing system based on electronic medical record

Abstract

The invention relates to the technical field of medical care, and provides an electronic medical record-based nursing system, which comprises a heterogeneous data integration module for acquiring multi-source heterogeneous electronic medical record data from a hospital HIS (hospital information system), wearable monitoring equipment and a nursing operation terminal and outputting cleaned and standardized data; the feature association module is used for receiving the cleaned and standardized data to generate a spatial-temporal feature association matrix and storing a feature topological graph; the dynamic decision-making module is used for processing the feature incidence matrix based on the deep space-time fusion network to generate a nursing decision-making vector; the scheme execution monitoring module is used for collecting nursing operation data flow in real time, calculating and executing deviation degree parameters and feeding back the parameters; the abnormity intervention module analyzes and executes deviation degree parameters to trigger the dynamic decision module to execute nursing scheme reconstruction; the multi-terminal cooperation module is used for converting the nursing decision vector into an equipment instruction set and feeding back an equipment state; and the secure storage module is used for encrypting and storing the feature topological graph and the decision knowledge graph and providing model parameter version management.

Description

Technical Field

[0001] The present invention relates to the technical field of medical care, and more specifically, to a nursing system based on electronic medical records. Background Art

[0002] In the field of medical care, with the continuous development of information technology, electronic medical record systems have gradually become popular, providing convenience for the storage and management of medical data. However, most existing nursing systems are limited to simple data recording and query functions, lacking the ability to deeply integrate and analyze multi-source heterogeneous data. In the actual nursing process, medical staff need to manually consult a large amount of scattered data, making it difficult to make nursing decisions quickly and accurately. In addition, there is a lack of real-time monitoring and dynamic adjustment mechanisms for the execution of nursing operations. Once a deviation from the standard operation occurs, it is difficult to detect and correct it in a timely manner, which may affect the quality of nursing and patient safety. At the same time, the security and privacy of nursing data also face many challenges, and there is a risk of being tampered with or leaked during data storage and transmission.

[0003] In the process of implementing the embodiments of the present invention, the present application discovers that there are at least the following problems or defects in the prior art: the existing nursing system has insufficient integration ability for multi-source heterogeneous data and cannot effectively mine the potential value in the data to assist nursing decisions; it lacks the function of real-time monitoring the execution of nursing operations and dynamically adjusting the nursing plan according to the deviation degree, making it difficult to meet the complex and changeable clinical nursing needs; the data security and privacy protection measures are not perfect enough to fully guarantee the security of patient information. Summary of the Invention

[0004] The present invention provides a nursing system based on electronic medical records, including: A heterogeneous data integration module, which is used to obtain multi-source heterogeneous electronic medical record data from the hospital HIS system, wearable monitoring devices and nursing operation terminals, clean and standardize the data, and then output it to the feature association module; A feature association module, which is used to receive the data after cleaning and standardization, generate a spatio-temporal feature association matrix, output the matrix to the dynamic decision-making module, and at the same time store the feature topology map in the secure storage module; A dynamic decision-making module, which is used to process the spatio-temporal feature association matrix based on a deep spatio-temporal fusion network, generate a nursing decision vector, and distribute the vector to the plan execution monitoring module and the multi-terminal collaboration module; A plan execution monitoring module, which is used to collect the nursing operation data stream in real time, calculate the execution deviation parameter, and feedback it to the abnormal intervention module and the dynamic decision-making module; An abnormal intervention module, which is used to analyze the execution deviation parameter, and trigger the dynamic decision-making module to reconstruct the nursing plan when the deviation degree exceeds the threshold; A multi - terminal collaboration module, which is used to convert the nursing decision vector into a device instruction set, arbitrate conflicting instructions and then distribute them to terminal devices, and feedback the device status to the dynamic decision - making module; A secure storage module, which is used to encrypt and store the feature topology graph and the decision knowledge graph, and provide model parameter version management for the dynamic decision - making module.

[0005] Further, the heterogeneous data integration module performs data cleaning including the following steps: Based on the Markov chain model, estimate the missing data through the previous state transition probability, historical mean and Gaussian noise; Perform piece - wise non - linear normalization on the complete data. When the data value is greater than or equal to the preset benchmark threshold, map it to the interval [0, 1]; when the data value is less than the benchmark threshold, map it to the interval [0, 0.5].

[0006] Further, the feature association module performs the following steps: Construct a three - dimensional feature tensor from the data after cleaning and standardization according to the dimensions of time slices, ward areas and physiological indicators; Calculate the spatio - temporal attention weight matrix based on the three - dimensional feature tensor, as shown in the following formula: , where, represents the query vector of the i - th time slice, represents the key vector of the j - th spatial position, represents the query projection matrix, represents the key projection matrix, represents the scaling factor, and the three - dimensional feature tensor , where 24 represents the number of 24 - hour time slices, W represents the total number of ward areas, and C = 12 represents the number of physiological indicator dimensions; After splicing the spatio - temporal attention weight matrix and the three - dimensional feature tensor, input them into a three - dimensional convolutional network for feature extraction.

[0007] Further, the three - dimensional convolutional network performs dynamic kernel update including the following steps: Real - time monitor the feature change amount of the three - dimensional feature tensor; Dynamically adjust the convolutional kernel weights through a gated recurrent unit, where the current weight is jointly determined by the previous moment's weight and the one - dimensional vector of the feature change amount.

[0008] Further, the processing flow of the deep spatio - temporal fusion network includes: Extract the physiological feature vector and the environmental parameter vector from the spatio - temporal feature correlation matrix respectively; Splice the physiological feature vector and the environmental parameter vector, and calculate the cross - modal correlation weight through a learnable parameter matrix; Based on the cross-modal correlation weights, perform multi-scale convolutional processing on the physiological feature vector, and extract features with different granularities using three convolutional kernel sizes: large, medium, and small respectively; Fuse the multi-scale features through a gating mechanism to generate the final nursing decision vector.

[0009] Furthermore, the execution monitoring module of the solution calculates the execution deviation parameter including: Decompose the actual nursing operation into an atomic event sequence and align it with the preset standard event sequence; Calculate the minimum path difference between the actual sequence and the standard sequence through the dynamic time warping algorithm; Based on the minimum path difference and the sequence lengths of the actual sequence and the standard sequence, calculate the window deviation through an exponential decay function to quantify the operation deviation.

[0010] Furthermore, when the deviation exceeds the threshold, triggering the dynamic decision module to reconstruct the nursing plan includes: Set the sliding window length and the time decay weight coefficient; Within the sliding window, accumulate the window deviation based on the time decay weight to obtain the cumulative deviation; When both the cumulative deviation and the current window deviation exceed the preset threshold, trigger the dynamic decision module to reconstruct the nursing plan.

[0011] Furthermore, the multi-terminal collaboration module converts the nursing decision vector into a device instruction set, including the following steps: Extract the instruction feature vector from the nursing decision vector and obtain the instruction delay time; Calculate the urgency component based on the delay time, calculate the importance component based on the instruction feature vector and the preset severity coefficient, and perform weighted summation to obtain the instruction priority; If there are concurrent conflicting instructions, select the instruction with shorter execution time or higher priority according to the priority difference.

[0012] Furthermore, fusing the multi-scale features through a gating mechanism to generate the final nursing decision vector includes: Calculate the orthogonality loss between the multi-scale feature matrices, and the orthogonality loss is the sum of the squares of the Frobenius norms of the transpose products of the feature matrices of each scale.

[0013] Furthermore, the total loss function of fusing the multi-scale features through a gating mechanism includes the weighted sum of the cross-entropy loss and the orthogonality loss, where the weight of the orthogonality loss is 0.01.

[0014] According to the above embodiments of the present invention, it has at least the following beneficial effects: The nursing system of the present invention can achieve the efficient integration and standardized processing of multi-source heterogeneous data, perform in-depth analysis on the data through a deep spatio-temporal fusion network, generate accurate nursing decision vectors, thereby providing scientific and reasonable nursing suggestions for medical staff, improving the accuracy and timeliness of nursing decisions. At the same time, the system can collect the nursing operation data stream in real time, calculate the execution deviation parameters and give feedback, trigger the dynamic reconstruction of the nursing plan, ensure the standardization and consistency of nursing operations, reduce nursing risks, and improve the quality of nursing.

[0015] In addition, the multi-terminal collaboration module in the system can convert the nursing decision vector into a device instruction set to realize the intelligent collaborative work of nursing devices and improve nursing efficiency. The secure storage module uses encryption technology to store the feature topology map and decision knowledge graph, and provides model parameter version management to ensure the security and integrity of data, which can provide strong support for the stable operation of the system and the reliable traceability of data. Brief Description of the Drawings

[0016] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown by way of example and not limitation, wherein:

[0017] Figure 1 It is a schematic structural diagram of a nursing system based on electronic medical records provided by an embodiment of the present invention. Detailed Embodiments

[0018] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and then implement the present invention, and not to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to be able to convey the scope of the present invention fully to those skilled in the art.

[0019] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, apparatus, device, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms: completely hardware, completely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0020] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0021] Below with reference to Figure 1 , Figure 1The structural schematic diagram of a nursing system based on electronic medical records provided by an embodiment of the present invention. As Figure 1 shown, a nursing system based on electronic medical records includes: a heterogeneous data integration module (1), which is used to obtain multi-source heterogeneous electronic medical record data from the hospital HIS system, wearable monitoring devices, and nursing operation terminals, clean and standardize the data, and then output it to the feature association module; a feature association module (2), which is used to receive the data after cleaning and standardization, generate a spatio-temporal feature association matrix, output the matrix to the dynamic decision-making module, and store the feature topology map in the secure storage module at the same time; a dynamic decision-making module (3), which is used to process the spatio-temporal feature association matrix based on a deep spatio-temporal fusion network, generate a nursing decision vector, and distribute the vector to the scheme execution monitoring module and the multi-terminal collaboration module; a scheme execution monitoring module (4), which is used to collect the nursing operation data stream in real time, calculate the execution deviation parameter, and feedback it to the abnormal intervention module and the dynamic decision-making module; an abnormal intervention module (5), which is used to analyze the execution deviation parameter, and trigger the dynamic decision-making module to reconstruct the nursing scheme when the deviation exceeds the threshold; a multi-terminal collaboration module (6), which is used to convert the nursing decision vector into a device instruction set, arbitrate conflicting instructions, and then distribute them to the terminal devices, and feedback the device status to the dynamic decision-making module; a secure storage module (7), which is used to encrypt and store the feature topology map and the decision knowledge graph, and provide model parameter version management for the dynamic decision-making module.

[0022] It should be noted that the nursing system of the present invention realizes the processing of electronic medical record data and the generation of nursing decisions through multiple interactive modules. The heterogeneous data integration module is responsible for obtaining multi-source heterogeneous electronic medical record data from the hospital's HIS system (hospital information system, a system for managing the daily operations of the hospital and patient information), wearable monitoring devices (such as smart bracelets, smart mattresses, etc., for real-time monitoring of patient vital signs and other data), and nursing operation terminals (devices used by medical staff to record nursing operation information, etc.), and standardizes these data and outputs them to the feature association module. After receiving the data that has been cleaned and standardized, the feature association module generates a spatio-temporal feature association matrix, which is used to describe the correlation of data in time and space, and outputs it to the dynamic decision-making module, while storing the feature topology graph describing the relationship between features in the secure storage module. The dynamic decision-making module processes the feature association matrix based on a deep spatio-temporal fusion network, generates a nursing decision vector, and distributes it to the scheme execution monitoring module and the multi-terminal collaboration module. The scheme execution monitoring module real-time collects the nursing operation data stream, calculates the execution deviation parameter (used to measure the deviation between the actual nursing operation and the standard operation), and feeds this parameter back to the abnormal intervention module and the dynamic decision-making module. The abnormal intervention module analyzes the execution deviation parameter, and when it finds that the deviation is too large, it triggers the dynamic decision-making module to execute nursing plan reconstruction to adjust the nursing plan. The multi-terminal collaboration module converts the nursing decision vector into a device instruction set (a set of instructions for controlling nursing devices), arbitrates and distributes it to the terminal devices, and feeds back the device status to the dynamic decision-making module. The secure storage module is responsible for encrypting and storing the feature topology graph and the decision knowledge graph (a graph structure for storing decision-related knowledge), and providing model parameter version management for the dynamic decision-making module to ensure the security and traceability of the data.

[0023] Specifically, the heterogeneous data integration module performs data cleaning including the following steps: Based on the Markov chain model, estimate the missing data through the previous state transition probability, historical mean, and Gaussian noise; Perform piecewise non-linear normalization on the complete data. When the data value is greater than or equal to the preset reference threshold, map it to the [0, 1] interval; when the data value is less than the reference threshold, map it to the [0, 0.5] interval.

[0024] When the heterogeneous data integration module performs data cleaning, it first estimates the missing data through a Markov chain model. A Markov chain is a mathematical model used to describe the transition probabilities of system states. In this system, it estimates the missing data by calculating the transition probability of the current state given the previous state, thus ensuring the integrity of the data. Next, a piecewise non-linear function is used to normalize the data. Normalization scales the data proportionally so that it falls within a small range, such as between 0 and 1. Here, the piecewise non-linear function uses different functional forms according to different intervals of the eigenvalue. When the eigenvalue is greater than or equal to a certain reference threshold, one functional form is used; when the eigenvalue is less than the reference threshold, another functional form is used. This processing method can better adapt to the distribution characteristics of different eigenvalues and improve the usability of the data. When the feature association module constructs a three-dimensional feature tensor, its dimensions are set to the number of 24-hour time slices, the total number of ward areas, and the number of physiological index dimensions. This means that the system divides a day into 24 time slices, considers all ward areas, and 12 physiological indicators, thus comprehensively describing the patient's state. When calculating the spatio-temporal attention weight matrix, the query vector, key vector, and corresponding projection matrix are used to calculate the weights through the scaled dot-product attention mechanism to highlight important spatio-temporal features. When the dynamic decision-making module processes the feature association matrix, it extracts the physiological feature vector and environmental parameter vector through a deep spatio-temporal fusion network, calculates the cross-modal association weight, performs multi-scale convolution processing, and performs gated fusion of multi-scale features, and finally generates a nursing decision vector. These steps comprehensively consider physiological and environmental factors, and through the fusion of multi-scale features, improve the accuracy of decision-making.

[0025] Preferably, in the heterogeneous data integration module, for the step of estimating missing data by the Markov chain model, the calculation method of the transition probability can be further optimized, such as introducing more historical data to improve the accuracy of the estimation. In the feature association module, for the construction of the three-dimensional feature tensor, the total number of ward areas can be considered to be dynamically adjusted according to the number of patients in different ward areas to more flexibly adapt to the actual operation of the hospital. In the dynamic decision-making module, for the processing flow of the deep spatio-temporal fusion network, a feature verification step can be added to verify the extracted feature vector before generating the nursing decision vector to ensure the accuracy and reliability of the features. In addition, in the scheme execution monitoring module, when calculating the execution deviation degree, more evaluation indicators can be introduced, such as the time efficiency of nursing operations, the accuracy of operations, etc., to more comprehensively evaluate the execution of nursing operations.

[0026] In some embodiments, the heterogeneous data integration module performing data cleaning includes: Step 1: Estimate the missing data through a Markov chain model, and the calculation formula is: , where represents the current state and the transition probability under the previous state , represents the eigenvalue at the previous moment represents the historical mean of the feature represents the random noise subject to Gaussian distribution represents the noise variance; Step 2: Normalize the data using a piecewise non-linear function. When is satisfied, execute: , When is satisfied, execute: , where k = 2.5 represents the steepness factor represents the feature reference threshold, and x represents the original feature value;

[0027] It should be noted that when the heterogeneous data integration module processes data, it will first perform data cleaning. This is to remove noise and fill in missing values in the data, thereby improving the quality and usability of the data. The first step of data cleaning is to estimate missing data through a Markov chain model. A Markov chain is a stochastic process model that assumes that the future state of the system depends only on the current state and is independent of past states. In this system, the Markov chain model can be used to predict missing data points based on existing data sequences. Specifically, the model calculates the transition probability of the current state under the previous state and combines the historical mean of the feature and the random noise of the Gaussian distribution to estimate the missing value. The Gaussian distribution is a common probability distribution, and its random noise is used to simulate the uncertainty in the data. The second step is to normalize the data using a piecewise non-linear function. Normalization is to transform the data into a specific range so that data with different features can be compared and processed on a unified scale. A piecewise non-linear function is a function that uses different function forms for different intervals of the input data, which can better adapt to the non-linear characteristics of the data and make data processing more flexible and accurate.

[0028] Specifically, when estimating missing data using the Markov chain model, the calculation of the transition probability is crucial. The transition probability represents the possibility of transitioning from one state to another. In this system, by analyzing the historical data sequence, a transition probability matrix between each state can be obtained. For example, if the current feature value is , and the feature value at the previous moment is , then the transition probability can be obtained by counting when Estimated by the frequency of occurrence. The historical mean of a feature is obtained by calculating the average value of the feature in historical data, which reflects the average level of the feature. Gaussian distributed random noise is used to simulate the random fluctuations in the data, and its variance Can be adjusted according to the degree of data fluctuation. In the piecewise non-linear function, the steepness factor k and the feature reference threshold Are two important parameters. The steepness factor k determines the steepness of the function, and a larger value of k will make the change of the function near the threshold more drastic. The feature reference threshold Is a value used to divide the data interval. When the original feature value x is greater than or equal to , one functional form is used; when x is less than , another functional form is used. This piecewise processing method can better adapt to the distribution characteristics of different feature values and improve the data processing effect.

[0029] Preferably, in the Markov chain model, more historical states can be introduced to estimate the missing data, rather than just the previous state. For example, the previous two or more states can be considered to calculate the transition probability, thereby improving the accuracy of the estimation. For the piecewise non-linear function, different functional forms and parameters can be selected according to the specific application scenario and data characteristics. For example, the value of the steepness factor k can be adjusted to make it more suitable for a specific feature distribution. In addition, other types of noise distributions can be introduced to replace the Gaussian distribution to better simulate the uncertainty in the data. For example, if the data fluctuations exhibit the characteristics of a spike-and-slab distribution, other distributions such as the Lévy distribution can be considered to replace the Gaussian distribution.

[0030] Furthermore, the feature association module performs the following steps: Construct a three-dimensional feature tensor from the data that has been cleaned and standardized, in terms of time slices, ward areas, and physiological index dimensions; Calculate a spatio-temporal attention weight matrix based on the three-dimensional feature tensor to quantify the importance of different spatio-temporal features; After concatenating the spatio-temporal attention weight matrix with the three-dimensional feature tensor, input it into a three-dimensional convolutional network for feature extraction.

[0031] In some embodiments, a three-dimensional feature tensor is constructed, where 24 represents the number of 24-hour time slices, W represents the total number of ward areas, and C = 12 represents the number of physiological index dimensions; Calculate the spatio-temporal attention weight matrix based on the tensor T: , where, represents the query vector for the i-th time slice, represents the key vector for the j-th spatial position, represents the query projection matrix, represents the key projection matrix, represents the scaling factor; After concatenating the weight matrix A and the tensor T, input them into a three-dimensional convolutional network, and the operation expression is: , where represents the channel dimension concatenation operation, represents the convolutional kernel size in the time, space, and feature dimensions, represents the sliding step size in each dimension;

[0032] It should be noted that the main function of the feature correlation module is to convert the standardized data into a correlation matrix that can reflect the spatio-temporal characteristics of the data and store it for subsequent use. This process involves two key steps: constructing a three-dimensional feature tensor and calculating the spatio-temporal attention weight matrix. The three-dimensional feature tensor is a multi-dimensional data structure that can simultaneously describe the distribution of data in the three dimensions of time, space, and features. The spatio-temporal attention weight matrix is used to measure the importance and correlation degree between different time slices and spatial positions. In this way, those spatio-temporal features that are more critical for nursing decisions can be highlighted, thereby providing more valuable information for subsequent decisions.

[0033] Specifically, when constructing the three-dimensional feature tensor, its dimensions are set to the number of 24-hour time slices, the total number of ward areas, and the number of physiological index dimensions. This means that the system divides a day into 24 time slices, considers all ward areas, and 12 physiological indexes, so as to comprehensively describe the patient's state. When calculating the spatio-temporal attention weight matrix, use the query vector, key vector, and the corresponding projection matrix, and calculate the weight through the scaled dot-product attention mechanism to highlight the important spatio-temporal features. The query vector and key vector are vectors extracted from the tensor, representing the features of the time slice and spatial position respectively. The projection matrix is used to map these vectors into a suitable feature space for effective calculation. The scaling factor is to prevent the problem of gradient disappearance caused by excessive numerical values during the calculation process, and usually takes the square root of the query vector dimension. In this way, a weight matrix that can reflect the correlation degree between different time slices and spatial positions can be obtained, and then provide guidance for subsequent convolutional operations.

[0034] Preferably, when constructing the three-dimensional feature tensor, the number of time slices, the division of ward areas, and the selection of physiological indicators can be flexibly adjusted according to the actual medical scenarios and data characteristics. For example, in some scenarios that require a finer time resolution, a day can be divided into more time slices, such as one time slice every half hour. For the division of ward areas, it can be dynamically adjusted according to the size of the ward, the number of patients, and the actual needs of nursing work. In the selection of physiological indicators, in addition to the common 12 physiological indicators, some special physiological indicators can also be added according to the monitoring requirements of specific diseases. When calculating the spatio-temporal attention weight matrix, different attention mechanisms can be tried, such as the multi-head attention mechanism, which can capture the correlations between spatio-temporal features from different perspectives, thereby improving the model's ability to understand complex data. In addition, the parameters of the projection matrix can be optimized, and their values can be adjusted through learning during the training process to better adapt to the distribution characteristics of the data.

[0035] In some embodiments, the three-dimensional convolutional network performing dynamic kernel update includes the following steps: Real-time monitor the feature change amount of the three-dimensional feature tensor; Dynamically adjust the convolutional kernel weights through a gated recurrent unit, where the current weight is jointly determined by the previous moment's weight and the one-dimensional vector of the feature change amount.

[0036] Specifically, the three-dimensional convolutional network performing dynamic kernel update includes: Monitor the feature change amount , where h = 24 represents the time dimension, w = W represents the spatial dimension, and c = 64 represents the number of feature channels; Adjust the convolutional kernel weights through a gated recurrent unit: , where, represents a gated recurrent network with 128 hidden units, represents the convolutional kernel weight at the previous moment, represents flattening into a one-dimensional vector.

[0037] It should be noted that the three-dimensional convolutional network plays a key role in the feature correlation module. It extracts more representative and discriminative features by performing convolutional operations on the spatio-temporal feature tensor. The dynamic kernel update mechanism is an important feature of this network. It can dynamically adjust the weights of the convolutional kernels according to the changes in the input features, thereby enhancing the network's adaptability to different input patterns. The update of the convolutional kernel weights is achieved through a gated recurrent unit (GRU). GRU is a neural network structure commonly used to process sequential data. It can remember the previous state information and adjust the current weights accordingly. This dynamic update mechanism enables the network to better capture the changes in spatio-temporal features and improve the accuracy and robustness of feature extraction.

[0038] Specifically, the dynamic kernel update mechanism first monitors the feature change amount, which is achieved by calculating the differences between the input features at different time steps. The dimensions of the feature change amount include the time dimension, the spatial dimension, and the number of feature channels, corresponding to the number of time slices, the number of ward areas, and the number of features respectively. Then, the gated recurrent unit (GRU) is used to adjust the convolutional kernel weights. GRU takes the feature change amount at the current moment and the convolutional kernel weights at the previous moment as inputs and outputs the updated convolutional kernel weights. There are multiple hidden units inside GRU, and the number of these hidden units can be set according to actual needs. For example, it is set to 128 in this system. In this way, the convolutional kernel weights can dynamically adapt to the changes in the input features, thereby improving the network's ability to extract spatio-temporal features. In addition, the update of the convolutional kernel weights also involves the operation of flattening the feature change amount into a one-dimensional vector because the input of GRU needs to be in the form of a one-dimensional vector, which can conveniently input the multi-dimensional feature change amount into GRU for processing.

[0039] Preferably, when monitoring the feature change amount, more context information can be introduced to improve the accuracy of weight update. For example, in addition to considering the feature changes at the current moment and the previous moment, the feature changes in the previous several moments can also be combined to calculate the feature change amount, thereby providing richer historical information for weight update. For the number of hidden units in the gated recurrent unit (GRU), it can be adjusted according to the complexity of the data and the training effect of the network. If the spatio-temporal feature changes in the data are relatively complex, the number of hidden units can be appropriately increased to enhance the network's memory ability and adaptability. In addition, other types of recurrent neural network structures can be tried to replace GRU, such as the long short-term memory network (LSTM), which has better performance in processing long sequence data. In the process of updating the convolutional kernel weights, a regularization mechanism such as weight decay can also be introduced to prevent overfitting during the weight update process, thereby improving the generalization ability of the network.

[0040] In some embodiments, the processing flow of the deep spatio-temporal fusion network includes: Extract the physiological feature vector and the environmental parameter vector from the spatio-temporal feature correlation matrix respectively; Concatenate the physiological feature vector and the environmental parameter vector, and calculate the cross-modal correlation weight through a learnable parameter matrix; Based on the cross-modal correlation weight, perform multi-scale convolution processing on the physiological feature vector, and extract features with different granularities using three convolution kernel sizes of large, medium, and small respectively; Fuse the multi-scale features through a gating mechanism to generate the final care decision vector.

[0041] Specifically, extract the physiological feature vector and the environmental parameter vector , where represents a 64-dimensional real vector space;

[0042] Calculate the cross-modal correlation weight: , where represents the vector concatenation operation, represents the learnable parameter matrix, represents the sigmoid activation function;

[0043] Perform multi-scale convolution processing: , , , where, represents the tensor product operation, kernel = 5 represents a 5×5×5 convolution kernel; is the feature tensor extracted by the 5×5×5 convolution kernel, is the feature tensor extracted by the 3×3×3 convolution kernel, is the feature tensor extracted by the 1×1×1 convolution kernel, is the cross-modal correlation weight.

[0044] The gating fusion of multi-scale features is: , , where, represents the fusion weight matrix, represents the element-wise multiplication operation, represents the feature concatenation operation, is the gating signal, is the output feature tensor.

[0045] It should be noted that the deep spatio-temporal fusion network is the core module in this system for processing the feature correlation matrix and generating the nursing decision vector. Through a series of complex processing steps, it fuses physiological features and environmental parameters, extracts multi-scale features, and performs gated fusion to finally generate an accurate nursing decision vector. This process involves multiple key steps, including the extraction of physiological feature vectors and environmental parameter vectors, the calculation of cross-modal correlation weights, multi-scale convolutional processing, and gated fusion, etc. The physiological feature vector and the environmental parameter vector respectively represent the physiological state of the patient and the parameters of the environment where the patient is located. Their fusion can provide more comprehensive information for nursing decisions. The cross-modal correlation weight is used to measure the degree of correlation between different modal data, while multi-scale convolutional processing can capture feature information at different scales, and gated fusion further optimizes the combination of features to ensure that the finally generated nursing decision vector has high accuracy and practicality.

[0046] Specifically, the deep spatio-temporal fusion network first extracts the physiological feature vector and the environmental parameter vector, and the dimensions of both vectors are 64-dimensional real vector spaces. The physiological feature vector contains physiological-related data such as the patient's vital signs and medical history, while the environmental parameter vector covers environmental factors such as the temperature, humidity, and light in the ward. Next, when calculating the cross-modal correlation weight, the physiological feature vector and the environmental parameter vector are combined through a vector concatenation operation, weighted by a learnable parameter matrix, and finally, a sigmoid activation function is used for non-linear transformation to obtain the cross-modal correlation weight. This weight reflects the correlation between physiological features and environmental parameters and helps with subsequent feature fusion. In the multi-scale convolutional processing stage, convolutional operations are performed on the feature tensor using convolutional kernels of different sizes (such as 5×5×5, 3×3×3, and 1×1×1) to extract features at different scales. The sizes and numbers of these convolutional kernels can be adjusted according to the characteristics of the actual data. Finally, in the gated fusion stage, a fusion weight matrix is used to weight and combine the multi-scale features, and element-wise multiplication operations are used for fusion to obtain the final feature representation, and then the nursing decision vector is generated.

[0047] Preferably, when extracting the physiological feature vector and the environmental parameter vector, different feature extraction methods can be selected according to the specific application scenario and data characteristics. For example, for the physiological feature vector, an autoencoder in deep learning can be introduced to extract more representative features. For the environmental parameter vector, data fusion techniques of sensor networks can be combined to improve the accuracy and reliability of environmental parameters. When calculating the cross-modal correlation weight, a more complex neural network structure, such as a multi-layer perceptron (MLP), can be tried to replace the simple learnable parameter matrix to better capture the complex relationship between physiological features and environmental parameters.

[0048] Furthermore, in multi-scale convolution processing, an attention mechanism can be introduced to enable the network to automatically learn the importance of features at different scales, thereby further improving the effect of feature extraction. In the gating fusion stage, different fusion strategies can be tried, such as attention-based fusion or graph neural network-based fusion, to better integrate multi-scale features and generate more accurate nursing decision vectors.

[0049] In some embodiments, the scheme execution monitoring module calculates the execution deviation parameter including: Decompose the actual nursing operation into an atomic event sequence and align it with the preset standard event sequence; Calculate the minimum path difference between the actual sequence and the standard sequence through the dynamic time warping algorithm; Based on the minimum path difference and the sequence lengths of the actual sequence and the standard sequence, calculate the window deviation through an exponential decay function to quantify the operation deviation.

[0050] In some embodiments, the scheme execution monitoring module calculates the execution deviation including: Decompose the nursing operation into an atomic event sequence and the standard sequence , where m represents the length of the actual sequence and n represents the length of the standard sequence; Calculate the minimum path difference by dynamic time warping: , where represents the i-th eigenvalue of the actual execution, represents the j-th eigenvalue of the standard, represents the minimum path difference between the actual sequence and the standard sequence, and i and j correspond to the positions in the actual sequence and the standard sequence respectively.

[0051] Calculate the window deviation: , where 0.5 represents the sensitivity adjustment coefficient, represents the maximum value of the sequence length, is the window deviation.

[0052] It should be noted that the core function of the solution execution monitoring module is to collect the nursing operation data stream in real time and calculate the execution deviation parameter, so as to quantitatively evaluate the execution of nursing operations. This process involves decomposing nursing operations into an atomic event sequence and a standard sequence, calculating the minimum path difference through the dynamic time warping algorithm, and then obtaining the window deviation. The atomic event sequence refers to the individual events that occur in sequence during the actual nursing operation, while the standard sequence is the ideal event sequence preset according to nursing specifications. The dynamic time warping algorithm is a method for measuring the similarity between two time series. It calculates the difference between them by finding the best matching path between the two sequences. The window deviation is an index that measures the degree of deviation of nursing operations from the standard based on this difference, which can provide a basis for subsequent abnormal intervention and nursing plan adjustment.

[0053] Specifically, in the solution execution monitoring module, nursing operations are decomposed into a series of atomic events, and these events are arranged in the order of actual occurrence to form an atomic event sequence, while the standard sequence is predefined according to nursing specifications. The dynamic time warping algorithm calculates the minimum path difference between these two sequences by comparing them, and this difference reflects the degree of deviation between the actual operation and the standard operation. During the calculation process, the algorithm will consider the characteristic values of each event, such as the time of nursing operation, operation type, etc., and determine the path by comparing the difference between the actually executed characteristic value and the standard characteristic value. The calculation of the window deviation further considers the cumulative effect of this difference, and balances the weight of the difference by setting a sensitivity adjustment coefficient, so as to obtain a quantitative index that can reflect the overall deviation degree of nursing operations. The closer the value of this index is to 1, the closer the nursing operation is to the standard operation; on the contrary, the smaller the value, the greater the degree of deviation.

[0054] Preferably, when decomposing nursing operations into an atomic event sequence, more event characteristics can be introduced, such as the duration of the operation, the executor of the operation, etc., to more comprehensively describe the process of nursing operations. For the dynamic time warping algorithm, different distance measurement methods, such as Euclidean distance, Manhattan distance, etc., can be tried to calculate the difference between characteristic values to find the most suitable measurement method for the current data characteristics. When calculating the window deviation, the sensitivity adjustment coefficient can be adjusted according to the actual nursing scenario and the tolerance of the deviation. For example, in a scenario with high requirements for nursing quality, the sensitivity adjustment coefficient can be appropriately reduced so that the window deviation can sensitively reflect even small deviations. In addition, the concept of a time window can be introduced to comprehensively evaluate nursing operations within a certain period of time, rather than only focusing on the deviation of individual events, so as to more accurately reflect the overall execution of nursing operations.

[0055] In some embodiments, the nursing plan reconstruction includes: Setting the sliding window length and the time decay weight coefficient; Within the sliding window, accumulating the window deviation degree based on the time decay weight to obtain the cumulative deviation degree; When both the cumulative deviation degree and the current window deviation degree exceed a preset threshold, trigger the dynamic decision module to reconstruct the nursing plan.

[0056] Specifically, the steps for the abnormal intervention module to trigger the reconstruction include: Set the sliding window length , the time decay constant hours; Calculate the cumulative deviation degree within the window: , where, represents the exponential decay weight, represents the deviation degree at the i-th moment, represents the cumulative deviation degree.

[0057] When both and are satisfied, trigger the plan reconstruction, where represents the square root of 6, and e represents the base of the natural logarithm;

[0058] It should be noted that the main function of the abnormal intervention module is to judge whether there is an abnormality in the nursing operation by analyzing the execution deviation degree parameter, and trigger the dynamic decision module to reconstruct the nursing plan when necessary. This process involves setting the sliding window length and the time decay constant, calculating the cumulative deviation degree within the window, and judging whether the nursing plan needs to be reconstructed according to the preset conditions. The sliding window length refers to the time range used to calculate the cumulative deviation degree, and the time decay constant is used to assign different weights to the deviation degrees at different time points to reflect the decay characteristics of the deviation degree over time. When the cumulative deviation degree exceeds the preset threshold, it indicates that there is a large deviation between the nursing operation and the standard plan, and at this time, the nursing plan needs to be reconstructed to ensure the standardization and effectiveness of the nursing operation.

[0059] Specifically, the abnormal intervention module first sets the sliding window length to 6 time units (such as 6 hours) and the time decay constant to 24 hours. This means that when calculating the cumulative deviation degree, the deviation degrees within the past 6 time units will be considered, and as time goes by, the influence of the deviation degree will gradually decrease. The calculation of the cumulative deviation degree is obtained by summing the product of the deviation degree at each time point and the corresponding exponential decay weight. The formula for the exponential decay weight is , where represents the current time point, Represents a past time point. When the cumulative deviation exceeds a preset threshold, such as 0.7 times the square root of 6, and within the past 6 time units, the average value of the deviation exceeds 0.9 times the decay coefficient, the reconstruction of the care plan is triggered. The settings of these parameters can be adjusted according to the actual care scenario and the tolerance of the deviation.

[0060] Preferably, when setting the sliding window length, it can be adjusted according to the complexity of the care operation and the requirement for real-time performance. For example, for care operations that require quick response, the sliding window length can be set shorter to detect abnormal situations more quickly. The time decay constant can also be adjusted according to the characteristics of the care operation. For example, for long-term care operations, the time decay constant can be appropriately increased to more smoothly reflect the change of the deviation. When calculating the cumulative deviation, more statistical methods can be introduced, such as moving average or weighted moving average, to more accurately reflect the trend of the deviation. In addition, the threshold of the trigger condition can be dynamically adjusted according to the importance and risk level of the care operation. For example, in high-risk care operations, a lower threshold can be set to more sensitively detect abnormal situations.

[0061] In some embodiments, the multi-terminal collaboration module converting the care decision vector into a device instruction set includes the following steps: Extract the instruction feature vector from the care decision vector and obtain the instruction delay time; Calculate the urgency component based on the delay time, calculate the importance component based on the instruction feature vector and the preset importance coefficient, and perform weighted summation to obtain the instruction priority; If there are concurrent conflicting instructions, select the instruction with shorter execution time or higher priority according to the priority difference.

[0062] Specifically, the steps for the multi-terminal collaboration module to perform instruction arbitration include: Collect the instruction feature vector and the delay time (unit: minute), where represents a 64-dimensional real vector space; Calculate the instruction priority: , , , where represents the learnable parameter vector, represents the preset importance coefficient, softmax represents the normalized exponential function, represents the urgency component, represents the importance component, is the instruction priority; Arbitration conflict instruction: If , select the instruction with the execution time-consuming , where represents the estimated time-consuming of the i-th instruction; otherwise, select the instruction with a larger P value; where is the priority of the i-th instruction, is the priority of the j-th instruction.

[0063] It should be noted that the core function of the multi-terminal collaboration module is to convert the nursing decision vector into a device instruction set, arbitrate and distribute the instructions of the terminal device, and at the same time feedback the device status to the dynamic decision module. This process involves collecting the instruction feature vector and the delay time, calculating the instruction priority, and arbitrating the conflict instructions according to the priority. The instruction feature vector is a vector describing the characteristics of the nursing decision vector, which is used to determine the execution priority of the instruction. The delay time refers to the time interval from the generation of the instruction to the actual execution of the instruction by the device. The calculation of the instruction priority is based on the characteristics of the instruction and the preset importance coefficient, and is normalized by the softmax function to ensure the reasonable execution order of the instructions.

[0064] Specifically, the multi-terminal collaboration module first collects the instruction feature vector and the delay time. The instruction feature vector is a 64-dimensional real vector, which contains the key feature information of the nursing decision vector. The delay time is in minutes, which reflects the timeliness of the instruction execution. When calculating the instruction priority, first calculate a basic priority according to the delay time. The shorter the delay time, the higher the basic priority. Then, weight the instruction feature vector by a learnable parameter vector and the preset importance coefficient to obtain the comprehensive priority. The preset importance coefficient is a value between 0 and 1, which is used to adjust the weights of the instruction feature vector and the delay time in the priority calculation. Finally, use the softmax function to normalize the comprehensive priorities of all instructions to obtain the final priority of each instruction. When arbitrating the conflict instructions, if the priority difference between two instructions is less than 0.1, select the instruction with the shorter execution time; otherwise, select the instruction with the higher priority.

[0065] Preferably, when collecting the instruction feature vectors, different feature extraction methods can be selected according to specific nursing scenarios and device types. For example, for complex nursing operations, a feature extraction network in deep learning can be introduced to more accurately extract instruction features. When calculating the instruction priority, different normalization methods can be tried, such as min-max normalization or Z-score normalization, to find the normalization method that best suits the current data characteristics. For the preset importance coefficient, it can be dynamically adjusted according to the importance and urgency of the nursing operation. For example, in an emergency nursing scenario, the weight of the delay time can be appropriately increased to ensure that the instruction can be executed quickly. When arbitrating conflicting instructions, in addition to considering the instruction priority and time consumption, the current state and load of the device can also be introduced as reference factors for arbitration to more reasonably allocate device resources and improve nursing efficiency.

[0066] In some embodiments, the steps for the security storage module to encrypt the knowledge graph include: Step 1: Divide the graph into subgraphs , where represents the total number of shards; Step 2: Perform double encryption on each subgraph: , where represents a 256-bit independent key, represents the bitwise XOR operation, represents the encrypted subgraph data, represents the symmetric encryption algorithm, represents the stream encryption algorithm.

[0067] Step 3: Generate an integrity verification label: , where represents the data concatenation operation, timestamp is the current Unix timestamp, represents the integrity verification label, represents a 256-bit independent key, represents the message authentication code algorithm based on the hash function.

[0068] It should be noted that the core function of the secure storage module is to encrypt and store the feature topology graph and the decision knowledge graph, and generate integrity verification tags to ensure the security and integrity of the data. This process involves splitting the graph into multiple subgraphs, double-encrypting each subgraph, and generating integrity verification tags based on the encrypted data. The feature topology graph is a graph structure that describes the relationships between features, and the decision knowledge graph stores the knowledge and rules related to care decisions. Double encryption is a method to enhance data security by encrypting data with two independent keys to ensure that even if one key is cracked, the data remains secure. The integrity verification tag is used to detect whether the data has been tampered with during storage and transmission, thus ensuring the integrity and reliability of the stored data.

[0069] Specifically, the secure storage module first splits the graph into 28 subgraphs, each subgraph containing a part of the graph data. This splitting method can improve the encryption efficiency and reduce the risk of a single subgraph being cracked. When double-encrypting each subgraph, two 256-bit independent keys are used to encrypt the subgraph through different encryption algorithms respectively. The first key is encrypted using the AES-256 algorithm, and the second key is XOR-operated after generating a hash value using the SHA-256 algorithm, thereby enhancing the complexity and security of the encryption. When generating the integrity verification tag, the encrypted subgraph data is concatenated with the current Unix timestamp, and then a hash value is generated through the SHA-256 algorithm. This hash value serves as the integrity verification tag for subsequent data integrity verification.

[0070] Preferably, when splitting the graph, the number of subgraphs can be dynamically adjusted according to the size and complexity of the graph. For example, for a larger graph, the number of subgraphs can be increased to improve the encryption efficiency and security. When selecting the encryption algorithm, more advanced encryption algorithms such as quantum encryption algorithms can be considered to cope with potential future security threats. For the generation of the integrity verification tag, more data elements such as device ID or user ID can be introduced to enhance the uniqueness and security of the tag. In addition, the encryption key and the generation algorithm of the integrity verification tag can be updated regularly to prevent security risks brought by long-term use of the same key or algorithm.

[0071] In some embodiments, fusing multi-scale features through a gating mechanism to generate the final care decision vector includes: Calculating the orthogonality loss between multi-scale feature matrices, where the orthogonality loss is the sum of the squares of the Frobenius norms of the transpose products of feature matrices at each scale; The total loss function for fusing multi-scale features through the gating mechanism includes the weighted sum of the cross-entropy loss and the orthogonality loss, where the weight of the orthogonality loss is 0.01.

[0072] Specifically, the orthogonal constraint imposed in the gating fusion step includes: Calculating the orthogonal loss of the feature matrix: , where, represents the square of the matrix Frobenius norm, represents the transposed matrix of the 5×5×5 convolutional feature;

[0073] The total loss function is: , where, y represents the true nursing plan label, represents the prediction result, and CrossEntropyLoss represents the cross-entropy loss function.

[0074] It should be noted that the orthogonal constraint imposed in the gating fusion step is to optimize the performance of the deep spatio-temporal fusion network and ensure that the feature matrix maintains good structural characteristics during the fusion process. The orthogonal constraint is a mathematical constraint condition that requires the rows or columns of the feature matrix to be orthogonal to each other, that is, their inner product is zero. This constraint can reduce the redundancy between features, enhance the independence and discriminative ability of features, and thus improve the generalization ability and stability of the model. The total loss function is a function used to measure the difference between the prediction result and the true label during the model training process. By adding the orthogonal constraint term to the loss function, while optimizing the model parameters, it can ensure that the feature matrix meets the orthogonality requirement.

[0075] Specifically, the calculation of the orthogonal constraint is obtained by multiplying the transpose of the feature matrix by itself, and then calculating the square of the Frobenius norm of the resulting matrix. The Frobenius norm is a matrix norm that measures the square root of the sum of the squares of all elements in the matrix. In this system, the orthogonal losses of the 5×5×5 convolutional feature, 3×3×3 convolutional feature, and 1×1×1 convolutional feature are calculated respectively, and these losses are added together to obtain the total orthogonal loss. The total loss function consists of two parts: the cross-entropy loss function and the orthogonal loss. The cross-entropy loss function is used to measure the difference between the predicted nursing decision vector and the true nursing plan label, while the orthogonal loss is used to ensure the orthogonality of the feature matrix. In the total loss function, the weight of the cross-entropy loss function is 1, and the weight of the orthogonal loss is 0.01. This weight setting makes the model pay attention to both the prediction accuracy and the orthogonality of the feature matrix during the optimization process.

[0076] Preferably, when calculating the orthogonal loss, the weight of the orthogonal loss can be dynamically adjusted according to the size of the feature matrix and the correlation between features. For example, if the feature matrix is large and the correlation between features is high, the weight of the orthogonal loss can be appropriately increased to impose the orthogonal constraint more strictly. In the total loss function, in addition to the cross-entropy loss function and the orthogonal loss, other types of loss functions, such as regularization terms, can be introduced to further improve the generalization ability of the model. For example, an L2 regularization term can be added to penalize the weights of the model and prevent overfitting. In addition, different optimization algorithms can be tried to minimize the total loss function, such as the Adam optimization algorithm or the RMSprop optimization algorithm, to improve the training efficiency and convergence speed of the model.

[0077] The above-mentioned various embodiments of the present invention have the following beneficial effects: Through the heterogeneous data integration module, the nursing system obtains multi-source heterogeneous electronic medical record data from the hospital HIS system, wearable monitoring devices, and nursing operation terminals, and outputs the data after cleaning and standardization, which can provide a data basis with a unified format for subsequent processing. The feature association module receives the data after cleaning and standardization, generates a spatio-temporal feature association matrix, and stores the feature topology map, which can mine the spatio-temporal association information in the data and provide a basis for accurate nursing decisions. The dynamic decision-making module processes the feature association matrix based on the deep spatio-temporal fusion network to generate a nursing decision vector, which can realize intelligent nursing decisions. The scheme execution monitoring module collects the nursing operation data stream in real time, calculates the execution deviation parameter, and feeds it back, which can timely detect the deviation in the nursing operation. The abnormal intervention module analyzes the execution deviation parameter and triggers the dynamic decision-making module to execute the nursing plan reconstruction, which can ensure the dynamic adjustment and optimization of the nursing plan. The multi-terminal collaboration module converts the nursing decision vector into a device instruction set and feeds back the device status, which can realize the collaborative work of nursing devices and improve the nursing efficiency. The secure storage module encrypts and stores the feature topology map and the decision knowledge graph and provides model parameter version management, which can ensure the security and reliability of the data.

[0078] In terms of data processing, the system uses a Markov chain model to estimate missing data and perform data normalization, which can improve the integrity and accuracy of the data. The feature association module can better capture the spatio-temporal features in the data by constructing a three-dimensional feature tensor and calculating the spatio-temporal attention weight matrix. The three-dimensional convolutional network performs dynamic kernel update, which can enhance the model's adaptability to feature changes. The processing flow of the deep spatio-temporal fusion network includes extracting physiological feature vectors and environmental parameter vectors, calculating cross-modal association weights, performing multi-scale convolutional processing, and gating and fusing multi-scale features, which can achieve deep fusion and effective utilization of multi-modal data. The scheme execution monitoring module uses dynamic time warping and window deviation calculation methods when calculating the execution deviation, which can more accurately evaluate the execution of nursing operations. The abnormal intervention module calculates the cumulative deviation within the window and triggers scheme reconstruction by setting the sliding window length and time decay constant, which can respond to abnormal situations in a timely manner and adjust the nursing plan. The multi-terminal collaboration module considers the instruction priority and conflict resolution mechanism when performing instruction arbitration, which can ensure the reasonable execution of instructions. The secure storage module encrypts and stores the knowledge graph using double encryption and integrity verification tags, which can improve the security and integrity of the data. The gating fusion step applies orthogonal constraints, which can optimize the performance of the model.

[0079] Furthermore, the storage medium of the embodiment of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.

[0080] The above description is only some preferred embodiments of the present invention and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A nursing system based on electronic medical records, characterized in that, Including: A heterogeneous data integration module, which is used to obtain multi-source heterogeneous electronic medical record data from the hospital HIS system, wearable monitoring devices and nursing operation terminals, clean and standardize the data, and then output it to the feature association module; A feature association module, which is used to receive the data after cleaning and standardization, generate a spatio-temporal feature association matrix, output the matrix to the dynamic decision-making module, and store the feature topology map in the secure storage module at the same time; A dynamic decision-making module, which is used to process the spatio-temporal feature association matrix based on a deep spatio-temporal fusion network, generate a nursing decision vector, and distribute the vector to the scheme execution monitoring module and the multi-terminal collaboration module; A scheme execution monitoring module, which is used to collect the nursing operation data stream in real time, calculate the execution deviation parameter, and feedback it to the abnormal intervention module and the dynamic decision-making module; An abnormal intervention module, which is used to analyze the execution deviation parameter, and trigger the dynamic decision-making module to reconstruct the nursing scheme when the deviation exceeds the threshold; A multi-terminal collaboration module, which is used to convert the nursing decision vector into a device instruction set, arbitrate conflicting instructions and then distribute them to the terminal devices, and feedback the device status to the dynamic decision-making module; A secure storage module, which is used to encrypt and store the feature topology map and the decision knowledge graph, and provide model parameter version management for the dynamic decision-making module.

2. The system according to claim 1, characterized in that The steps for the heterogeneous data integration module to perform data cleaning include: Based on the Markov chain model, estimate the missing data through the previous state transition probability, historical mean and Gaussian noise; Perform piecewise non-linear normalization on the complete data. When the data value is greater than or equal to the preset reference threshold, map it to the [0,1] interval; when the data value is less than the reference threshold, map it to the [0,0.5] interval.

3. The system according to claim 1, characterized in that, The steps for the feature association module to perform include: Construct a three-dimensional feature tensor from the data after cleaning and standardization according to the time slice, ward area and physiological index dimensions; Calculate the spatio-temporal attention weight matrix based on the three-dimensional feature tensor, as shown in the following formula: ， Among them, represents the query vector of the i-th time slice, represents the key vector of the j-th spatial position, represents the query projection matrix, represents the key projection matrix, represents the scaling factor, the three-dimensional feature tensor , where 24 represents the number of 24-hour time slices, W represents the total number of ward areas, and C = 12 represents the number of physiological index dimensions; After splicing the spatio-temporal attention weight matrix and the three-dimensional feature tensor, input them into a three-dimensional convolutional network for feature extraction.

4. The system according to claim 3, wherein The steps for the three-dimensional convolutional network to perform dynamic kernel update include: Monitor the feature change amount of the three-dimensional feature tensor in real time; Dynamically adjust the convolutional kernel weight through a gated recurrent unit, where the current weight is jointly determined by the previous moment weight and the one-dimensional vector of the feature change amount.

5. The system according to claim 1, characterized in that, The processing flow of the deep spatio-temporal fusion network includes: Extract the physiological feature vector and the environmental parameter vector from the spatio-temporal feature association matrix respectively; Splice the physiological feature vector and the environmental parameter vector, and calculate the cross-modal association weight through a learnable parameter matrix; Based on the cross-modal association weight, perform multi-scale convolution processing on the physiological feature vector, and extract features with different granularities using three convolution kernel sizes of large, medium and small respectively; Fuse the multi-scale features through a gating mechanism to generate the final nursing decision vector.

6. The system according to claim 1, wherein The steps for the scheme execution monitoring module to calculate the execution deviation parameter include: Decompose the actual nursing operation into an atomic event sequence and align it with the preset standard event sequence; Calculate the minimum path difference between the actual sequence and the standard sequence through the dynamic time warping algorithm; Based on the minimum path difference and the sequence lengths of the actual sequence and the standard sequence, calculate the window deviation through an exponential decay function to quantify the operation deviation.

7. The system according to claim 1, characterized in that When the deviation exceeds the threshold, triggering the dynamic decision module to reconstruct the care plan includes: Set the sliding window length and the time decay weight coefficient; Within the sliding window, accumulate the window deviation based on the time decay weight to obtain the cumulative deviation; When both the cumulative deviation and the current window deviation exceed the preset threshold, trigger the dynamic decision module to reconstruct the care plan.

8. The system according to claim 1, wherein The multi-terminal collaboration module converting the care decision vector into a device instruction set includes the following steps: Extract the instruction feature vector from the care decision vector and obtain the instruction delay time; Calculate the urgency component based on the delay time, calculate the importance component based on the instruction feature vector and the preset importance coefficient, and perform weighted summation to obtain the instruction priority; If there are concurrent conflicting instructions, select the instruction with shorter execution time or higher priority according to the priority difference.

9. The system according to claim 5, wherein Fusing multi-scale features through a gating mechanism to generate the final care decision vector includes: Calculate the orthogonality loss between the multi-scale feature matrices, and the orthogonality loss is the sum of the squares of the Frobenius norms of the transposed products of the feature matrices of each scale.

10. The system according to claim 9, characterized in that, The total loss function for fusing multi-scale features through the gating mechanism includes the weighted sum of the cross-entropy loss and the orthogonality loss, where the weight of the orthogonality loss is 0.01.

Images