Early warning system for patients with critical illness based on MEWS score

Abstract

This invention discloses an early warning system for critically ill patients based on the MEWS score, comprising a data acquisition module, a data preprocessing module, a module for establishing an early warning model for critically ill patients, a module for searching the optimal parameters of the model, and a specific operation module. This invention relates to the field of patient condition early warning technology, specifically to an early warning system for critically ill patients based on the MEWS score. This scheme improves the model's feature learning ability and accuracy, accelerates training convergence speed, and enhances the model's generalization ability by designing dense layers, batch normalization layers, random inactivation layers, and classification layers. Furthermore, it improves the algorithm's search efficiency and avoids getting trapped in local optima by employing search initialization, initial search point generation strategies, corrected probability distribution variance, corrected probability distribution, and a balanced long and short step size search function, thereby enhancing the algorithm's search capability.

Description

Technical Field

[0001] This invention relates to the field of patient condition early warning technology, specifically to an early warning system for critically ill patients based on the MEWS score. Background Technology

[0002] An early warning system for critically ill patients based on the MEWS score utilizes machine learning and data analysis techniques. It builds a model based on the MEWS score and patient data to assess the severity of the patient's condition, enabling early detection and intervention for critically ill patients and providing timely and accurate warning information for medical staff. However, traditional early warning systems for critically ill patients suffer from problems such as weak feature learning ability, slow training convergence speed, weak generalization ability, and low accuracy due to the difficulty in capturing complex patterns in the data; traditional model parameter search algorithms also suffer from low search efficiency, susceptibility to getting trapped in local optima, and weak search capability. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an early warning system for critically ill patients based on the MEWS score. Addressing the problems of traditional critically ill patient warning systems, such as weak feature learning ability, slow training convergence speed, weak generalization ability, and low accuracy due to the difficulty in capturing complex patterns in data, this solution improves the model's feature learning ability and accuracy, accelerates training convergence speed, and enhances generalization ability by designing dense layers, batch normalization layers, random deactivation layers, and classification layers. Furthermore, addressing the problems of low search efficiency, susceptibility to local optima, and weak search ability in traditional model parameter search algorithms, this solution improves the algorithm's search efficiency, avoids getting trapped in local optima, and enhances search ability through search initialization, initial search point generation strategies, correction of probability distribution variance, correction of probability distribution, and a balanced long and short step size search function.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an early warning system for critically ill patients based on MEWS score, including a data acquisition module, a data preprocessing module, a module for establishing an early warning model for critically ill patients, a module for searching the optimal parameters of the model, and a specific operation module;

[0005] The data acquisition module collects the patient's vital signs data, laboratory test data, medical history, medical operation records, and critical condition warning category data. The vital signs data include heart rate, respiratory rate, body temperature, systolic blood pressure, and level of consciousness. The laboratory test data includes blood, urine, and imaging test results. The medical operation records include operation records of infusion, medication, surgery, and nursing measures.

[0006] The data preprocessing module processes the collected data through data cleaning, feature selection, data transformation, MEWS score calculation, and feature and label setting, specifically including the following:

[0007] Data cleaning: The collected data is cleaned, missing values ​​are filled with the mean, median and mode; categorical features are filled with the most frequent category; outliers are deleted and replaced with the mean, median and mode; and duplicate samples and features are removed.

[0008] Feature selection: Select features that are highly relevant to the early warning and prediction of critically ill patients, and remove irrelevant and redundant features to reduce the impact of irrelevant and redundant features on the model;

[0009] Data transformation: The original data is transformed to create new feature data. Health level feature data is created based on vital sign data, treatment activity feature data is created based on medical operation record data, and physical fitness feature data is created based on laboratory test data and medical history data. At the same time, patient data is standardized and normalized to ensure that patient data has the same scale and range.

[0010] The MEWS score is calculated by collecting vital sign data, including heart rate, respiratory rate, body temperature, systolic blood pressure, and level of consciousness. A scoring strategy is created and scores are assigned. The corresponding scores are added together to obtain a total score, and the corresponding MEWS score for the patient is calculated.

[0011] Set features and labels, and set MEWS score, vital signs data, laboratory test data, medical history, medical operation records, health level feature data, treatment activity feature data, and physical strength feature data as feature data; set critical illness warning category as label data.

[0012] Furthermore, the module for establishing an early warning model for critically ill patients specifically includes the following components: a dense layer, a batch normalization layer, a random inactivation layer, and a classification layer.

[0013] The design model structure consists of four layers: a dense layer, a batch normalization layer, a random deactivation layer, and a classification layer.

[0014] The design incorporates dense layers that perform linear transformations to capture complex patterns in the data. It comprises 512 neurons in total, arranged in three layers with 170, 170, and 172 neurons respectively, as shown below:

[0015] ;

[0016] Where re(·) represents the ReLU activation function, x1 represents the dense layer input, ω1, ω2, ω3, β1, β2, β3 represent the weights and biases of the dense layer, and y1 represents the output of the dense layer;

[0017] Design a batch normalization layer to normalize the input data, as shown below:

[0018] ;

[0019] Where y2 is the output of the normalization layer, x2 represents the input of the batch normalization layer, E(x2) represents the mean of the input, var(x2) represents the variance of the input, ε represents the division-by-zero smoothing term, γ represents the scaling parameter, and Γ represents the bias parameter.

[0020] Design a random deactivation layer, randomly shutting down a portion of neurons to maintain robust learning, as shown below:

[0021] ;

[0022] Where y3 represents the output of the random deactivation layer, x3 represents the input of the random deactivation layer, rand(0,1) represents a random number between 0 and 1, and P dropout This represents the random inactivation rate. Setting the inactivation rate to 0.2 means that neurons are randomly inactivated with a 20% probability to ensure proper fitting.

[0023] The design incorporates a classification layer. By calculating the probability associated with the severity warning category, the original scores are transformed into a probability distribution, thereby achieving the final prediction, as shown below:

[0024] ;

[0025] Where y5 represents the classification warning result, x5 represents the classification layer input, j represents the j-th warning category, k represents the index of the warning category, and K represents the total number of warning categories. ω is the probability that input data x5 is classified as warning category j. f1 ω f2 β f1 β f2 This represents the weights and biases of the classification layer.

[0026] Furthermore, the optimal parameter search module for the model searches for optimal parameters by designing an initial search point generation strategy, designing a corrected probability distribution variance, designing a corrected probability distribution, and designing a long and short step size balancing search function. Specifically, this includes the following:

[0027] Search initialization, setting parameters to optimize the search space, and setting the negative of the error of the early warning model for critically ill patients as the individual fitness value of the parameter;

[0028] The initial search point generation strategy is designed to generate Q search points and construct a parameter search cluster, as shown below:

[0029] ;

[0030] in, Let represent the position of the q-th initial search point, up represent the upper boundary of the search space, and low represent the lower boundary of the search space. The origin parameter is represented by q, where q represents the index of the generated point, and Q represents the maximum value of the generated point.

[0031] The corrected probability distribution variance is represented as follows:

[0032] ;

[0033] Where τ represents the power law exponent, sin(·) represents the sine function, and Γ(·) represents the gamma function;

[0034] The modified probability distribution is designed as follows:

[0035] ;

[0036] Where Dbution[] represents a random number obtained from a modified probability distribution, U represents a random number sampled from a Gaussian distribution, and O represents a Gaussian distribution. Random number samples in;

[0037] Design a long-short step balance search function, expressed as follows:

[0038] ;

[0039] in, and Let X represent the absolute values ​​of two random numbers obtained from the corrected probability distribution, t represent the number of searches, and X represent the number of searches. t+1 X represents the position found by the long and short step size balanced search function in the (t+1)th search. t X represents the position found by the long and short step balance search function in the t-th search. best X represents the position with the highest current fitness value. worst X represents the position with the minimum fitness value. ave Indicates the average position of the initial search cluster;

[0040] The model parameters are optimized through iterative search. A fitness threshold is pre-set, and the fitness values ​​of each point in the search cluster are calculated. The positions with the highest and lowest fitness values ​​are selected and set. A position search using a balanced long and short step search function is performed on each point. The fitness value of the searched points is calculated, and the point with the highest fitness value is selected. If the fitness value of this point is greater than the fitness threshold, it is updated to the position with the highest fitness value. Simultaneously, the point with the lowest fitness value is selected. If the fitness value of this point is less than the fitness threshold, it is updated to the position with the lowest fitness value. If more than 75% of the points in the same parallel search of the search cluster have fitness values ​​greater than the fitness threshold, the parameters of the position with the highest fitness value are selected as the optimal parameters and the search stops. If the maximum number of searches is reached, the parameters are re-initialized and searched again; otherwise, the iterative search continues.

[0041] Furthermore, the specific operating module is that the model, based on the input characteristic data of the patients to be monitored, derives the early critical illness warning category of the patients and provides warning information; when the early critical illness warning category given by the model reaches the high level, the system issues an alarm, prompting medical staff to take emergency medical intervention measures, increase monitoring frequency, adjust treatment plans, and increase vigilance, providing medical staff with reference and decision support, discovering and handling dangerous situations of patients, ensuring patient safety, and returning the data to the model for continuous updating and improvement of the model.

[0042] The beneficial effects achieved by the present invention using the above solution are as follows:

[0043] (1) In view of the problems of weak feature learning ability, slow training convergence speed, weak generalization ability and low accuracy of traditional critical illness patient early warning system due to difficulty in capturing complex patterns in data, this solution helps the model to better adapt to data features by designing dense layer, batch normalization layer, random deactivation layer and classification layer, thereby improving the feature learning ability and accuracy of the model, accelerating the training convergence speed and improving the generalization ability of the model.

[0044] (2) In view of the problems of low search efficiency, easy to get trapped in local optima and weak search ability of traditional model parameter search algorithms, this scheme improves the search efficiency of the algorithm, avoids getting trapped in local optima and improves the search ability of the algorithm by search initialization, designing initial search point generation strategy, designing corrected probability distribution variance, designing corrected probability distribution and designing long and short step size balanced search function. Attached Figure Description

[0045] Figure 1 A schematic diagram of an early warning system for critically ill patients based on MEWS score provided by the present invention;

[0046] Figure 2 This is a schematic diagram of the data preprocessing module;

[0047] Figure 3 A schematic diagram of a module for establishing an early warning model for critically ill patients;

[0048] Figure 4 This is a schematic diagram of the module for searching the optimal parameters of the model.

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0051] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0052] Example 1, see Figure 1 The present invention provides an early warning system for critically ill patients based on MEWS score, including a data acquisition module, a data preprocessing module, a module for establishing an early warning model for critically ill patients, a module for searching the optimal parameters of the model, and a specific operation module;

[0053] The data acquisition module collects the patient's vital signs data, laboratory test data, medical history, medical operation records, and critical illness warning category data;

[0054] The data preprocessing module performs data cleaning, feature selection, and data transformation on the raw data, calculates the MEWS score, and calibrates the data.

[0055] The module for establishing an early warning model for critically ill patients constructs an early warning model for critically ill patients by designing a dense layer, a batch normalization layer, a random inactivation layer, and a classification layer.

[0056] The optimal parameter search module searches for the optimal parameters of the model by designing an initial search point generation strategy, designing a corrected probability distribution variance, designing a corrected probability distribution, and designing a long and short step size balancing search function.

[0057] The specific operating model is that the module determines the critical condition warning category of the patient based on the input characteristic information of the patient that needs to be monitored, and provides the corresponding warning information.

[0058] Example 2, see Figure 1 This embodiment is based on the above embodiment. The data acquisition module collects the patient's vital signs data, laboratory test data, medical history, medical operation records, and critical illness warning category data. The vital signs data include heart rate, respiratory rate, body temperature, systolic blood pressure, and level of consciousness. The laboratory test data includes blood, urine, and imaging test results. The medical operation records include operation records of infusion, medication, surgery, and nursing measures.

[0059] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The data preprocessing module processes the collected data through data cleaning, feature selection, data transformation, calculation of MEWS scores, and setting features and labels. Specifically, it includes the following:

[0060] Data cleaning: The collected data is cleaned, missing values ​​are filled with the mean, median and mode; categorical features are filled with the most frequent category; outliers are deleted and replaced with the mean, median and mode; and duplicate samples and features are removed.

[0061] Feature selection: Select features that are highly relevant to the early warning and prediction of critically ill patients, and remove irrelevant and redundant features to reduce the impact of irrelevant and redundant features on the model;

[0062] Data transformation: The original data is transformed to create new feature data. Health level feature data is created based on vital sign data, treatment activity feature data is created based on medical operation record data, and physical fitness feature data is created based on laboratory test data and medical history data. At the same time, patient data is standardized and normalized to ensure that patient data has the same scale and range.

[0063] The MEWS score is calculated by collecting vital sign data, including heart rate, respiratory rate, body temperature, systolic blood pressure, and level of consciousness. A scoring strategy is created and scores are assigned. The corresponding scores are added together to obtain a total score, and the corresponding MEWS score for the patient is calculated.

[0064] Set features and labels, and set MEWS score, vital signs data, laboratory test data, medical history, medical operation records, health level feature data, treatment activity feature data, and physical strength feature data as feature data; set critical illness warning category as label data.

[0065] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. The module for establishing an early warning model for critically ill patients specifically includes the following: It constructs an early warning model for critically ill patients by designing a dense layer, a batch normalization layer, a random inactivation layer, and a classification layer.

[0066] The design model structure consists of four layers: a dense layer, a batch normalization layer, a random deactivation layer, and a classification layer.

[0067] The design incorporates dense layers that perform linear transformations to capture complex patterns in the data. It comprises 512 neurons in total, arranged in three layers with 170, 170, and 172 neurons respectively, as shown below:

[0068] ;

[0069] Where re(·) represents the ReLU activation function, x1 represents the dense layer input, ω1, ω2, ω3, β1, β2, β3 represent the weights and biases of the dense layer, and y1 represents the output of the dense layer;

[0070] Design a batch normalization layer to normalize the input data, as shown below:

[0071] ;

[0072] Where y2 is the output of the normalization layer, x2 represents the input of the batch normalization layer, E(x2) represents the mean of the input, var(x2) represents the variance of the input, ε represents the division-by-zero smoothing term, γ represents the scaling parameter, and Γ represents the bias parameter.

[0073] Design a random deactivation layer, randomly shutting down a portion of neurons to maintain robust learning, as shown below:

[0074] ;

[0075] Where y3 represents the output of the random deactivation layer, x3 represents the input of the random deactivation layer, rand(0,1) represents a random number between 0 and 1, and P dropout This represents the random inactivation rate. Setting the inactivation rate to 0.2 means that neurons are randomly inactivated with a 20% probability to ensure proper fitting.

[0076] The design incorporates a classification layer. By calculating the probability associated with the severity warning category, the original scores are transformed into a probability distribution, thereby achieving the final prediction, as shown below:

[0077] ;

[0078] Where y5 represents the classification warning result, x5 represents the classification layer input, j represents the j-th warning category, k represents the index of the warning category, and K represents the total number of warning categories. ω is the probability that input data x5 is classified as warning category j. f1 ω f2 β f1 β f2 This represents the weights and biases of the classification layer.

[0079] By performing the above operations, this solution addresses the problems of weak feature learning ability, slow training convergence speed, weak generalization ability, and low accuracy in traditional critical illness patient early warning systems due to the difficulty in capturing complex patterns in the data. This is achieved by designing dense layers, batch normalization layers, random deactivation layers, and classification layers to help the model better adapt to data features, thereby improving the model's feature learning ability and accuracy, accelerating the training convergence speed, and improving the model's generalization ability.

[0080] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. The optimal parameter search module searches for the optimal parameters of the model by designing an initial search point generation strategy, designing a corrected probability distribution variance, designing a corrected probability distribution, and designing a long and short step balance search function. Specifically, it includes the following:

[0081] Search initialization, setting parameters to optimize the search space, and setting the negative of the error of the early warning model for critically ill patients as the individual fitness value of the parameter;

[0082] The initial search point generation strategy is designed to generate Q search points and construct a parameter search cluster, as shown below:

[0083] ;

[0084] in, Let represent the position of the q-th initial search point, up represent the upper boundary of the search space, and low represent the lower boundary of the search space. The origin parameter is represented by q, where q represents the index of the generated point, and Q represents the maximum value of the generated point.

[0085] The corrected probability distribution variance is represented as follows:

[0086] ;

[0087] Where τ represents the power law exponent, sin(·) represents the sine function, and Γ(·) represents the gamma function;

[0088] The modified probability distribution is designed as follows:

[0089] ;

[0090] Where Dbution[] represents a random number obtained from a modified probability distribution, U represents a random number sampled from a Gaussian distribution, and O represents a Gaussian distribution. Random number samples in;

[0091] Design a long-short step balance search function, expressed as follows:

[0092] ;

[0093] in, and Let X represent the absolute values ​​of two random numbers obtained from the corrected probability distribution, t represent the number of searches, and X represent the number of searches. t+1 X represents the position found by the long and short step size balanced search function in the (t+1)th search. t X represents the position found by the long and short step balance search function in the t-th search. best X represents the position with the highest current fitness value. worst X represents the position with the minimum fitness value. ave Indicates the average position of the initial search cluster;

[0094] The model parameters are optimized through iterative search. A fitness threshold is pre-set, and the fitness values ​​of each point in the search cluster are calculated. The positions with the highest and lowest fitness values ​​are selected and set. A position search using a balanced long and short step search function is performed on each point. The fitness value of the searched points is calculated, and the point with the highest fitness value is selected. If the fitness value of this point is greater than the fitness threshold, it is updated to the position with the highest fitness value. Simultaneously, the point with the lowest fitness value is selected. If the fitness value of this point is less than the fitness threshold, it is updated to the position with the lowest fitness value. If more than 75% of the points in the same parallel search of the search cluster have fitness values ​​greater than the fitness threshold, the parameters of the position with the highest fitness value are selected as the optimal parameters and the search stops. If the maximum number of searches is reached, the parameters are re-initialized and searched again; otherwise, the iterative search continues.

[0095] By performing the above operations, this solution addresses the problems of low search efficiency, susceptibility to local optima, and weak search capability in traditional model parameter search algorithms. It improves the algorithm's search efficiency, avoids getting trapped in local optima, and enhances its search capability through search initialization, designing an initial search point generation strategy, designing a corrected probability distribution variance, designing a corrected probability distribution, and designing a balanced long and short step size search function.

[0096] Example 6, see Figure 1 The specific operating module is that the model determines the early critical illness warning category of the patient based on the input characteristic data of the patient that needs to be monitored, and provides warning information; when the early critical illness warning category given by the model reaches the high level, the system issues an alarm, prompting medical staff to take emergency medical intervention measures, increase monitoring frequency, adjust treatment plan, and increase vigilance, providing medical staff with reference and decision support, discovering and handling dangerous situations of patients, ensuring patient safety, and returning data to the model for continuous updating and improvement of the model.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0099] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An early warning system for critically ill patients based on MEWS scores, characterized in that: It includes a data acquisition module, a data preprocessing module, a module for establishing an early warning model for critically ill patients, a module for searching the optimal parameters of the model, and a specific operation module; The data acquisition module collects the patient's vital signs data, laboratory test data, medical history, medical operation records, and critical illness warning category data; The data preprocessing module performs data cleaning, feature selection, and data transformation on the raw data, calculates the MEWS score, and calibrates the data. The module for establishing an early warning model for critically ill patients constructs an early warning model for critically ill patients by designing a dense layer, a batch normalization layer, a random inactivation layer, and a classification layer. The optimal parameter module for the search model searches for optimal parameters by designing an initial search point generation strategy, designing a corrected probability distribution variance, designing a corrected probability distribution, and designing a long and short step size balancing search function. The specific operating model is that the module determines the critical condition warning category of the patient based on the input characteristic information of the patient that needs to be monitored, and provides the corresponding warning information.

2. The early warning system for critically ill patients based on MEWS score according to claim 1, characterized in that: The optimal parameter search module searches for optimal model parameters by designing an initial search point generation strategy, a corrected probability distribution variance, a corrected probability distribution, and a balanced long and short step size search function. Specifically, this includes the following: Search initialization, setting parameters to optimize the search space, and setting the negative of the error of the early warning model for critically ill patients as the individual fitness value of the parameter; The initial search point generation strategy is designed to generate Q search points and construct a parameter search cluster, as shown below: ； in, Let represent the position of the q-th initial search point, up represent the upper boundary of the search space, and low represent the lower boundary of the search space. The origin parameter is represented by q, where q represents the index of the generated point, and Q represents the maximum value of the generated point. The corrected probability distribution variance is represented as follows: ； Where τ represents the power law exponent, sin(·) represents the sine function, and Γ(·) represents the gamma function; The modified probability distribution is designed as follows: ； Where Dbution[] represents a random number obtained from a modified probability distribution, U represents a random number sampled from a Gaussian distribution, and O represents a Gaussian distribution. Random number samples in; Design a long-short step balance search function, expressed as follows: ； in, and Let X represent the absolute values ​​of two random numbers obtained from the corrected probability distribution, t represent the number of searches, and X represent the number of searches. t+1 X represents the position found by the long and short step size balanced search function in the (t+1)th search. t X represents the position found by the long and short step balance search function in the t-th search. best X represents the position with the highest current fitness value. worst X represents the position with the minimum fitness value. ave Indicates the average position of the initial search cluster; The model parameters are optimized through iterative search. A fitness threshold is pre-set, and the fitness values ​​of each point in the search cluster are calculated. The positions with the highest and lowest fitness values ​​are selected and set. A position search using a balanced long and short step search function is performed on each point. The fitness value of the searched points is calculated, and the point with the highest fitness value is selected. If the fitness value of this point is greater than the fitness threshold, it is updated to the position with the highest fitness value. Simultaneously, the point with the lowest fitness value is selected. If the fitness value of this point is less than the fitness threshold, it is updated to the position with the lowest fitness value. If more than 75% of the points in the same parallel search of the search cluster have fitness values ​​greater than the fitness threshold, the parameters of the position with the highest fitness value are selected as the optimal parameters and the search stops. If the maximum number of searches is reached, the parameters are re-initialized and searched again; otherwise, the iterative search continues.

3. The early warning system for critically ill patients based on MEWS score according to claim 1, characterized in that: The module for establishing an early warning model for critically ill patients specifically includes the following components: a dense layer, a batch normalization layer, a random inactivation layer, and a classification layer. The design model structure consists of four layers: a dense layer, a batch normalization layer, a random deactivation layer, and a classification layer. The design incorporates dense layers that perform linear transformations to capture complex patterns in the data. It comprises 512 neurons in total, arranged in three layers with 170, 170, and 172 neurons respectively, as shown below: ； Where re(·) represents the ReLU activation function, x1 represents the dense layer input, ω1, ω2, ω3, β1, β2, β3 represent the weights and biases of the dense layer, and y1 represents the output of the dense layer; Design a batch normalization layer to normalize the input data, as shown below: ； Where y2 is the output of the normalization layer, x2 represents the input of the batch normalization layer, E(x2) represents the mean of the input, var(x2) represents the variance of the input, ε represents the division-by-zero smoothing term, γ represents the scaling parameter, and Γ represents the bias parameter. Design a random deactivation layer, randomly shutting down a portion of neurons to maintain robust learning, as shown below: ； Where y3 represents the output of the random deactivation layer, x3 represents the input of the random deactivation layer, rand(0,1) represents a random number between 0 and 1, and P dropout This represents the random inactivation rate. Setting the inactivation rate to 0.2 means that neurons are randomly inactivated with a 20% probability to ensure proper fitting. The design incorporates a classification layer. By calculating the probability associated with the severity warning category, the original scores are transformed into a probability distribution, thereby achieving the final prediction, as shown below: ； Where y5 represents the classification warning result, x5 represents the classification layer input, j represents the j-th warning category, k represents the index of the warning category, and K represents the total number of warning categories. ω is the probability that input data x5 is classified as warning category j. f1 ω f2 β f1 β f2 This represents the weights and biases of the classification layer.

4. The early warning system for critically ill patients based on MEWS score according to claim 1, characterized in that: The data preprocessing module processes the collected data through data cleaning, feature selection, data transformation, MEWS score calculation, and feature and label setting, specifically including the following: Data cleaning: The collected data is cleaned, missing values ​​are filled with the mean, median and mode; for categorical features, the most frequent category is used for filling; for obvious outliers, they are deleted and replaced with the mean, median and mode; the data is deduplicated, and duplicate samples and features are deleted. Feature selection: Select features that are highly relevant to the early warning and prediction of critically ill patients, and remove irrelevant and redundant features to reduce the impact of irrelevant and redundant features on the model; Data transformation: The original data is transformed to create new feature data. Health level feature data is created based on vital sign data, treatment activity feature data is created based on medical operation record data, and physical fitness feature data is created based on laboratory test data and medical history data. At the same time, patient data is standardized and normalized to ensure that patient data has the same scale and range. The MEWS score is calculated by collecting vital sign data, including heart rate, respiratory rate, body temperature, systolic blood pressure, and level of consciousness. A scoring strategy is created and scores are assigned. The corresponding scores are added together to obtain a total score, and the corresponding MEWS score for the patient is calculated. Set features and labels, and set MEWS score, vital signs data, laboratory test data, medical history, medical operation records, health level feature data, treatment activity feature data, and physical strength feature data as feature data; set critical illness warning category as label data.

5. The early warning system for critically ill patients based on MEWS score according to claim 1, characterized in that: The data acquisition module collects the patient's vital signs data, laboratory test data, medical history, medical operation records, and critical condition warning category data. The vital signs data include heart rate, respiratory rate, body temperature, systolic blood pressure, and level of consciousness. The laboratory test data includes blood, urine, and imaging test results. The medical operation records include operation records of infusion, medication, surgery, and nursing measures.

6. The early warning system for critically ill patients based on MEWS score according to claim 1, characterized in that: The specific operating module is that the model determines the early critical illness warning category of the patient based on the input characteristic data of the patient to be monitored, and provides warning information; when the early critical illness warning category given by the model reaches the high level, the system issues an alarm, prompting medical staff to take emergency medical intervention measures, increase monitoring frequency, adjust treatment plan, and increase vigilance, providing medical staff with reference and decision support, discovering and handling dangerous situations of patients, ensuring patient safety, and returning data to the model for continuous updating and improvement of the model.

Images