Self-optimized edge intelligent early-stage cerebral apoplexy prediction method and self-optimized edge intelligent early-stage cerebral apoplexy prediction system

Abstract

The invention belongs to the field of cerebral apoplexy prediction, and relates to a self-optimization edge intelligent cerebral apoplexy early prediction method and system, and the method comprises the steps: S1, collecting the marked data of cerebral apoplexy, and carrying out the preprocessing of the marked data, and obtaining a marked feature set D; s2, training the prediction model according to D to obtain a latest prediction model; initializing a pseudo mark feature set; s3, collecting data in real time and preprocessing the data to obtain a feature ft; s4, inputting ft into the latest prediction model to obtain a prediction result; combining the feature ft and a prediction result thereof to obtain a pseudo-mark feature sample ft '; s5, updating the newest pseudo-mark feature set according to ft'and judging whether a retraining condition is met or not, and if yes, executing the step S6; otherwise, returning to the step S3; s6, retraining the latest prediction model in combination with D and the updated pseudo-mark feature set Ft ', updating the feature set Ft' according to the latest prediction model to obtain a latest pseudo-mark feature set, and returning to the step S3; according to the method, the performance of the prediction model is improved by self-optimizing the pseudo-mark feature set.

Description

Technical Field

[0001] The present invention belongs to the technical field of early stroke onset prediction, and specifically relates to a self-optimizing edge intelligent early stroke prediction method and system. Background Art

[0002] Stroke, commonly known as "stroke," is a common cerebrovascular disease characterized by high morbidity, high disability and mortality rates, high recurrence rates, and a significant economic burden. According to the World Health Organization, stroke is the second leading cause of death worldwide, with elderly patients accounting for over 60% of all cases. Stroke has subtle pre-symptoms, irreversible and devastating damage, and is difficult to treat. This makes early identification and intervention crucial in managing stroke.

[0003] In recent years, existing technical solutions have achieved stroke classification prediction and risk assessment by using cloud platform architecture and traditional machine learning algorithms to build stroke prediction models, and then using medical clinical data for model training.

[0004] Existing technical solutions have achieved certain results in the field of early stroke onset prediction technology, but the following problems still exist: 1) High-quality labeled stroke data is scarce. The cost of clinical labeling of stroke data is high, the ratio of positive and negative samples in the labeled data is seriously unbalanced, and there are large differences in characteristics between the general population in the community and the clinical high-risk population, as well as between people in different regions. This makes it difficult to initially construct a high-quality, universal labeled stroke data set, which seriously affects the accuracy of prediction and generalization effect. 2) Poor real-time prediction, high resource requirements and costs. The prediction system based on medical imaging examination results, although relatively accurate in judgment, has high resource requirements and long prediction cycle intervals, and cannot meet the needs of continuous stroke early warning and monitoring for the general public. 3) High risk of privacy leakage. The prediction system based on cloud platforms and online websites aggregates sensitive health data to remote central nodes, which is prone to privacy leakage. Summary of the Invention

[0005] To solve the above-mentioned problems in the prior art, the present invention adopts a self-optimizing edge intelligent early stroke prediction method, comprising:

[0006] S1. Collect labeled stroke data to construct an initial stroke dataset, preprocess the initial stroke dataset, and obtain a labeled feature set;

[0007] S2. Build a stroke prediction model, train the stroke prediction model based on the labeled feature set, and obtain the latest trained stroke prediction model; initialize the pseudo-labeled feature set to obtain the latest pseudo-labeled feature set;

[0008] S3, collect the current data of the monitored object in real time, pre-process the collected data, and obtain real-time feature f t ;

[0009] S4, the real-time feature f t Input the latest trained stroke prediction model to get the prediction result; combined with the real-time feature f t And its prediction results, get the pseudo-label feature sample f t ';

[0010] S5, based on the pseudo-labeled feature sample f t 'Update the latest pseudo-label feature set, and determine whether the retraining condition is met based on the updated pseudo-label feature set. If so, execute step S6; otherwise, use the updated pseudo-label feature set as the latest pseudo-label feature set and return to step S3;

[0011] S6. Retraining the most recently trained stroke prediction model by combining the labeled feature set and the updated pseudo-labeled feature set to obtain a retrained stroke prediction model; and using the retrained stroke prediction model as the most recently trained stroke prediction model;

[0012] S7. Update the updated pseudo-label feature set according to the latest trained stroke prediction model to obtain the latest pseudo-label feature set, and return to step S3.

[0013] On the other hand, the present invention adopts a self-optimizing edge intelligent early stroke prediction system to implement the above-mentioned prediction method, including two major subsystems: edge prediction nodes and edge learning nodes, which are deployed in the edge nodes close to the monitored object.

[0014] Beneficial effects:

[0015] 1. The present invention divides the prediction system into an edge prediction node subsystem and an edge learning node subsystem, which interactively and iteratively implement new data collection, prediction and pseudo-labeling, data set update, prediction model retraining and edge redeployment update, so that the prediction system can continuously absorb new data and optimize the model, forming an intelligent learning and self-optimization mechanism, effectively improving the prediction accuracy and generalization; 2. The present invention constructs a pseudo-label optimization and screening mechanism to update the pseudo-label feature set, realizing large-capacity, high-quality, and continuous expansion and optimization of the pseudo-label feature set, effectively improving the scarcity, category imbalance and poor quality of labeled stroke data, and further improving the performance of the prediction model; 3. The present invention designs and deploys the prediction system at the edge nodes of home environments and community service stations, greatly reducing the risk of privacy leakage; 4. The present invention designs a prediction model restoration strategy to improve the robustness of the prediction model; 5. The present invention predicts stroke based on physiological characteristics, behavioral signs, and health record data collected by sensing devices deployed in the home environment, realizing low-resource, non-perceptual, and short-cycle interval real-time early warning monitoring of early onset stroke. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention provides an example diagram of a self-optimizing edge intelligent early stroke prediction system architecture;

[0017] Figure 2 The present invention provides a self-optimizing edge intelligent early stroke prediction method step flow chart;

[0018] Figure 3 The present invention provides a schematic diagram of the importance weights of feature items in a self-optimized edge intelligence stroke early prediction method;

[0019] Figure 4 The present invention provides a self-optimizing edge intelligent stroke early prediction method prediction model evaluation and maintenance flow chart;

[0020] Figure 5 The present invention provides a self-optimizing edge intelligent stroke early prediction method pseudo-marker screening and elimination flow chart. DETAILED DESCRIPTION

[0021] The following will be combined with the accompanying drawings in the embodiments of the present invention to describe the technical solutions of the present invention in an example application. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] like Figure 2As shown, the present invention adopts a self-optimizing edge intelligent early stroke prediction method, comprising the following steps:

[0023] S1: Collect labeled stroke data from recent years to construct an initial stroke dataset; preprocess the initial stroke dataset to obtain a labeled feature set;

[0024] Preferably, the labeled stroke data in recent years include: clinical diagnosis data of stroke in the past 10 years and medical record data of stroke screening in the past 15 years.

[0025] The preprocessing of the initial stroke data set includes: standard normalization, abnormality removal, and feature item dimensionality reduction.

[0026] Preferably, the standard normalization method includes: converting unordered binary values ​​to 1 or 2, and retaining the original values ​​for non-ordered binary values; determining the L1 norm based on the overall range of the feature item values ​​of each data sample, and then dividing the value of each feature item by the L1 norm. Table 1 below shows examples of data before and after standard normalization.

[0027] Table 1 Examples of data before and after normalization

[0028]

[0029] Preferably, the feature item dimensionality reduction method uses the fusion chi-square test (Chi2), decision tree feature weight (DTW), and variance analysis (AF) to construct a feature item composite score index; Figure 3 As shown in the figure, the importance score of each feature item in the initial data set is calculated, and the 10 feature items with the highest importance scores (age, gender, drinking history, smoking history, heart rate, pulse, blood pressure, blood sugar, sleep quality, and activity status) are selected to form a labeled feature set.

[0030] Preferably, the abnormal data cleaning method uses an isolation forest algorithm to identify abnormal samples in the initial data set and then remove them.

[0031] S2. Build a stroke prediction model, train the stroke prediction model according to the labeled feature set, and obtain the latest trained stroke prediction model M0; initialize the pseudo-labeled feature set to obtain the latest pseudo-labeled feature set F0;

[0032] The pseudo-mark feature set F0 includes: pseudo-mark update counter C 伪-添加 、C 伪-替换 、C 伪-剔除 , the pseudo-label update counter C of the pseudo-label feature set F0 伪-添加 、C 伪-替换 、C伪-剔除 Both are 0.

[0033] Constructing a stroke prediction model includes selecting one or a combination of multiple algorithms from alternative machine learning algorithms. The alternative machine learning algorithms include random forest, support vector machine (SVM), K nearest neighbor (KNN), Gaussian Bayes (GS), logistic regression, recurrent neural network (RNN), and XGBoost.

[0034] S3, collect the current data of the monitored object in real time through the data acquisition equipment, pre-process the collected data to obtain real-time feature f t ;

[0035] Preferably, the data collection equipment includes: wearable sensing devices (smart bracelets, armbands), environmental sensing devices (millimeter wave radar sensors, depth sensors), and information reporting devices (mobile phones, tablets).

[0036] The data of the monitored subjects include: physiological signs, behavioral characteristics, and health records; among them, physiological signs include: heart rate, pulse, blood pressure, and blood sugar; behavioral characteristics include: sleep quality and activity status; health record collection includes: age, gender, drinking history, and smoking history.

[0037] The preprocessing of the collected data includes: caching, filtering, outlier cleaning, and standard normalization of the collected real-time data.

[0038] Preferably, for physiological sign data, the number of cached data samples is 5; for behavioral characteristic data and health record data, new data is directly used to overwrite old data.

[0039] Preferably, a 5-point sliding weighted mean filter is performed on the physiological sign data to reduce noise interference.

[0040] S4, the real-time feature f t Input the latest trained stroke prediction model to get the prediction result; combined with the real-time feature f t And its prediction results, get the pseudo-label feature sample f t ';

[0041] When the real-time feature f t When the data is collected for the first time (i.e., real-time feature f1), the real-time feature f1 is input into the newly trained stroke prediction model M0 to obtain the prediction result; the pseudo-labeled feature sample f1′ is obtained by combining the real-time feature f1 and its prediction result.

[0042] Stroke prediction results include: binary pseudo-label results (using the numbers 1 and 0 to indicate whether the monitored subject has stroke or not, respectively) and their pseudo-label confidence (ranging from 0.5 to 1.0, with larger values ​​representing higher confidence). Reported pseudo-label feature samples include: pseudo-labels (numbers 1 or 0), their confidence, and real-time features.

[0043] S5, based on the pseudo-labeled feature sample f t 'Update the latest pseudo-label feature set, and determine whether the retraining condition is met based on the updated pseudo-label feature set. If so, execute step S6; otherwise, use the updated pseudo-label feature set as the latest pseudo-label feature set and return to step S3;

[0044] When the pseudo-labeled feature sample f t When ′ is a sample constructed from the data collected for the first time (i.e., the pseudo-labeled feature sample f1′), the latest pseudo-labeled feature set F0 is updated according to the pseudo-labeled feature sample f1′, and whether the retraining condition is met is determined according to the updated pseudo-labeled feature set F1′. If so, step S6 is executed to retrain the stroke prediction model M0 in combination with the labeled feature set and the updated pseudo-labeled feature set F1′; otherwise, return to step S3.

[0045] like Figure 5 As shown, according to the pseudo-label feature sample f t ′Update the latest pseudo-label feature set including:

[0046] Set the positive sample judgment threshold HT 伪正 And the negative sample judgment threshold HT 伪负 , calculated on the latest pseudo-labeled feature set F t-1 The pseudo-labeled feature samples f t ′ pseudo-label t The number of samples with the same feature N t-1 , if the number of feature samples N t-1 If the value is greater than the judgment threshold of the corresponding category, it means that the pseudo-label feature set of the corresponding category is full, and the pseudo-label feature sample f is calculated. t ′ and pseudo-labeled feature set F t-1 The pseudo-label is label t The training gain score of the pseudo-labeled feature sample is t ′ has a higher training gain score than the pseudo-labeled feature set F t-1 Any pseudo-label in t The pseudo-labeled feature samples, then the pseudo-labeled feature set F t-1 The pseudo-labeled feature sample with the smallest training gain score is replaced by the pseudo-labeled feature sample f t ′, and the pseudo-labeled feature set Ft-1 The pseudo-mark update counter C 伪-替换 Accumulate 1; otherwise, the pseudo-labeled feature sample f t ′Add pseudo-label feature set F t-1 , and the pseudo-labeled feature set F t-1 The pseudo-mark update counter C 伪-添加 Accumulate 1.

[0047] Positive sample judgment threshold HT 伪正 And the negative sample judgment threshold HT 伪负 Calculated according to the following formula:

[0048] HT 伪正 =ε p ×(N 正 +N 负 )-N 正

[0049] HT 伪负 =ε p ×(N 正 +N 负 )-N 负

[0050] Among them, N 正 、N 负 are the number of positive and negative samples in the labeled feature set, ε p is the coefficient of determination.

[0051] Preferably, the determination coefficient ε p Take 5.

[0052] Pseudo-labeled feature samples f t The training gain score of ′ is calculated according to the following formula:

[0053]

[0054] Among them, score t is the pseudo-labeled feature sample f t ′’s training gain score, c t is the pseudo-labeled feature sample f t 'Confidence level, label t is the pseudo-labeled feature sample f t ′’s pseudo-label, N is the set The total number of feature samples in the set is the labeled feature set D t The characteristic sample of is the pseudo-label feature set F t-1 The pseudo-label is label tfeature samples, H is the number of feature items of pseudo-labeled feature samples, c j Representing a collection The confidence of the jth feature sample in (the confidence of the labeled sample is set to 1), w i Representing a collection The importance weight of the i-th feature item of each feature sample in, f′ t,i Represents the pseudo-labeled feature sample f t The value of the i-th characteristic term of ′, Representing a collection The value of the i-th feature item of the j-th feature sample in .

[0055] Similarly, the pseudo-label feature sample f in the above formula t ′ is replaced by the pseudo-label feature set F t-1 The feature samples in the pseudo-label feature set F t-1 The training gain score of the feature samples in .

[0056] According to the updated pseudo-label feature set F t 'Determining whether the retraining conditions are met includes:

[0057] Set the percentage x, calculate the labeled feature set and the updated pseudo-labeled feature set F t ′’s total number of feature samples, and determines the updated pseudo-label feature set F t ′’s pseudo-tag update counter C 伪-添加 、C 伪-替换 、C 伪-剔除 Whether the sum of exceeds x% of the total number of feature samples, if so, the retraining condition is met and the updated pseudo-labeled feature set F t ′’s pseudo-tag update counter C 伪-添加 、C 伪-替换 、C 伪-剔除 Reset to 0; otherwise, the retraining condition is not met.

[0058] Further preferably, the triggering threshold is dynamically adjusted, and its value is positively correlated with the AUC value of the latest prediction model: x = b x +ε x ×AUC; where b x is the base value of percentage, ε x is the coefficient of the AUC value, and AUC represents the AUC value of the latest trained stroke prediction model.

[0059] Preferably, the base value b of the percentage x Take 1, the coefficient of AUC value ε x Take 10.

[0060] S6. Combine the labeled feature set and the updated pseudo-labeled feature set F t 'The latest trained stroke prediction model M k-1 Retraining is performed to obtain a retrained stroke prediction model, and the retrained stroke prediction model is used as the latest trained stroke prediction model;

[0061] like Figure 4 As shown, the latest trained stroke prediction model M k-1 Retraining includes:

[0062] Combine the labeled feature set and the updated pseudo-labeled feature set F t ′, get the overall feature set;

[0063] The overall feature set is divided into a training set and a test set. The latest trained stroke prediction model M is trained based on the training set. k-1 Retrain and obtain the retrained stroke prediction model M k ; The ratio of training set to test set is 7:3, and k is the number of retraining times of the stroke prediction model;

[0064] Evaluate the stroke prediction model M based on the test set k-1 、M k The quality of the stroke prediction model M k-1 The quality of the stroke prediction model M is better than that of the k , then the stroke prediction model M k Restored to stroke prediction model M k-1 , the stroke prediction model M k The parameters of the stroke prediction model M are restored k-1 Parameters.

[0065] Evaluate the stroke prediction model M based on the test set k-1 、M k The quality of the stroke prediction model M is evaluated using a 13-fold cross-validation method based on the test set. k-1 、M k The true positive rate (TPR) and false positive rate (FPR) of the stroke prediction model M k The true rate is greater than that of the stroke prediction model M k-1 And the false positive rate is lower than that of the stroke prediction model M k-1 When the stroke prediction model M k The performance quality is satisfied; otherwise, the stroke prediction model M k If the performance quality is not satisfied, the stroke prediction model M k Restored to stroke prediction model M k-1 .

[0066] S7, based on the latest trained stroke prediction model M k For the updated pseudo-labeled feature set F t ′ is updated to obtain the latest pseudo-label feature set F t , and return to step S3 for the next prediction.

[0067] like Figure 5 As shown, for the updated pseudo-label feature set F t 'Updates include:

[0068] The updated pseudo-labeled feature set F t ' Input the latest trained stroke prediction model M k , get the pseudo-label feature set F t The updated prediction results and confidence of each pseudo-label feature sample in ′; set the positive sample pseudo-label quality judgment threshold and the negative sample pseudo-label quality judgment threshold, and judge the pseudo-label feature set F one by one t Is the updated confidence of each pseudo-label feature sample in ′ less than the quality judgment threshold of the corresponding pseudo-label? If so, the corresponding pseudo-label feature sample is removed and the pseudo-label feature set F t ′’s pseudo-tag update counter C 伪-剔除 Accumulate 1 to get the latest pseudo-labeled feature set.

[0069] As the ratio of positive and negative samples in the data set changes, the threshold for determining the reliability of pseudo-labeling is dynamically adjusted, namely:

[0070] The threshold for judging the quality of pseudo-labeling of positive samples is: b q,正 +ε q ×(N 正 +N 伪正 ) / (N 正 +N 负 +N 伪正 +N 伪负 ).

[0071] The threshold for judging the quality of negative sample pseudo-labeling is: b q,负 +ε q ×(N 负 +N 伪负 ) / (N 正 +N 负 +N 伪正 +N 伪负 ).

[0072] In the above formula, b q,正 、b q,负 are the basic values ​​for quality judgment of positive samples and negative samples, respectively, q is the quality determination coefficient, N 正 、N 负are the number of positive and negative samples in the labeled feature set; N 伪正 、N 伪负 are the number of positive and negative samples in the latest pseudo-labeled feature set, respectively.

[0073] Preferably, the basic value b for quality judgment of positive samples and negative samples is q,正 、b q,负 Take 0.58 and 0.62 respectively, and the quality determination coefficient ε q Take 0.1.

[0074] On the other hand, Figure 1 As shown, the present invention adopts a self-optimizing edge intelligent early stroke prediction system to implement the above-mentioned prediction method, including: an edge prediction node subsystem and an edge learning node subsystem; the edge prediction node subsystem and the edge learning node subsystem are directly deployed in the edge nodes close to the monitored object: the edge prediction node is deployed in the family living room, and the edge learning node is deployed in the family living room or community service station.

[0075] The edge learning node subsystem includes:

[0076] The initial dataset construction module is used to collect labeled stroke data from recent years to construct the initial stroke dataset;

[0077] The initial data preprocessing module is used to preprocess the initial stroke data set to obtain a labeled feature set;

[0078] Prediction model building module, used to build a stroke prediction model;

[0079] The pseudo-label screening and feature set updating module stores the pseudo-label feature set and is used to update the stored pseudo-label feature set according to the pseudo-label feature samples reported by the edge prediction node subsystem;

[0080] Specifically, according to the binary classification predicted label value of the newly reported pseudo-label feature sample, the corresponding positive or negative sample pseudo-label feature set is selected as the update object. When it is not full, the newly reported pseudo-label feature sample is added to it. When it is full, the newly reported pseudo-label feature sample replaces the pseudo-label feature sample with the smallest training gain score.

[0081] The prediction model training module includes: a prediction model initial training module and a prediction model retraining module. The prediction model initial training module is used to perform initial training on the stroke prediction model based on the labeled feature set. The prediction model retraining module is used to retrain the stroke prediction model based on the labeled feature set and the pseudo-label feature set stored in the pseudo-label screening and feature set updating module.

[0082] The model evaluation and maintenance module is used to evaluate the quality of the trained stroke prediction model and the pseudo-label feature samples in the pseudo-label feature set stored in the pseudo-label screening and feature set update module. If the quality of the trained stroke prediction model does not meet the requirements, the stroke prediction model restoration is triggered; if the quality of the pseudo-label feature samples in the pseudo-label feature set does not meet the requirements, the pseudo-label feature samples are eliminated.

[0083] The edge prediction node subsystem includes:

[0084] The real-time data acquisition module collects data of the monitored objects in real time through data acquisition devices deployed in the living rooms of residents' homes;

[0085] Real-time data preprocessing module, used to preprocess the collected data to obtain real-time features;

[0086] The prediction model update module is used for edge deployment and real-time updating of the latest trained stroke prediction model uploaded by the edge learning node subsystem;

[0087] Preferably, the prediction model update module monitors the update notification issued by the edge learning node subsystem; after receiving the notification, the prediction model edge after the current round of self-optimization iterative training is deployed to the edge prediction node subsystem.

[0088] The prediction and pseudo-labeling module is used to predict the real-time features according to the stroke prediction model deployed by the prediction model update module, obtain the prediction results, combine the real-time features and their prediction results, obtain the pseudo-labeled feature samples, and report the pseudo-labeled feature samples to the edge learning node subsystem.

[0089] The output display module visualizes the two-class stroke prediction results and confidence levels.

[0090] Preferably, the information is displayed in the form of pictures, tables or text on mobile phones, tablets or dedicated display terminals.

[0091] The above embodiments further illustrate the objectives, technical solutions, and advantages of the present invention. It should be understood that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements to the present invention that fall within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A self-optimizing edge intelligence early stroke prediction method, characterized in that: include: S1. Collect labeled stroke data to construct an initial stroke dataset, preprocess the initial stroke dataset, and obtain a labeled feature set; S2. Build a stroke prediction model, train the stroke prediction model based on the labeled feature set, and obtain the latest trained stroke prediction model; initialize the pseudo-labeled feature set to obtain the latest pseudo-labeled feature set; S3, collect the current data of the monitored object in real time, pre-process the collected data, and obtain real-time feature f t ; S4, the real-time feature f t Input the latest trained stroke prediction model to get the prediction result; combined with the real-time feature f t And its prediction results, get the pseudo-label feature sample f t '; S5, based on the pseudo-labeled feature sample f t 'Update the latest pseudo-label feature set, and determine whether the retraining condition is met based on the updated pseudo-label feature set. If so, execute step S6; otherwise, use the updated pseudo-label feature set as the latest pseudo-label feature set and return to step S3; S6. Retraining the most recently trained stroke prediction model by combining the labeled feature set and the updated pseudo-labeled feature set to obtain a retrained stroke prediction model; and using the retrained stroke prediction model as the most recently trained stroke prediction model; S7. Update the updated pseudo-label feature set according to the latest trained stroke prediction model to obtain the latest pseudo-label feature set, and return to step S3.

2. The self-optimizing edge intelligence early stroke prediction method according to claim 1, characterized in that: The pseudo-mark feature set includes: pseudo-mark update counter C 伪-添加 、C 伪-替换 、C 伪-剔除 ; According to the pseudo-label feature sample f t ′Update the latest pseudo-label feature set F t-1 include: Set the positive sample judgment threshold HT 伪正 And the negative sample judgment threshold HT 伪负 , calculated on the latest pseudo-labeled feature set F t-1 The pseudo-labeled feature samples f t ′ pseudo-label t The number of samples with the same feature N t-1 , if the number of feature samples N t-1 If it is greater than the judgment threshold of the corresponding category, the pseudo-label feature sample f is calculated t ′ and pseudo-labeled feature set F t-1 The pseudo-label is label t The training gain score of the pseudo-labeled feature sample is t ′ has a higher training gain score than the pseudo-labeled feature set F t-1 Any pseudo-label in t The pseudo-labeled feature samples, then the pseudo-labeled feature set F t-1 The pseudo-labeled feature sample with the smallest training gain score is replaced by the pseudo-labeled feature sample f t ′, and the pseudo-labeled feature set F t-1 The pseudo-mark update counter C 伪-替换 Accumulate 1; otherwise, the pseudo-labeled feature sample f t ′Add pseudo-label feature set F t-1 , and the pseudo-labeled feature set F t-1 The pseudo-mark update counter C 伪-添加 Accumulate 1.

3. The self-optimizing edge intelligence early stroke prediction method according to claim 2, characterized in that: Positive sample judgment threshold HT 伪正 =ε p ×(N 正 +N 负 )-N 正 , negative sample judgment threshold HT 伪负 =ε p ×(N 正 ++N 负 )-N 负 ; Among them, N 正 、N 负 are the number of positive and negative samples in the labeled feature set, ε p is the coefficient of determination.

4. The self-optimizing edge intelligence early stroke prediction method according to claim 2, characterized in that: Pseudo-labeled feature samples f t ′ training gain score Among them, score t is the pseudo-labeled feature sample f t ′’s training gain score, c t is the pseudo-labeled feature sample f t 'Confidence level, label t is the pseudo-labeled feature sample f t ′’s pseudo-label, N is the set The total number of feature samples in the set Label the labeled features in the set t The characteristic sample of is the pseudo-label feature set F t-1 The pseudo-label is label t feature samples, H is the number of feature items of pseudo-labeled feature samples, c j Representing a collection The confidence of the j-th feature sample in w i Representing a collection The importance weight of the i-th feature item in the feature sample, f′ t,i Represents the pseudo-labeled feature sample f t The value of the i-th characteristic term of ′, Representing a collection The value of the i-th feature item of the j-th feature sample in .

5. The self-optimizing edge intelligence early stroke prediction method according to claim 1, characterized in that: The pseudo-mark feature set includes: pseudo-mark update counter C 伪-添加 、C 伪-替换 、C 伪-剔除 ; Determining whether the retraining conditions are met based on the updated pseudo-labeled feature set includes: Set the percentage x, calculate the total number of feature samples of the marked feature set and the updated pseudo-marked feature set, and determine the pseudo-mark update counter C of the updated pseudo-marked feature set. 伪-添加 、C 伪-替换 、C 伪-剔除 Whether the sum exceeds x% of the total number of feature samples, if so, the retraining condition is met, and the pseudo-label update counter C of the updated pseudo-label feature set is used 伪-添加 、C 伪-替换 、C 伪-剔除 Reset to 0; otherwise, the retraining condition is not met.

6. The self-optimizing edge intelligence early stroke prediction method according to claim 5, characterized in that: Percentage x = b x +ε x ×AUC; where b x is the base value of percentage, ε x is the coefficient of the AUC value, and AUC represents the AUC value of the latest trained stroke prediction model.

7. The self-optimizing edge intelligence early stroke prediction method according to claim 1, characterized in that: The latest trained stroke prediction model M k-1 Retraining includes: Combine the labeled feature set and the updated pseudo-labeled feature set to obtain the overall feature set; The overall feature set is divided into a training set and a test set. The latest trained stroke prediction model M is trained based on the training set. k-1 Retrain and obtain the retrained stroke prediction model M k ; Where k is the number of retraining times of the stroke prediction model; Evaluate the stroke prediction model M based on the test set k-1 、M k The quality of the stroke prediction model M k-1 The quality of the stroke prediction model M is better than that of the k , then the stroke prediction model M k Restored to stroke prediction model M k-1 .

8. The self-optimizing edge intelligence early stroke prediction method according to claim 1, characterized in that: The pseudo-mark feature set includes the pseudo-mark update counter C 伪-添加 、C 伪-替换 、C 伪-剔除 ; For the updated pseudo-labeled feature set F t 'Updates include: The updated pseudo-labeled feature set F t ' Input the latest trained stroke prediction model M k , get the pseudo-label feature set F t The updated prediction results and confidence of each pseudo-label feature sample in ′; set the positive sample pseudo-label quality judgment threshold and the negative sample pseudo-label quality judgment threshold, and judge the pseudo-label feature set F one by one t Is the updated confidence of each pseudo-label feature sample in ′ less than the quality judgment threshold of the corresponding pseudo-label? If so, the corresponding pseudo-label feature sample is removed and the pseudo-label feature set F t ′’s pseudo-tag update counter C 伪-剔除 Accumulate 1 to get the latest pseudo-label feature set F t .

9. The self-optimizing edge intelligence early stroke prediction method according to claim 8, characterized in that: The calculation formula for the positive sample pseudo-label quality judgment threshold is: b q,正 +ε q ×(N 正 +N 伪正 ) / (N 正 +N 负 +N 伪正 +N 伪负 ), the calculation formula for the negative sample pseudo-label quality judgment threshold is: b q,负 +ε q ×(N 负 +N 伪负 ) / (N 正 +N 负 +N 伪正 +N 伪负 ); where b q,正 、b q,负 are the basic values ​​for quality judgment of positive samples and negative samples, respectively, q is the quality determination coefficient, N 正 、N 负 are the number of positive and negative samples in the labeled feature set, N 伪正 、N 伪负 They are the updated pseudo-label feature sets F t The number of positive and negative samples in ′.

10. A self-optimizing edge intelligent early stroke prediction system, used to implement the prediction method according to any one of claims 1 to 9, characterized in that: include: Edge prediction node subsystem and edge learning node subsystem; The edge prediction node subsystem and edge learning node subsystem are directly deployed in the edge nodes close to the monitored object; The edge learning node subsystem includes: The initial dataset construction module is used to collect labeled stroke data to construct the initial stroke dataset; The initial data preprocessing module is used to preprocess the initial stroke data set to obtain a labeled feature set; Prediction model building module, used to build a stroke prediction model; The pseudo-label screening and feature set updating module stores the pseudo-label feature set and is used to update the stored pseudo-label feature set according to the pseudo-label feature samples reported by the edge prediction node subsystem; The prediction model training module includes: a prediction model initial training module and a prediction model retraining module. The prediction model initial training module is used to perform initial training on the stroke prediction model based on the labeled feature set. The prediction model retraining module is used to retrain the stroke prediction model based on the labeled feature set and the pseudo-label feature set stored in the pseudo-label screening and feature set updating module. The model evaluation and maintenance module is used to evaluate the quality of the trained stroke prediction model and the pseudo-label feature samples in the pseudo-label feature set stored in the pseudo-label screening and feature set update module. If the quality of the trained stroke prediction model does not meet the requirements, the stroke prediction model restoration is triggered; if the quality of the pseudo-label feature samples in the pseudo-label feature set does not meet the requirements, the pseudo-label feature samples are eliminated. The edge prediction node subsystem includes: Real-time data acquisition module, used to collect data of monitored objects in real time; Real-time data preprocessing module, used to preprocess the collected data to obtain real-time features; The prediction model update module is used for edge deployment and real-time updating of the latest trained stroke prediction model uploaded by the edge learning node subsystem; The prediction and pseudo-labeling module is used to predict the real-time features according to the stroke prediction model deployed by the prediction model update module, obtain the prediction results, combine the real-time features and their prediction results, obtain the pseudo-labeled feature samples, and report the pseudo-labeled feature samples to the edge learning node subsystem.

Images