Service system for rapidly evaluating and screening multi-region risks of cerebral apoplexy

Abstract

The invention discloses a service system for rapid assessment and screening of multi-region risks of cerebral apoplexy, and relates to the technical field of assessment and screening of cerebral apoplexy, and the system comprises an information classification module which is used for receiving multiple pieces of symptom information of a cerebral apoplexy patient, inputting a symptom classification table, outputting a face image classification identifier, a limb image classification identifier, a voice data classification identifier and a physiological parameter classification identifier; the weight generation module is used for retrieving a first cerebral apoplexy patient abnormal state sample set meeting the illness duration and complication set of the cerebral apoplexy patient, distributing the first cerebral apoplexy patient abnormal state sample set in the brain partition map and generating attention weight distribution; and the risk prompting module is used for retrieving a second cerebral apoplexy patient abnormal state sample set meeting various classification identifiers, distributing the second cerebral apoplexy patient abnormal state sample set in the brain partition map, extracting brain partitions of which the distribution number of second samples is greater than or equal to a distribution number threshold value, and executing risk prompting. The technical problem that the information richness degree of the risk assessment result of the stroke patient is insufficient in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of stroke assessment and screening, and particularly relates to a service system for rapid assessment and screening of stroke multi-region risks. BACKGROUND

[0002] In recent years, machine learning technology has been increasingly applied in the field of risk assessment of stroke patients. Existing technologies usually collect physiological parameters, image data and other risk indicators of patients, and construct a unified prediction model for overall risk assessment. For example, the patents with publication numbers CN119517393A and CN109864745A both disclose risk assessment models based on machine learning, but they are mostly based on the collection of risk data of stroke patients, and the application models are trained generally, so that the risk assessment results are insufficient in information richness, resulting in low reference value. SUMMARY

[0003] The application provides a service system for rapid assessment and screening of stroke multi-region risks, which is used to solve the technical problem that the general prompt limited to overall risk probability in the prior art results in insufficient information richness of the risk assessment results of stroke patients.

[0004] In view of the above problems, the application provides a service system for rapid assessment and screening of stroke multi-region risks, which comprises:

[0005] An information classification module is configured to receive a facial image, a limb image, voice data and physiological parameters of a stroke patient, input a symptom classification table, and output a facial image classification identifier, a limb image classification identifier, a voice data classification identifier and a physiological parameter classification identifier.

[0006] A weight generation module is configured to retrieve a first abnormal state sample set of stroke patients that meets the disease duration and complication set of the stroke patient, distribute the sample set in a brain sub-region graph, extract the proportion of the number of first samples, and generate an attention weight distribution.

[0007] A risk prompt module is configured to retrieve a second abnormal state sample set of stroke patients that meets the facial image classification identifier, the limb image classification identifier, the voice data classification identifier and the physiological parameter classification identifier, distribute the sample set in a brain sub-region graph, extract brain sub-regions with a second sample distribution number greater than or equal to a distribution number threshold, and perform a risk prompt, wherein the second sample distribution number is a distribution number corrected by the attention weight distribution product.

[0008] One or more technical solutions provided in the application have at least the following technical effects or advantages:

[0009] The present application achieves the goal of collecting the duration of illness and complication sets that represent the individualized status data of stroke patients, and distributing attention weights on brain regions; combining the symptom data and attention weights of stroke patients, and outputting the statistical data of the regions with relatively high abnormal frequencies of similar users; the present application realizes the partition evaluation, and the individual status of the patient is taken into account during the partition evaluation, thereby achieving the technical effect of improving the information richness of the risk assessment results, which is conducive to assisting the medical end in the risk management of stroke patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A schematic diagram of the structure of a service system for rapid multi-regional stroke risk assessment and screening provided in an embodiment of the present application;

[0012] Figure 2 This is a schematic diagram of the execution steps of a risk prompt module in a service system for rapid multi-region stroke risk assessment and screening provided in an embodiment of the present application.

[0013] In the accompanying drawings, the components represented by the reference numerals are described as follows:

[0014] Information classification module 100, weight generation module 200, risk warning module 300. DETAILED DESCRIPTION

[0015] This application provides a service system for rapid assessment and screening of multi-regional stroke risks, which is used to solve the technical problem that the existing technology is limited to general prompts of overall risk probability, resulting in insufficient information richness in the risk assessment results of stroke patients.

[0016] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] It should be noted that the terms "including" and "having" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0018] Example 1, as Figure 1 As shown, the present application provides a service system for rapid assessment and screening of multi-regional stroke risk, wherein the system includes:

[0019] The information classification module 100 is used to receive facial images, limb images, voice data and physiological parameters of a stroke patient, input a symptom classification table, and output a facial image classification identifier, a limb image classification identifier, a voice data classification identifier and a physiological parameter classification identifier;

[0020] Existing stroke risk assessment systems lack a unified symptom mapping mechanism when receiving multi-source heterogeneous data (such as facial, limb, voice, and physiological parameters), resulting in insufficient data collaborative processing capabilities. Traditional methods usually analyze various types of data independently, and are unable to establish a standardized association path between cross-modal symptoms and brain abnormalities, resulting in the omission or misassociation of key symptom features. Especially when faced with unstructured data, the raw data cannot be directly used for risk modeling and requires manual secondary labeling, which seriously hinders assessment efficiency. Distributed symptom information cannot support the unified identification input required for subsequent brain region risk focusing, becoming the first technical obstacle to refined assessment.

[0021] The information classification module 100 in the system provided in the embodiment of the present application includes:

[0022] Receive facial images, limb images, voice data and physiological parameters of a stroke patient, input a symptom classification table, and output facial image classification identifiers, limb image classification identifiers, voice data classification identifiers and physiological parameter classification identifiers, wherein the symptom classification table has embedded facial image symptom classifiers, limb image symptom classifiers, voice data symptom classifiers and physiological parameter symptom classifiers;

[0023] Processing the facial image through a facial image symptom classifier to obtain a facial image classification identifier;

[0024] Processing the limb image through a limb image symptom classifier to obtain a limb image classification identifier;

[0025] Processing the voice data through a voice data symptom classifier to obtain a classification identifier for the voice data;

[0026] Processing the physiological parameter by a physiological parameter symptom classifier to obtain the physiological parameter classification identifier;

[0027] The step of processing the facial image by a facial image symptom classifier to obtain the facial image classification identifier includes:

[0028] Obtaining, through a user terminal, a facial symptom set of a stroke patient, wherein the facial symptom set of the stroke patient comprises facial symptoms of the first stroke patient to facial symptoms of the Nth stroke patient;

[0029] Collecting a first stroke patient facial image dataset including the first stroke patient's facial symptoms, and constructing a first dataset using 1 as an identifier, where 1 represents facial symptoms of the first stroke patient;

[0030] collecting a second stroke patient facial image dataset that does not include the first stroke patient's facial symptoms, and constructing a second dataset using 0 as an identifier, where 0 represents the absence of the first stroke patient's facial symptoms;

[0031] Based on the first data set and the second data set, training a first stroke patient facial symptom classifier until an N-th stroke patient facial symptom classifier is trained;

[0032] The first stroke patient facial symptom classifiers up to the Nth stroke patient facial symptom classifiers are integrated to obtain the facial image symptom classifier.

[0033] In an embodiment of the present application, facial images, limb images, voice data and physiological parameters of a stroke patient are received, and a symptom classification table is input for classification and identification, wherein the symptom classification table is embedded with a facial image symptom classifier, a limb image symptom classifier, a voice data symptom classifier and a physiological parameter symptom classifier.

[0034] Taking the face classifier as an example, a neural network is used to build a face classifier.

[0035] With the informed consent of the user, a facial symptom set of stroke patients is legally collected through the user terminal, wherein the facial symptom set of stroke patients consists of the facial symptoms of the first stroke patient to the facial symptoms of the Nth stroke patient.

[0036] A first stroke patient facial image dataset including facial symptoms of a first stroke patient is collected, and a first dataset is constructed with 1 as an identifier, where 1 represents facial symptoms of the first stroke patient.

[0037] A second stroke patient facial image dataset that does not include the facial symptoms of the first stroke patient is collected, and a second dataset is constructed with 0 as the identifier, where 0 represents the absence of the first stroke patient's facial symptoms.

[0038] A similar method is used to perform acquisition integration until the third to 2Nth data sets are acquired and constructed.

[0039] A neural network was used to construct a first stroke patient facial symptom classifier. The first stroke patient facial symptom classifier consists of three layers: an input layer with one node receiving facial image data, a hidden layer with 128 nodes activated by the Reluctance Unit (ReLU) function, and an output layer with one node activated by the Sigmoid function to output the classification results. A binary cross-entropy loss function was used. Supervised iterative training was performed using the first and second datasets until the classifier converged (i.e., the output classification accuracy exceeded 95%), indicating that the first stroke patient facial symptom classifier training was complete.

[0040] Using a similar training approach to the first stroke patient facial symptom classifier, the second to N stroke patient facial symptom classifiers are trained. The trained first to N stroke patient facial symptom classifiers are integrated to obtain a facial image symptom classifier.

[0041] Similarly, the same construction and training method as the facial image symptom classifier is used to construct and train the limb image symptom classifier, voice data symptom classifier and physiological parameter symptom classifier to obtain the trained limb image symptom classifier, voice data symptom classifier and physiological parameter symptom classifier.

[0042] The facial image is processed by a facial image symptom classifier to obtain a facial image classification identifier.

[0043] The limb images are processed by a limb image symptom classifier to obtain a limb image classification identifier.

[0044] The speech data is processed by a speech data symptom classifier to obtain a speech data classification identifier.

[0045] The physiological parameters are processed by a physiological parameter symptom classifier to obtain a physiological parameter classification identifier.

[0046] Finally, the facial image classification identifier, limb image classification identifier, voice data classification identifier and physiological parameter classification identifier obtained by classification are output.

[0047] The symptom classification table is used to achieve automatic identification conversion of multimodal data, converting heterogeneous data such as raw images and speech into standardized classification identifiers. This processing not only retains key symptom characteristics, but also eliminates the interference of data source differences on subsequent analysis, providing a structured input foundation for cross-modal symptom collaboration. Compared with traditional feature extraction, this mechanism significantly improves the robustness of data processing. For example, it can still output stable identifiers for ambiguous speech or low-quality limb images, avoiding delays caused by manual intervention. The extracted identification system is more directly connected to brain partition association analysis, laying a foundation for accurate and reliable data in the future.

[0048] The weight generation module 200 is used to retrieve a first stroke patient abnormal state sample set that meets the stroke patient's illness duration and complication set, distribute it in the brain area map, extract the first sample distribution quantity ratio, and generate an attention weight distribution.

[0049] Existing technologies use static sample sets for risk assessment, ignoring the dynamic impact of the duration of illness and the combination of complications on the evolution of brain damage. For example, the same facial paralysis symptom may correspond to different brain region risk distributions in patients in the acute and chronic phases. Traditional models do not quantify such individual differences, resulting in sample distributions that fail to reflect the true state. High-risk brain regions in patients with complications may be diluted by universal samples, and disease-specific risk signals may be submerged. The lack of an individualized weighting mechanism causes subsequent risk analysis to deviate from the patient's actual pathological process, and regional assessments are inaccurate.

[0050] The weight generation module 200 in the system provided in the embodiment of the present application includes:

[0051] Based on the illness duration and the illness duration deviation threshold, a first search constraint is constructed; based on the complication set, a complication intersection-over-union ratio threshold is configured and set as a second search constraint;

[0052] configuring a logical AND relationship for the first search constraint and the second search constraint to generate a search constraint;

[0053] Collecting the first stroke patient abnormal state sample set that meets the search constraint, distributing it on a brain region map, extracting the first sample distribution quantity ratio, and generating an attention weight distribution;

[0054] The method includes collecting the first stroke patient abnormal state sample set that meets the search constraints, distributing the sample set on the brain region map, extracting the first sample distribution quantity ratio, and generating the attention weight distribution, which includes:

[0055] Extracting a data set of abnormal brain regions from the first stroke patient abnormal state sample set, distributing the data on the brain region map, and obtaining a first brain region distribution number percentage to a Qth brain region distribution number percentage;

[0056] Configure the quantity ratio threshold through the user end;

[0057] Delete the brain region distribution quantity ratios that are less than or equal to the quantity ratio threshold from the first brain region distribution quantity ratios to the Qth brain region distribution quantity ratios, to obtain the first reconstructed brain region distribution quantity ratios to the Mth reconstructed brain region distribution quantity ratios;

[0058] The first reconstructed brain region distribution quantity ratio is traversed until the Mth reconstructed brain region distribution quantity ratio is obtained, and the ratio is added to the reconstructed brain region distribution quantity ratio to obtain the attention weight distribution.

[0059] In an embodiment of the present application, a first retrieval constraint is constructed based on the duration of illness and in combination with a deviation threshold for the duration of illness. The deviation of the duration of illness reflects the similarity between the duration of illness and the duration of illness data in the sample library. The deviation of the duration of illness = |duration of illness - duration of illness in the sample library|. The smaller the deviation, the more similar it is to the sample library data, and the more likely the status is to be similar. The deviation threshold for the duration of illness refers to the threshold of the absolute difference between the duration of illness and the duration of illness in the sample library. For example, the deviation threshold for the duration of illness is set to 3 months, and the deviation of the duration of illness is less than 3 months, which is considered to meet the first constraint.

[0060] Based on the complication set, configure a complication IoU threshold as the second search constraint. The IoU reflects the similarity between a complication and the complication in the sample database. The IoU = (complications ∩ complications in the sample database) ÷ (complications ∪ complications in the sample database). A larger IoU indicates a more similar complication to the complication data in the sample database, and a more likely similar status. For example, the IoU threshold is set to 60%, and complications with an IoU greater than 60% are considered to meet the second search constraint.

[0061] A logical AND relationship is configured for the first search constraint and the second search constraint to generate a search constraint, that is, a search constraint is considered to be satisfied if both the first search constraint and the second search constraint are satisfied.

[0062] We searched the first set of stroke patient abnormality samples and extracted a dataset of abnormal brain regions that met the search constraints. We then distributed and annotated these regions on a standard medical brain region map, obtaining the distribution percentages for the first, and continuing through the Qth, brain regions. The distribution percentage for each brain region is calculated as: the number of abnormal regions divided by the number of brain regions.

[0063] Configure the percentage threshold through the user. This threshold is a pre-set percentage value. A percentage below the threshold is less likely to occur frequently and have a smaller impact over time. For example, the percentage threshold is set to 0.1.

[0064] The brain region distribution counts from the first brain region distribution count to the Qth brain region distribution count are deleted if they are less than or equal to the count threshold, obtaining the first reconstructed brain region distribution counts through the Mth reconstructed brain region distribution counts. Deleting less frequently occurring regions makes subsequent weight calculations more accurate.

[0065] The first reconstructed brain part area distribution quantity proportion is traversed until the Mth reconstructed brain part area distribution quantity proportion, and a ratio is obtained by adding the reconstructed brain part area distribution quantity proportions. The first attention weight = the first reconstructed brain part area distribution quantity ÷ the sum of the reconstructed brain part area distribution quantity proportions, for example, the first reconstructed brain part area distribution quantity is 20, and the sum of the reconstructed brain part area distribution quantity proportions is 100, then the first attention weight = 20 ÷ 100 = 0.2. Similarly, the second to Mth attention weights are calculated, and the attention weight distribution is integrated. A greater weight is given to a higher number of occurrences, which can more accurately adapt to a specific state.

[0066] In the present application, based on the disease duration deviation threshold and the complication intersection ratio, a retrieval constraint is dynamically constructed, and the distribution weight of the first sample set in the brain part area graph is extracted. This mechanism converts individual differences into attention quantitative values of each brain area, for example, the temporal lobe area obtains a higher weight due to the high frequency of similar disease course samples. Compared with the static model, the attention weight distribution can accurately adapt the sample set to the current patient state, and significantly improves the individual adaptability.

[0067] The risk prompt module 300 is configured to retrieve a second brain stroke patient abnormal state sample set satisfying the face image classification identifier, the limb image classification identifier, the voice data classification identifier, and the physiological parameter classification identifier, distribute the second sample set in a brain part area graph, and extract a brain part area with a second sample distribution quantity greater than or equal to a distribution quantity threshold. The risk prompt is executed, wherein the second sample distribution quantity is a distribution quantity corrected by the product of the attention weight distribution.

[0068] The traditional method directly matches the sample distribution according to the symptoms to execute the risk prompt, without correcting the deviation of the sample set itself. When the symptom identifier corresponds to multiple brain areas, the system cannot distinguish the primary and secondary risk sources, resulting in overload of the prompt information or omission of the key area. For example, dysarthria may be related to the basal ganglia area or the frontal lobe area, but if the frontal lobe area data is missing in the sample set, the traditional method will incorrectly focus on the basal ganglia area. This uncorrected coarse-grained analysis makes the risk assessment lose the brain area specificity and cannot support accurate decision-making.

[0069] The execution steps of the risk prompt module 300 in the system provided by the embodiments of the present application are as follows Figure 2 , including:

[0070] The first classification identifier is extracted from the face image classification identifier, the limb image classification identifier, the voice data classification identifier, and the physiological parameter classification identifier.

[0071] The first brain abnormality partition set commonly triggered in history is retrieved as a constraint.

[0072] The first classification identifier is used as a constraint to retrieve a first set of brain abnormality partitions that are commonly triggered by historical information, including:

[0073] Taking the first classification identifier as a constraint, the historical common abnormal trigger frequency ratio of the brain regions is counted and set as the brain region support set;

[0074] Based on the brain region support set, extracting brain regions whose support is greater than or equal to a support threshold, and adding them to the first brain abnormality region set;

[0075] Until the Yth brain abnormality partition set that is commonly triggered by the history is retrieved with the Yth classification identifier as a constraint;

[0076] Extracting an initial sample distribution quantity based on the first brain abnormality partition set to the Yth brain abnormality partition set and the second stroke patient abnormal state sample set distributed on the brain partition map;

[0077] Based on the attention weight distribution, multiplying the initial sample distribution quantity to generate the second sample distribution quantity;

[0078] sorting the brain regions according to the second sample distribution quantity from large to small to obtain a brain region sequence;

[0079] Based on the brain region sequence, risk prompting is performed.

[0080] In the embodiment of the present application, the first classification identifier to the Yth classification identifier are extracted from the facial image classification identifier, the limb image classification identifier, the voice data classification identifier and the physiological parameter classification identifier.

[0081] Using the first classification identifier as a constraint, we retrieve the set of first brain abnormality partitions that have been jointly triggered in the past. Jointly triggered brain abnormality partitions are those that are triggered by more than one abnormal state. These partitions are more dangerous and more likely to reflect high-frequency abnormalities.

[0082] Taking the first classification identifier as a constraint, the historical common abnormal trigger frequency ratios of brain regions are counted, and the frequency ratios are integrated to form the brain region support set.

[0083] Based on the brain region support set, extract brain regions whose support is greater than or equal to a support threshold. The support threshold is a pre-set threshold reflecting the support of a brain region. A higher support indicates a greater risk for that brain region, while a lower support indicates a lower risk and a lesser impact. For example, the support threshold is set to 0.2. Brain regions whose support is greater than or equal to the support threshold are added to the first set of abnormal brain regions.

[0084] Repeat the addition until the Yth set of brain abnormality partitions that are commonly triggered by the history are retrieved with the Yth classification identifier as a constraint.

[0085] According to the first brain abnormality partition set to the Yth brain abnormality partition set and the second stroke patient abnormal state sample set, which are distributed on the brain partition map, the initial sample distribution number is extracted. The initial sample distribution number includes the abnormal state number distribution corresponding to each brain abnormal partition.

[0086] Based on the attention weight distribution, the initial sample distribution is multiplied to generate the second sample distribution. The second sample distribution is calculated as follows: the initial sample distribution is calculated as the number of samples multiplied by the attention weight. For example, if the initial sample distribution is 20 and the attention weight is 0.6, the second sample distribution is calculated as 20 × 0.6 = 12.

[0087] Sort the brain regions from largest to smallest based on the number of samples in the second distribution to obtain a brain region sequence. The higher the brain region in the sequence, the greater the risk and the more attention it requires.

[0088] Risk warnings are implemented based on the brain region sequence. For example, for abnormalities in brain regions near the front of the sequence, a voice prompt is used: "High risk, please pay attention to abnormal conditions and seek medical attention promptly." For abnormalities in brain regions near the back of the sequence, a text prompt is used: "Please pay attention to your health and seek medical examinations promptly."

[0089] In this application, the number of second sample distributions matching the symptom markers is multiplied and corrected by focusing on the weight distribution, and then the brain regions exceeding the threshold are screened for prompts. This achieves the effect of dual optimization of risk positioning: on the one hand, weight correction compensates for sample set bias, such as increasing the weight of brain regions with low sample volume but high weight to adapt to individual differences; on the other hand, a distribution number threshold is used to filter weakly associated brain regions. The final output brain region sequence accurately reflects the strength of the association between symptoms and brain regions, significantly improving the accuracy of risk positioning and individual adaptability compared to the general prompts of traditional methods.

[0090] In summary, the embodiments of the present application have at least the following technical effects:

[0091] This application proposes a service system for rapid multi-regional stroke risk assessment and screening. By integrating multimodal symptom collaborative analysis of facial images, limb images, voice, and physiological parameters, combined with a dynamic weight correction mechanism based on duration and complications, the system significantly improves the accuracy of stroke risk positioning and adaptability to individual differences. Compared with traditional methods, the technical solution provided by this application significantly breaks through the limitations of single-dimensional assessment: First, a classifier is used to finely identify and classify multi-source heterogeneous data, constructing a correlation mapping between cross-modal symptoms and brain region abnormalities, solving the problem that traditional models are insufficient for collaborative analysis of complex symptoms; second, a sample set is generated by constraining the intersection and union ratio of illness duration deviation thresholds and complications, and brain region distribution weights are extracted, achieving quantitative integration of individual patient characteristics. This allows the assessment results to break away from the homogeneous output of static models and more accurately adapt to individual patient conditions; further, based on the number of weight-corrected second sample distributions, highly correlated brain regions are identified. Combined with a brain region serialization risk warning mechanism, the general overall risk assessment is transformed into a targeted warning for specific functional regions, refining the warning results.

[0092] This application achieves the technical effect of accurate stroke risk positioning. By collecting the duration of illness and complication sets that represent the individual status data of stroke patients, attention weights are distributed to brain regions. Combined with the symptom data and attention weights of stroke patients, the regions with relatively high abnormal frequencies of similar users are statistically output. This application realizes partition evaluation, and the individual status of the patient is taken into account during the partition evaluation, which significantly improves the individual adaptability of risk assessment and achieves the technical effect of increasing the information richness of risk assessment results, which is conducive to assisting the medical end in controlling the risks of stroke patients.

[0093] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0094] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0095] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A service system for rapid multi-regional stroke risk assessment and screening, characterized by: include: An information classification module is used to receive facial images, limb images, voice data and physiological parameters of a stroke patient, input a symptom classification table, and output a facial image classification identifier, a limb image classification identifier, a voice data classification identifier and a physiological parameter classification identifier; A weight generation module is used to retrieve a first stroke patient abnormal state sample set that meets the stroke patient's illness duration and complication set, distribute it on the brain area map, extract the first sample distribution quantity ratio, and generate an attention weight distribution; The risk warning module is used to retrieve a second set of abnormal state samples of stroke patients that meet the facial image classification identifier, the limb image classification identifier, the voice data classification identifier and the physiological parameter classification identifier, distribute them in a brain region map, extract brain regions whose second sample distribution quantity is greater than or equal to a distribution quantity threshold, and execute risk warnings, wherein the second sample distribution quantity is the distribution quantity corrected by the product of the attention weight distribution.

2. The system according to claim 1, wherein The method receives facial images, limb images, voice data and physiological parameters of a stroke patient, inputs a symptom classification table, and outputs facial image classification identifiers, limb image classification identifiers, voice data classification identifiers and physiological parameter classification identifiers, wherein the symptom classification table is embedded with a facial image symptom classifier, a limb image symptom classifier, a voice data symptom classifier and a physiological parameter symptom classifier, including: Processing the facial image through a facial image symptom classifier to obtain a facial image classification identifier; Processing the limb image through a limb image symptom classifier to obtain a limb image classification identifier; Processing the voice data through a voice data symptom classifier to obtain a classification identifier of the voice data; The physiological parameter is processed by a physiological parameter symptom classifier to obtain the physiological parameter classification identifier.

3. The system according to claim 2, wherein: Processing the facial image using a facial image symptom classifier to obtain a facial image classification identifier includes: Obtaining, through a user terminal, a facial symptom set of a stroke patient, wherein the facial symptom set of the stroke patient comprises facial symptoms of the first stroke patient to facial symptoms of the Nth stroke patient; Collecting a first stroke patient facial image dataset including the first stroke patient's facial symptoms, and constructing a first dataset using 1 as an identifier, where 1 represents facial symptoms of the first stroke patient; collecting a second stroke patient facial image dataset that does not include the first stroke patient's facial symptoms, and constructing a second dataset using 0 as an identifier, where 0 represents the absence of the first stroke patient's facial symptoms; Based on the first data set and the second data set, training a first stroke patient facial symptom classifier until an N-th stroke patient facial symptom classifier is trained; The first stroke patient facial symptom classifiers up to the Nth stroke patient facial symptom classifiers are integrated to obtain the facial image symptom classifier.

4. The system according to claim 1, wherein: Retrieve the first abnormal state sample set of stroke patients that meets the stroke patient's illness duration and complication set, distribute it on the brain area map, extract the proportion of the first sample distribution, and generate the attention weight distribution, including: Based on the illness duration and the illness duration deviation threshold, a first search constraint is constructed; based on the complication set, a complication intersection-over-union ratio threshold is configured and set as a second search constraint; configuring a logical AND relationship for the first search constraint and the second search constraint to generate a search constraint; The first stroke patient abnormal state sample set that meets the retrieval constraint is collected and distributed on a brain area map, the first sample distribution quantity ratio is extracted, and an attention weight distribution is generated.

5. The system according to claim 4, wherein: Collecting the first stroke patient abnormal state sample set that meets the search constraint and distributing it on a brain region map, extracting the first sample distribution quantity ratio, and generating an attention weight distribution, including: Extracting a data set of abnormal brain regions from the first stroke patient abnormal state sample set, distributing the data on the brain region map, and obtaining a first brain region distribution number percentage to a Qth brain region distribution number percentage; Configure the quantity ratio threshold through the user end; Delete the brain region distribution quantity ratios that are less than or equal to the quantity ratio threshold from the first brain region distribution quantity ratios to the Qth brain region distribution quantity ratios, to obtain the first reconstructed brain region distribution quantity ratios to the Mth reconstructed brain region distribution quantity ratios; The first reconstructed brain region distribution quantity ratio is traversed until the Mth reconstructed brain region distribution quantity ratio is obtained, and the ratio is added to the reconstructed brain region distribution quantity ratio to obtain the attention weight distribution.

6. The system according to claim 1, wherein: Retrieving a second stroke patient abnormal state sample set that meets the facial image classification identifier, the limb image classification identifier, the voice data classification identifier, and the physiological parameter classification identifier, distributed in a brain region map, and extracting the second sample distribution quantity, including: Extracting a first classification identifier to a Yth classification identifier from the facial image classification identifier, the limb image classification identifier, the voice data classification identifier, and the physiological parameter classification identifier; Using the first classification identifier as a constraint, searching for a first set of brain abnormality partitions that are commonly triggered by historical events; Until the Yth brain abnormality partition set that is commonly triggered by the history is retrieved with the Yth classification identifier as a constraint; Extracting an initial sample distribution quantity based on the first brain abnormality partition set to the Yth brain abnormality partition set and the second stroke patient abnormal state sample set distributed on the brain partition map; Based on the attention weight distribution, the initial sample distribution quantity is multiplied to generate the second sample distribution quantity.

7. The system according to claim 6, wherein: Using the first classification identifier as a constraint, searching for a first set of brain abnormality partitions that are commonly triggered by history includes: Taking the first classification identifier as a constraint, the historical common abnormal trigger frequency ratio of the brain regions is counted and set as the brain region support set; Based on the brain region support set, brain regions with support greater than or equal to a support threshold are extracted and added to the first brain abnormality region set.

8. The system according to claim 1, wherein: Execution risk warnings, including: sorting the brain regions according to the second sample distribution quantity from large to small to obtain a brain region sequence; Based on the brain region sequence, risk prompting is performed.

Images