A stroke behavior detection method based on BEFAST judgment principle

Abstract

The application discloses a kind of based on BEFAST determination principle's cerebral apoplexy behavior detection method, comprising the following steps: according to BEFAST determination principle collection human state's data set;The data collected is preprocessed;Build double neural network model based on deep learning;The neural network is trained using the data after preprocessing;The network model established by deep learning is used to detect the measured personnel, and the cerebral apoplexy symptom is detected;According to the detection result of network model, and the physical condition of measured personnel is reasonably suggested.The application combines deep learning technology with BEFAST determination principle, uses double neural network to judge the probability of cerebral apoplexy, compared with current technology, from more obvious feature, reduces the implementation difficulty;From the angle of multi-dimensional feature combination, greatly reduces the probability of misjudgment and omission.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical technology and computer vision, and particularly relates to a cerebral stroke detection method based on a BEFAST judgment principle. BACKGROUND

[0002] Cerebral stroke, commonly known as stroke, is called the "number one killer" of the health of the Chinese people, and its comprehensive harm has surpassed cardiovascular diseases and malignant tumors, with high morbidity, high recurrence rate, and high mortality and disability rate. It is a group of acute cerebrovascular diseases caused by sudden blockage or rupture of cerebral blood vessels, leading to ischemic and anoxic injury of local brain tissue, and is mainly divided into two types: ischemic stroke and hemorrhagic stroke, commonly known as "cerebral infarction" and "cerebral hemorrhage", which has great harm to human health.

[0003] In order to detect and treat cerebral stroke early, the medical community has proposed the "BEFAST" judgment principle. Among them, "B" - Balance, which refers to the sudden loss of balance or coordination; "E" - Eyes, which refers to sudden difficulty in seeing; "F" - Face, which refers to sudden facial asymmetry; "A" - Arms, which refers to sudden weakness or numbness of the arms, usually appearing on one side of the body; "S" - Speech, which refers to sudden speech or articulatory disorders; "T" - Time, which emphasizes that if any of the above symptoms occurs within a short period of time, it may indicate a cerebral stroke, and immediate medical attention is required, with each hour of delay causing the brain to age irreversibly by 3.6 years.

[0004] In recent years, with the rapid development of deep learning technology and its application in various fields, medical intelligent research based on deep learning has received increasing attention, and technologies and products for detecting cerebral stroke using artificial intelligence technology have also emerged. According to the search, the invention patent with Chinese patent number CN112489793A discloses a cerebral stroke risk patient early warning system, which detects the brain electrical signals of the measured person through a deep learning method, compares them with the established patient brain electrical signal database, and judges the cerebral stroke probability. The invention patent with Chinese patent number CN111419237A discloses a cerebral stroke hand movement function Carroll score prediction method, which is also a detection method based on deep learning technology, and its feature extraction focuses on hand movement posture to calculate the Carroll score and predict the cerebral stroke probability. In addition, there are numerous cerebral stroke prediction methods based on deep learning, such as detection technologies or products targeting upper limb movement function, speech characteristics, blood lipids, and medical history, which have been put into use in various medical institutions.

[0005] However, the comprehensive deep learning detection method for the most authoritative and most prominent stroke phenomenon detection-BEFAST judgment principle method is still in a blank state, and the above-mentioned single or non-prominent feature detection method, or the feature extraction is complex, such as blood lipids and various medical histories, etc. ; Or the single feature has high misjudgment rate. Therefore, a deep learning stroke detection method based on the BEFAST judgment principle is needed. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art and provide a stroke detection method based on the BEFAST judgment principle, which realizes intuitive, prominent, convenient and multi-dimensional stroke detection, reduces the detection difficulty and improves the detection accuracy.

[0007] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0008] A stroke detection method based on the BEFAST judgment principle, comprising the following steps:

[0009] A, collecting the data set of the human state according to the BEFAST judgment principle;

[0010] B, preprocessing the collected data;

[0011] C, building a neural network model MobileNetV3-Small and AGCN-Transformer based on deep learning;

[0012] D, training the neural network using the preprocessed data;

[0013] E, detecting the measured personnel through the network model established by deep learning to detect the symptoms of stroke;

[0014] F, giving reasonable suggestions according to the detection results of the network model and the physical condition of the measured personnel.

[0015] Optionally, the data set of the human state collected according to the BEFAST judgment principle includes B-Balance, limb balance data; F-Face, facial symmetry data; A-Arms, arm lifting and holding data.

[0016] Optionally, the B-Balance, limb balance data is image data of the balance / tilt degree of the limbs of stroke patients and normal people during walking. Because stroke patients often have sudden loss of balance or coordination, feel unstable when walking, and tilt to one side of the body; Or become clumsier than usual when doing fine motor movements.

[0017] Optionally, the F-Face, facial symmetry data is image data of the degree of facial symmetry / one-side skew of the stroke patient and normal person. Because stroke patients often have sudden facial asymmetry. Generally, face the patient and observe whether the nasolabial sulcus on both sides of the patient is symmetrical or whether one side is shallower; then make the patient smile or show teeth, and observe whether one side of the mouth is skewed.

[0018] Optionally, the A-Arms, arm lifting and holding data is image data of the stroke patient and normal person holding the arms lifted. Because stroke patients often have sudden weakness or numbness in one side of the body. Lift both arms and observe whether they can be lifted to the same height; if they can, further observe whether both arms can be held for 10 seconds. If one side of the arm cannot be lifted or falls down in advance, attention should be paid.

[0019] Optionally, the preprocessing of the collected data includes: scaling the data set collected in step A; classifying the cropped data set; calibrating the classified data according to requirements; and normalizing the classified and calibrated picture data

[0020] Optionally, the scaling of the data collected in step A refers to scaling all pictures in proportion to reach a unified length and width size.

[0021] Optionally, the unified size is 224x224, and the proportional scaling method is to fix the long side and fill the short side:

[0022] h scale =h size / max(h,w)

[0023] w scale =h scale

[0024] h dst =h scale *h

[0025] w dst =w scale *w

[0026] padding_size=abs(h dst -w dst ) / 2

[0027] wherein (h, w) is the original size of the picture, (h scale , w scale ) is the scaling ratio, (h dst , w dst) is the scaled picture size, and padding_size is the number of pixels needed to fill on each side of the short side. According to the calculation result (h dst , w dst ), the scaling is realized by using the bilinear interpolation method; if the original image has different length and width and needs to be filled, 0 values of the size of padding_size are added on each side of the short side.

[0028] Optionally, the classification of the cropped data set refers to classifying the picture data into six categories of normal body balance data, stroke patient body balance data, normal facial symmetry data, stroke patient facial symmetry data, normal arm lifting and holding data, and stroke patient arm lifting and holding data according to two classification standards of normal people and stroke patients, body balance data, facial symmetry data, and arm lifting and holding data.

[0029] Optionally, the calibration of the classified data according to requirements refers to skeleton calibration of human body limbs in Balance image data and Arms image data, and converting picture measured targets and background data into skeleton data.

[0030] Optionally, the normalization processing of the classified and calibrated picture data refers to normalization of picture pixels, and transforms.ToTensor() and transform.Normolize() functions are used to realize the normalization:

[0031] Among them, transform.ToTensor() first changes the storage mode of pixel data from (h, w, c) to (c, h, w) form, and calculates:

[0032] transforms.ToTensor(x)=x / 255

[0033] The data is normalized to the range of [0, 1]; then transform.Normolize() standardizes the data in the range of [0, 1] to data with a mean of 0 and a standard deviation of 1, so as to speed up the convergence speed:

[0034] transforms.Normolize(x)=(x-mean(x)) / std(x)

[0035] mean(x) is the mean of x, and std(x) is the standard deviation of x.

[0036] Optionally, the neural network model MobileNetV3-Small based on deep learning is built, and the structure thereof includes 11 Bottleneck layers, one standard convolution layer, and two pointwise convolution layers.

[0037] The bottleneck layer structure includes a 1x1 convolution layer for dimensionality expansion, a 3x3 or 5x5 convolution layer for feature extraction, and a 1x1 convolution layer for dimensionality reduction, which is mainly used to realize picture feature extraction.

[0038] The standard convolution layer is a two-dimensional convolution operation with a step of 2, a kernel size of 3x3, and an output channel of 16, which is used as a receptive field to perform preliminary feature extraction.

[0039] The pointwise convolution layer is a two-dimensional convolution operation with a step of 1 and a kernel size of 1x1, which is used as a fully connected layer to fuse inter-channel feature information.

[0040] The network uses two activation functions to introduce nonlinearity and improve network accuracy: RELU6 and Hardswish, which are calculated as follows:

[0041]

[0042]

[0043] The network also uses SE modules to automatically obtain the importance of each feature channel through learning, then enhances useful features and suppresses features that are not useful for the current task according to the importance, which is calculated as follows:

[0044]

[0045] where the AdaptiveAvgPool2d function takes the average value of all pixels in each channel, compresses the input feature map size to 1x1, and keeps the dimension (channel number) unchanged;

[0046] The Conv2d(in_channel, in_channel / reduction, kernel_size=1) function uses a 1x1 convolution to compress the input feature map dimension to 1 / reduction, and the RELU() function is:

[0047]

[0048] The Conv2d(in_channel / reduction, in_channel, kernel_size=1) function uses a 1x1 convolution to expand the input feature map dimension by a factor of reduction, and the h_sigmoid() function replaces the general sigmoid() function, which is calculated as follows:

[0049] h_sigmoid(x) = RELU6(x+3) / 6

[0050] Optionally, the construction is based on a deep learning neural network model AGCN-Transformer, which uses skeleton data as the input of the network model, processes the original skeleton data containing only position information to obtain second-order speed information for splicing, realizes the expansion and improvement of the initial spatial information of the node, and then performs normalization preprocessing on the spliced data, and then performs BN operation on the input data before inputting into the network, constructs a space-time graph of the skeleton sequence in space and time, and performs subsequent space graph convolution operation according to the adjacency matrix of the adaptive graph.

[0051] The model first combines a GCN basic unit and a TCN basic unit combined with an attention mechanism designed and constructed by an additional dropout layer to form an adaptive space-time graph convolution unit module.

[0052] In the shallow stage of the network, 3-layer space-time graph convolution unit modules are used to extract node spatial features and short-time time features, and residual connection structures are used in each space-time convolution unit module. The output feature channel number of the first three space-time convolution unit modules is 64. In the spatial dimension, that is, in the node dimension of the feature tensor, GAP is performed, so that the action features no longer contain the spatial information of the human structure.

[0053] The array is changed in dimension from left to right and from top to bottom in chronological order to obtain the representation of time in two-dimensional space, that is, continuous time and discontinuous time are both in the space adjacency of two-dimensional representation. The changed feature map is input into the Transformer network to extract time information, and the Transformer network structure can be composed of SwinMLP structure. Among them, the Swin-MLP structure about the channel number C of the first stage hidden layer and the hyperparameter setting of the number of layers in each stage is: C=96, layer numbers={2, 2, 2}, head numbers={3, 6, 3}.

[0054] After feature extraction by the multi-layer encoder layer of the Transformer network, the output feature is input into the full connection layer and connected to the softmax function for classification.

[0055] Optionally, the pre-processed Faces and Balance, Arms data are used to train neural networks MobileNetV3-Small and AGCN-Transformer, respectively. For MobileNetV3-Small: the pre-processed Faces data are subjected to a standard convolution layer to preliminarily extract feature maps, 11 Bottleneck structure layers to deeply extract features, a pooling layer to reduce dimension, and two pointwise convolution layers to fuse channel feature information of the feature maps and output classification results. For AGCN-Transformer: the pre-processed Balance and Arms data converted into skeleton data are subjected to 3 spatio-temporal graph convolution unit modules, a Transformer network, and a full connection layer and a softmax function for processing and calculation, to determine whether the subject is a cerebral apoplexy patient according to whether the subject's pace is stable and whether the subject's limbs move in coordination.

[0056] Optionally, the network model established by deep learning is used to detect the subject, and the step of detecting the symptoms of cerebral apoplexy is as follows:

[0057] (1) collecting image data of the subject;

[0058] (2) pre-processing the image data to obtain standard normalized image data and skeleton data with a size of 224x224;

[0059] (3) inputting the image data into the trained neural networks MobileNetV3-Small and AGCN-Transforme for calculation;

[0060] (4) obtaining results by network detection to determine the cerebral apoplexy condition.

[0061] Optionally, in the detection of the symptoms of cerebral apoplexy, if any one of the three features "B", "F", and "A" of the subject is similar to the symptoms of a cerebral apoplexy patient and is determined to be the same symptom, the subject is determined to be a cerebral apoplexy risk population; if any two of the three features are determined to be the same symptom, the subject is determined to be a high-probability cerebral apoplexy patient; and if all the three features are determined to be the same symptom, the subject is determined to be an emergency cerebral apoplexy patient.

[0062] Optionally, according to the detection results of the network model and the physical condition of the subject, reasonable suggestions are given. From the perspective of the detection results, if the subject is detected to be a cerebral apoplexy risk population, the subject is suggested to go to a hospital for examination as soon as possible, and information of a nearby hospital is provided; if the subject is detected to be a high-probability cerebral apoplexy patient, in addition to the above measures, 120 dialing service is provided; and if the subject is detected to be an emergency cerebral apoplexy patient, 120 is automatically dialed to provide emergency rescue service for the subject. In addition, from the perspective of the physical condition of the subject, cerebral apoplexy prevention measures or recovery suggestions are given.

[0063] The application combines deep learning technology and BEFAST judgment principle, adopts double neural networks to judge the probability of suffering from stroke, and compared with the current technology, judges from more obvious features, reduces the implementation difficulty, judges from the angle of multi-dimensional feature combination, and greatly reduces the misjudgment and omission probability.

[0064] The application fuses a space-time graph convolution network AGCN and a transformer structure, a time convolution network extracts time features in a local range, the transformer extracts long-range time information, and the deficiency of the previous time convolution network in long-range time information extraction is made up. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 A flow chart of the method of the application.

[0066] Figure 2 A structure diagram of the neural network model MobileNetV3-Small of the application.

[0067] Figure 3 A structure diagram of the neural network model AGCN-Transformer of the application. DETAILED DESCRIPTION

[0068] The technical scheme of the application will be further described in detail below in combination with the drawings, but the protection scope of the application is not limited to the following description.

[0069] As shown in the figure, a stroke detection method based on the BEFAST judgment principle comprises the following steps: Figure 1 A. According to the BEFAST judgment principle, the data set of the human state is collected:

[0070]

[0071] ​Three features of normal people and stroke patients are collected, namely B-Balance, body balance data; F-Face, face symmetry data; and A-Arms, arm lifting and holding data. Specifically, the B-Balance, body balance data is image data of the balance / tilt degree of the body of the stroke patients and normal people during walking. Because stroke patients often have sudden loss of balance or coordination, feel unstable when walking, and tilt to one side of the body; or become clumsier than usual when doing fine movements. The F-Face, face symmetry data is image data of the face symmetry / one-side tilt degree of the stroke patients and normal people. Because stroke patients often have sudden facial asymmetry. Generally, face the patient, observe whether the nasolabial sulcus on both sides of the patient is symmetrical, and whether one side is shallower; then make the patient smile or show teeth, and observe whether one side of the mouth is crooked. The A-Arms, arm lifting and holding data is image data of the stroke patients and normal people holding the arms horizontally after lifting. Because stroke patients often have sudden weakness or numbness in one side of the body. Lift both arms and see if they can be lifted to the same height; if they can, further observe whether both arms can be lifted for 10 seconds. If one side of the arm cannot be lifted or falls down in advance, attention should be paid.

[0072] B. Preprocessing the collected data:

[0073] This step includes scaling the data set collected in step A; classifying the cropped data set; calibrating the classified data according to requirements; and normalizing the classified and calibrated picture data.

[0074] Specifically, the scaling of the data collected in step A refers to scaling all pictures by the same proportion to achieve a unified length and width size. The unified size is 224x224, and the equal proportion scaling method is to fix the long side and fill the short side:

[0075] h scale =h size / max(h,w)

[0076] w scale =h scale

[0077] h dst =h scale *h

[0078] w dst =w scale *w

[0079] padding_size=abs(h dst -w dst) / 2

[0080] where (h, w) is the original size of the picture, (h scale , w scale ) is the scaling ratio, (h dst , w dst ) is the picture size after equal scaling, and padding_size is the number of pixels to be filled on each side of the short side. dst dst According to the calculated (h dst , w dst ), the scaling is realized by using the bilinear interpolation method; if the original image has different length and width and needs to be filled, 0 values of the size of padding_size are added on each side of the short side.

[0081] The classification of the cropped data set refers to dividing the picture data into six categories of normal body balance data, stroke patient body balance data, normal facial symmetry data, stroke patient facial symmetry data, normal arm lifting and holding data, and stroke patient arm lifting and holding data according to two classification standards of normal people and stroke patients, body balance data, facial symmetry data, and arm lifting and holding data.

[0082] The calibration of the classified data according to the requirements refers to the skeleton calibration of the human body limbs in the Balance image data and the Arms image data, and the conversion of the picture measured target and background data into skeleton data.

[0083] The normalization processing of the classified and calibrated picture data refers to the normalization of the picture pixels, which is realized by using the transforms.ToTensor() and transform.Normolize() functions:

[0084] Among them, transform.ToTensor() first changes the storage mode of pixel data from (h, w, c) to (c, h, w) form, and calculates:

[0085] transforms.ToTensor(x)=x / 255

[0086] The data is normalized to the range of [0, 1]; then transform.Normolize() standardizes the data in the range of [0, 1] to data with a mean of 0 and a standard deviation of 1, so as to speed up the convergence speed:

[0087] transforms.Normolize(x)=(x-mean(x)) / std(x)

[0088] mean(x) is the mean of x, and std(x) is the standard deviation of x.

[0089] C. Building a neural network model MobileNetV3-Small and AGCN-Transformer based on deep learning:

[0090] Specifically, the neural network model MobileNetV3-Small based on deep learning has a structure including 11 Bottleneck layers, one standard convolution layer, and two pointwise convolution layers, as shown in Figure 2 .

[0091] The Bottleneck layer structure includes a 1x1 convolution layer for dimensionality increase, a 3x3 or 5x5 convolution layer for feature extraction, and a 1x1 convolution layer for dimensionality reduction. This structure is mainly used to realize picture feature extraction.

[0092] The standard convolution layer is a two-dimensional convolution operation with a stride of 2, a convolution kernel size of 3x3, and an output channel of 16, which is used as a receptive field to perform preliminary feature extraction.

[0093] The pointwise convolution layer is a two-dimensional convolution operation with a stride of 1 and a convolution kernel size of 1x1, which is used as a fully connected layer to fuse feature information between channels.

[0094] This network uses two activation functions to introduce nonlinearity and improve network accuracy: RELU6 and Hardswish, which are calculated as follows:

[0095]

[0096]

[0097] Meanwhile, this network uses SE modules to automatically obtain the importance of each feature channel through learning, and then enhances useful features and suppresses features that are not useful for the current task according to the importance. The calculation method is as follows:

[0098]

[0099] where the AdaptiveAvgPool2d function takes the average value of all pixels in each channel, compresses the input feature map size to 1x1, and keeps the dimension (channel number) unchanged;

[0100] The Conv2d(in_channel, in_channel / reduction, kernel_size=1) function uses a 1x1 convolution to compress the input feature map dimension to 1 / reduction, and the RELU() function is:

[0101]

[0102] The function of Conv2d(in_channel / reduction, in_channel, kernel_size=1) expands the input feature map dimension by 1x1 convolution to reduction times, and the function of h_sigmoid() replaces the general sigmoid() function, which is calculated as follows:

[0103] h_sigmoid(x) = RELU6(x+3) / 6

[0104] Specifically, the neural network model AGCN-Transformer based on deep learning is built, which uses skeleton data as the input of the network model, processes the original skeleton data containing only position information to obtain second-order velocity information for splicing, realizes the expansion and perfection of the initial spatial information of the node. Then the spliced data is normalized and pretreated, and then the input data is subjected to BN operation before being sent into the network. The skeleton sequence is constructed into a space-time graph in space and time, and the adjacency matrix of the adaptive graph is constructed to perform subsequent space graph convolution operation.

[0105] The model first combines a GCN basic unit and a TCN basic unit combined with an additional dropout layer designed and constructed to form an adaptive space-time graph convolution unit module.

[0106] In the shallow stage of the network, 3 space-time graph convolution unit modules are used to extract the spatial features and short-time temporal features of the nodes, and residual connection structure is adopted for each space-time convolution unit module. The output feature channel number of the first three space-time convolution unit modules is 64. In the spatial dimension, i.e., in the node dimension of the feature tensor, GAP is performed to make the action features no longer contain the spatial information of the human structure.

[0107] The array is changed in dimension from left to right and from top to bottom in chronological order to obtain the representation of time in two-dimensional space, i.e., continuous time and discontinuous time can exist in space adjacency in two-dimensional representation. The changed feature map is sent into the Transformer network to extract the time information, and the Transformer network structure can be composed of SwinMLP structure. Among them, the Swin-MLP structure about the channel number C of the first stage hidden layer and the hyperparameter setting of the number of layers in each stage are: C=96, layer numbers={2, 2, 2}, head numbers={3, 6, 3}.

[0108] After the feature extraction by the multiple encoder layers of the Transformer network, the output feature is sent into the fully connected layer and connected to the softmax function for classification.

[0109] D. training the neural network using the preprocessed data:

[0110] Specifically, the neural network MobileNetV3-Small and AGCN-Transformer are trained using the preprocessed Faces and Balance, Arms data, respectively. For MobileNetV3-Small: the preprocessed Faces data is first extracted through a standard convolution layer, then deep features are extracted through 11 Bottleneck structure layers, and then the dimension is reduced through a pooling layer. Finally, the feature information of each channel of the feature map is fused through two pointwise convolution layers, and the classification result is output. For AGCN-Transformer: the preprocessed Balance and Arms data converted into skeleton data are processed and calculated through 3 layers of spatio-temporal graph convolution unit modules, Transformer network, fully connected layer and softmax function. According to whether the measured person's pace is stable and whether the limbs move in coordination, it is determined whether the person is a cerebral apoplexy patient.

[0111] E. detecting the measured person through the network model established by deep learning to detect the symptoms of cerebral apoplexy:

[0112] Specifically, the network model established by deep learning is used to detect the measured person to detect the symptoms of cerebral apoplexy, and the steps are:

[0113] (1) collecting image data of the measured person;

[0114] (2) pre-processing the image data to obtain standard normalized image data and skeleton data with a size of 224x224;

[0115] (3) inputting the image data into the trained neural network MobileNetV3-Small and AGCN-Transforme for calculation;

[0116] (4) obtaining the detection result through the network to determine the cerebral apoplexy condition.

[0117] Specifically, the detection of cerebral apoplexy symptoms, any one of the three features of the measured person "B", "F", and "A" is similar to the symptoms of cerebral apoplexy patients and is judged as the same symptom, then the measured person is judged as a cerebral apoplexy risk population; any two is judged as a high probability cerebral apoplexy patient; if all three symptoms are judged as the same symptoms of cerebral apoplexy, the measured person is judged as an emergency cerebral apoplexy patient.

[0118] F. giving reasonable suggestions according to the detection result of the network model and the physical condition of the measured person:

[0119] The detection result according to the network model and the physical condition of the measured person are used to give reasonable suggestions, from the detection result, if the detected person is a stroke risk group, the suggestion of going to the hospital for examination as soon as possible is given, and the information of the nearby hospital is provided; if the detected person is a high-probability stroke patient, in addition to the above measures, the 120 dialing service is provided; if the detected person is an emergency stroke patient, 120 is automatically dialed, and emergency rescue service is provided for the measured person. In addition, from the physical condition of the measured person, stroke prevention measures or recovery suggestions are given.

[0120] The present application combines deep learning technology and BEFAST judgment principle, adopts double neural network to judge the stroke probability, compared with the current technology, judges from more obvious features, reduces the implementation difficulty; judges from the angle of multi-dimensional feature combination, greatly reduces the misjudgment and omission probability.

[0121] The above is the preferred embodiment of the present application, it should be understood that the present application is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein by the above teaching or related technical or knowledge. The modification and change made by the person skilled in the art without departing from the spirit and scope of the present application shall be within the protection scope of the appended claims of the present application.

Claims

1. A method for detecting stroke based on BEFAST judgment principle, characterized in that, The method comprises the following steps: Step A: obtaining three features of normal people and stroke patients according to the BEFAST judgment principle to obtain a data set, the data including Balance body balance data images, Face facial symmetry data images and Arms arm lifting and holding data images; Step B: preprocessing the collected data set to obtain preprocessed data; Step C: building a neural network model MobileNetV3-Small and AGCN-Transformer based on deep learning, the AGCN-Transformer fusing a spatio-temporal graph convolutional network AGCN and a transformer structure, the time convolutional network extracting time features in a local range and the transformer extracting long-range time information, the MobileNetV3-Small being used for processing the Face facial symmetry data images and the AGCN-Transformer being used for processing the Balance body balance data images and the Arms arm lifting and holding data images; Step D: training the neural network model MobileNetV3-Small and AGCN-Transformer using the preprocessed data; Step E: detecting the detected personnel by the neural network model MobileNetV3-Small and AGCN-Transformer established by deep learning and outputting the detected stroke symptom behaviors.

2. The stroke detection method based on the BEFAST judgment principle according to claim 1, characterized in that: the Balance body balance data images include image data of the balance / tilt degree of the limbs of the stroke patients and normal people during walking; the Face facial symmetry data images are image data of the facial symmetry / one-side tilt degree of the stroke patients and normal people; the Arms arm lifting and holding data images are image data of the arm lifting and holding horizontally by the stroke patients and normal people. 3.The stroke detection method based on BEFAST decision rule of claim 1, wherein, The preprocessing of the collected data comprises: scaling the pictures in the data set collected in step A by the same proportion to reach a unified length-width size; classifying the scaled data set; labeling the classified data according to requirements; and normalizing the classified and labeled picture data. 4.The stroke detection method based on BEFAST decision rule of claim 3, wherein, The uniform size is The equal scaling method is long side fixed and short side filled. where (h, w) is the original size of the picture, (h scale ,w scale ) is the scaling ratio, (h dst ,w dst ) is the picture size after equal scaling, and padding_size is the number of pixels needed to be filled on each side of the short side, and the scaling is realized by using the bilinear interpolation method according to the calculated (h dst ,w dst ); for example, when the original image needs to be filled due to the inequality of the length and width, the value of 0 is added to the pixels of the size of padding_size on each side of the short side. 5.The stroke detection method based on BEFAST decision rule of claim 3, wherein, The classification of the scaled data set refers to classifying the body balance data, facial symmetry data and arm lifting and holding data picture data into six categories of normal body balance data, stroke patient body balance data, normal facial symmetry data, stroke patient facial symmetry data, normal arm lifting and holding data and stroke patient arm lifting and holding data according to two classification standards of normal people and stroke patients respectively. 6.The stroke detection method based on BEFAST decision rule of claim 3, wherein, The calibration of the classified data according to requirements refers to skeleton calibration of the human body limbs in the limb balance data image and the arm lifting and holding data image, and converting the image measured target and background data into skeleton data. 7.The stroke detection method based on BEFAST decision rule of claim 3, wherein, The normalization processing of the classified and calibrated picture data refers to normalizing the picture pixels, and using transforms.ToTensor() and transform.Normolize() functions to realize: where transform.ToTensor() first changes the storage mode of pixel data from (h, w, c) to (c, h, w) form, and through the calculation: Normalize the data to the range [0, 1]; then transform.Normolize() standardizes the data in the range [0, 1] to data with mean 0 and standard deviation 1 to speed up the convergence: mean(x) is the mean value of x, and std(x) is the standard deviation of x. 8.The stroke detection method based on BEFAST decision rule of claim 1, wherein, The neural network model MobileNetV3-Small based on deep learning is built, and the structure includes 11 Bottleneck layers, one standard convolution layer and two pointwise convolution layers; Bottleneck layer structure includes one layer Convolution is used to increase dimensionality, one layer or Convolution is used for feature extraction and one layer The convolution is used for dimensionality reduction; this structure is mainly used to extract image features. The standard convolution layer is a two-dimensional convolution operation with a step of 2, a convolution kernel size of , and 16 output channels, which is used as a receptive field for preliminary feature extraction. The point-wise convolution layer is a two-dimensional convolution operation with a step of 1 and a kernel size of fuses the feature information between channels as a full connection layer. The neural network model MobileNetV3-Small uses two kinds of activation functions to introduce nonlinearity and improve network accuracy: RELU6 and Hardswish, and the calculation method is: In the above formula, x is the image feature data input into the activation function; Meanwhile, the neural network model MobileNetV3-Small uses the SE module to automatically obtain the importance of each feature channel through learning, and then improves useful features and suppresses features that are not useful for the current task according to the importance, and the calculation method is: wherein, The function acts to average all pixels over each channel, compressing the input feature map size to and the dimension, number of channels, remains unchanged; The function adopts Convolution of the input feature map dimension is compressed to The RELU() function is: The function adopts Convolution will expand the input feature map dimension to times, The function replaces the general The calculation method is: ； The neural network model AGCN-Transformer based on deep learning is built, which uses skeleton data as the input of the neural network model AGCN-Transformer, processes the original skeleton data containing only position information to obtain second-order speed information for splicing, realizes the expansion and improvement of the initial spatial information of the node, then normalizes the spliced data, and then sends the input data into the neural network model AGCN-Transformer before the BN operation, constructs the adjacency matrix of the spatial and temporal graph according to the adaptive graph, and performs subsequent spatial graph convolution operation; The neural network model AGCN-Transformer first combines a designed GCN basic unit combined with an attention mechanism and a TCN basic unit combined with an additional dropout layer to form an adaptive spatio-temporal graph convolution unit module, namely a GCN module; In the shallow stage of the network, 3 spatio-temporal graph convolution unit modules are used to extract node spatial features and short-time temporal features, and residual connection structure is used in each spatio-temporal convolution unit module, the output feature channel number of the first three spatio-temporal convolution unit modules is 64, and GAP is performed in the spatial dimension, i.e. the node dimension of the feature tensor, so that the action features no longer contain spatial information of the human body structure. The feature information extracted by the AGCN module is arrayed in a dimension change manner from left to right and from top to bottom in chronological order to obtain a representation of time on a two-dimensional space, that is, both continuous time and discontinuous time exist in spatial adjacency in the two-dimensional representation. The changed feature map is input into the Transformer module to extract time information. The structure of the Transformer module is composed of the SwinMLP structure, wherein the Swin-MLP structure is about the channel number C of the first stage hidden layer and the hyperparameter setting of the number of layers of each stage: , , ; After the feature extraction by the multi-layer encoder layer of the Transformer module, the output feature is sent to the full connection layer and the softmax function for classification. 9.The stroke detection method based on BEFAST decision rule of claim 8, wherein, The preprocessed data is used to train the neural network models MobileNetV3-Small and AGCN-Transformer respectively, For the neural network model MobileNetV3-Small: the pre-processed Faces facial symmetry data image is subjected to a layer of standard convolution to preliminarily extract a feature map, 11 layers of bottleneck structure to deeply extract features, a pooling layer to reduce dimension, and two layers of pointwise convolution layers to fuse channel feature information of the feature map and output a classification result; For the neural network model AGCN-Transformer: the pre-processed Balance limb balance data image and Arms arm lifting and holding data image converted into skeleton data are subjected to 3 layers of spatio-temporal graph convolution unit modules, a Transformer network, a fully connected layer and a softmax function for processing and calculation, and according to whether the measured person's pace is stable and whether the limbs are coordinated, the judgment of whether the person is a cerebral apoplexy patient is realized. 10.The stroke detection method based on BEFAST decision rule of claim 1, wherein, The network model established by deep learning is used to detect the measured person, detect the symptoms and behaviors of cerebral apoplexy, and the steps are: (1) collecting image data of the measured person; (2) the image data is preprocessed to obtain standard normalized image data and skeleton data of the same size; (3) inputting the pre-processed image data into the trained neural network model MobileNetV3-Small and AGCN-Transformer for calculation; (4) obtaining the detection result by the network, and judging the prediction result of the symptoms and behaviors of cerebral apoplexy; The detection of the behaviors of cerebral apoplexy prediction includes: if any one of the three features of the measured person is similar to the symptoms of the cerebral apoplexy patient and is judged as the same symptom, the measured person is determined as a cerebral apoplexy risk population; if any two are determined as high-probability cerebral apoplexy patients; if all three symptoms are determined as the same symptoms of cerebral apoplexy, the measured population is determined as an emergency cerebral apoplexy patient; if the measured person is detected as a cerebral apoplexy risk population, the suggestion of going to the hospital for examination as soon as possible is given, and the information of the nearby hospital is provided; if the measured person is detected as a high-probability cerebral apoplexy patient, in addition to the above measures, the fast 120 dialing service is provided; if the measured person is detected as an emergency cerebral apoplexy patient, 120 is automatically dialed, and emergency rescue service is provided for the measured person.

Images