Intensive care unit early warning system based on artificial intelligence

Abstract

The invention relates to the technical field of medical monitoring, in particular to an intensive care unit early warning system based on artificial intelligence, which comprises a data acquisition module, a data transmission module, a data center, a physical sign alarm module, a health prediction module, a risk early warning module and an auxiliary decision making module, in order to solve the problem of high false alarm rate for monitoring vital sign abnormity of a critical patient, an abnormity detection scoring algorithm is adopted in a sign alarm module, and various vital sign data monitored in real time are scored comprehensively through a risk threshold value, an influence factor and an alarm correct rate; in order to solve the problem that the future vital sign trend of a critical patient is difficult to accurately predict, a health prediction model based on Transform-BiGRU is designed, physical sign features, statistical features and medical features are comprehensively learned by using Transform, and the ability of capturing a bidirectional dependency relationship is further enhanced through BiGRU, so that the accuracy of model prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical monitoring technology, and in particular to an intensive care unit early warning system based on artificial intelligence. Background Art

[0002] With the development of medical technology, the level of patient monitoring in intensive care units has been significantly improved; however, traditional monitoring methods have limitations in data processing and real-time early warning; the introduction of artificial intelligence technology, especially machine learning and deep learning algorithms, has provided new solutions for intensive care units; by analyzing a large amount of clinical data, artificial intelligence can help doctors promptly detect subtle changes in the condition of critically ill patients, thereby providing early warning; however, the current artificial intelligence-based intensive care unit early warning system faces a high false alarm rate when monitoring abnormal vital signs of critically ill patients, and it is difficult to accurately predict the future trend of critically ill patients' vital signs, thus affecting the allocation of manpower and medical resources. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an intensive care unit early warning system based on artificial intelligence. In view of the high false alarm rate problem when monitoring abnormal vital signs of critically ill patients, the present invention proposes a vital sign alarm module, and adopts an abnormal detection scoring algorithm in the vital sign alarm module. The vital sign data are scored by risk threshold, influencing factor and alarm accuracy. When the total score threshold or the single score threshold is exceeded, an alarm is issued, and a preliminary review is conducted by the management personnel, and then the rapid response team is mobilized for timely rescue. In view of the problem that it is difficult to accurately predict the future trend of vital signs of critically ill patients, thereby affecting the allocation of manpower and medical resources, the present invention proposes a health prediction model based on Transformer-BiGRU, which combines vital sign features, statistical features and medical features, uses Transformer to learn the vital sign data of critically ill patients, and further captures the bidirectional dependency through BiGRU to improve the accuracy of model prediction.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intensive care unit warning system based on artificial intelligence, which specifically includes: a data acquisition module, a data transmission module, a data center, a vital sign alarm module, a health prediction module, a risk warning module and an auxiliary decision module, the output end of the data acquisition module is connected to the input end of the data transmission module, the output end of the data transmission module is connected to the input end of the data center, the output end of the data center is respectively connected to the input end of the health prediction module and the input end of the vital sign alarm module, the output end of the vital sign alarm module is connected to the input end of the auxiliary decision module, the output end of the health prediction module is connected to the input end of the risk warning module, and the output end of the risk warning module is connected to the input end of the auxiliary decision module;

[0005] The data acquisition module collects vital sign data;

[0006] The data transmission module transmits vital sign data in real time;

[0007] The data center uses a big data platform to store vital signs data;

[0008] The vital sign alarm module adopts an abnormal detection and scoring algorithm to monitor the vital sign data in real time and generate alarm information and alarm when abnormal;

[0009] The health prediction module establishes a health prediction model based on Transformer-BiGRU;

[0010] The risk warning module uses the health prediction model to establish a trend chart of vital signs for each critically ill patient in the next 24 hours and identify high-risk periods;

[0011] The auxiliary decision-making module summarizes the vital sign data, alarm information and vital sign trend chart of critically ill patients, and helps doctors establish risk assessment scales according to different diseases of critically ill patients;

[0012] The data acquisition module collects the vital signs data of critically ill patients through a data collector and a life support medical device: the data collector is connected to the life support medical device through an RS32 serial port and a network port, and the vital signs data of the critically ill patients are obtained by parsing the communication protocol of the life support medical device;

[0013] The vital signs data include respiratory data, blood oxygen data, blood pressure data, body temperature data and heart rate data;

[0014] The data transmission module transmits the vital signs data to the data center via Bluetooth, 5G, and Wifi technology;

[0015] The data center is deployed in the cloud, and a Hadoop big data platform is built and deployed in the cloud to store the vital signs data in a distributed manner and call them through an API interface.

[0016] The specific steps of the anomaly detection scoring algorithm are as follows:

[0017] D1: Define the upper and lower risk thresholds for each vital sign data. If each vital sign data exceeds the corresponding upper and lower risk thresholds, 1 point will be awarded.

[0018] D2: Define the impact factors of the accumulated time and each vital sign data, multiply the scores of each vital sign data within the accumulated time calculated by D1 by the corresponding impact factors to obtain the product result;

[0019] D3: Count the alarm accuracy rates of each vital sign data, multiply the product result in D2 by the alarm accuracy rate, and obtain the individual scores of each vital sign data;

[0020] D4: Define the total score threshold and the individual score threshold of each vital sign data, add up the individual scores of each vital sign data to get the total score, and generate an alarm message when the total score exceeds the total score threshold or the individual score of a vital sign data exceeds the corresponding individual score threshold within the accumulated time;

[0021] D5: Establish a rapid response team for emergency response and send the alarm information to the management personnel. The management personnel will conduct a preliminary review based on the alarm information, identify and eliminate false alarms, and after confirming that it is a real alarm, send the alarm information to the rapid response team.

[0022] The health prediction model in the health prediction module includes an input layer, a Transformer encoder layer, a BiGRU decoder layer and an output layer. The specific steps of establishing a health prediction model based on Transformer-BiGRU in the health prediction module are as follows:

[0023] S1: Remove outliers that are out of range in each vital sign data, use linear interpolation to supplement missing data, then take the average of each vital sign data in hours, and connect the hourly averages of each vital sign data to obtain the vital sign characteristics;

[0024] S2: Extract statistical features and medical features: Statistical features include the mean, standard deviation, maximum, minimum and median of each vital sign data in the past 3 hours and the past 5 hours; medical features include the product of heart rate and blood pressure, the ratio of heart rate to blood pressure, cardiac output and absolute continuous difference of heart rate;

[0025] S3: construct physical sign embedding vector, statistical feature embedding vector and medical feature embedding vector according to physical sign features, statistical features and medical features respectively;

[0026] S4: Concatenate the physical sign embedding vector, the statistical feature embedding vector, and the medical feature embedding vector to form an input embedding vector. The specific formula is as follows:

[0027] ;

[0028] in, is the input embedding vector, For the connection operation, For embedded operations, For physical characteristics, is a statistical feature, For medical characteristics;

[0029] S5: The input embedding vector is put into the encoder of the Transformer. The encoder of the Transformer contains a position embedding vector, a multi-head self-attention module and a feedforward network. The position embedding vector is added to the input embedding vector. The input embedding vector is learned using the multi-head self-attention module. Finally, the output of the multi-head self-attention module is sent to the feedforward network in sequence. The Transformer output is obtained through linear transformation and ReLU activation function. The specific formula is as follows:

[0030] ;

[0031] ;

[0032] in, Indicates that The position embedding vector, represents the result of adding the position embedding vector to the input embedding vector, Represents the operation result of the multi-head attention module, Respectively The query vector, key vector and value vector of Indicates The calculation result of each head, is the weight matrix;

[0033] S6: Input the Transformer output to the BiGRU, which includes a forward GRU unit and a reverse GRU unit. The hidden states of the forward GRU unit and the reverse GRU unit are merged through a connection operation to obtain the output of the BiGRU. The specific formula is as follows:

[0034] ;

[0035] in, Represents the first A hidden state, represents the forward GRU unit, represents the reverse GRU unit, Represents the output of Transformer vectors;

[0036] S7: Send the output of BiGRU to the fully connected layer for linear transformation, use the mean square error as the loss function, accumulate the loss values ​​of each vital sign data and take the average to get the final loss value, which is used in the back propagation algorithm and optimizes the Transformer. The specific formula is as follows:

[0037] ;

[0038] ;

[0039] in, Indicates The predicted value of each vital sign data item, is the operation of the fully connected layer, is the output of BiGRU, is the final loss value, is the number of items of vital signs data, Indicates The true value of each vital sign data item, is the mean square error loss function.

[0040] The risk warning module uses a health prediction model to construct a respiratory trend graph, blood oxygen trend graph, blood pressure trend graph, body temperature trend graph and heart rate trend graph for each critically ill patient in the next 24 hours, and identifies the time period when the corresponding respiratory data, blood oxygen data, blood pressure data, body temperature data and heart rate data exceed the upper and lower risk thresholds as the patient's high-risk period, and notifies the medical staff to pay special attention to the high-risk period of the critically ill patient.

[0041] The beneficial results achieved by the present invention using the above scheme are as follows:

[0042] (1) The present invention proposes an intensive care unit early warning system based on artificial intelligence, which realizes timely early warning and advance notification of abnormal vital signs of patients in the intensive care unit; the present invention collects and transmits various vital sign data of critically ill patients in real time, and stores massive vital sign data through a big data platform; further designs a vital sign alarm module, and adopts an abnormality detection and scoring algorithm in the vital sign alarm module to timely detect abnormalities in the vital signs of critically ill patients and issue an alarm; at the same time, a health prediction model based on Transformer-BiGRU is established to achieve accurate prediction of the trend of vital signs of critically ill patients; relying on the health prediction model, a vital sign trend chart is constructed for each patient and high-risk time periods are identified, thereby strengthening early warning and preparation for potential health risks of critically ill patients; finally, the vital sign data, alarm information and vital sign trend chart of critically ill patients are summarized to assist doctors in establishing risk assessment scales according to different diseases of critically ill patients.

[0043] (2) In view of the high false alarm rate problem faced when monitoring abnormal vital signs of critically ill patients, the present invention proposes a vital sign alarm module and adopts an abnormality detection scoring algorithm in the vital sign alarm module. The real-time monitored vital sign data are scored according to the risk threshold, influencing factor and alarm accuracy. When the total score threshold or the single score threshold is exceeded, an alarm is issued. The management personnel conduct a preliminary review of the alarm information and then mobilize the rapid response team to rescue the critically ill patients in time, thereby reducing the false alarm probability of abnormal vital signs of critically ill patients and optimizing the allocation of human and medical resources.

[0044] (3) In order to address the problem of difficulty in accurately predicting the future trends of critically ill patients’ vital signs, which in turn affects the allocation of human and medical resources, a health prediction model based on Transformer-BiGRU was designed. Transformer was used to comprehensively learn the physical, statistical, and medical features of vital sign data, and BiGRU was used to further enhance the ability to capture bidirectional dependencies, thereby improving the accuracy of model predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a module diagram of an intensive care unit early warning system based on artificial intelligence proposed by the present invention;

[0046] Figure 2 This is a model diagram of the health prediction model based on Transformer-BiGRU proposed in the present invention.

[0047] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0049] Example 1, see Figure 1 , Figure 1 This is a module diagram of an intensive care unit early warning system based on artificial intelligence. In this embodiment, an intensive care unit early warning system based on artificial intelligence is implemented, which specifically includes:

[0050] A data acquisition module, a data transmission module, a data center, a vital sign alarm module, a health prediction module, a risk warning module and an auxiliary decision module, wherein the output end of the data acquisition module is connected to the input end of the data transmission module, the output end of the data transmission module is connected to the input end of the data center, the output end of the data center is respectively connected to the input end of the health prediction module and the input end of the vital sign alarm module, the output end of the vital sign alarm module is connected to the input end of the auxiliary decision module, the output end of the health prediction module is connected to the input end of the risk warning module, and the output end of the risk warning module is connected to the input end of the auxiliary decision module;

[0051] The data collection module collects vital sign data;

[0052] The data transmission module transmits vital signs data in real time;

[0053] The data center department uses a big data platform to store vital signs data;

[0054] The vital sign alarm module uses an abnormal detection and scoring algorithm to monitor vital sign data in real time and generate alarm information and alarm when abnormalities occur;

[0055] The health prediction module establishes a health prediction model based on Transformer-BiGRU;

[0056] The risk warning module uses the health prediction model to establish a trend chart of vital signs for each critically ill patient in the next 24 hours and identify high-risk periods;

[0057] The decision support module summarizes the vital signs data, alarm information and vital signs trend charts of critically ill patients, helping doctors to establish risk assessment scales based on the different types of diseases of critically ill patients.

[0058] Embodiment 2: Based on the above embodiment, this embodiment realizes data collection, transmission and storage. The specific contents are as follows:

[0059] The data acquisition module collects the vital signs data of critically ill patients through a data collector and life support medical equipment: the data collector is connected to the life support medical equipment through the RS32 serial port and the network port, and the vital signs data of critically ill patients are obtained by parsing the communication protocol of the life support medical equipment;

[0060] Life support medical equipment includes monitors, electrocardiographs, ventilators, blood purification equipment, extracorporeal membrane oxygenation devices, infusion pumps, and counterpulsation pumps;

[0061] Vital signs data include respiratory data, blood oxygen data, blood pressure data, body temperature data and heart rate data;

[0062] Respiratory data include respiratory rate, inspiratory volume, tidal volume, oxygen concentration, inspiratory time, expiratory time, inspiratory pressure, expiratory pressure, breathing pattern, airflow time, oxygen input pressure, average respiratory pressure, and inspiration-expiration ratio;

[0063] Blood oxygen data includes blood oxygen concentration, blood oxygen saturation, arterial oxygen partial pressure, venous oxygen partial pressure, pulse oxygen saturation, and blood oxygen content;

[0064] Blood pressure data include average blood pressure, peak blood pressure, low blood pressure, invasive average blood pressure, invasive systolic blood pressure, invasive diastolic blood pressure, non-invasive average blood pressure, non-invasive systolic blood pressure, and non-invasive diastolic blood pressure;

[0065] Body temperature data includes skin temperature, oral temperature, and axillary temperature;

[0066] Heart rate data include total heart rate, heart rate, average heart rate, fastest heart rate, slowest heart rate, RR period, heart rate variability, atrial premature beats, atrial tachycardia, ventricular premature beats, and ventricular tachycardia;

[0067] The data transmission module transmits vital signs data to the data center through Bluetooth, 5G, and Wifi technologies. When transmitting vital signs data, data frame verification is used to ensure integrity. In addition, in order to protect patient privacy, all data transmission uses symmetric encryption technology to ensure information security.

[0068] The data center is deployed in the cloud. The Hadoop big data platform is built and deployed in the cloud. The vital signs data are distributedly stored and called through the API interface. A patient ID is established as a unique identifier for each critically ill patient, and each vital sign data is matched with the patient ID one by one. At the same time, it is integrated with the patient's medical history information and patient personal information data system to achieve data interconnection and deep integration.

[0069] Example 3: Based on the above example, this example implements alarming of abnormal vital sign data of critically ill patients. The specific contents are as follows:

[0070] The vital sign alarm module uses an abnormal detection and scoring algorithm to monitor vital sign data in real time and generate alarm information and issue an alarm when an abnormality occurs. The specific steps of the abnormal detection and scoring algorithm are as follows:

[0071] D1: define the upper and lower risk thresholds for each vital sign data. When each vital sign data exceeds the corresponding upper and lower risk thresholds, 1 point is scored. In this embodiment, the upper risk threshold of the respiratory rate is set to 24 times / minute, and the lower risk threshold is set to 8 times / minute.

[0072] D2: define the cumulative time and the influencing factors of each vital sign data, multiply the scores of each vital sign data within the cumulative time calculated by D1 by the corresponding influencing factors, and obtain the product result, wherein the value of the influencing factor is a positive value; in this embodiment, the cumulative time is defined as 5 minutes, and the influencing factor of the respiratory rate is set to 1.5;

[0073] D3: Count the alarm accuracy rates of each vital sign data based on past alarm information, multiply the product result in D2 by the alarm accuracy rate, and obtain the individual scores of each vital sign data; in this embodiment, the alarm accuracy rate of respiratory rate is 100%;

[0074] D4: define the total score threshold and the individual score threshold of each vital sign data, add up the individual scores of each vital sign data to get the total score, and generate an alarm message when the total score exceeds the total score threshold or the individual score of a certain vital sign data exceeds the corresponding individual score threshold within the accumulated time. The alarm message includes the personal information of the critically ill patient, the ICU ward number, and the vital sign data within the last hour, and highlights the relevant vital sign data that triggers the alarm; in this embodiment, the total score threshold is set to 5, and the individual score threshold of each vital sign data is set to 3. Within 5 minutes, the critically ill patient's respiratory rate exceeds 24 times / minute twice in a row, so the individual score of the respiratory rate is 2*1.5*100%=3, reaching the individual score threshold, so the alarm is triggered;

[0075] D5: Establish a rapid response team for emergency treatment and send the alarm information to the management personnel. The management personnel will conduct a preliminary review based on the alarm information, identify and eliminate false alarms, and after confirming that it is a real alarm, send the alarm information to the rapid response team. After receiving the message, the rapid response team will rush to the scene and provide assistance to the relevant critically ill patients.

[0076] Example 4, see Figure 2 , Figure 2 This is a model diagram of a health prediction model based on Transformer-BiGRU. This embodiment establishes a health prediction model based on the above embodiment. The specific contents are as follows:

[0077] The health prediction model in the health prediction module includes an input layer, a Transformer encoder layer, a BiGRU decoder layer and an output layer. The specific steps of establishing a health prediction model based on Transformer-BiGRU in the health prediction module are as follows:

[0078] S1: Remove outliers that exceed the value range in each vital sign data, use linear interpolation to supplement missing data, then take the average of each vital sign data in hours, and connect the hourly averages of each vital sign data according to the patient ID to obtain the vital sign characteristics. In this embodiment, respiratory rate, average respiratory pressure, average heart rate, fastest heart rate, slowest heart rate, RR period, average blood pressure, skin temperature, and blood oxygen saturation are selected as key vital sign data to train and predict the health prediction model;

[0079] S2: Extract statistical features and medical features: Statistical features include the mean, standard deviation, maximum, minimum and median of each vital sign data in the past 3 hours and the past 5 hours; medical features include the product of heart rate and blood pressure, the ratio of heart rate and blood pressure, cardiac output and absolute continuous difference of heart rate, where cardiac output = heart rate * (invasive systolic pressure - invasive diastolic pressure) / invasive average blood pressure, and absolute continuous difference of heart rate is the sum of the absolute values ​​of continuous changes in heart rate in one minute;

[0080] S3: construct physical sign embedding vector, statistical feature embedding vector and medical feature embedding vector according to physical sign features, statistical features and medical features respectively;

[0081] S4: Concatenate the physical sign embedding vector, the statistical feature embedding vector, and the medical feature embedding vector to form an input embedding vector. The specific formula is as follows:

[0082] ;

[0083] in, is the input embedding vector, For the connection operation, For embedded operations, For physical characteristics, is a statistical feature, For medical characteristics;

[0084] S5: The input embedding vector is put into the encoder of the Transformer. The encoder of the Transformer contains a position embedding vector, a multi-head self-attention module and a feedforward network. The position embedding vector is added to the input embedding vector. The input embedding vector is learned using the multi-head self-attention module. Finally, the output of the multi-head self-attention module is sent to the feedforward network in sequence. The Transformer output is obtained through linear transformation and ReLU activation function. The specific formula is as follows:

[0085] ;

[0086] ;

[0087] in, Indicates that The position embedding vector, represents the result of adding the position embedding vector to the input embedding vector, Represents the operation result of the multi-head attention module, Respectively The query vector, key vector and value vector of Indicates The calculation result of each head, is the weight matrix;

[0088] The multi-head attention module in this embodiment has 12 heads;

[0089] S6: Input the Transformer output to the BiGRU, which includes a forward GRU unit and a reverse GRU unit. The hidden states of the forward GRU unit and the reverse GRU unit are merged through a connection operation to obtain the output of the BiGRU. The specific formula is as follows:

[0090] ;

[0091] in, Represents the first A hidden state, represents the forward GRU unit, represents the reverse GRU unit, Represents the output of Transformer vectors;

[0092] S7: The output of BiGRU is sent to the fully connected layer for linear transformation. The mean square error is used as the loss function. The loss values ​​of respiratory frequency, average respiratory pressure, average heart rate, fastest heart rate, slowest heart rate, RR period, average blood pressure, skin temperature, and blood oxygen saturation are accumulated and averaged to obtain the final loss value. The loss value is used in the back propagation algorithm and the Transformer is optimized. The specific formula is as follows:

[0093] ;

[0094] ;

[0095] in, Indicates The predicted value of each vital sign data item, is the operation of the fully connected layer, is the output of BiGRU, is the final loss value, is the number of items of vital signs data, Indicates The true value of each vital sign data item, is the mean square error loss function.

[0096] Example 5: Based on the above example, this example realizes the identification of the high-risk period of the patient in the next 24 hours. The specific contents are as follows:

[0097] The risk warning module uses the health prediction model to construct a respiratory trend graph, blood oxygen trend graph, blood pressure trend graph, body temperature trend graph and heart rate trend graph for each critically ill patient in the next 24 hours, and identifies the time period when the corresponding respiratory data, blood oxygen data, blood pressure data, body temperature data and heart rate data exceed the upper and lower risk thresholds as the patient's high-risk period, and notifies the medical staff in charge to pay special attention to the high-risk period of the critically ill patient;

[0098] In this embodiment, the respiratory trend graph includes the respiratory rate trend and the average respiratory pressure trend, the blood oxygen trend graph includes the blood oxygen saturation trend, the blood pressure trend graph includes the blood pressure average value trend, the body temperature data includes the skin temperature trend, and the heart rate data includes the average heart rate trend, the fastest heart rate trend, the slowest heart rate trend and the RR period trend.

[0099] Embodiment 6: Based on the above embodiments, this embodiment realizes assisting doctors in making decisions. The specific contents are as follows:

[0100] The decision-making support module summarizes the vital signs data, alarm information and vital signs trend charts of critically ill patients, helping doctors to establish risk assessment scales based on the different types of diseases of critically ill patients;

[0101] In this embodiment, for critically ill patients with acute pulmonary embolism, the vital signs data of respiratory rate, oral body temperature, non-invasive systolic blood pressure, blood oxygen saturation, total heart rate, and heart rate are used to assist doctors in establishing the Geneva assessment scale and sPESI assessment scale for clinical possibility assessment of acute pulmonary embolism, and combined with the corresponding vital signs trend chart, a preliminary judgment is made on the patient's future health trend.

Claims

1. An artificial intelligence-based early warning system for intensive care units, Features: It includes data acquisition module, data transmission module, data center, vital sign alarm module, health prediction module, risk warning module and decision support module; The data acquisition module collects vital sign data; The data transmission module transmits vital sign data in real time; The data center uses a big data platform to store vital signs data; The vital sign alarm module adopts an abnormal detection and scoring algorithm to monitor the vital sign data in real time and generate alarm information and alarm when abnormal; The health prediction module establishes a health prediction model based on Transformer-BiGRU; The risk warning module uses the health prediction model to establish a vital sign trend chart for each critically ill patient and identify high-risk periods; The auxiliary decision-making module summarizes the vital sign data, alarm information and vital sign trend chart of critically ill patients to help doctors establish a risk assessment scale.

2. An artificial intelligence-based intensive care unit early warning system according to claim 1, Features: The specific steps of the anomaly detection scoring algorithm are as follows: D1: Define the upper and lower risk thresholds for each vital sign data. If each vital sign data exceeds the corresponding upper and lower risk thresholds, 1 point will be awarded. D2: Define the impact factors of the accumulated time and each vital sign data, multiply the scores of each vital sign data within the accumulated time calculated by D1 by the corresponding impact factors to obtain the product result; D3: Count the alarm accuracy rates of each vital sign data, multiply the product result in D2 by the alarm accuracy rate, and obtain the individual scores of each vital sign data; D4: Define the total score threshold and the individual score threshold of each vital sign data, add up the individual scores of each vital sign data to get the total score, and generate an alarm message when the total score exceeds the total score threshold or the individual score of a vital sign data exceeds the corresponding individual score threshold within the accumulated time; D5: Establish a rapid response team for emergency response and send the alarm information to the management personnel. The management personnel will conduct a preliminary review based on the alarm information, identify and eliminate false alarms, and after confirming that it is a real alarm, send the alarm information to the rapid response team.

3. According to claim 1, an artificial intelligence-based intensive care unit early warning system, Features: The health prediction model includes an input layer, a Transformer encoder layer, a BiGRU decoder layer and an output layer. The specific steps of the health prediction module to establish a health prediction model based on Transformer-BiGRU are as follows: S1: Remove outliers that are out of range in each vital sign data, use linear interpolation to supplement missing data, then take the average of each vital sign data in hours, and connect the hourly averages of each vital sign data to obtain the vital sign characteristics; S2: Extract statistical and medical features; S3: construct physical sign embedding vector, statistical feature embedding vector and medical feature embedding vector according to physical sign features, statistical features and medical features respectively; S4: Concatenate the physical sign embedding vector, the statistical feature embedding vector and the medical feature embedding vector to form an input embedding vector; S5: Input the input embedding vector into the Transformer encoder to obtain the Transformer output; S6: Input the Transformer output to the BiGRU, which includes a forward GRU unit and a reverse GRU unit. The hidden states of the forward GRU unit and the reverse GRU unit are merged through a connection operation to obtain the output of the BiGRU. S7: Send the output of BiGRU to the fully connected layer for linear transformation, use the mean square error as the loss function, accumulate the loss values ​​of each vital sign data and take the average to get the final loss value, which is used in the back propagation algorithm and optimizes the Transformer.

4. An artificial intelligence-based intensive care unit early warning system according to claim 3, Features: The statistical features include the mean, standard deviation, maximum value, minimum value and median of each vital sign data in the past 3 hours and the past 5 hours.

5. According to claim 3, an artificial intelligence-based intensive care unit early warning system, Features: The medical characteristics include the product of heart rate and blood pressure, the ratio of heart rate and blood pressure, cardiac output and absolute continuous difference of heart rate.

Images