A stroke emergency identification method and system combining TOPSIS and artificial intelligence

Abstract

The invention discloses a stroke emergency recognition method and system combining TOPSIS and artificial intelligence, belonging to the technical field of stroke early warning, and solves the problem that the existing method cannot accurately capture facial feature points in an emergency state, resulting in insufficient response speed and accuracy when judging by the TOPSIS method. The method comprises real-time acquisition of patient physiological parameter information, recognition of facial parameter information in a preprocessing set based on a multi-view feature positioning model, generation of facial expression recognition results, and analysis of the facial expression recognition results and the preprocessing set by a stroke analysis model, and calculation of the patient's stroke risk level; the invention pre-constructs a stroke analysis model based on TOPSIS and artificial intelligence, and the stroke analysis model cooperates with a multi-view feature positioning model and a flexible trigger mechanism, so that the stroke analysis model quickly responds to the recognition results of the multi-view feature positioning model and the flexible trigger mechanism, and the multi-view feature positioning model can analyze facial parameter information from multiple angles.

Description

Technical Field

[0001] The present invention belongs to the technical field of stroke early warning, and in particular relates to a stroke emergency recognition method and system combining TOPSIS with artificial intelligence. Background Art

[0002] Panvascular disease is a group of vascular system diseases with atherosclerosis as a common pathological feature. It harms important organs such as the heart, brain, kidneys, and limbs. As an important type of disease, stroke endangers human health. Stroke, also known as infarction or cerebrovascular accident, is mainly divided into two categories: ischemic stroke and hemorrhagic stroke, which are the well-known "cerebral infarction" and "cerebral hemorrhage". Ischemic stroke is caused by blockage of cerebral blood vessels, which prevents blood from flowing into the brain, causing ischemia and hypoxia of brain tissue and damage. Hemorrhagic stroke is caused by rupture of cerebral blood vessels, blood flows into brain tissue or ventricles, compressing brain tissue and causing damage. Stroke is characterized by high morbidity, high disability rate and high mortality rate, which seriously affects people's quality of life and threatens people's life and health.

[0003] After a stroke occurs, brain cells will die irreversibly due to lack of blood and oxygen for 6 minutes, and about 1.9 million brain cells will die for every minute of delayed treatment. Therefore, rapid identification of stroke and immediate hospitalization for professional treatment can minimize brain tissue damage and protect the patient's neurological function.

[0004] Chinese patent CN113936799A discloses a method for rapid stroke identification that combines TOPSIS with artificial intelligence. First, the user's arms are raised horizontally to determine whether there is a unilateral muscle strength decline, and the muscle strength difference is graded by the angle between the arms; then the degree of slurred speech of the user is analyzed by artificial intelligence speech recognition and combined with the error rate calculation method; then the position coordinates of the facial feature points are detected by the artificial intelligence facial feature point recognition model, and the degree of facial droop is calculated by the TOPSIS method. Finally, the risk of stroke is quickly identified by the TOPSIS method based on the comprehensive muscle strength difference classification, slurred speech and facial droop. However, the existing artificial intelligence facial feature point recognition model cannot accurately capture facial feature points in an emergency state, resulting in insufficient response speed and accuracy when judging by the TOPSIS method. To address the above problems, we propose a stroke emergency identification method and system that combines TOPSIS with artificial intelligence. Summary of the invention

[0005] The purpose of the present invention is to address the shortcomings of the prior art and provide a stroke emergency identification method and system combining TOPSIS and artificial intelligence, which solves the problem that the existing artificial intelligence facial feature point recognition model cannot accurately capture facial feature points under emergency conditions, resulting in insufficient response speed and accuracy when judging through the TOPSIS method.

[0006] The present invention is implemented by combining TOPSIS with an artificial intelligence stroke emergency identification method, the stroke emergency identification method combining TOPSIS with artificial intelligence comprising:

[0007] Collecting the patient's physiological parameter information in real time, preprocessing the patient's physiological parameter information based on a flexible trigger mechanism to obtain a preprocessing set, wherein the patient's physiological parameter information includes the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information;

[0008] Traversing the preprocessing set, identifying the facial parameter information in the preprocessing set based on the multi-view feature positioning model, and generating facial expression recognition results, wherein the facial expression recognition results include facial symmetry, facial droop, and expression classification results;

[0009] Pre-build a stroke analysis model based on TOPSIS and artificial intelligence, use a web crawler to obtain a modeling sample set, use a flexible trigger mechanism combined with a multi-view feature positioning model to pre-process the modeling sample set, divide the modeling sample set into a training set and a test set, use the training set and the test set to iteratively train the stroke analysis model, and output a converged stroke analysis model;

[0010] The facial expression recognition results and preprocessing sets are loaded in real time. The stroke analysis model recognizes and analyzes the facial expression recognition results and preprocessing sets and calculates the patient's stroke risk level.

[0011] Based on the patient's stroke risk level, it is determined whether it exceeds the preset risk threshold. If it exceeds the preset risk threshold, a stroke emergency instruction is triggered.

[0012] Preferably, the method for preprocessing patient physiological parameter information based on the flexible trigger mechanism includes:

[0013] Load the patient's physiological parameter information, identify the vital sign monitoring information in the patient's physiological parameter information, generate a vital sign monitoring set, and perform missing value processing on the vital sign monitoring set;

[0014] Based on the analytic hierarchy process, a judgment matrix of physical sign indicators in the physical sign monitoring center is constructed, and the weights of the single group of physical sign indicators in the physical sign monitoring center are determined by combining the judgment matrix of physical sign indicators.

[0015] Among them, the judgment matrix of physical sign indicators is expressed as:

[0016] (1)

[0017] in, Represents the judgment matrix of physical sign indicators, For physical indicators and physical indicators The importance between

[0018] The indicator weight of a single group of physical signs is calculated by the following formula:

[0019] (2)

[0020] (3)

[0021] (4)

[0022] in, Represents the indicator weight of a single group of physical signs indicators, is the maximum eigenvalue of the judgment matrix, is the number of physical signs, represents the normalized value of the physical sign index, is the initial weight of a single group of physical signs, Respectively represent the accumulated values ​​of the rows and columns of the judgment matrix;

[0023] Obtaining the indicator weight of a single group of vital sign indicators, and determining the flexible trigger threshold and event integral threshold of the single group of vital sign indicators based on the indicator weight of the single group of vital sign indicators;

[0024] Traverse the vital sign monitoring set, extract the vital sign monitoring information exceeding the event integral threshold in the vital sign monitoring set, and generate an abnormal monitoring set;

[0025] Based on the isolation forest anomaly detection algorithm, anomaly detection is performed on the anomaly monitoring set within the collection period, and the flexible trigger integral value of the anomaly monitoring set within the collection period is calculated;

[0026] The flexible trigger integral value is calculated by the following formula:

[0027] (5)

[0028] in, is the flexible trigger integral value, Indicates that a single group of physical signs indicators The event integration threshold at time, Indicates the current moment, is the collection cycle, Indicates the number of events exceeding the event integration threshold during the acquisition period;

[0029] Determine whether the flexible trigger integral value exceeds the flexible trigger threshold. If it exceeds the flexible trigger threshold, trigger the flexible trigger mechanism and upload the flexible trigger integral value of the abnormal monitoring set within the collection period.

[0030] Preferably, the multi-view feature localization model is based on a cascaded pose regression model. The reduce_domain function is introduced into the cascaded pose regression model to crop the facial parameter information and narrow the facial detection target area. The Ghost module is also introduced into the multi-view feature localization model to perform a linear transformation on the feature map already generated by the cascaded pose regression model to generate a complete feature map. After the Ghost module, CONVTranspose deconvolution is introduced to restore the size of the feature map;

[0031] The loss function of the multi-view feature localization model is defined as:

[0032] (6)

[0033] where is the number of facial sagging points in the current image in, represents the predicted value of the pixel point feature, represents the true value of the pixel point feature, are the penalty parameter for reducing the angle influence and the classification parameter for the facial symmetry degree of the sample in the cascaded pose regression model, respectively.

[0034] Preferably, the method for identifying the facial parameter information in the preprocessing set based on the multi-view feature localization model specifically includes:

[0035] Load the facial parameter information in the preprocessing set. The multi-view feature localization model corrects the angle of the facial parameter information based on the cascaded pose regression algorithm combined with the isolated forest algorithm to obtain an angle adaptation set;

[0036] Obtain the angle adaptation set, encode the local texture feature values of the angle adaptation set based on the cascaded pose regression algorithm to generate a binarized feature vector, and integrate at least one group of binarized feature vectors to obtain a binarized feature set;

[0037] Load the binarized feature set, and the reduce_domain function crops the binarized feature set to narrow the facial detection target area to obtain at least one group of target area anchor boxes;

[0038] Based on the cascaded pose regression algorithm combined with the Ghost module, perform a linear transformation on the target area anchor box to generate a complete feature map;

[0039] CONVTranspose deconvolution restores the size of the feature map to obtain a feature restoration map, evaluates the facial sagging degree of the feature restoration map through the Bayesian criterion, and calculates the facial symmetry degree in combination with the Euclidean distance algorithm.

[0040] Preferably, the method for iteratively training the stroke analysis model using the training set and the test set specifically includes:

[0041] Load the TOPSIS model, set the loss function, activation function, training rounds and hyperparameters of the TOPSIS model;

[0042] Load the training set, use the training set to perform unsupervised pre-training on the data preprocessing module, introduce conditional random fields during training, and output the data processing set;

[0043] Obtain a data processing set, iteratively train the matrix construction module and the distance calculation module through the data processing set combined with adversarial training, calculate the gradient of the loss function of the TOPSIS model for the data processing set, perturb the data processing set along the direction of the gradient, and generate adversarial samples;

[0044] Load the adversarial sample, project the adversarial sample into the data processing set based on the projected gradient descent method, use the data processing set to integrate the data preprocessing module, matrix construction module, distance calculation module, and risk level calculation module in the TOPSIS model, and output a converged stroke analysis model;

[0045] Load the test set, use the test set as input, execute the stroke analysis model, the stroke analysis model recognizes and analyzes the test set, outputs the test results, and determines whether the test results meet the preset accuracy threshold. If they meet the preset accuracy threshold, output the converged stroke analysis model.

[0046] Preferably, the stroke analysis model uses the TOPSIS model as the initial model, and the TOPSIS model includes a data preprocessing module, a matrix construction module, a distance calculation module, and a risk level calculation module, wherein the data preprocessing module includes three hierarchical analysis layers, and the hierarchical analysis layers are connected via the TCP protocol;

[0047] When pre-building a stroke analysis model based on TOPSIS and artificial intelligence, the data pre-processing module in the TOPSIS model is improved, and the CRNN network and the generative adversarial network are introduced into the data pre-processing module. The CRNN network is used to extract follow-up voice information, continuously map the high-dimensional input follow-up voice information to the low-dimensional feature space, and output the low-dimensional semantic feature vector. The hierarchical analysis layer normalizes the semantic feature vector. The generative adversarial network consists of a generator and a discriminator. The generator is used to extract the feature vectors of facial parameter information and vital sign monitoring information, and the facial parameter information and vital sign monitoring information are weighted and classified based on the discriminator, and the facial feature vector and vital sign monitoring vector are input into the corresponding hierarchical analysis layer.

[0048] The matrix construction module is used to construct a weighted decision matrix using patient basic information and patient medical record information as prior information, combined with facial feature vectors, vital sign monitoring vectors, and semantic feature vectors.

[0049] Preferably, the method of the stroke analysis model for facial expression recognition results and preprocessing set recognition analysis specifically includes:

[0050] Load the facial expression recognition results and preprocessing set, import the facial expression recognition results and preprocessing set into the data preprocessing module, the CRNN network in the data preprocessing module continuously maps the high-dimensional input follow-up voice information to the low-dimensional feature space, generates the adversarial network to extract the feature vectors of the facial droop degree, facial symmetry degree, and physical sign monitoring information, and classifies the facial droop degree, facial symmetry degree, and physical sign monitoring information weights based on the discriminator, and integrates the semantic feature vector, facial feature vector, physical sign monitoring vector, and basic information vector;

[0051] Acquire semantic feature vectors, facial feature vectors, vital sign monitoring vectors, and basic information vectors, construct feature vector matrices based on the semantic feature vectors, facial feature vectors, vital sign monitoring vectors, and basic information vectors, and perform normalization processing on the feature vector matrix by the hierarchical analysis layer to obtain a normalized matrix;

[0052] Load the normalized matrix. The matrix construction module constructs a feature decision matrix based on the feature weights corresponding to the semantic feature vector, facial feature vector, vital sign monitoring vector, and basic information vector.

[0053] Obtain a feature decision matrix, generate an ideal solution and a negative ideal solution based on the feature decision matrix, and a distance calculation module calculates the distance between the normalized matrix and the ideal solution and the negative ideal solution based on the Euclidean distance algorithm;

[0054] The risk level calculation module loads the distance between the normalized matrix and the ideal solution and the negative ideal solution, calculates the relative closeness based on the distance between the ideal solution and the negative ideal solution, and uses the relative closeness as the patient's stroke risk level.

[0055] Preferably, when the discriminator is used to discriminate and classify the degree of facial droop, the degree of facial symmetry, and the weight of physical sign monitoring information, the discriminator discrimination formula is expressed as:

[0056] (7)

[0057] (8)

[0058] in, Represents the discriminator's output result. is the number of characteristic indicators, They are the feature input matrix and adversarial vector matrix respectively. represents the number of times the discriminator counteracts disturbances, represents the classification function of the discriminator, is the discriminator perturbation factor, is the feature weight, is the initial weight of the features based on principal component analysis, is the correlation coefficient of the feature;

[0059] The relative closeness calculated based on the distance between the ideal solution and the negative ideal solution is expressed by the following formula:

[0060] (9)

[0061] in, Relative closeness, are the distances between the normalized matrix and the ideal solution and the negative ideal solution, respectively.

[0062] On the other hand, the present invention also provides a stroke emergency recognition system combining TOPSIS and artificial intelligence, the stroke emergency recognition system combining TOPSIS and artificial intelligence comprising:

[0063] An information collection module is used to collect the patient's physiological parameter information in real time, and pre-process the patient's physiological parameter information based on a flexible trigger mechanism to obtain a pre-processed set;

[0064] The expression recognition module is used to traverse the preprocessing set, identify the facial parameter information in the preprocessing set based on the multi-view feature positioning model, and generate facial expression recognition results;

[0065] The risk identification module is used to pre-build a stroke analysis model based on TOPSIS and artificial intelligence, load facial expression recognition results and pre-processing sets in real time, and the stroke analysis model identifies and analyzes the facial expression recognition results and pre-processing sets to calculate the patient's stroke risk level;

[0066] The emergency trigger module determines whether the patient's stroke risk level exceeds the preset risk threshold based on the patient's stroke risk level. If the preset risk threshold is exceeded, the stroke emergency instruction is triggered.

[0067] Preferably, the information collection module includes:

[0068] Basic information collection unit, used to collect basic patient information and patient medical record information;

[0069] A facial acquisition unit, used for acquiring facial parameter information;

[0070] A vital sign collection unit, used for collecting vital sign monitoring information;

[0071] The preprocessing unit is used to load the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information, and to preprocess the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information.

[0072] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0073] In an embodiment of the present invention, a stroke analysis model based on TOPSIS and artificial intelligence is pre-constructed, and the stroke analysis model cooperates with a multi-view feature positioning model and a flexible trigger mechanism, so that the stroke analysis model can quickly respond to the recognition results of the multi-view feature positioning model and the flexible trigger mechanism, thereby improving the recognition accuracy and efficiency of the stroke analysis model. The multi-view feature positioning model can analyze facial parameter information from multiple angles, thereby overcoming the problem that the existing artificial intelligence facial feature point recognition model cannot accurately capture facial feature points under emergency conditions, resulting in insufficient response speed and accuracy when judging through the TOPSIS method.

[0074] In the embodiment of the present invention, the patient's physiological parameter information is preprocessed based on the flexible trigger mechanism, so that the warning threshold and judgment criteria can be dynamically adjusted according to the changes in the patient's physiological parameters, thereby improving the accuracy of identifying stroke precursors. The flexible trigger mechanism helps to reduce false alarms and missed alarms, ensuring that alarms can be issued in time at critical moments. The flexible trigger mechanism cooperates with the stroke analysis model, which not only improves the data quality and model performance of the model, but also can perform parameter flow on abnormal monitoring sets within the acquisition cycle, thereby reducing system load.

[0075] In the embodiment of the present invention, the degree of facial droop in the feature recovery image is evaluated by the Bayesian criterion, and the degree of facial symmetry is calculated in combination with the Euclidean distance algorithm. The Bayesian criterion evaluates the degree of facial droop in the feature recovery image by calculating the prior probability and the posterior probability. Combining prior knowledge and observation data, the Bayesian algorithm can accurately update the probability distribution of events and provide a scientific basis for the degree of facial droop. Combining the Bayesian criterion with the Euclidean distance algorithm can more comprehensively evaluate the degree of facial droop and symmetry. The Bayesian criterion provides probability distribution updates, while the Euclidean distance accurately quantifies the symmetry differences. The two complement each other to improve the accuracy of the evaluation.

[0076] In an embodiment of the present invention, a multi-view feature localization model is provided. The multi-view feature localization model is based on a cascaded posture regression model. A reduce_domain function is introduced in the cascaded posture regression model to crop facial parameter information. A Ghost module is also introduced to perform a linear transformation on the feature map that has been generated by the cascaded posture regression model. By combining features from different perspectives, the model can more accurately capture subtle changes in the face, thereby improving the recognition accuracy of stroke symptoms. By cropping irrelevant facial parameter information, the amount of calculation is reduced, thereby improving the operating efficiency of the stroke analysis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the implementation flow of the stroke emergency identification method combining TOPSIS and artificial intelligence provided by the present invention.

[0078] Figure 2 The figure shows a schematic diagram of the implementation process of the method for preprocessing patient physiological parameter information based on the flexible trigger mechanism.

[0079] Figure 3 The present invention shows a schematic diagram of the implementation process of the method for preprocessing concentrated facial parameter information based on multi-view feature positioning model recognition.

[0080] Figure 4 The figure shows a schematic diagram of the implementation process of the iterative training method for the stroke analysis model using a training set and a test set.

[0081] Figure 5 A schematic diagram of the implementation process of the stroke analysis model for facial expression recognition results and preprocessing set recognition and analysis methods is shown.

[0082] Figure 6 The structural diagram of the stroke emergency identification system combining TOPSIS and artificial intelligence is shown. DETAILED DESCRIPTION

[0083] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0084] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0085] The existing artificial intelligence facial feature point recognition model cannot accurately capture facial feature points under emergency conditions, resulting in insufficient response speed and accuracy when judging through the TOPSIS method. To address the above problems, we proposed a stroke emergency recognition method and system that combines TOPSIS and artificial intelligence. When the method is implemented, the patient's physiological parameter information is first collected in real time, and the patient's physiological parameter information is preprocessed based on a flexible trigger mechanism. Then, the preprocessing set is traversed, and the facial parameter information in the preprocessing set is identified based on a multi-view feature positioning model to generate facial expression recognition results. At the same time, a stroke analysis model based on TOPSIS and artificial intelligence is pre-built. The stroke analysis model analyzes the facial expression recognition results and the preprocessing set recognition and calculates the patient's stroke risk level. In an embodiment of the present invention, a stroke analysis model based on TOPSIS and artificial intelligence is pre-constructed, and the stroke analysis model cooperates with a multi-view feature positioning model and a flexible trigger mechanism, so that the stroke analysis model can quickly respond to the recognition results of the multi-view feature positioning model and the flexible trigger mechanism, thereby improving the recognition accuracy and efficiency of the stroke analysis model. The multi-view feature positioning model can analyze facial parameter information from multiple angles, thereby overcoming the problem that the existing artificial intelligence facial feature point recognition model cannot accurately capture facial feature points under emergency conditions, resulting in insufficient response speed and accuracy when judging through the TOPSIS method.

[0086] The embodiment of the present invention provides a stroke emergency identification method combining TOPSIS and artificial intelligence. Figure 1 The figure shows a schematic diagram of the implementation process of the stroke emergency identification method combining TOPSIS and artificial intelligence. The stroke emergency identification method combining TOPSIS and artificial intelligence specifically includes:

[0087] Step S10, collecting the patient's physiological parameter information in real time, and preprocessing the patient's physiological parameter information based on a flexible trigger mechanism to obtain a preprocessing set.

[0088] It should be noted that the patient's physiological parameter information includes but is not limited to patient basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information.

[0089] Among them, the basic information of patients includes the patient's name, gender, age, home address, stroke type, education level, contact information, and occupation, and the patient's medical history information includes hospitalization information, surgery information, rehabilitation and physical therapy information, and medical records. The vital signs monitoring information includes heart rate, blood pressure, body temperature, respiratory rate, blood oxygen saturation, isokinetic muscle strength test parameters, and handheld dynamometer parameters.

[0090] Step S20, traversing the preprocessing set, identifying facial parameter information in the preprocessing set based on the multi-view feature positioning model, and generating a facial expression recognition result, wherein the facial expression recognition result includes facial symmetry, facial droop, and expression classification result;

[0091] Step S30, pre-constructing a stroke analysis model based on TOPSIS and artificial intelligence, using a web crawler to obtain a modeling sample set, using a flexible trigger mechanism combined with a multi-view feature positioning model to pre-process the modeling sample set, dividing the modeling sample set into a training set and a test set, using the training set and the test set to iteratively train the stroke analysis model, and outputting a converged stroke analysis model;

[0092] Step S40, loading the facial expression recognition results and the preprocessing set in real time, and the stroke analysis model recognizes and analyzes the facial expression recognition results and the preprocessing set to calculate the patient's stroke risk level;

[0093] Step S50, judging whether the patient's stroke risk level exceeds a preset risk threshold based on the patient's stroke risk level;

[0094] Step S60: If the risk exceeds the preset risk threshold, a stroke emergency instruction is triggered.

[0095] If the preset risk threshold is not exceeded, the patient's stroke risk level is saved.

[0096] In this embodiment, the preset risk threshold is set to 0.55-0.6.

[0097] In an embodiment of the present invention, a stroke analysis model based on TOPSIS and artificial intelligence is pre-constructed, and the stroke analysis model cooperates with a multi-view feature positioning model and a flexible trigger mechanism, so that the stroke analysis model can quickly respond to the recognition results of the multi-view feature positioning model and the flexible trigger mechanism, thereby improving the recognition accuracy and efficiency of the stroke analysis model. The multi-view feature positioning model can analyze facial parameter information from multiple angles, thereby overcoming the problem that the existing artificial intelligence facial feature point recognition model cannot accurately capture facial feature points under emergency conditions, resulting in insufficient response speed and accuracy when judging through the TOPSIS method.

[0098] The embodiment of the present invention provides a method for preprocessing patient physiological parameter information based on a flexible trigger mechanism. Figure 2 The schematic diagram of the implementation process of the method for preprocessing the patient's physiological parameter information based on the flexible trigger mechanism is shown. The method for preprocessing the patient's physiological parameter information based on the flexible trigger mechanism specifically includes:

[0099] Step S101, loading the patient's physiological parameter information, identifying the vital sign monitoring information in the patient's physiological parameter information, generating a vital sign monitoring set, and performing missing value processing on the vital sign monitoring set;

[0100] In this embodiment, the missing value processing method for the vital sign monitoring set can be a deletion method, a filling method, or a multiple interpolation method, wherein the deletion method can be a list deletion method or a pairwise deletion method.

[0101] Step S102, constructing a judgment matrix of physical sign indicators in the physical sign monitoring set based on the hierarchical analysis method, and determining the indicator weights of a single group of physical sign indicators in the physical sign monitoring set in combination with the judgment matrix of the physical sign indicators;

[0102] Among them, the judgment matrix of physical sign indicators is expressed as:

[0103] (1)

[0104] in, Represents the judgment matrix of physical sign indicators, For physical indicators and physical indicators The importance between

[0105] The indicator weight of a single group of physical signs is calculated by the following formula:

[0106] (2)

[0107] (3)

[0108] (4)

[0109] in, Represents the indicator weight of a single group of physical signs indicators, is the maximum eigenvalue of the judgment matrix, is the number of physical sign indicators. In this embodiment, the number of physical sign indicators is 2-10. represents the normalized value of the physical sign index, is the initial weight of a single group of physical sign indicators. The initial weight of a single group of physical sign indicators can be determined by expert consultation method and entropy value method. Respectively represent the accumulated values ​​of the rows and columns of the judgment matrix;

[0110] Step S103, obtaining the indicator weight of the single group of vital sign indicators, and determining the flexible trigger threshold and event integral threshold of the single group of vital sign indicators based on the indicator weight of the single group of vital sign indicators;

[0111] It should be noted that when determining the event integral threshold, a continuous multi-point estimation method can be used to determine it. The physical sign indicators are continuously segmented at multiple points based on the indicator weights to obtain multiple binary variables. The relative risk ratio of each segmentation point is then calculated respectively. The event integral threshold is determined by observing the trend changes in the relative risk ratio of the segmentation points. The event integral threshold refers to the abnormal events corresponding to a single group of physical sign indicators. For example, when monitoring the patient's heart rate signs, during the collection period, the event integral threshold refers to the number of abnormal heart rate tachycardia, bradycardia or abnormal rhythm. If the number of abnormal heart rate exceeds the event integral threshold during the collection period, the abnormal monitoring set is uploaded.

[0112] Step S104, traversing the vital sign monitoring set, extracting vital sign monitoring information exceeding the event integral threshold in the vital sign monitoring set, and generating an abnormal monitoring set;

[0113] Step S105, performing anomaly detection on the abnormal monitoring set within the collection period based on the isolation forest anomaly detection algorithm, and calculating the flexible trigger integral value of the abnormal monitoring set within the collection period;

[0114] The flexible trigger integral value is calculated by the following formula:

[0115] (5)

[0116] in, is the flexible trigger integral value, Indicates that a single group of physical signs indicators The event integration threshold at time, Indicates the current moment, is the collection cycle, Indicates the number of events exceeding the event integration threshold during the acquisition period;

[0117] Step S106, determining whether the flexible trigger integral value exceeds the flexible trigger threshold;

[0118] Step S107: If the flexible trigger threshold is exceeded, the flexible trigger mechanism is triggered, and the flexible trigger integral value of the abnormal monitoring set within the collection period is uploaded.

[0119] If the flexible trigger threshold is not exceeded, the Markov prediction algorithm is used to perform parameter lapse on the abnormal monitoring set within the acquisition cycle, thereby reducing the system load;

[0120] In the embodiment of the present invention, the patient's physiological parameter information is preprocessed based on the flexible trigger mechanism, so that the warning threshold and judgment criteria can be dynamically adjusted according to the changes in the patient's physiological parameters, thereby improving the accuracy of identifying stroke precursors. The flexible trigger mechanism helps to reduce false alarms and missed alarms, ensuring that alarms can be issued in time at critical moments. The flexible trigger mechanism cooperates with the stroke analysis model, which not only improves the data quality and model performance of the model, but also can perform parameter flow on abnormal monitoring sets within the acquisition cycle, thereby reducing system load.

[0121] The embodiment of the present invention provides a method for identifying facial parameter information in a preprocessing set based on a multi-view feature positioning model. Figure 3 The present invention shows a schematic diagram of the implementation process of the method for identifying facial parameter information in a preprocessing set based on a multi-view feature positioning model. The method for identifying facial parameter information in a preprocessing set based on a multi-view feature positioning model specifically includes:

[0122] Step S201, loading the facial parameter information in the preprocessing set, and using the multi-view feature positioning model to correct the angle of the facial parameter information based on the cascaded posture regression algorithm combined with the isolation forest algorithm to obtain an angle adaptation set;

[0123] In this embodiment, the cascaded posture regression algorithm can effectively deal with the problem of locating the feature points of the face in various postures, especially in the case of large posture deflection angles. In this embodiment, the cascaded posture regression algorithm is combined with the isolation forest algorithm to correct the angle of facial parameter information, which can overcome the problem of difficulty in collecting facial expressions of stroke patients during the onset of the disease, and combined with the isolation forest algorithm, the robustness of the model to complex scenes and abnormal data can be further improved. As an integrated learning algorithm, the isolation forest algorithm can effectively identify and process outliers and noise data by constructing multiple random trees and calculating the average isolation degree of samples.

[0124] Step S202, obtaining an angle adaptation set, encoding local texture feature values ​​of the angle adaptation set based on a cascaded attitude regression algorithm, generating a binary feature vector, and integrating at least one set of binary feature vectors to obtain a binary feature set;

[0125] Step S203, loading the binary feature set, using the reduce_domain function to crop the binary feature set, reduce the face detection target area, and obtain at least one set of target area anchor frames;

[0126] Step S204, performing linear transformation on the target area anchor frame based on the cascaded posture regression algorithm combined with the Ghost module to generate a complete feature map;

[0127] It should be noted that the Ghost module generates more feature maps through low-cost operations, thereby enhancing the model's feature extraction capabilities. This helps to more accurately identify and locate target areas in complex scenes.

[0128] In step S205, CONVTranspose deconvolution restores the size of the feature map to obtain a feature recovery map. The facial droop degree of the feature recovery map is evaluated by the Bayesian criterion, and the facial symmetry degree is calculated in combination with the Euclidean distance algorithm. The Bayesian criterion evaluates the facial droop degree in the feature recovery map by calculating the prior probability and the posterior probability. Combining prior knowledge and observation data, the Bayesian algorithm can accurately update the probability distribution of events and provide a scientific basis for the degree of facial droop. In the embodiment of the present invention, the Bayesian criterion is combined with the Euclidean distance algorithm to more comprehensively evaluate the degree of facial droop and symmetry. The Bayesian criterion provides probability distribution updates, while the Euclidean distance accurately quantifies the symmetry differences. The two complement each other to improve the accuracy of the evaluation.

[0129] In this embodiment, the multi-view feature localization model is based on the cascaded posture regression model. The reduce_domain function is introduced in the cascaded posture regression model to crop the facial parameter information and reduce the facial detection target area. The multi-view feature localization model also introduces a Ghost module to perform a linear transformation on the feature map generated by the cascaded posture regression model to generate a complete feature map. The CONVTranspose deconvolution is introduced after the Ghost module to restore the size of the feature map.

[0130] The loss function of the multi-view feature localization model is defined as:

[0131] (6)

[0132] in, For the current image Number of sagging points in the mid-face, Represents the predicted value of pixel feature, Represents the true value of the pixel feature, They are respectively a penalty parameter for reducing the angle influence of the cascaded posture regression model and a classification parameter for the sample facial symmetry degree. In this embodiment, the penalty parameter for reducing the angle influence of the cascaded posture regression model can be 0.1-0.5, and the classification parameter for the sample facial symmetry degree can be 0.1-0.8.

[0133] In the embodiment of the present invention, the degree of facial droop in the feature recovery image is evaluated by the Bayesian criterion, and the degree of facial symmetry is calculated in combination with the Euclidean distance algorithm. The Bayesian criterion evaluates the degree of facial droop in the feature recovery image by calculating the prior probability and the posterior probability. Combining prior knowledge and observation data, the Bayesian algorithm can accurately update the probability distribution of events and provide a scientific basis for the degree of facial droop. Combining the Bayesian criterion with the Euclidean distance algorithm can more comprehensively evaluate the degree of facial droop and symmetry. The Bayesian criterion provides probability distribution updates, while the Euclidean distance accurately quantifies the symmetry differences. The two complement each other to improve the accuracy of the evaluation.

[0134] In an embodiment of the present invention, a multi-view feature localization model is provided. The multi-view feature localization model is based on a cascaded posture regression model. A reduce_domain function is introduced in the cascaded posture regression model to crop facial parameter information. A Ghost module is also introduced to perform a linear transformation on the feature map that has been generated by the cascaded posture regression model. By combining features from different perspectives, the model can more accurately capture subtle changes in the face, thereby improving the recognition accuracy of stroke symptoms. By cropping irrelevant facial parameter information, the amount of calculation is reduced, thereby improving the operating efficiency of the stroke analysis model.

[0135] The embodiment of the present invention provides a method for iteratively training a stroke analysis model using a training set and a test set. Figure 4 The following is a schematic diagram of the implementation process of the iterative training method for the stroke analysis model using the training set and the test set. The method of iterative training the stroke analysis model using the training set and the test set specifically includes:

[0136] Step S301, load the TOPSIS model, set the loss function, activation function, training rounds and hyperparameters of the TOPSIS model;

[0137] In this embodiment, the loss function of the data preprocessing module in the TOPSIS model can be a binary cross entropy loss function, and the activation function can be a ReLU function, while the loss function of the matrix construction module, the distance calculation module, and the risk level calculation module can be a cross entropy loss function, and the activation function can be a Sigmoid function. The training rounds are 100-120 times, and the learning rate hyperparameter is set to 0.002.

[0138] Step S302, loading a training set, using the training set to perform unsupervised pre-training on a data pre-processing module, introducing a conditional random field during training, and outputting a data processing set;

[0139] It should be noted that when training the stroke analysis model, the training set and the test set are obtained by using a web crawler. The web crawler technology refers to a program or script that automatically captures Internet information according to certain rules, and extracts the required data, such as pictures, texts, videos, etc., from it, and then saves or further processes the data. In this embodiment, the web crawler technology can be the Scrapy crawler framework, which uses an asynchronous request processing method, greatly improving the crawling speed and enabling large-scale crawling projects to be realized.

[0140] Step S303, obtaining a data processing set, iteratively training the matrix construction module and the distance calculation module through the data processing set combined with adversarial training, calculating the gradient of the loss function of the TOPSIS model to the data processing set, perturbing the data processing set along the direction of the gradient, and generating adversarial samples;

[0141] Step S304, loading the adversarial sample, projecting the adversarial sample into the data processing set based on the projected gradient descent method, using the data processing set to integrate the data preprocessing module, matrix construction module, distance calculation module, and risk level calculation module in the TOPSIS model, and outputting a converged stroke analysis model;

[0142] Step S305, loading the test set, taking the test set as input, executing the stroke analysis model, the stroke analysis model recognizes and analyzes the test set, and outputs the test result;

[0143] Step S306, determining whether the test result meets a preset accuracy threshold;

[0144] In this embodiment, the preset accuracy threshold of the stroke analysis model may be 0.9-0.95.

[0145] Step S307: if the preset accuracy threshold is met, output a converged stroke analysis model.

[0146] If it does not meet the preset accuracy threshold, the process returns to step S302 to continue iterative training of the stroke analysis model.

[0147] In this embodiment, the stroke analysis model uses the TOPSIS model as the initial model. The TOPSIS model includes a data preprocessing module, a matrix construction module, a distance calculation module, and a risk level calculation module. The data preprocessing module includes three hierarchical analysis layers, and the hierarchical analysis layers are connected through the TCP protocol.

[0148] When pre-building a stroke analysis model based on TOPSIS and artificial intelligence, the data pre-processing module in the TOPSIS model is improved, and the CRNN network and the generative adversarial network are introduced into the data pre-processing module. The CRNN network is used to extract follow-up voice information, continuously map the high-dimensional input follow-up voice information to the low-dimensional feature space, and output the low-dimensional semantic feature vector. The hierarchical analysis layer normalizes the semantic feature vector. The generative adversarial network consists of a generator and a discriminator. The generator is used to extract the feature vectors of facial parameter information and vital sign monitoring information, and the facial parameter information and vital sign monitoring information are weighted and classified based on the discriminator, and the facial feature vector and vital sign monitoring vector are input into the corresponding hierarchical analysis layer.

[0149] The matrix construction module is used to construct a weighted decision matrix using patient basic information and patient medical record information as prior information, combined with facial feature vectors, vital sign monitoring vectors, and semantic feature vectors.

[0150] It should be noted that CRNN (Convolutional Recurrent Neural Network) combines the advantages of convolutional neural network (CNN) and recurrent neural network (RNN), making it perform well in processing sequence data. In speech recognition tasks, CRNN first uses CNN to extract local features, and then uses RNN to process time series information, thereby efficiently extracting meaningful features from speech signals.

[0151] The embodiment of the present invention provides a method for analyzing facial expression recognition results and preprocessing set recognition using a stroke analysis model. Figure 5 The figure shows a schematic diagram of the implementation process of the stroke analysis model for facial expression recognition results and pre-processing set recognition and analysis methods, and the stroke analysis model for facial expression recognition results and pre-processing set recognition and analysis methods specifically include:

[0152] Step S401, loading facial expression recognition results and preprocessing sets, importing facial expression recognition results and preprocessing sets into a data preprocessing module, wherein a CRNN network in the data preprocessing module continuously maps high-dimensional input follow-up voice information to a low-dimensional feature space, generates an adversarial network to extract feature vectors of facial droop, facial symmetry, and physical sign monitoring information, and based on a discriminator, classifies and discriminates the facial droop, facial symmetry, and physical sign monitoring information weights, and integrates to obtain a semantic feature vector, a facial feature vector, a physical sign monitoring vector, and a basic information vector;

[0153] Step S402, obtaining a semantic feature vector, a facial feature vector, a vital sign monitoring vector, and a basic information vector, constructing a feature vector matrix based on the semantic feature vector, the facial feature vector, the vital sign monitoring vector, and the basic information vector, and normalizing the feature vector matrix by a hierarchical analysis layer to obtain a normalized matrix;

[0154] In the embodiment of the present invention, a multimodal feature vector matrix is ​​constructed by fusing semantic feature vectors, facial feature vectors, vital sign monitoring vectors and basic information vectors, so that the multimodal feature vector matrix can fully capture the physiological and psychological characteristics of patients and provide rich data support for emergency stroke identification.

[0155] Step S403, loading the normalized matrix, the matrix construction module constructs a feature decision matrix based on the feature weights corresponding to the semantic feature vector, the facial feature vector, the vital sign monitoring vector, and the basic information vector;

[0156] It should be noted that in the feature decision matrix, the weight allocation of different feature vectors is crucial. By analyzing historical data through machine learning algorithms, feature weights can be optimized to improve the accuracy and reliability of stroke risk prediction.

[0157] Step S404, obtaining a feature decision matrix, generating an ideal solution and a negative ideal solution based on the feature decision matrix, and a distance calculation module calculating the distance between the normalized matrix and the ideal solution and the negative ideal solution based on a Euclidean distance algorithm;

[0158] In the embodiment of the present invention, based on the discriminator's classification of the degree of facial droop, the degree of facial symmetry, and the weight of physical sign monitoring information, a semantic feature vector, a facial feature vector, a physical sign monitoring vector, and a basic information vector are integrated to avoid the influence of subjective and objective factors brought about by a single weight calculation method, making the evaluation results of the stroke analysis model more accurate and objective.

[0159] Step S405, the risk level calculation module loads the normalized matrix and the distance between the ideal solution and the negative ideal solution, calculates the relative closeness based on the distance between the ideal solution and the negative ideal solution, and uses the relative closeness as the patient's stroke risk level.

[0160] It should be noted that the patient's stroke risk level can be 0-5, among which level 0 is healthy, level 1 is low-risk group, level 2 is medium-low risk group, level 3 is medium-risk group, level 4 is medium-high risk group, and level 5 is high-risk group.

[0161] In this embodiment, when the discriminator classifies the degree of facial droop, the degree of facial symmetry, and the weight of the physical sign monitoring information, the discriminator's discriminant formula is expressed as:

[0162] (7)

[0163] (8)

[0164] in, Represents the discriminator's output result. is the number of characteristic indicators. In this embodiment, the number of characteristic indicators may be 3. They are the feature input matrix and adversarial vector matrix respectively. represents the number of times the discriminator counteracts disturbances, represents the classification function of the discriminator, is the discriminator perturbation factor, which is 0.002. is the feature weight, is the initial weight of the features based on principal component analysis, is the correlation coefficient of the feature, and the feature correlation coefficient is 0.02-0.3;

[0165] The relative closeness calculated based on the distance between the ideal solution and the negative ideal solution is expressed by the following formula:

[0166] (9)

[0167] in, Relative closeness, are the distances between the normalized matrix and the ideal solution and the negative ideal solution, respectively.

[0168] In the embodiment of the present invention, the relative closeness of the distance calculation based on the ideal solution and the negative ideal solution can effectively improve the accuracy of stroke risk assessment. This method provides a more accurate risk level classification by quantifying the gap between various indicators of the patient and the ideal state, which helps to identify high-risk patients at an early stage.

[0169] The embodiment of the present invention provides a stroke emergency recognition system combining TOPSIS and artificial intelligence. Figure 6 The structural diagram of the stroke emergency recognition system combining TOPSIS and artificial intelligence is shown. The stroke emergency recognition system combining TOPSIS and artificial intelligence specifically includes:

[0170] The information acquisition module 100 is used to acquire the patient's physiological parameter information in real time, and pre-process the patient's physiological parameter information based on a flexible trigger mechanism to obtain a pre-processed set;

[0171] The expression recognition module 200 is used to traverse the preprocessing set, identify the facial parameter information in the preprocessing set based on the multi-view feature positioning model, and generate a facial expression recognition result;

[0172] The risk identification module 300 is used to pre-build a stroke analysis model based on TOPSIS and artificial intelligence, load facial expression recognition results and pre-processing sets in real time, and the stroke analysis model recognizes and analyzes the facial expression recognition results and pre-processing sets to calculate the patient's stroke risk level;

[0173] The emergency trigger module 400 determines whether the patient's stroke risk level exceeds a preset risk threshold based on the patient's stroke risk level, and triggers a stroke emergency instruction if the stroke risk level exceeds the preset risk threshold.

[0174] It should be noted that the information collection module 100, the expression recognition module 200, the risk identification module 300, and the emergency trigger module 400 are connected by 5G or Bluetooth communication. The stroke emergency identification system combining TOPSIS and artificial intelligence provided in the embodiment of the present invention corresponds to the stroke emergency identification method combining TOPSIS and artificial intelligence, which will not be elaborated here.

[0175] In this embodiment, the information collection module 100 includes:

[0176] Basic information collection unit 110, used to collect basic patient information and patient medical record information;

[0177] A facial acquisition unit 120, used for acquiring facial parameter information;

[0178] A vital sign collection unit 130, used for collecting vital sign monitoring information;

[0179] The preprocessing unit 140 is used to load the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information, and preprocess the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information.

[0180] In the embodiment of the present invention, the information acquisition module 100 is composed of a basic information acquisition unit 110, a facial acquisition unit 120, a vital sign acquisition unit 130, and a preprocessing unit 140. The basic information acquisition unit 110, the facial acquisition unit 120, and the vital sign acquisition unit 130 cooperate to collect patient physiological parameter information in many aspects, thereby providing sufficient data support for emergency stroke identification. The facial acquisition unit 120 can be a facial animation acquisition device or a facial expression capture device, and the vital sign acquisition unit 130 can be a vital sign monitor or a heart rate monitor.

[0181] In summary, the present invention provides a stroke emergency identification method and system combining TOPSIS and artificial intelligence. In an embodiment of the present invention, a stroke analysis model based on TOPSIS and artificial intelligence is pre-constructed, and the stroke analysis model cooperates with a multi-view feature positioning model and a flexible trigger mechanism, so that the stroke analysis model can quickly respond to the recognition results of the multi-view feature positioning model and the flexible trigger mechanism, thereby improving the recognition accuracy and efficiency of the stroke analysis model. The multi-view feature positioning model can analyze facial parameter information from multiple angles, thereby overcoming the problem that the existing method artificial intelligence facial feature point recognition model cannot accurately capture facial feature points in an emergency state, resulting in insufficient response speed and accuracy when judging by the TOPSIS method.

[0182] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0183] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A stroke emergency recognition method combining TOPSIS and artificial intelligence, characterized in that: include: Collecting the patient's physiological parameter information in real time, preprocessing the patient's physiological parameter information based on a flexible trigger mechanism to obtain a preprocessing set, wherein the patient's physiological parameter information includes the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information; Traverse the preprocessing set, identify the facial parameter information in the preprocessing set based on the multi-view feature positioning model, and generate facial expression recognition results, where the facial expression recognition results include facial symmetry, facial droop, and expression classification results. The multi-view feature positioning model is based on the cascaded posture regression model, introduces the reduce_domain function to crop the facial parameter information, the Ghost module performs linear transformation on the feature map generated by the cascaded posture regression model, and introduces CONVTranspose deconvolution to restore the size of the feature map; The facial parameter information in the recognition preprocessing set includes: The multi-view feature localization model corrects the angle of facial parameter information based on the cascaded posture regression algorithm combined with the isolation forest algorithm to obtain the angle adaptation set; Based on the cascaded attitude regression algorithm, the local texture feature values ​​of the angle adaptation set are encoded to obtain a binary feature set; The reduce_domain function crops the binary feature set, reduces the target area for face detection, and obtains at least one set of target area anchor boxes; Perform linear transformation on the target area anchor box to generate a complete feature map; CONVTranspose deconvolution restores the size of the feature map, evaluates the degree of facial droop in the feature restoration map through the Bayesian criterion, and calculates the degree of facial symmetry in combination with the Euclidean distance algorithm; Pre-build a stroke analysis model based on TOPSIS and artificial intelligence, use a web crawler to obtain a modeling sample set, use a flexible trigger mechanism combined with a multi-view feature positioning model to pre-process the modeling sample set, divide the modeling sample set into a training set and a test set, use the training set and the test set to iteratively train the stroke analysis model, and output a converged stroke analysis model. The stroke analysis model uses the TOPSIS model as the initial model, including a data pre-processing module, a matrix construction module, a distance calculation module, and a risk level calculation module. The CRNN network and the generative adversarial network are introduced into the data pre-processing module. The facial expression recognition results and preprocessing sets are loaded in real time. The stroke analysis model recognizes and analyzes the facial expression recognition results and preprocessing sets and calculates the patient's stroke risk level. Based on the patient's stroke risk level, it is determined whether it exceeds the preset risk threshold. If it exceeds the preset risk threshold, a stroke emergency instruction is triggered.

2. The stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in claim 1, characterized in that: The method for preprocessing patient physiological parameter information based on the flexible trigger mechanism includes: Load the patient's physiological parameter information, identify the vital sign monitoring information in the patient's physiological parameter information, generate a vital sign monitoring set, and perform missing value processing on the vital sign monitoring set; Based on the analytic hierarchy process, a judgment matrix of physical sign indicators in the physical sign monitoring center is constructed, and the weights of the single group of physical sign indicators in the physical sign monitoring center are determined by combining the judgment matrix of physical sign indicators. Among them, the judgment matrix of physical sign indicators is expressed as: （1） in, Represents the judgment matrix of physical sign indicators, For physical indicators and physical indicators The importance between The indicator weight of a single group of physical signs is calculated by the following formula: （2） （3） （4） in, Represents the indicator weight of a single group of physical signs indicators, is the maximum eigenvalue of the judgment matrix, is the number of physical signs, represents the normalized value of the physical sign index, is the initial weight of a single group of physical signs, Respectively represent the accumulated values ​​of the rows and columns of the judgment matrix; Obtaining the indicator weight of a single group of vital sign indicators, and determining the flexible trigger threshold and event integral threshold of the single group of vital sign indicators based on the indicator weight of the single group of vital sign indicators; Traverse the vital sign monitoring set, extract the vital sign monitoring information exceeding the event integral threshold in the vital sign monitoring set, and generate an abnormal monitoring set; Based on the isolation forest anomaly detection algorithm, anomaly detection is performed on the anomaly monitoring set within the collection period, and the flexible trigger integral value of the anomaly monitoring set within the collection period is calculated; The flexible trigger integral value is calculated by the following formula: （5） in, is the flexible trigger integral value, Indicates that a single group of physical signs indicators The event integration threshold at time, Indicates the current moment, is the collection cycle, Indicates the number of events exceeding the event integration threshold during the acquisition period; Determine whether the flexible trigger integral value exceeds the flexible trigger threshold. If it exceeds the flexible trigger threshold, trigger the flexible trigger mechanism and upload the flexible trigger integral value of the abnormal monitoring set within the collection period.

3. The stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in claim 1, characterized in that: The multi-view feature localization model is based on the cascaded posture regression model. The reduce_domain function is introduced in the cascaded posture regression model to crop facial parameter information and reduce the facial detection target area. The multi-view feature localization model also introduces a Ghost module to perform a linear transformation on the feature map generated by the cascaded posture regression model to generate a complete feature map. The CONVTranspose deconvolution is introduced after the Ghost module to restore the size of the feature map. The loss function of the multi-view feature localization model is defined as: （6） in, For the current image Number of sagging points in the mid-face, Represents the predicted value of pixel feature, Represents the true value of the pixel feature, They are respectively the penalty parameter for reducing the angle influence of the cascaded posture regression model and the classification parameter for the degree of facial symmetry of the sample.

4. The stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in claim 1, characterized in that: The method for iteratively training the stroke analysis model using a training set and a test set specifically includes: Load the TOPSIS model, set the loss function, activation function, training rounds and hyperparameters of the TOPSIS model; Load the training set, use the training set to perform unsupervised pre-training on the data preprocessing module, introduce conditional random fields during training, and output the data processing set; Obtain a data processing set, iteratively train the matrix construction module and the distance calculation module through the data processing set combined with adversarial training, calculate the gradient of the loss function of the TOPSIS model for the data processing set, perturb the data processing set along the direction of the gradient, and generate adversarial samples; Load the adversarial sample, project the adversarial sample into the data processing set based on the projected gradient descent method, use the data processing set to integrate the data preprocessing module, matrix construction module, distance calculation module, and risk level calculation module in the TOPSIS model, and output a converged stroke analysis model; Load the test set, use the test set as input, execute the stroke analysis model, the stroke analysis model recognizes and analyzes the test set, outputs the test results, and determines whether the test results meet the preset accuracy threshold. If they meet the preset accuracy threshold, output the converged stroke analysis model.

5. The stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in claim 4, characterized in that: The stroke analysis model uses the TOPSIS model as the initial model. The TOPSIS model includes a data preprocessing module, a matrix construction module, a distance calculation module, and a risk level calculation module. The data preprocessing module includes three hierarchical analysis layers, and the hierarchical analysis layers are connected through the TCP protocol. When pre-building a stroke analysis model based on TOPSIS and artificial intelligence, the data pre-processing module in the TOPSIS model is improved, and the CRNN network and the generative adversarial network are introduced into the data pre-processing module. The CRNN network is used to extract follow-up voice information, continuously map the high-dimensional input follow-up voice information to the low-dimensional feature space, and output the low-dimensional semantic feature vector. The hierarchical analysis layer normalizes the semantic feature vector. The generative adversarial network consists of a generator and a discriminator. The generator is used to extract the feature vectors of facial parameter information and vital sign monitoring information, and the facial parameter information and vital sign monitoring information are weighted and classified based on the discriminator, and the facial feature vector and vital sign monitoring vector are input into the corresponding hierarchical analysis layer. The matrix construction module is used to construct a weighted decision matrix using patient basic information and patient medical record information as prior information, combined with facial feature vectors, vital sign monitoring vectors, and semantic feature vectors.

6. The stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in claim 5, characterized in that: The method of the stroke analysis model for facial expression recognition results and preprocessing set recognition analysis specifically includes: Load the facial expression recognition results and preprocessing set, import the facial expression recognition results and preprocessing set into the data preprocessing module, the CRNN network in the data preprocessing module continuously maps the high-dimensional input follow-up voice information to the low-dimensional feature space, generates the adversarial network to extract the feature vectors of the facial droop degree, facial symmetry degree, and physical sign monitoring information, and classifies the facial droop degree, facial symmetry degree, and physical sign monitoring information weights based on the discriminator, and integrates the semantic feature vector, facial feature vector, physical sign monitoring vector, and basic information vector; Acquire semantic feature vectors, facial feature vectors, vital sign monitoring vectors, and basic information vectors, construct feature vector matrices based on the semantic feature vectors, facial feature vectors, vital sign monitoring vectors, and basic information vectors, and perform normalization processing on the feature vector matrix by the hierarchical analysis layer to obtain a normalized matrix; Load the normalized matrix. The matrix construction module constructs a feature decision matrix based on the feature weights corresponding to the semantic feature vector, facial feature vector, vital sign monitoring vector, and basic information vector. Obtain a feature decision matrix, generate an ideal solution and a negative ideal solution based on the feature decision matrix, and a distance calculation module calculates the distance between the normalized matrix and the ideal solution and the negative ideal solution based on the Euclidean distance algorithm; The risk level calculation module loads the distance between the normalized matrix and the ideal solution and the negative ideal solution, calculates the relative closeness based on the distance between the ideal solution and the negative ideal solution, and uses the relative closeness as the patient's stroke risk level.

7. The stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in claim 6, characterized in that: When the discriminator classifies the degree of facial droop, the degree of facial symmetry, and the weight of the physical sign monitoring information, the discriminator's discriminant formula is expressed as: （7） （8） in, Represents the discriminator's output result. is the number of characteristic indicators, They are the feature input matrix and adversarial vector matrix respectively. represents the number of times the discriminator counteracts disturbances, represents the classification function of the discriminator, is the discriminator perturbation factor, is the feature weight, is the initial weight of the features based on principal component analysis, is the correlation coefficient of the feature; The relative closeness calculated based on the distance between the ideal solution and the negative ideal solution is expressed by the following formula: （9） in, Relative closeness, are the distances between the normalized matrix and the ideal solution and the negative ideal solution, respectively.

8. A stroke emergency identification system combining TOPSIS and artificial intelligence, used to implement the stroke emergency identification method combining TOPSIS and artificial intelligence as claimed in any one of claims 1 to 7, characterized in that: The stroke emergency recognition system combining TOPSIS and artificial intelligence includes: An information collection module is used to collect the patient's physiological parameter information in real time, and pre-process the patient's physiological parameter information based on a flexible trigger mechanism to obtain a pre-processed set; The expression recognition module is used to traverse the preprocessing set, identify the facial parameter information in the preprocessing set based on the multi-view feature positioning model, and generate facial expression recognition results; The risk identification module is used to pre-build a stroke analysis model based on TOPSIS and artificial intelligence, load facial expression recognition results and pre-processing sets in real time, and the stroke analysis model identifies and analyzes the facial expression recognition results and pre-processing sets to calculate the patient's stroke risk level; The emergency trigger module determines whether the patient's stroke risk level exceeds the preset risk threshold based on the patient's stroke risk level. If the preset risk threshold is exceeded, the stroke emergency instruction is triggered.

9. The stroke emergency recognition system combining TOPSIS and artificial intelligence as claimed in claim 8, characterized in that: The information collection module comprises: Basic information collection unit, used to collect basic patient information and patient medical record information; A facial acquisition unit, used for acquiring facial parameter information; A vital sign collection unit, used for collecting vital sign monitoring information; The preprocessing unit is used to load the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information, and to preprocess the patient's basic information, facial parameter information, vital sign monitoring information, follow-up voice information, and patient medical record information.

Images