Cerebral stroke data management analysis method and system based on large model technology

Abstract

The invention discloses a cerebral apoplexy data management analysis method and system based on a large model technology, and particularly relates to the technical field of cerebral apoplexy data management. The method comprises the following steps: generating test images under different noise and interference conditions by collecting noise-free and interference-free real image data, and constructing a plurality of groups of test data sets in a grading manner; inputting the data sets into an analysis model to obtain focus identification and risk assessment results, and assessing the prediction reliability of the model by analyzing error accumulation and accuracy fluctuation conditions of the model under different noise and interference conditions; furthermore, the prediction results are classified according to the accuracy, and particularly for the prediction results with incomplete accuracy, the storage and isolation requirements of the prediction results are judged through fine analysis in a fixed time period, so that the robustness of the model under different input conditions is remarkably improved, and the prediction accuracy of the model is improved. Errors caused by noise or interference are prevented from being spread to a data management system, and accurate analysis and robust management of cerebral apoplexy data are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of stroke data management, and in particular to a stroke data management and analysis method and system based on large model technology. Background Art

[0002] Stroke data management and analysis based on big model technology mainly refers to the use of artificial intelligence models such as deep learning to systematically manage and analyze large amounts of complex stroke-related data, thereby supporting clinical diagnosis, treatment plan optimization, and rehabilitation prediction. These data come from a wide range of sources, including patients' clinical images, genetic information, medical history records, and treatment responses. Through big model technology, this information can be efficiently processed, classified, and interpreted, helping doctors discover potential etiological associations and individualized treatment strategies, thereby improving diagnosis and treatment outcomes and patients' recovery rates.

[0003] In specific applications, large model technology can automatically identify key features in stroke data, such as lesion areas and blood flow characteristics in images, to predict the progression of the disease and provide risk assessments. In addition, this type of technology can also identify differences in the efficacy of specific treatments for different patients through training and learning of massive data, thereby assisting doctors in developing treatment plans that are more suitable for individual patients. Overall, stroke data management and analysis based on large model technology can not only improve the processing efficiency of medical data, but also provide a scientific basis for personalized medicine and promote the development of precision medicine in the field of stroke.

[0004] The prior art has the following deficiencies:

[0005] Large models may be extremely sensitive to changes in data input and may produce extreme prediction results when encountering slight input anomalies or noise interference. This phenomenon is particularly dangerous in stroke image recognition, because artifacts accidentally generated by imaging equipment, imaging noise, or image blur caused by patient non-cooperation may cause the model to output incorrect risk assessments. In addition, the prediction results of large models are often input into other data management systems for comprehensive analysis and long-term tracking. If the model makes incorrect predictions due to input interference, these errors may propagate to the entire data management system. For example, incorrect lesion identification results will affect the patient's medical records and make subsequent data analysis inaccurate. At the same time, the accumulation of erroneous data will affect the decision-making model for risk prediction or treatment optimization for other patients. Summary of the invention

[0006] The purpose of the present invention is to provide a stroke data management and analysis method and system based on large model technology to overcome the shortcomings of the background technology.

[0007] In order to achieve the above object, the present invention provides the following technical solution: a method for managing and analyzing stroke data based on large model technology, comprising the following steps:

[0008] S1: Collect noise-free and interference-free image data of real stroke patients, and generate test images with different noise levels and interference types based on the collected image data;

[0009] S2: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple test data sets, and classify the data sets according to the degree of noise and interference;

[0010] S3: Input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them;

[0011] S4: According to the abnormal accumulation degree of model prediction error and the fluctuation of lesion identification accuracy under different noise and interference conditions, the accuracy of prediction results of the target stroke data analysis model under different conditions was evaluated;

[0012] S5: Analyze the accuracy of the prediction results of the model under different conditions according to the target stroke data, divide the model prediction results into accurate prediction results, incompletely accurate prediction results and inaccurate prediction results, and perform corresponding processing;

[0013] S6: For incompletely accurate prediction results, further analyze the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and determine whether to isolate and store the prediction results of the target stroke data analysis model based on the analysis results.

[0014] Preferably, in S2, test data with different noise levels and different interference levels are combined separately to generate a data set, and different levels of noise are combined with different degrees of interference in pairs, including: mild noise + mild interference, moderate noise + moderate interference, and heavy noise + heavy interference. The combined data set is divided into different levels according to the degree of influence of noise and interference.

[0015] Preferably, in S4, after analyzing the abnormal accumulation degree of model prediction error under different noise and interference conditions, a prediction error abnormal accumulation index is generated, and the method for obtaining the prediction error abnormal accumulation index is:

[0016] Under noise-free and interference-free conditions, the mean of each error index is calculated as the reference value to obtain the reference error mean vector μ; the reference vector is expressed as μ = [μ 位置误差 ,μ 风险评分误差 ] T, T is the matrix transpose; calculate the covariance matrix Σ of the error indicators under various noise and interference conditions. The diagonal elements of the covariance matrix represent the variance of each error indicator, and the off-diagonal elements represent the covariance between different error indicators. The covariance matrix is:

[0017]

[0018] For each noise and interference condition, calculate the error vector E i The difference E from the reference vector μ i -μ, substitute into the Mahalanobis distance formula to calculate the forecast error anomaly accumulation index, the expression is: Where SD is the abnormal cumulative index of forecast error, Σ -1 is the inverse matrix of the covariance matrix Σ of the error vector.

[0019] Preferably, in S4, after analyzing the fluctuation of the lesion recognition accuracy under different noise and interference conditions, a lesion recognition accuracy fluctuation index is generated, and the method for obtaining the lesion recognition accuracy fluctuation index is:

[0020] Assume that under different noise and interference conditions, the fluctuation value of the lesion recognition accuracy is a time series {W s}, s is time, arrange the conditional sequence so that the noise and interference intensity gradually increase, select the smoothing factor α to control the degree of smoothing, and the formula of the exponential smoothing method is: S s =α·W s +(1-α)·S s-1 ; Among them, S s is the smooth fluctuation value under the conditions at the sth moment, W s is the actual lesion recognition accuracy value under the condition of the sth moment, S s-1 is the smoothed fluctuation value under the previous condition, and the smoothed value S1 under the initial condition is set to the actual accuracy value under the first condition, that is, S1=W1; starting from the second condition, the smoothed fluctuation value of each condition is gradually calculated according to the exponential smoothing calculation formula until the fluctuation values ​​of all conditions are smoothed, and the smoothed value of the last condition is used as the fluctuation index of lesion recognition accuracy.

[0021] Preferably, in S4, the prediction error abnormal accumulation index and the lesion identification accuracy fluctuation index are converted into a comprehensive feature vector, and the comprehensive feature vector is used as the input of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the accuracy value label of the prediction results of the target stroke data analysis model under different conditions as the prediction target, and takes minimizing the sum of the prediction errors of the accuracy value labels of the prediction results of all target stroke data analysis models under different conditions as the training target. The machine learning model is trained until the sum of the prediction errors converges, and the model training is stopped. The accuracy value of the prediction results of the target stroke data analysis model under different conditions is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0022] Preferably, in S5, according to the accuracy of the prediction results of the target stroke data analysis model under different conditions, the model prediction results are divided into accurate prediction results, incompletely accurate prediction results and inaccurate prediction results, and corresponding processing is performed, specifically:

[0023] Comparing the accuracy values ​​of the prediction results of the acquired target stroke data analysis model under different conditions with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy values ​​of the prediction results of the target stroke data analysis model under different conditions with the first standard threshold and the second standard threshold respectively;

[0024] If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is greater than the second standard threshold, it means that the prediction results of the target stroke data analysis model under different conditions are highly accurate, and a high-accuracy prediction signal is generated at this time, and the model prediction results are divided into accuracy prediction results, which are directly used for clinical decision support;

[0025] If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is medium, and a medium accuracy prediction signal is generated at this time, and the model prediction result is classified as a medium accuracy prediction result, which needs to be further confirmed;

[0026] If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is less than the first standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is low. At this time, a low-accuracy prediction signal is generated, and the model prediction result is classified as a low-accuracy prediction result, which is not directly used for clinical decision-making.

[0027] Preferably, in S6, for the incompletely accurate prediction results, the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period is further analyzed, and it is determined whether to isolate and store the prediction results of the target stroke data analysis model according to the analysis results, specifically:

[0028] Assume that within a fixed time period T, the accuracy value sequence of the medium accuracy prediction results is {A t}, where t = 1, 2, ..., n represents the prediction accuracy value of different conditions in the time period, and the prediction accuracy change rate R in the time period is calculated AC , indicating whether the accuracy of the medium-accuracy forecast remains stable, by quantifying its volatility using the standard deviation: in, is the average accuracy value within the time period; set the fluctuation threshold θ to determine the accuracy change rate R AC Whether it exceeds the normal range; if the prediction accuracy change rate is greater than the fluctuation threshold, it is considered that the accuracy fluctuates greatly and the model is not stable enough under medium conditions; otherwise, it is considered that the accuracy remains stable and needs to be isolated and stored. The judgment formula is expressed as: Where I = 1 means that isolated storage is required, and I = 0 means that isolated storage is not required; the isolated storage formula is expressed as: S 隔离 ={A t |I=1,t∈T}; where S 隔离 Represents a collection of medium-accuracy prediction results that require isolated storage.

[0029] The present invention also provides a stroke data management and analysis system based on the large model technology, including a data acquisition module, a data classification module, a model testing module, a model prediction accuracy evaluation module, an accuracy classification module and a management module;

[0030] Data acquisition module: collects noise-free and interference-free image data of real stroke patients, and generates test images with different noise levels and interference types based on the collected image data;

[0031] Data classification module: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple test data sets, and classify the data sets according to the degree of noise and interference;

[0032] Model testing module: input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them;

[0033] Model prediction accuracy evaluation module: Based on the abnormal accumulation of model prediction errors and fluctuations in lesion identification accuracy under different noise and interference conditions, the module evaluates the accuracy of prediction results of the target stroke data analysis model under different conditions;

[0034] Accuracy classification module: According to the target stroke data, the accuracy of the model's prediction results under different conditions is analyzed, and the model's prediction results are divided into accurate prediction results, incompletely accurate prediction results, and inaccurate prediction results, and corresponding processing is performed;

[0035] Management module: For incompletely accurate prediction results, further analysis is performed on the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and based on the analysis results, it is determined whether the prediction results of the target stroke data analysis model are to be stored in isolation.

[0036] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0037] 1. This invention introduces a multi-level noise and interference test data set, and by gradually evaluating the error accumulation index and accuracy fluctuation index of the model under these conditions, it can accurately identify the abnormal performance of the model under complex input conditions. Combined with the training and accuracy analysis of the machine learning model, the prediction results are classified into three accuracy levels: high, medium, and low, and the medium accuracy results are further analyzed for volatility, which realizes the refinement of the prediction results, thereby effectively avoiding the risk of error propagation to other data management systems.

[0038] 2. The present invention can hierarchically store and isolate the model output results according to the prediction accuracy level under different conditions, ensuring that low-accuracy and unstable prediction results will not be directly used for clinical decision support, thereby improving the reliability of data management. At the same time, the isolated storage mechanism of this method can also help the model identify unstable prediction results in long-term applications, and improve the robustness and overall accuracy of the model through feedback optimization iteration, so as to better support the comprehensive analysis and risk assessment of stroke data, and improve the scientificity and safety of patient management. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0040] Figure 1 The figure is a flow chart of the method of the present invention.

[0041] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 creative work are within the scope of protection of the present invention.

[0043] Example 1, please refer to Figure 1 and Figure 2 As shown, the stroke data management and analysis method based on the large model technology described in this embodiment includes the following steps:

[0044] S1: Collect noise-free and interference-free image data of real stroke patients, and generate test images with different noise levels and interference types based on the collected image data;

[0045] S2: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple test data sets, and classify the data sets according to the degree of noise and interference;

[0046] S3: Input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them;

[0047] S4: According to the abnormal accumulation degree of model prediction error and the fluctuation of lesion identification accuracy under different noise and interference conditions, the accuracy of prediction results of the target stroke data analysis model under different conditions was evaluated;

[0048] S5: Analyze the accuracy of the prediction results of the model under different conditions according to the target stroke data, divide the model prediction results into accurate prediction results, incompletely accurate prediction results and inaccurate prediction results, and perform corresponding processing;

[0049] S6: For incompletely accurate prediction results, further analyze the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and determine whether to isolate and store the prediction results of the target stroke data analysis model based on the analysis results.

[0050] In S1, noise-free and interference-free image data of real stroke patients are collected, and test images with different noise levels and interference types are generated based on the collected image data, specifically:

[0051] Use high-resolution MRI (magnetic resonance imaging) or CT (computed tomography) equipment to obtain image data, ensuring that the resolution and clarity of the image meet clinical diagnostic standards. Ensure that the environment is stable during image acquisition, including calibration of the imaging equipment, avoiding external electromagnetic interference, and ensuring that the patient's posture is fixed to reduce motion blur. These controls help obtain clear, noise-free images and provide standardized raw data for subsequent addition of noise and interference.

[0052] Based on the noise-free image, different types of noise are simulated through algorithms. Common noise types include: Gaussian noise: adding noise that simulates electronic signal fluctuations to the image, often used to simulate slight interference caused by signal fluctuations in imaging equipment. Salt and pepper noise: adding random black and white pixels to simulate image spots caused by equipment failure or external signal interference. This noise is especially common in low-quality imaging equipment. Poisson noise: randomly distributed noise perturbations of image pixel values ​​simulate noise caused by low light conditions or image magnification. Noise level control: set multiple noise intensity levels for each noise, such as low, medium, and high. Generate multiple sets of noise images with different degrees of noise by controlling the noise amplitude and frequency to test the performance of the model under different noise intensities.

[0053] On the basis of non-distracted images, different image processing techniques are used to generate a variety of common interferences to simulate the complex situations that may be encountered in clinical imaging. Common interference types include: Image blur: Introduce slight Gaussian blur or motion blur to simulate the image blur caused by the patient's movement during imaging. Control the degree of blur to cover mild blur (slightly blurred edges) to severe blur (severe loss of details). Simulate uneven exposure of the imaging device or changes in lighting conditions by adjusting the overall brightness of the image. Generate images with different brightness deviations, from underexposure (low brightness) to overexposure (high brightness) to test the robustness of the model under different brightness conditions. Change the image contrast to simulate abnormal contrast settings of the imaging device. Generate images of different levels with reduced contrast (darker images, unclear details) and enhanced contrast (clear image layers but loss of details). Add simulated streaks, halos and other artifacts to the image to simulate artifact interference caused by hardware problems of the imaging device or metal implants in the patient's body (such as dental fillings, prostheses, etc.). These artifacts usually appear in specific areas and affect the accurate identification of lesions.

[0054] Single noise / interference test dataset: Save images generated with different noise levels and interference types as separate test datasets to evaluate the sensitivity of the large model to a single type of noise or interference. Combined noise and interference test dataset: Add noise and interference to some noise-free and interference-free images to generate a mixed test dataset. For example, combine moderate Gaussian noise with mild blur, heavy salt and pepper noise with brightness deviation, etc. to simulate more complex real-world situations.

[0055] Provide standard annotations of the lesion area for each generated test image (annotated by professional physicians or generated by automatic annotation systems) as a reference standard for model evaluation. Perform quality verification on the generated test dataset to ensure that the intensity of noise and interference meets the preset standards and that the annotations of the lesion area are accurate, so as to ensure that the dataset can truly simulate the interference conditions in clinical imaging.

[0056] According to the noise intensity and interference type, the data is stored and clearly labeled in a hierarchical manner so that data with different noise levels and interference combinations can be flexibly called for model evaluation later. The data set is classified and managed by noise type, noise intensity, interference type, and interference intensity, and the original noise-free and interference-free high-quality images are attached for easy comparison with the original images.

[0057] S2: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple sets of test data sets, and grade the data sets according to the degree of noise and interference.

[0058] Based on the original noise-free, interference-free image, apply different levels of noise types to generate multiple test images. For example: Mild noise combination: Add low-level Gaussian noise, salt and pepper noise, and Poisson noise to the original image to generate a mild noise combination. Moderate noise combination: Add medium-level Gaussian noise, salt and pepper noise, and Poisson noise to the original image to generate a moderate noise combination. Heavy noise combination: Add a higher level of noise to the original image to generate a heavy noise combination. Interference combination generation: Generate different degrees of interference type test images based on the original image, such as: Mild interference combination: Add mild blur or slight brightness deviation to the original image, etc. Moderate interference combination: Add more obvious brightness deviation, blur or contrast adjustment, etc.; Heavy interference combination: Add severe image blur, overexposure brightness deviation, etc. to simulate extreme imaging conditions.

[0059] Datasets were generated by combining test data with different noise levels (mild, moderate, and severe) and different levels of interference. These single-factor datasets facilitate comparison of the effects of noise and interference. Cross-combined datasets of noise and interference were generated, combining different levels of noise with different degrees of interference, including: Mild noise + mild interference: mild Gaussian noise and mild blur are added to the original image. Moderate noise + moderate interference: moderate salt and pepper noise and obvious brightness deviation are added to the original image. Heavy noise + heavy interference: high level Poisson noise and severe blur are added to the original image.

[0060] According to the degree of influence of noise and interference, the combined data set is divided into different levels. The common classifications are: Level 1 (mild): The image quality changes slightly, but the key features (such as the lesion area) are still clear and easy for the model to recognize. Level 2 (moderate): The image quality is significantly reduced, but the lesion area is still recognizable. This level simulates the more common imaging quality deviations. Level 3 (severe): The image is severely blurred or highly noisy, and the lesion area is difficult to identify. It is usually the most challenging condition.

[0061] The datasets are named and labeled according to the noise and interference intensity, clearly identifying the noise type, intensity, and interference conditions of the image. For example, "moderate Gaussian noise + mild blur" or "heavy salt and pepper noise + moderate brightness deviation", etc., so that the source and characteristics of the data can be quickly identified in subsequent analysis.

[0062] Store each level of data set separately to ensure that the classification of noise and interference combinations is clear and independent of each other, so that they can be flexibly retrieved during testing. Store according to the combined level of noise and interference to form different folders or database structures. Add metadata tags (such as noise type, interference intensity, classification level) to each test image and establish a database index to facilitate the rapid call of specific combination image sets when evaluating the robustness of the model later.

[0063] The image quality under different noise and interference combinations was checked by visual and statistical methods to ensure that the identifiability of the lesion area met the expected grading standards. Manual inspection was performed when necessary to ensure that the quality differences between different levels were significant. After data generation, a small-scale model test dataset was used to confirm the impact of different noise and interference combinations on model predictions and verify the rationality of the combined dataset.

[0064] S3: Input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them.

[0065] Load the graded test data set into the model evaluation environment to ensure that the data set covers different noise levels and interference types to ensure the comprehensiveness of the evaluation. Preprocess the image data according to the model input requirements to ensure that the image size, resolution, color channels and other formats are consistent with the model input. For example, unify the image size, adjust the pixel value range (such as normalizing to [0,1] or [-1,1] interval), etc. Set key parameters for model operation, such as batch size, processing speed, etc., to ensure that the model can efficiently process each set of test data during the evaluation process.

[0066] The test data set is input into the model in batches to avoid problems such as memory overflow caused by inputting too much data at one time. Each batch of data is input in order according to the grade (mild, moderate, severe) and type (noise, interference) to ensure that data with different interference conditions can be evaluated independently. During the operation of the model, the prediction results of each image are automatically obtained. For the stroke analysis model, the output results generally include the following two parts: Lesion recognition results: The model automatically detects and marks the lesion area, and outputs the location information (such as pixel coordinates), size and shape information of the lesion. The lesion area is usually marked in the form of a mask or a bounding box. Based on the model's recognition results of lesion features and clinical experience data, a risk score or risk level (such as low, medium, high) is output to predict the patient's condition progression or the possibility of further deterioration.

[0067] For each image, record the key parameters such as the location, size and shape of the lesion output by the model, and save the information in a visual way (for example, annotate it on the image) for subsequent verification. The records include: Coordinate information: the center coordinates of the lesion area, the coordinates of the four corners of the bounding box, or the pixel area of ​​the mask. Region size: the area or number of pixels of the lesion area, which is convenient for comparison under different noise and interference conditions. Region shape: such as the aspect ratio of the lesion, shape characteristics (circular, elliptical, irregular shape), etc. Record the model's risk score for each image, or the classification result (low risk, medium risk, high risk), and store it in the database for subsequent analysis. The recorded details include: Risk score: the specific score output by the model, which is used to quantify the risk level. The risk level given by the model according to the scoring criteria, such as by setting a threshold to divide the score into different risk levels. If the model supports output uncertainty estimates or credibility scores, the uncertainty should also be recorded to evaluate the predictive stability of the model under noise and interference conditions.

[0068] Compare the prediction results of each image with noise or interference with the prediction results of the original noise-free image to detect the errors introduced by noise or interference. The specific comparison contents include: Lesion area deviation: Compare the offset and area changes of the lesion area in the noisy image and the original image, and record the differences in position and size between the two. Compare the changes in risk scores under different noise and interference conditions, and record the changes in score deviation and risk level to evaluate the sensitivity of the model to input interference.

[0069] Generate a structured record report of the lesion identification and risk assessment results for each image, including the location information, shape, size, risk score, risk level, etc. Save the visualization results of each test image, including the lesion area marking map and risk level map, and attach the noise / interference level information for subsequent analysis.

[0070] S4: Evaluate the accuracy of the prediction results of the target stroke data analysis model under different conditions based on the abnormal accumulation degree of model prediction error and the fluctuation of lesion identification accuracy under different noise and interference conditions.

[0071] Calculate the accumulation of model prediction errors under different noise and interference conditions to quantify the cumulative impact of various conditions on the model output results. Prediction errors can include lesion area location errors and risk score errors. Set an abnormal accumulation threshold for prediction errors (such as more than 2 or 3 times the error under the original noise-free condition) to determine whether the model has produced abnormal cumulative errors under specific noise and interference conditions. For each noise or interference condition, cumulatively record the prediction errors, such as calculating the sum of the average errors of all test images under that condition. For lesion location errors, the average distance deviation between the predicted lesion center and the true lesion center in all images can be calculated; for risk score errors, the average risk score deviation of all test images is calculated. When the degree of error accumulation significantly exceeds the abnormal threshold, it is marked as abnormal accumulation. This marking can help identify model weaknesses under specific conditions and record the specific conditions for abnormal error accumulation for subsequent optimization.

[0072] After analyzing the abnormal accumulation degree of model prediction error under different noise and interference conditions, the prediction error abnormal accumulation index is generated. The method for obtaining the prediction error abnormal accumulation index is as follows:

[0073] Under noise-free and interference-free conditions, the mean of each error index is calculated as the benchmark value to obtain the benchmark error mean vector μ; for example, if the lesion location error (such as IoU difference) and risk score error are considered, the benchmark vector can be expressed as μ = [μ 位置误差 ,μ 风险评分误差 ] T, T is the matrix transpose; calculate the covariance matrix Σ of the error indicators under various noise and interference conditions. The diagonal elements of the covariance matrix represent the variance of each error indicator, and the off-diagonal elements represent the covariance between different error indicators. The covariance matrix is:

[0074]

[0075] For each noise and interference condition, calculate the error vector E i The difference E from the reference vector μ i -μ, substitute into the Mahalanobis distance formula to calculate the forecast error anomaly accumulation index, the expression is: Where SD is the abnormal cumulative index of forecast error, Σ -1 is the inverse matrix of the covariance matrix Σ of the error vector, taking into account the correlation between error indicators.

[0076] The larger the prediction error accumulation index, the more the prediction error of the target stroke data analysis model deviates from the baseline value under specific noise or interference conditions. This usually indicates that the performance of the model under this condition is unstable, the degree of error accumulation is significant, and it may lead to reduced accuracy of lesion identification and risk assessment. Therefore, a larger abnormal accumulation index reflects that the model has a large deviation when dealing with noise and interference of a specific type or intensity, and cannot reliably provide accurate support for clinical decision-making. At this point, the model may need to be further optimized to enhance its robustness under this specific condition.

[0077] On the contrary, if the prediction error abnormal accumulation index is small, it means that the prediction error of the model under the noise or interference condition is close to the baseline condition, indicating that the model's performance under this condition is relatively stable. A smaller abnormal accumulation index means that noise or interference has less impact on the model, and the model can maintain a high prediction accuracy under a variety of input conditions. In this case, the model has good generalization ability and robustness, and can be used more reliably for clinical analysis. It can still provide accurate lesion identification and risk assessment results when facing different noise and interference scenarios.

[0078] The fluctuation of model accuracy is measured by the intersection over union (IoU) of lesion recognition. The mean and variance of IoU under each noise or interference condition are calculated to evaluate the recognition stability of the model under different conditions. The mean and variance of IoU are calculated under each condition. If the variance is large, it means that the accuracy under this condition is highly volatile and the model recognition is unstable; if the variance is small, it means that the model has strong robustness and consistency under this condition. The IoU variance under noise-free conditions is set as the benchmark. If the fluctuation variance under a specific condition exceeds the benchmark by several times, it is considered that this condition has a significant impact on the accuracy of the model prediction results.

[0079] After analyzing the fluctuation of lesion recognition accuracy under different noise and interference conditions, the lesion recognition accuracy fluctuation index is generated. The method for obtaining the lesion recognition accuracy fluctuation index is as follows:

[0080] Assume that under different noise and interference conditions, the fluctuation value of the lesion recognition accuracy is a time series {W s}, s is time (i.e., accuracy under each condition). Arrange the condition sequence so that the noise and interference intensity gradually increase (such as mild noise, mild blur, moderate noise, etc.) so that the exponential smoothing method can be applied. Select the smoothing factor α (usually 0<α<1) to control the degree of smoothing. A larger α value pays more attention to the latest fluctuation data, while a smaller α value pays more attention to historical fluctuations, and the smoothing effect is stronger. The formula for exponential smoothing is: S s =α·W s +(1-α)·S s-1 ; Among them, S s is the smooth fluctuation value under the conditions at the sth moment, W s is the actual lesion recognition accuracy value under the condition of the sth moment, S s-1 is the smoothed fluctuation value under the previous condition. The smoothed value S1 under the initial condition can be set to the actual accuracy value under the first condition, that is, S1=W1. Starting from the second condition, the smoothed fluctuation value of each condition is calculated step by step according to the above formula until the fluctuation values ​​of all conditions are smoothed. The smoothed value of the last condition is used as the lesion recognition accuracy fluctuation index and marked as HK, which represents the overall fluctuation trend of the model under all conditions.

[0081] The larger the fluctuation index of lesion recognition accuracy, the greater the variation in the accuracy of the model under different noise and interference conditions, showing higher volatility. This usually means that the model has poor adaptability to different conditions, lacks robustness, and cannot stably identify the lesion area. As the fluctuation index increases, the accuracy of the model under specific conditions may drop significantly, affecting its reliable analysis and assessment of the patient's condition. High volatility may lead to insufficient consistency of the model in clinical applications, and further optimization is needed to cope with various input interference conditions.

[0082] On the contrary, the smaller the fluctuation index of lesion recognition accuracy, the more stable the recognition accuracy of the model under different conditions, and the smaller the fluctuation. This shows that the model has strong robustness and can maintain high recognition accuracy under changing noise or interference conditions. A low fluctuation index means that the model has good adaptability to different conditions, and can provide consistent and reliable prediction results in practical applications, which is suitable for a variety of clinical scenarios.

[0083] The prediction error abnormal accumulation index and the lesion identification accuracy fluctuation index are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as the input of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the accuracy value label of the prediction results of the target stroke data analysis model under different conditions as the prediction target, and takes minimizing the sum of the prediction errors of the accuracy value labels of the prediction results of all target stroke data analysis models under different conditions as the training target. The machine learning model is trained until the sum of the prediction errors converges and the model training is stopped. The accuracy value of the prediction results of the target stroke data analysis model under different conditions is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0084] The method for obtaining the accuracy value of the prediction results of the target stroke data analysis model under different conditions is as follows: from the comprehensive feature vector training data of the trained machine learning model, the corresponding function expression is obtained: LW=F(SD, HK); wherein F is the output function of the model, SD is the abnormal accumulation index of prediction error, HK is the fluctuation index of lesion identification accuracy, and LW is the accuracy value of the prediction results of the target stroke data analysis model under different conditions.

[0085] S5: According to the target stroke data, the accuracy of the prediction results of the model under different conditions is analyzed, and the model prediction results are divided into accurate prediction results, incompletely accurate prediction results and inaccurate prediction results, and corresponding processing is performed.

[0086] Comparing the accuracy values ​​of the prediction results of the acquired target stroke data analysis model under different conditions with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy values ​​of the prediction results of the target stroke data analysis model under different conditions with the first standard threshold and the second standard threshold respectively;

[0087] If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is greater than the second standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is high, and a high-accuracy prediction signal is generated at this time, and the model prediction results are divided into accuracy prediction results, which are directly used for clinical decision support;

[0088] If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is medium, and a medium accuracy prediction signal is generated at this time, and the model prediction result is classified as a medium accuracy prediction result for further confirmation or secondary verification to reduce potential errors;

[0089] If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is less than the first standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is low. At this time, a low-accuracy prediction signal is generated, and the model prediction results are classified as low-accuracy prediction results. There is a large error, and it is not recommended to be used directly for clinical decision-making.

[0090] S6: For incompletely accurate prediction results, further analyze the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and determine whether to isolate and store the prediction results of the target stroke data analysis model based on the analysis results.

[0091] Assume that within a fixed time period T, the accuracy value sequence of the medium accuracy prediction results is {A t}, where t = 1, 2, ..., n represents the prediction accuracy value of different conditions in the time period. Calculate the prediction accuracy change rate R in the time period AC , indicating whether the accuracy of the medium-accuracy forecast remains stable, by quantifying its volatility using the standard deviation: in, is the average accuracy value within the time period; set the fluctuation threshold θ to determine the accuracy change rate R AC Whether it exceeds the normal range. The fluctuation threshold can be set according to the average fluctuation of historical data.

[0092] If the rate of change of the prediction accuracy is greater than the fluctuation threshold, it is considered that the accuracy fluctuates greatly and the model is not stable enough under moderate conditions; otherwise, the accuracy is considered to remain stable.

[0093] If the fluctuation threshold is exceeded, the medium-accuracy prediction results within the time period are not stable enough and need to be stored in isolation. This judgment can be expressed by the formula: Here, I=1 indicates that isolated storage is required, and I=0 indicates that isolated storage is not required.

[0094] Use the above formula to determine whether to store in isolation. If I = 1, the medium-accuracy prediction results within the current time period T are stored separately to avoid these results from directly participating in clinical decision-making or subsequent data analysis. The isolation storage formula is expressed as: S 隔离 ={A t |I=1,t∈T}; where S 隔离 Represents a collection of medium-accuracy prediction results that require isolated storage.

[0095] In this embodiment, noise-free, interference-free image data of real stroke patients are first collected, and test images with different noise levels and interference types are generated based on this. Then, these test images are combined with the original noise-free data to generate multiple test data sets, and are graded according to the degree of noise and interference. These graded data sets are input into the target stroke data analysis model, and the lesion identification results and risk assessment results of the model are obtained and recorded. Then, the prediction accuracy of the model is evaluated by analyzing the abnormal accumulation degree of prediction error and the fluctuation of recognition accuracy under different noise and interference conditions. According to the prediction accuracy, the model output results are divided into accuracy prediction, incomplete accuracy prediction and inaccuracy prediction, and different accuracy results are processed accordingly. Finally, for the incomplete accuracy prediction results, the accuracy fluctuation is further analyzed within a fixed time period, and it is determined whether these data need to be isolated and stored according to the results to ensure the stability of data quality and model evaluation.

[0096] Embodiment 2, a stroke data management and analysis system based on large model technology described in this embodiment includes a data acquisition module, a data classification module, a model testing module, a model prediction accuracy evaluation module, an accuracy classification module and a management module;

[0097] Data acquisition module: collects noise-free and interference-free image data of real stroke patients, and generates test images with different noise levels and interference types based on the collected image data;

[0098] Data classification module: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple test data sets, and classify the data sets according to the degree of noise and interference;

[0099] Model testing module: input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them;

[0100] Model prediction accuracy evaluation module: Based on the abnormal accumulation of model prediction errors and fluctuations in lesion identification accuracy under different noise and interference conditions, the module evaluates the accuracy of prediction results of the target stroke data analysis model under different conditions;

[0101] Accuracy classification module: According to the target stroke data, the accuracy of the model's prediction results under different conditions is analyzed, and the model's prediction results are divided into accurate prediction results, incompletely accurate prediction results, and inaccurate prediction results, and corresponding processing is performed;

[0102] Management module: For incompletely accurate prediction results, further analysis is performed on the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and based on the analysis results, it is determined whether the prediction results of the target stroke data analysis model are to be stored in isolation.

[0103] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0104] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0105] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0106] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A stroke data management and analysis method based on large model technology, characterized in that: The following steps are involved: S1: Collect noise-free and interference-free image data of real stroke patients, and generate test images with different noise levels and interference types based on the collected image data; S2: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple test data sets, and classify the data sets according to the degree of noise and interference; S3: Input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them; S4: According to the abnormal accumulation degree of model prediction error and the fluctuation of lesion identification accuracy under different noise and interference conditions, the accuracy of prediction results of the target stroke data analysis model under different conditions was evaluated; S5: Analyze the accuracy of the prediction results of the model under different conditions according to the target stroke data, divide the model prediction results into accurate prediction results, incompletely accurate prediction results and inaccurate prediction results, and perform corresponding processing; S6: For incompletely accurate prediction results, further analyze the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and determine whether to isolate and store the prediction results of the target stroke data analysis model based on the analysis results.

2. The method for managing and analyzing stroke data based on large model technology according to claim 1, characterized in that: In S2, test data with different noise levels and different interference levels are combined separately to generate data sets, and different levels of noise are combined with different degrees of interference in pairs, including: mild noise + mild interference, moderate noise + moderate interference, and heavy noise + heavy interference. The combined data sets are divided into different levels according to the degree of influence of noise and interference.

3. The method for managing and analyzing stroke data based on large model technology according to claim 2 is characterized in that: In S4, the prediction error anomaly accumulation degree of the model under different noise and interference conditions is analyzed to generate a prediction error anomaly accumulation index. The method for obtaining the prediction error anomaly accumulation index is as follows: Under noise-free and interference-free conditions, the mean of each error index is calculated as the reference value to obtain the reference error mean vector μ; the reference vector is expressed as μ = [μ 位置误差 ,μ 风险评分误差 ] T , T is the matrix transpose; calculate the covariance matrix Σ of the error indicators under various noise and interference conditions. The diagonal elements of the covariance matrix represent the variance of each error indicator, and the off-diagonal elements represent the covariance between different error indicators. The covariance matrix is: For each noise and interference condition, calculate the error vector E i The difference E from the reference vector μ i -μ, substitute into the Mahalanobis distance formula to calculate the forecast error anomaly accumulation index, the expression is: Where SD is the abnormal cumulative index of forecast error, Σ -1 is the inverse matrix of the covariance matrix Σ of the error vector.

4. The method for managing and analyzing stroke data based on large model technology according to claim 3 is characterized in that: In S4, the fluctuation of the lesion recognition accuracy under different noise and interference conditions is analyzed to generate a lesion recognition accuracy fluctuation index. The method for obtaining the lesion recognition accuracy fluctuation index is as follows: Assume that under different noise and interference conditions, the fluctuation value of the lesion recognition accuracy is a time series {W s }, s is time, arrange the conditional sequence so that the noise and interference intensity gradually increase, select the smoothing factor α to control the degree of smoothing, and the formula of the exponential smoothing method is: S s =α·W s +(1-α)·S s-1 ; Among them, S s is the smooth fluctuation value under the conditions at the sth moment, W s is the actual lesion recognition accuracy value under the condition of the sth moment, S s-1 is the smoothed fluctuation value under the previous condition, and the smoothed value S1 under the initial condition is set to the actual accuracy value under the first condition, that is, S1=W1; starting from the second condition, the smoothed fluctuation value of each condition is gradually calculated according to the exponential smoothing calculation formula until the fluctuation values ​​of all conditions are smoothed, and the smoothed value of the last condition is used as the fluctuation index of lesion recognition accuracy.

5. The method for managing and analyzing stroke data based on large model technology according to claim 4 is characterized in that: In S4, the prediction error abnormal accumulation index and the lesion identification accuracy fluctuation index are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as the input of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the accuracy value label of the prediction results of the target stroke data analysis model under different conditions as the prediction target, and takes minimizing the sum of the prediction errors of the accuracy value labels of the prediction results of all target stroke data analysis models under different conditions as the training target. The machine learning model is trained until the sum of the prediction errors converges, and the model training is stopped. The accuracy value of the prediction results of the target stroke data analysis model under different conditions is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

6. The method for managing and analyzing stroke data based on large model technology according to claim 5 is characterized in that: In S5, according to the accuracy of the prediction results of the target stroke data analysis model under different conditions, the model prediction results are divided into accurate prediction results, incompletely accurate prediction results and inaccurate prediction results, and corresponding processing is performed, specifically: Comparing the accuracy values ​​of the prediction results of the acquired target stroke data analysis model under different conditions with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy values ​​of the prediction results of the target stroke data analysis model under different conditions with the first standard threshold and the second standard threshold respectively; If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is greater than the second standard threshold, it means that the prediction results of the target stroke data analysis model under different conditions are highly accurate, and a high-accuracy prediction signal is generated at this time, and the model prediction results are divided into accuracy prediction results, which are directly used for clinical decision support; If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is medium, and a medium accuracy prediction signal is generated at this time, and the model prediction result is classified as a medium accuracy prediction result, which needs to be further confirmed; If the accuracy value of the prediction results of the target stroke data analysis model under different conditions is less than the first standard threshold, it means that the accuracy of the prediction results of the target stroke data analysis model under different conditions is low. At this time, a low-accuracy prediction signal is generated, and the model prediction result is classified as a low-accuracy prediction result, which is not directly used for clinical decision-making.

7. The method for managing and analyzing stroke data based on large model technology according to claim 1, characterized in that: In S6, for the incompletely accurate prediction results, the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period is further analyzed, and whether to isolate and store the prediction results of the target stroke data analysis model is determined according to the analysis results, specifically: Assume that within a fixed time period T, the accuracy value sequence of the medium accuracy prediction results is {A t }, where t = 1, 2, ..., n represents the prediction accuracy value of different conditions in the time period, and the prediction accuracy change rate R in the time period is calculated AC , indicating whether the accuracy of the medium-accuracy forecast remains stable, by quantifying its volatility using the standard deviation: in, is the average accuracy value within the time period; set the fluctuation threshold θ to determine the accuracy change rate R AC Whether it exceeds the normal range; if the prediction accuracy change rate is greater than the fluctuation threshold, it is considered that the accuracy fluctuates greatly and the model is not stable enough under medium conditions; otherwise, it is considered that the accuracy remains stable and needs to be isolated and stored. The judgment formula is expressed as: Where I = 1 means that isolated storage is required, and I = 0 means that isolated storage is not required; the isolated storage formula is expressed as: S 隔离 ={A t |I=1,t∈T}; where S 隔离 Represents a collection of medium-accuracy prediction results that require isolated storage.

8. A stroke data management and analysis system based on large model technology, used to implement a stroke data management and analysis method based on large model technology as described in any one of claims 1 to 7, characterized in that: It includes data collection module, data classification module, model testing module, model prediction accuracy assessment module, accuracy classification module and management module; Data acquisition module: collects noise-free and interference-free image data of real stroke patients, and generates test images with different noise levels and interference types based on the collected image data; Data classification module: Combine the test data with different noise levels and image interference in the test image with the original noise-free data to generate multiple test data sets, and classify the data sets according to the degree of noise and interference; Model testing module: input the test data set into the target stroke data analysis model, obtain the lesion identification results and risk assessment results output by the model, and record them; Model prediction accuracy evaluation module: Based on the abnormal accumulation of model prediction errors and fluctuations in lesion identification accuracy under different noise and interference conditions, the module evaluates the accuracy of prediction results of the target stroke data analysis model under different conditions; Accuracy classification module: According to the target stroke data, the accuracy of the model's prediction results under different conditions is analyzed, and the model's prediction results are divided into accurate prediction results, incompletely accurate prediction results, and inaccurate prediction results, and corresponding processing is performed; Management module: For incompletely accurate prediction results, further analysis is performed on the accuracy of the prediction results of the target stroke data analysis model under different conditions within a fixed time period, and based on the analysis results, it is determined whether to isolate and store the prediction results of the target stroke data analysis model.

Images