Intelligent diagnosis and risk assessment method and system for cerebral apoplexy related to atrial fibrillation

Abstract

The invention discloses an intelligent diagnosis and risk assessment method and system for cerebral apoplexy related to atrial fibrillation, and relates to the technical field of atrial fibrillation detection.The method comprises the steps that multi-dimensional data of atrial fibrillation patients in a clinical information system are integrated, a structured database is generated, and a standardized data set is output; a standardized data set is adopted to train a first machine learning model, and model performance is optimized through parameter joint search and a training set-verification set convergence dynamic monitoring mechanism; performing cross validation on a feature weight sorting result in the optimized diagnosis model and a clinical index risk association degree calculated by a second machine learning model to generate an interaction map; and based on clinical event data containing timestamps in the structured database, adopting a third machine learning model to extract time sequence features, and combining with a survival probability analysis model to generate a risk assessment report. According to the invention, through an intelligent model adjusting and optimizing mechanism, multi-dimensional medical data are effectively integrated, and the recognition precision of the atrial fibrillation related cerebral apoplexy is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of atrial fibrillation detection, and particularly relates to an intelligent diagnosis and risk assessment method and system for atrial fibrillation-related stroke. Background Art

[0002] Atrial fibrillation (AF) is a core risk factor for stroke. However, traditional assessment methods rely on a single indicator (such as the CHA2DS2-VASc score) or doctor experience, with a misdiagnosis rate as high as 30%, and the dynamic changes in risk cannot be quantified. With the explosion of medical data (electronic medical records, laboratory indicators, medication records, etc.), the integration and standardization of multi-source heterogeneous data have become technical difficulties. Problems such as data missing, outliers, and temporal chaos severely restrict the clinical application of artificial intelligence models, and a systematic solution is urgently needed.

[0003] In the prior art, the intelligent diagnosis model for AF-related stroke lacks a dynamic optimization mechanism and multi-dimensional data collaborative verification, resulting in insufficient diagnostic accuracy and unreliable results. The specific manifestations are as follows:

[0004] Limited model accuracy: Traditional single models (such as logistic regression) have rough parameter tuning (only relying on grid search or random search), cannot dynamically monitor the convergence state of the training set and the validation set, resulting in a high risk of overfitting and generally low AUC, and it is difficult to meet the high-precision requirements of clinical practice.

[0005] Limitations of static assessment: Existing methods only evaluate risks based on static indicators (such as single detection values), do not integrate temporal data such as medication records and review results, and cannot predict the dynamic time window of stroke occurrence (such as the high-risk period in the next 6-12 months), resulting in a lag in the intervention timing. Summary of the Invention

[0006] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide an AI-based urban traffic flow prediction and dynamic signal optimization method and system, aiming to effectively integrate multi-dimensional medical data through an intelligent model optimization mechanism, greatly improve the recognition accuracy of AF-related stroke, and provide high-reliability support for clinical decision-making.

[0007] The first aspect of the present invention provides an intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke, and the method includes:

[0008] Integrate multi-dimensional data of atrial fibrillation patients in the clinical information system, generate a structured database through data quality verification, outlier removal, and missing value filling, and output a standardized data set;

[0009] Train the first machine learning model using the standardized dataset, optimize the model performance through parameter joint search and training set-validation set convergence dynamic monitoring mechanism, lock the model parameters when the validation set reaches the preset diagnostic accuracy threshold, and output the optimized diagnostic model;

[0010] Cross-validate the feature weight ranking results in the optimized diagnostic model with the clinical index risk correlation calculated by the second machine learning model to generate a visual risk factor interaction map;

[0011] Repeat the division of the standardized dataset, train at least three different types of machine learning models respectively, and generate a comparison and verification result by calculating the average diagnostic efficacy and confidence interval of each model;

[0012] Based on the clinical event data with timestamps in the structured database, use the third machine learning model to extract time series features, and generate a risk assessment report including risk level stratification and prognosis time window in combination with the survival probability analysis model.

[0013] According to one aspect of the above technical solution, the data quality verification includes:

[0014] Execute column deletion for fields with missing value ratio exceeding the preset threshold;

[0015] For outliers in laboratory test indicators, perform truncation processing based on the medical reference value range;

[0016] For clinical event data with inconsistent timestamp formats, convert it to a unified time axis record through time alignment rules.

[0017] According to one aspect of the above technical solution, the steps of training the first machine learning model using the standardized dataset, optimizing the model performance through parameter joint search and training set-validation set convergence dynamic monitoring mechanism, and locking the model parameters when the validation set reaches the preset diagnostic accuracy threshold and outputting the optimized diagnostic model include:

[0018] Use the grid search method to perform combined traversal on the tree depth, learning rate, and regularization parameters of the first machine learning model;

[0019] Real-time calculate the accuracy difference between the training set and the validation set, and terminate the search and lock the current parameter combination when the difference continuously shrinks to the convergence threshold.

[0020] According to one aspect of the above technical solution, the first machine learning model is a gradient boosting decision tree model, and the area under the receiver operating characteristic curve of its optimized diagnostic model in the validation set is not less than 0.95.

[0021] According to one aspect of the above technical solution, the second machine learning model is a logistic regression model;

[0022] The step of cross - validating the sorted result of feature weights in the optimized diagnostic model with the risk correlation degree of clinical indicators calculated by the second machine learning model to generate a visual risk factor interaction map includes:

[0023] Calculate the odds ratio of each clinical indicator in the logistic regression model, and screen out the significant indicators whose absolute value of the odds ratio is greater than 1 and the confidence interval does not contain 1;

[0024] Perform a rank correlation analysis on the significant indicators and the sorted result of feature weights, and retain the indicators whose correlation coefficient exceeds the preset threshold to generate the visual risk factor interaction map.

[0025] According to one aspect of the above - mentioned technical solution, the third machine learning model is a recurrent neural network;

[0026] The step of extracting time - series features based on the clinical event data containing timestamps in the structured database by using a third machine learning model and generating a risk assessment report including risk - level stratification and prognosis time window in combination with a survival probability analysis model includes:

[0027] Obtain the clinical event data containing timestamps in the structured database, and use a recurrent neural network to divide the clinical event data into multiple clinical time series according to a preset time window, and output the corresponding time - series feature vectors;

[0028] Receive the time - series feature vectors and static clinical indicators through a survival probability analysis model, calculate the risk ratio and survival probability, and generate a risk assessment report including time - window prediction.

[0029] According to one aspect of the above - mentioned technical solution, the step of receiving the time - series feature vectors and static clinical indicators through a survival probability analysis model, calculating the risk ratio and survival probability, and generating a risk assessment report including time - window prediction includes:

[0030] Perform dimensionality compression on the time - series feature vectors to generate time - series embedding representations;

[0031] Concatenate the time - series embedding representations with the static clinical indicators after standardization processing to form a combined input vector;

[0032] Calculate the risk ratio of each clinical indicator through a COX proportional hazards model, and estimate the survival probability based on the partial likelihood function;

[0033] Dynamically divide the time windows corresponding to the low - risk period, medium - risk period, and high - risk period according to the survival probability decay rate;

[0034] Generate a visual survival curve and a risk time - window distribution map, and mark the clinical intervention suggestions corresponding to each stage.

[0035] The second aspect of the present invention is to provide an intelligent diagnosis and risk assessment system for atrial fibrillation-related stroke, which is applied to the method described in the above technical solution. The system includes:

[0036] A data processing module, configured to integrate multi-dimensional data of atrial fibrillation patients in the clinical information system, generate a structured database through data quality verification, outlier removal, and missing value filling, and output a standardized data set;

[0037] A model training module, configured to train a first machine learning model using the standardized data set, optimize the model performance through a parameter joint search and a training set-validation set convergence dynamic monitoring mechanism, lock the model parameters when the validation set reaches a preset diagnostic accuracy threshold, and output an optimized diagnostic model;

[0038] A first verification module, configured to perform cross-verification on the feature weight ranking results in the optimized diagnostic model and the risk correlation degree of clinical indicators calculated by a second machine learning model, and generate a visual risk factor interaction map;

[0039] A second verification module, configured to repeatedly divide the standardized data set, train at least three different types of machine learning models respectively, and generate a comparative verification result by calculating the average diagnostic efficacy and confidence interval of each model;

[0040] A report output module, configured to extract time series features using a third machine learning model based on the clinical event data including timestamps in the structured database, and generate a risk assessment report including risk level stratification and prognosis time window in combination with a survival probability analysis model.

[0041] The third aspect of the present invention is to provide a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method described in the above technical solution is implemented.

[0042] The fourth aspect of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the method described in the above technical solution is implemented.

[0043] Compared with the prior art, the beneficial effects of adopting the intelligent diagnosis and risk assessment method and system for atrial fibrillation-related stroke shown in the present invention are as follows:

[0044] The present invention integrates multi-dimensional data of atrial fibrillation patients in a clinical information system, generates a structured database through data quality verification, outlier elimination, and missing value filling, and outputs a standardized data set. The standardized data set is used to train a first machine learning model, and the model performance is optimized through a parameter joint search and training set-validation set convergence dynamic monitoring mechanism. When the validation set reaches a preset diagnostic accuracy threshold, the model parameters are locked, and an optimized diagnostic model is output. The cross-validation is performed on the sorting result of the feature weights in the optimized diagnostic model and the risk correlation degree of the clinical indicators calculated by the second machine learning model to generate a visual risk factor interaction map. The standardized data set is repeatedly divided, and at least three different types of machine learning models are trained respectively. The comparative verification result is generated by calculating the average diagnostic efficacy and confidence interval of each model. Based on the clinical event data containing time stamps in the structured database, the third machine learning model is used to extract time series features, and a risk assessment report including risk level stratification and prognosis time window is generated in combination with a survival probability analysis model. Through an intelligent model tuning mechanism, the present invention effectively integrates multi-dimensional medical data, greatly improves the recognition accuracy of atrial fibrillation-related strokes, and provides high-reliability support for clinical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and / or additional aspects and advantages of the present invention will become obvious and easy to understand from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0046] Figure 1 It is a schematic flowchart of a method for intelligent diagnosis and risk assessment of atrial fibrillation-related strokes shown in an embodiment of the present invention;

[0047] Figure 2 It is a structural block diagram of a system for intelligent diagnosis and risk assessment of atrial fibrillation-related strokes shown in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] To make the objectives, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention is provided in conjunction with the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0050] Example 1

[0051] Please refer to Figure 1 , the first embodiment of the present invention provides an intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke, and the method includes:

[0052] Step S10, integrate multi-dimensional data of atrial fibrillation patients in the clinical information system, generate a structured database through data quality verification, outlier removal, and missing value filling, and output a standardized data set.

[0053] In this embodiment, the data quality verification includes:

[0054] Exclude the entire column for fields where the proportion of missing values exceeds a preset threshold;

[0055] For outliers in laboratory test indicators, perform truncation processing based on the medical reference value range;

[0056] For clinical event data with inconsistent timestamp formats, convert it to a unified time axis record through time alignment rules.

[0057] Specifically, in this embodiment, multi-dimensional data of atrial fibrillation patients obtains comprehensive data of atrial fibrillation patients from the hospital electronic medical record system, laboratory information management system, image archiving system, etc., including demographic information (such as age, gender), test results (such as coagulation function indicators, liver and kidney function values, electrocardiogram signals), medication records (such as anticoagulant drug usage), imaging reports (such as brain CT results), and follow-up records. Among them, when generating a structured database, it includes missing value processing, outlier correction, timestamp unification, and standardization processing.

[0058] More specifically, missing value processing identifies fields (such as certain special gene detection items) in the data table with a missing rate exceeding a preset threshold (such as 30%), and directly deletes the entire column to ensure the integrity of the remaining data. Outlier correction truncates the detected values outside the range according to the medical standard reference range (such as the normal value of platelet count is 150 - 450×109 / L). For example, values higher than 450 are corrected to 450, and values lower than 150 are corrected to 150. Timestamp unification converts different formats of time data (such as "2023-05-01" and "01 / 05 / 2023") into a unified "year-month-day" format, and aligns event records in chronological order to ensure the coherence of subsequent time series analysis. Standardization processing includes standardizing numerical data (such as prothrombin time, creatinine value) to eliminate dimensional differences. For example, the Z-score method is used to convert the data into a distribution with a mean of 0 and a standard deviation of 1 to ensure the comparability of different indicators in model training.

[0059] Step S20: Train the first machine learning model using the standardized data set, optimize the model performance through parameter joint search and dynamic monitoring mechanism of training set - validation set convergence, lock the model parameters when the validation set reaches the preset diagnostic accuracy threshold, and output the optimized diagnostic model.

[0060] In this embodiment, the steps of training the first machine learning model using the standardized data set, optimizing the model performance through parameter joint search and dynamic monitoring mechanism of training set - validation set convergence, locking the model parameters when the validation set reaches the preset diagnostic accuracy threshold, and outputting the optimized diagnostic model include:

[0061] Use the grid search method to perform combined traversal on the tree depth, learning rate, and regularization parameters of the first machine learning model;

[0062] Calculate the accuracy difference between the training set and the validation set in real time. When the difference continuously shrinks to the convergence threshold, terminate the search and lock the current parameter combination.

[0063] Among them, the first machine learning model is a gradient - boosting decision tree model, and the area under the receiver operating characteristic curve of its optimized diagnostic model in the validation set is not less than 0.95.

[0064] Specifically, in this embodiment, when outputting the diagnostic model, it includes model selection and optimization, parameter joint search, dynamic monitoring of convergence, and model locking and output.

[0065] More specifically, for model selection and initialization, a gradient - boosting decision tree is selected as the basic model because it is good at processing mixed - type data (numerical and categorical) and has the ability to analyze feature importance. Parameter joint search (grid search) exhausts all possible parameter combinations within the preset parameter space (for example, the tree depth range is from 3 to 8 layers, and the learning rate range is from 0.01 to 0.3), trains the model respectively and records the performance. During the dynamic monitoring of convergence, during each parameter combination training process, the accuracy curves of the training set and the validation set are plotted in real time. When the accuracy difference between the two remains stable within a very small range (such as within 1%) for multiple consecutive iterations (such as 3 times), it is determined that the model has converged, and the search is stopped to avoid overfitting. For model locking and output, by selecting the parameter combination with the highest validation set accuracy and stable convergence, the model parameters are saved as the final diagnostic model for subsequent analysis.

[0066] Step S30: Cross - validate the feature weight ranking result in the optimized diagnostic model with the clinical index risk correlation calculated by the second machine learning model to generate a visual risk factor interaction map.

[0067] In this embodiment, the second machine learning model is a logistic regression model;

[0068] The steps of cross - validating the sorted result of feature weights in the optimized diagnostic model with the risk correlation degree of clinical indicators calculated by a second machine learning model to generate a visualized risk factor interaction map include:

[0069] Calculate the odds ratio of each clinical indicator in the logistic regression model, and screen out the significant indicators whose absolute value of the odds ratio is greater than 1 and the confidence interval does not contain 1;

[0070] Perform a rank - correlation analysis on the significant indicators and the sorted result of feature weights, and retain the indicators whose correlation coefficient exceeds the preset threshold to generate the visualized risk factor interaction map.

[0071] Specifically, in this embodiment, when generating the visualized risk factor interaction map, it includes feature weight extraction, second - model risk correlation degree calculation, cross - validation and screening, and visualization generation.

[0072] More specifically, feature weight extraction includes obtaining the feature importance scores of each clinical indicator from the gradient - boosting decision tree model (for example, the weight of prothrombin time is 0.2, and the weight of age is 0.15), and sorting them from high to low. The second - model risk correlation degree calculation calculates the odds ratio (OR value) of each clinical indicator by using the logistic regression model. For example, the OR value of age is 1.2 (indicating that for every 1 - year increase in age, the stroke risk increases by 20%), and its 95% confidence interval is calculated. Cross - validation and screening include performing a correlation analysis (such as Spearman rank - correlation coefficient) between the feature importance ranking and the OR value, and screening out the indicators that are significant and highly correlated in both methods (for example, prothrombin time ranks among the top 3 in both models, and the correlation coefficient exceeds 0.8). Visualization generation displays the key risk factors through an interactive chart. For example, in a bubble chart, the size of the bubble represents the feature weight, the depth of the color represents the OR value, and the horizontal and vertical axes display different statistical indicators, helping doctors intuitively understand the influence degree and confidence level of each factor.

[0073] Step S40: Repeatedly divide the standardized data set, train at least three different types of machine learning models respectively, and generate a comparative verification result by calculating the average diagnostic efficacy and confidence interval of each model.

[0074] Specifically, in this embodiment, generating the comparative verification result of each model includes setting the data - set division strategy, multi - model training, calculation of efficacy indicators, and visualization generation.

[0075] More specifically, the setting of the dataset partitioning strategy includes randomly partitioning the standardized data into a training set and a test set in a ratio of 8:2, and repeating this process 100 times to reduce the impact of randomness. Multi-model training is carried out by selecting three models: gradient boosting decision tree, deep neural network (DNN), and support vector machine (SVM), training them respectively under different data partitions, and recording the AUC value each time. The calculation of performance indicators includes taking the average of the AUC results of 100 iterations and calculating the 95% confidence interval (for example, the average AUC of the gradient boosting decision tree is 0.95, and the confidence interval is 0.93 - 0.97). Visual comparison is carried out by generating a forest plot, where each horizontal line in the plot represents the mean and confidence interval of the AUC of a model, intuitively showing the performance differences of each model and verifying the superiority of the gradient boosting decision tree.

[0076] Step S50: Based on the clinical event data containing timestamps in the structured database, use a third machine learning model to extract temporal features, and combine with a survival probability analysis model to generate a risk assessment report including risk level stratification and prognosis time window.

[0077] In this embodiment, the third machine learning model is a recurrent neural network;

[0078] The steps of using a third machine learning model to extract temporal features based on the clinical event data containing timestamps in the structured database and combining with a survival probability analysis model to generate a risk assessment report including risk level stratification and prognosis time window include:

[0079] Obtain the clinical event data containing timestamps in the structured database, and use a recurrent neural network to divide the clinical event data into multiple clinical time series according to a preset time window, and output corresponding temporal feature vectors;

[0080] Receive the temporal feature vectors and static clinical indicators through the survival probability analysis model, calculate the hazard ratio and survival probability, and generate a risk assessment report including time window prediction.

[0081] Among them, the steps of receiving the temporal feature vectors and static clinical indicators through the survival probability analysis model, calculating the hazard ratio and survival probability, and generating a risk assessment report including time window prediction include:

[0082] Perform dimensionality compression on the temporal feature vectors to generate temporal embedding representations;

[0083] Perform feature concatenation on the temporal embedding representations and the standardized static clinical indicators to form a combined input vector;

[0084] Calculate the hazard ratio of each clinical indicator through the COX proportional hazards model and estimate the survival probability based on the partial likelihood function;

[0085] Dynamically divide the time windows corresponding to the low-risk period, medium-risk period, and high-risk period according to the decay rate of the survival probability;

[0086] Generate a visual survival curve and a distribution map of the risk time window, and mark the clinical intervention suggestions corresponding to each stage.

[0087] Specifically, in this embodiment, generating a risk assessment report includes time series data processing, time series feature extraction, survival analysis modeling, risk staging, and report output.

[0088] More specifically, time series data processing includes slicing the patient's clinical event data (such as medication records, review results) according to a fixed time window (such as every 3 months) to form time series segments. Time series feature extraction processes the time series by using a recurrent neural network (such as a gated recurrent unit GRU) to capture the patterns of events changing over time and outputs a feature vector representing the time series dynamics. Survival analysis modeling splices the time series features with static indicators (such as age, gender), inputs them into the COX proportional hazards model, and calculates the hazard ratio of each indicator (such as the hazard ratio of the INR value is 1.5, indicating that an abnormal INR value increases the risk by 50%) and the survival probability curve. Risk staging and reporting include dividing the risk period according to the decline rate of the survival probability (such as the survival probability ≥ 80% is the low-risk period, lasting for the next 6 months; 50% - 80% is the medium-risk period, lasting for 6 - 12 months; < 50% is the high-risk period, lasting for more than 12 months), generating a PDF report containing the staging results, survival curve, and intervention suggestions (such as suggesting a weekly review of the coagulation function during the high-risk period), or a report in other data formats.

[0089] Compared with the prior art, the intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke shown in this embodiment has the beneficial effects that:

[0090] In this embodiment, by integrating multi-dimensional data of atrial fibrillation patients in the clinical information system, a structured database is generated through data quality verification, outlier elimination, and missing value filling, and a standardized data set is output. The standardized data set is used to train the first machine learning model, and the model performance is optimized through a parameter joint search and a dynamic monitoring mechanism for the convergence of the training set - validation set. When the validation set reaches the preset diagnostic accuracy threshold, the model parameters are locked, and an optimized diagnostic model is output. The sorting result of the feature weights in the optimized diagnostic model is cross-validated with the risk correlation degree of the clinical indicators calculated by the second machine learning model to generate a visual risk factor interaction map. The standardized data set is repeatedly divided, and at least three different types of machine learning models are trained respectively. The comparative verification results are generated by calculating the average diagnostic efficacy and confidence interval of each model. Based on the clinical event data containing timestamps in the structured database, the third machine learning model is used to extract time series features, and a risk assessment report including risk level stratification and prognosis time window is generated in combination with the survival probability analysis model. Through the intelligent model tuning mechanism, this embodiment effectively integrates multi-dimensional medical data, greatly improves the recognition accuracy of atrial fibrillation-related stroke, and provides high-reliability support for clinical decision-making.

[0091] Embodiment 2

[0092] Please refer to Figure 2 , the second embodiment of the present invention provides an intelligent diagnosis and risk assessment system for atrial fibrillation-related stroke, which is applied to the method described in the above embodiment. The system includes:

[0093] A data processing module 10, configured to integrate multi-dimensional data of atrial fibrillation patients in the clinical information system, generate a structured database through data quality verification, outlier elimination, and missing value filling, and output a standardized data set;

[0094] A model training module 20, configured to use the standardized data set to train the first machine learning model, optimize the model performance through a parameter joint search and a dynamic monitoring mechanism for the convergence of the training set - validation set, lock the model parameters when the validation set reaches the preset diagnostic accuracy threshold, and output an optimized diagnostic model;

[0095] A first verification module 30, configured to cross-validate the sorting result of the feature weights in the optimized diagnostic model with the risk correlation degree of the clinical indicators calculated by the second machine learning model to generate a visual risk factor interaction map;

[0096] A second verification module 40, configured to repeatedly divide the standardized data set, train at least three different types of machine learning models respectively, and generate comparative verification results by calculating the average diagnostic efficacy and confidence interval of each model;

[0097] A report output module 50, configured to extract temporal features based on the clinical event data including timestamps in the structured database, and generate a risk assessment report including risk level stratification and prognostic time window in combination with a survival probability analysis model.

[0098] Compared with the prior art, the intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke shown in this embodiment has the following beneficial effects:

[0099] In this embodiment, by integrating multi-dimensional data of atrial fibrillation patients in the clinical information system, a structured database is generated through data quality verification, outlier removal, and missing value filling, and a standardized data set is output; the first machine learning model is trained using the standardized data set, and the model performance is optimized through a parameter joint search and training set-validation set convergence dynamic monitoring mechanism. When the validation set reaches the preset diagnostic accuracy threshold, the model parameters are locked, and an optimized diagnostic model is output; the feature weight ranking result in the optimized diagnostic model is cross-validated with the clinical index risk correlation degree calculated by the second machine learning model to generate a visual risk factor interaction map; the standardized data set is repeatedly divided, and at least three different types of machine learning models are trained respectively, and a comparison and verification result is generated by calculating the average diagnostic efficacy and confidence interval of each model; based on the clinical event data including timestamps in the structured database, temporal features are extracted using the third machine learning model, and a risk assessment report including risk level stratification and prognostic time window is generated in combination with a survival probability analysis model. Through the intelligent model tuning mechanism in this embodiment, multi-dimensional medical data is effectively integrated, the recognition accuracy of atrial fibrillation-related stroke is greatly improved, and high-reliability support is provided for clinical decision-making.

[0100] Embodiment 3

[0101] The third embodiment of the present invention provides a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method described in any of the above embodiments is implemented.

[0102] Embodiment 4

[0103] The fourth embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method described in any of the above embodiments is implemented.

[0104] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc., mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0105] The above-described embodiments only express several implementation manners of the present invention. The descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be subject to the appended claims.

Claims

1. An intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke, characterized in that, The method includes: Integrating multi-dimensional data of atrial fibrillation patients in the clinical information system, generating a structured database through data quality verification, outlier removal, and missing value filling, and outputting a standardized data set; Training a first machine learning model using the standardized data set, optimizing the model performance through parameter joint search and dynamic monitoring mechanism of training set-validation set convergence, locking the model parameters when the validation set reaches the preset diagnostic accuracy threshold, and outputting an optimized diagnostic model; Cross-verifying the feature weight ranking results in the optimized diagnostic model with the risk correlation degree of clinical indicators calculated by a second machine learning model to generate a visual risk factor interaction map; Repeatedly dividing the standardized data set, training at least three different types of machine learning models respectively, and generating a comparison verification result by calculating the average diagnostic efficacy and confidence interval of each model; Based on the clinical event data containing timestamps in the structured database, using a third machine learning model to extract time series features, and generating a risk assessment report including risk level stratification and prognosis time window in combination with a survival probability analysis model.

2. The intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke according to claim 1, wherein The data quality verification includes: Performing column deletion on fields with a missing value ratio exceeding a preset threshold; Truncating outliers in laboratory test indicators based on the medical reference value range; Converting clinical event data with inconsistent timestamp formats into a unified time axis record through time alignment rules.

3. The intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke according to claim 1, characterized in that The steps of training a first machine learning model using the standardized data set, optimizing the model performance through parameter joint search and dynamic monitoring mechanism of training set-validation set convergence, locking the model parameters when the validation set reaches the preset diagnostic accuracy threshold, and outputting an optimized diagnostic model include: Using the grid search method to perform combined traversal on the tree depth, learning rate, and regularization parameters of the first machine learning model; Calculating the accuracy difference between the training set and the validation set in real time, and terminating the search and locking the current parameter combination when the difference continuously shrinks to the convergence threshold.

4. The atrial fibrillation-related stroke intelligent diagnosis and risk assessment method according to claim 3, wherein The first machine learning model is a gradient boosting decision tree model, and the area under the receiver operating characteristic curve of its optimized diagnostic model in the validation set is not less than 0.

95.

5. The intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke according to claim 1, characterized in that The second machine learning model is a logistic regression model; The steps of cross-verifying the feature weight ranking results in the optimized diagnostic model with the risk correlation degree of clinical indicators calculated by a second machine learning model to generate a visual risk factor interaction map include: Calculating the odds ratio of each clinical indicator in the logistic regression model, and screening significant indicators with an absolute value of the odds ratio greater than 1 and a confidence interval not containing 1; Performing rank correlation analysis on the significant indicators and the feature weight ranking results, and retaining the indicators with a correlation coefficient exceeding the preset threshold to generate the visual risk factor interaction map.

6. The intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke according to claim 1, wherein The third machine learning model is a recurrent neural network; The steps of using a third machine learning model to extract time series features based on the clinical event data containing timestamps in the structured database, and generating a risk assessment report including risk level stratification and prognosis time window in combination with a survival probability analysis model include: Obtain the clinical event data containing timestamps in the structured database, and use a recurrent neural network to obtain multiple clinical time series by dividing the clinical event data according to a preset time window, and output corresponding time series feature vectors; Receive the time series feature vectors and static clinical indicators through a survival probability analysis model, calculate the hazard ratio and survival probability, and generate a risk assessment report including time window prediction.

7. The intelligent diagnosis and risk assessment method for atrial fibrillation-related stroke according to claim 6, characterized in that The steps of receiving the time series feature vectors and static clinical indicators through a survival probability analysis model, calculating the hazard ratio and survival probability, and generating a risk assessment report including time window prediction include: Perform dimensionality compression on the time series feature vectors to generate time series embedding representations; Perform feature splicing on the time series embedding representations and the static clinical indicators after standardization processing to form a joint input vector; Calculate the hazard ratio of each clinical indicator through a COX proportional hazards model, and estimate the survival probability based on the partial likelihood function; Dynamically divide the time windows corresponding to the low-risk period, medium-risk period, and high-risk period according to the survival probability decay rate; Generate a visual survival curve and a risk time window distribution map, and mark the clinical intervention suggestions corresponding to each stage.

8. An intelligent diagnosis and risk assessment system for atrial fibrillation-related stroke, characterized in that, Applied to the method according to any one of claims 1-7, the system includes: A data processing module, configured to integrate multi-dimensional data of atrial fibrillation patients in a clinical information system, generate a structured database through data quality verification, outlier removal, and missing value filling, and output a standardized data set; A model training module, configured to train a first machine learning model using the standardized data set, optimize the model performance through a parameter joint search and training set-validation set convergence dynamic monitoring mechanism, lock the model parameters when the validation set reaches a preset diagnostic accuracy threshold, and output an optimized diagnostic model; A first verification module, configured to perform cross-verification on the feature weight ranking results in the optimized diagnostic model and the clinical indicator risk correlation calculated by a second machine learning model, and generate a visual risk factor interaction map; A second verification module, configured to repeatedly divide the standardized data set, train at least three different types of machine learning models respectively, and generate a comparison verification result by calculating the average diagnostic efficacy and confidence interval of each model; A report output module, configured to extract time series features based on the clinical event data containing timestamps in the structured database, and generate a risk assessment report including risk level stratification and prognosis time window in combination with a survival probability analysis model.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1-7.

10. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, When the processor executes the computer program, it implements the method according to any one of claims 1-7.

Images