Diabetic symptom ICAS prediction method and system based on machine learning

Abstract

The invention relates to a diabetic symptomatic ICAS prediction method and system based on machine learning. The method comprises the following steps: collecting an HRMRI image sample of a target patient; performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features; performing image omics feature extraction and screening on the HRMRI image samples to obtain image omics features; based on an XGBoost algorithm, training the high-risk plaque features and the radiomics features to obtain a prediction model; and performing symptomatic ICAS attack prediction on a to-be-predicted patient through the prediction model to obtain a prediction result. The accuracy and the stability of the clinical model in predicting the new symptomatic intracranial atherosclerotic disease of the diabetic patient can be improved, and the clinical model adapts to individual specificity.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method and system for predicting symptomatic ICAS in diabetic patients based on machine learning. Background Art

[0002] Diabetes mellitus (DM) is associated with an increased risk of symptomatic intracranial atherosclerosis stenosis (ICAS). Previous studies have mainly focused on exploring clinical risk factors for predicting intracranial ischemic events in diabetic patients, such as gender, age, duration of diabetes, and blood glucose control level. However, the prediction models constructed based on traditional clinical risk factors have low accuracy and lack stability and individual specificity.

[0003] High-resolution vessel wall MR imaging (HRMRI) is a non-invasive and preferred method for evaluating intracranial atherosclerosis. It can perform qualitative and quantitative analysis of plaques and is used to assess the risk of cerebrovascular events in diabetic patients. However, the traditional analysis results based on HRMRI images are subjective, and the traditional image features provide little information. More potential image feature information needs to be mined to help better predict the risk of new symptomatic ICAS in diabetic patients.

[0004] Therefore, there is an urgent need in the art for a method to explore potential associations with disease states to detect and identify symptomatic intracranial atherosclerotic disease in patients with DM. Summary of the invention

[0005] The present invention provides a method and system for predicting symptomatic ICAS in diabetic patients based on machine learning, so as to solve the defects of the prior art.

[0006] The present invention provides a method for predicting symptomatic ICAS in diabetic patients based on machine learning, comprising: S1: Collect HRMRI image samples of target patients; S2: performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features; S3: extracting and screening radiomics features from the HRMRI image samples to obtain radiomics features; S4: Based on the XGBoost algorithm, the high-risk plaque features and the imaging genomics features are trained to obtain a prediction model; S5: Predicting the onset of symptomatic ICAS in the patient to be predicted by using the prediction model to obtain a prediction result.

[0007] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, the target patient in step S1 is a diabetic patient with intracranial atherosclerotic disease who is screened according to a preset first screening factor.

[0008] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, the first screening factor includes: Patients with a history of stroke, cardioembolic stroke, responsible vessel occlusion, non-atherosclerotic intracranial artery stenosis, history of stent implantation or endarterectomy, substandard imaging quality, missing clinical information, or a history of cerebral ischemic events.

[0009] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, step S1 further includes: S11: The target patients with ischemic events were classified into the basic symptomatic group, and the target patients without ischemic events were classified into the basic asymptomatic group; S12: performing propensity score matching on the basic symptomatic group and the basic asymptomatic group by using a greedy nearest neighbor algorithm to obtain a comparison symptomatic group and a comparison asymptomatic group.

[0010] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, the high-risk plaque features in step S2 include: The plaque load at the narrowest part of the blood vessel is expressed as: ; in, To calculate the plaque burden at the narrowest part of the vessel, is the area of ​​the blood vessel at its narrowest point, is the lumen area at the narrowest point; Vascular remodeling index, the expression of the vascular remodeling index is: ; in, To calculate the vascular remodeling index, is the vascular area at the site of stenosis, is the vessel area of ​​the reference vessel; Intraplaque hemorrhage, which means a plaque showing a signal intensity 1.5 times higher than that of the adjacent muscle on T1-weighted images; The plaque surface is irregular, which means that there is a discontinuity in the surface of the plaque immediately adjacent to the lumen.

[0011] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, step S3 further comprises: S31: performing plaque marking on the HRMRI image sample; S32: extracting the volume of interest based on the annotated plaque outline; S33: extracting radiomics features from the volume of interest to obtain a first image feature; S34: Screening the first image features to obtain radiomics features.

[0012] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, step S34 further comprises: S341: adjusting the second screening factors including plaque load and glycated hemoglobin respectively, and screening the first image features by linear regression analysis to obtain second image features related to symptomatic plaques; S342: Processing the second image feature by using a Z score to obtain a standardized third image feature; S343: Filter the third image feature by using the LASSO operator to obtain radiomics features.

[0013] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, step S4 further comprises: S41: Establish parameter dictionary and create multiple XGBClassifier() classifiers; S42: establishing multiple data sets based on the high-risk plaque features and the radiomics features; S43: performing 10-fold cross validation on multiple XGBClassifier() classifiers through multiple data sets respectively, and adjusting the parameters of the training learner to converge the XGBClassifier() classifiers to obtain multiple nested models; wherein each set of data sets corresponds to each XGBClassifier() classifier one by one; S44: Combine multiple nested models to obtain a prediction model.

[0014] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, the nested model in step S43 includes: A first nested model, wherein the first nested model is trained using high-risk plaque features as a data set; a second nested model, wherein the second nested model is trained using high-risk plaque characteristics and a second screening factor as a data set; The third nested model, the first nested model is trained with high-risk plaque characteristics, the second screening factor, and imaging omics characteristics as data sets.

[0015] According to a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided by the present invention, after step S4, the method further includes: A global explanation method is used to provide personalized explanations for each feature of the prediction model.

[0016] The present invention also provides a system for predicting symptomatic ICAS in diabetic patients based on machine learning, comprising: Acquisition module: used for acquiring HRMRI image samples of target patients; wherein the target patients are diabetic patients with intracranial atherosclerotic diseases who are screened according to a preset first screening factor; Analysis module: used for performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features; Extraction module: used for extracting and screening radiomics features from the HRMRI image samples to obtain radiomics features; Training module: used to train the high-risk plaque features and the imaging genomics features based on the XGBoost algorithm to obtain a prediction model; Prediction module: used to predict the onset of symptomatic ICAS for the patient to be predicted by using the prediction model obtained through training in the training module to obtain a prediction result.

[0017] The present invention provides a prediction method and system for symptomatic ICAS in diabetic patients based on machine learning, aiming to develop an interpretable ML model that combines traditional plaque features with clinical features of diabetes and imaging features based on HRMRI to detect and identify symptomatic intracranial atherosclerotic plaques in DM patients. In addition, the present invention evaluates the contribution of each feature in the prediction model through the Shapley Additive Explanation (SHAP) value to explain the importance of the feature and verify the interpretability of the model.

[0018] Based on the analysis of high-risk plaque characteristics of intracranial atherosclerosis in traditional MRI images (first nested model), the present invention combines some clinical features of diabetes (plaque load and HbA1c value) to form a second nested model for predicting intracranial atherosclerosis in diabetic patients. In view of the subjectivity and less information of traditional image analysis, the combination of the two may ignore some imaging information that is more important for prediction, and does not consider the influence of other clinical factors related to diabetes on the prediction results, which will cause the prediction results of the model to be out of clinical practice or some prediction results to be difficult to explain. To this end, the present invention uses ischemic time as a classification standard, further divides the target population of the training model into a symptomatic group and an asymptomatic group, and uses some diabetes and atherosclerosis risk factors to match the propensity scores of the two groups using a greedy nearest neighbor algorithm to reduce the impact on subsequent imaging genomics feature data mining. By performing imaging genomics feature analysis and mining on the HRMRI images of the two groups of patients, 15 optimal imaging genomics features are screened out through linear regression and LASSO regression and combined with the second nested model to form a final third nested model for predicting symptomatic ICAS in diabetic patients. By using the global interpretation method and the SHAP value to evaluate the contribution of each feature in the final prediction model, it was found that the contribution of the clinical features of diabetes (HbA1c value) and the imaging features after mining and screening ranked high, which is not only consistent with the prediction performance of the model (the performance of the third nested model learning these features is higher than that of the first and second nested models), but also consistent with clinical practice (i.e., individual differences, diversified risk factors, etc.), indicating that the prediction model (third nested model) of the present invention has stronger interpretability and transparency, and better stability. This shows that the prediction model of the present invention can avoid the "black box" problem to a large extent, and is more accurately used to predict the occurrence of new symptomatic intracranial atherosclerotic diseases in diabetic patients, and is suitable for promotion and use in clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 A flow chart of a method for predicting symptomatic ICAS in diabetic patients based on machine learning provided in an embodiment of the present invention; Figure 2 A Manhattan plot for displaying the association p-values ​​between symptomatic plaques and imaging omics features in diabetic patients provided by an embodiment of the present invention; Figure 3An ROC diagram of the diagnostic performance of the model of the present invention in distinguishing symptomatic plaques in diabetic patients provided in an embodiment of the present invention; Figure 4 A schematic diagram of the importance ranking of features for identifying symptomatic ICAS patients provided by an embodiment of the present invention; Figure 5 A schematic diagram of the prediction results of symptomatic ICAS in diabetic patients provided by an embodiment of the present invention; Figure 6 A schematic diagram of the prediction results of asymptomatic ICAS in diabetic patients provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. 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. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0022] The embodiments of the present invention are described below with reference to the accompanying drawings.

[0023] like Figure 1 As shown, the present invention provides a method for predicting symptomatic ICAS in diabetic patients based on machine learning, comprising: S1: Collect HRMRI image samples of the target patient.

[0024] In step S1, eligible patients without contraindications to MRI were scanned using a 3.0T scanner (Philips Healthcare, Best, The Netherlands) equipped with a dedicated 8-channel phased array carotid coil and a 15-channel phased array head coil.

[0025] High-resolution vessel wall MRI was performed using a T1-weighted sequence (volume isotropic fast spin-echo acquisition, VISTA) with the following parameters: repetition time / echo time (TR / TE), 800 / 21 ms; field of view (FOV), 180 mm × 180 mm × 40 mm; flip angle, 90°; acquisition matrix, 300 × 300; number of excitations (NEX), 1; slice thickness, 0.6 mm; and voxel volume, 0.6 mm × 0.6 mm × 0.6 mm.

[0026] Wherein, the target patient in step S1 is a diabetic patient with intracranial atherosclerotic disease who is screened according to a preset first screening factor.

[0027] Wherein, the first screening factor includes: Patients with a history of stroke, cardioembolic stroke, responsible vessel occlusion, non-atherosclerotic intracranial artery stenosis, history of stent implantation or endarterectomy, substandard imaging quality, missing clinical information, or a history of cerebral ischemic events.

[0028] In step S1, we first collected intracranial high-resolution magnetic resonance imaging images of diabetic patients suspected of having intracranial atherosclerotic disease. Patients with a history of stroke, evidence of cardioembolic stroke, occlusion of the responsible vessel, non-atherosclerotic intracranial arterial stenosis (including moyamoya disease, dissection, or vasculitis), a history of stent implantation or endarterectomy, poor image quality, incomplete clinical information, or a cerebral ischemic event 4 weeks ago were excluded.

[0029] Wherein, step S1 also includes: S11: Patients with ischemic events among the target patients are classified into a basic symptomatic group, and patients without ischemic events among the target patients are classified into a basic asymptomatic group. S12: Propensity score matching is performed on the basic symptomatic group and the basic asymptomatic group using a greedy nearest neighbor algorithm to obtain a comparison symptomatic group and a comparison asymptomatic group.

[0030] In order to enhance the comparability of subsequent training data sets, the present invention divides the included population into symptomatic group and asymptomatic group according to whether there is an ischemic event (stroke or transient ischemic attack). Subsequently, in order to minimize the factors that may affect the imaging characteristics, the greedy nearest neighbor algorithm is used for 1:1 propensity score matching, matching age, gender, and traditional risk factors such as hypertension, dyslipidemia, and smoking.

[0031] In a specific embodiment, this embodiment recruited 311 diabetic patients who underwent intracranial high-resolution magnetic resonance imaging due to suspected intracranial atherosclerotic disease, among which the exclusion criteria were patients with a history of previous stroke, cardiogenic stroke, responsible vessel occlusion, non-atherosclerotic intracranial artery stenosis (including moyamoya disease, dissection or vasculitis), a history of stent implantation or endarterectomy, poor image quality, incomplete clinical information, and a cerebral ischemic event 4 weeks ago. The greedy nearest neighbor algorithm was used for 1:1 propensity score matching, matching age, gender, and traditional risk factors such as hypertension, dyslipidemia, and smoking.

[0032] Among them, the diagnostic criteria for diabetes mellitus are that diabetes mellitus can be diagnosed if at least one of the following is met: fasting blood glucose level ≥126 mg / dl (7.0 mmol / L); plasma glucose ≥11.1 mmol / L 2 hours after an oral glucose load of 75 g; random blood glucose ≥200 mg / dl (11.1 mmol / L), accompanied by symptoms of diabetes mellitus; glycosylated hemoglobin ≥6.5% (48 mmol / dl).

[0033] Finally, in this example, a total of 59 diabetic patients with symptomatic intracranial atherosclerotic disease and 59 diabetic patients without symptomatic ICAS were matched for subsequent analysis, and patient information such as age, gender, previous history of hypertension, history of hyperlipidemia, smoking history, duration of diabetes, previous medication history, glycated hemoglobin level, fasting blood glucose level, homocysteine, blood lipid profile, and white blood cell level were subsequently collected from the hospital inpatient information system.

[0034] S2: Performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features.

[0035] After acquiring the HRMRI image of the patient in step S1, in step S2, the image is reconstructed in the cross-sectional view mainly at the maximum stenosis to perform traditional high-risk plaque feature analysis, and the lumen area and the blood vessel area are measured by outlining the lumen and blood vessel boundaries.

[0036] Wherein, the high-risk plaque features in step S2 include: The plaque load at the narrowest part of the blood vessel is expressed as: ; in, To calculate the plaque burden at the narrowest part of the vessel, The area of ​​the blood vessel at its narrowest point, is the lumen area at the narrowest point; Vascular remodeling index, the expression of the vascular remodeling index is: ; in, To calculate the vascular remodeling index, is the vascular area at the site of stenosis, is the vessel area of ​​the reference vessel; Intraplaque hemorrhage, which means a plaque showing a signal intensity 1.5 times higher than that of the adjacent muscle on T1-weighted images; The plaque surface is irregular, which means that there is a discontinuity in the surface of the plaque immediately adjacent to the lumen.

[0037] After outlining the lumen and vascular boundaries, the plaque characteristics can be calculated and analyzed, including plaque burden: calculated by the above formula, reflecting the proportion of the plaque in the vascular wall; vascular remodeling index (RI): determined by dividing the vascular area of ​​the stenotic lesion by the vascular area of ​​the reference vessel, used to evaluate the remodeling of the vascular wall; intraplaque hemorrhage (IPH): on T1-weighted images, plaques that show a signal intensity 1.5 times higher than that of the adjacent muscle are considered to have IPH; plaque surface irregularity: defined as the discontinuity of the plaque adjacent to the luminal surface, which may be related to the stability and rupture risk of the plaque.

[0038] In a specific embodiment provided by the present invention, among diabetic patients, the plaque burden of symptomatic ICAS patients was higher than that of asymptomatic ICAS patients (77.55 ± 11.84 vs 71.95 ± 12.94, p = 0.02), and the median remodeling index of symptomatic ICAS patients was higher than that of asymptomatic ICAS patients (0.96 vs 0.88, p = 0.03). In addition, the proportion of traditional high-risk plaque features was higher in symptomatic ICAS patients, and the specific proportions of the two groups were as follows: intraplaque hemorrhage (32.30% vs 37.29%), surface irregularity (28.81% vs 33.90%), and positive remodeling (30.51% vs 45.76%).

[0039] S3: Extracting and screening radiomics features from the HRMRI image samples to obtain radiomics features.

[0040] Wherein, step S3 further comprises: S31: performing plaque marking on the HRMRI image sample.

[0041] S32: Extracting a volume of interest based on the marked plaque outline.

[0042] S33: Perform radiomics feature extraction on the volume of interest to obtain a first image feature.

[0043] In steps S31 to S33, radiologists first used the open source software ITK-SNAP (version 3.8.0) to annotate plaques without knowing the clinical results of the patients. Then, they precisely delineated the volume of interest (VOI) layer by layer on the T1-VISIA images. Finally, the radiomics features were extracted using the widely used open source Python package PyRadiomics.

[0044] Specifically, seven categories of radiomics features can be extracted from the VOI defined in the T1-VISIA image, including first-order statistical features, shape-based features, gray-level co-occurrence matrix (GLCM) features, gray-level run length matrix (GLRLM) features, gray-level size interval matrix (GLSZM) features, neighborhood gray-level difference matrix (NGTDM) features, and gray-level dependency matrix (GLDM) features. In addition, filters were applied to the original images to generate filtered images, which were also used for feature extraction. Finally, a total of 1,197 features were extracted from the HRMRI images of each patient.

[0045] S34: Screening the first image features to obtain radiomics features.

[0046] Wherein, step S34 further comprises: S341: The second screening factors including plaque burden and glycated hemoglobin are adjusted respectively, and the first image features are screened by linear regression analysis to obtain second image features related to symptomatic plaques.

[0047] Furthermore, linear regression analysis is a statistical method for evaluating the relationship between an independent variable (i.e., the second screening factor described above) and a dependent variable (i.e., the presence or absence of symptomatic ICAS). Through linear regression analysis, it is possible to identify which first image features are significantly associated with the second screening factor, and then to consider that these features are associated with symptomatic plaques. In step S341, the first image features of the initial imaging feature set extracted from the HRMRI image sample are screened based on the linear regression method, and the first image features that are significantly associated with the second screening factor can be screened to form a second image feature set.

[0048] S342: Process the second image feature by using a Z score to obtain a standardized third image feature.

[0049] Furthermore, the Z score (also called the standard score or Z value) is a statistical method used to convert raw data into a standard normal distribution. The Z score processing helps to eliminate the dimensional differences and distribution differences between different features so that they can be compared and analyzed on the same scale.

[0050] In step S342, after Z-score processing, the second image feature is converted into a standardized third image feature, and the standardized feature has better stability and interpretability in subsequent model training.

[0051] S343: Filter the third image feature by using the LASSO operator to obtain radiomics features.

[0052] Furthermore, LASSO (Least Absolute Shrinkage and Selection Operator) is a regression analysis method that simultaneously performs feature selection and parameter estimation by adding an L1 regularization term to the loss function. The L1 regularization term causes the regression coefficients of some features to become zero, thereby achieving the effect of feature selection.

[0053] In step S343, the present invention uses the LASSO operator to screen the standardized third image features. By adjusting the regularization parameter of LASSO, the strictness of feature selection can be controlled. Finally, features with non-zero regression coefficients are screened out, and the selected features are considered to be the imaging features most relevant to symptomatic ICAS.

[0054] In order to study the relationship between symptomatic ICAS and 1197 radiomics features in diabetic patients, the present invention performed a linear regression analysis in step S34, and corresponding S341 to S343, using plaque type (symptomatic plaque and asymptomatic plaque) as a categorical predictor variable, and taking asymptomatic plaque as a reference group.

[0055] In step S341, in order to further control the potential confounding factors affecting plaques, the present invention adjusted the plaque burden and glycosylated hemoglobin (HbA1c) level. Finally, through multivariate linear regression analysis, the present invention identified 58 significant imaging genomic features associated with symptomatic plaques.

[0056] In the subsequent steps S342 to S343, in order to further optimize the selection of imaging features, the 58 selected imaging features were normalized by Z scores, and the least absolute shrinkage and selection operator (LASSO) regression was applied to further select the features. The LASSO algorithm optimizes the log-likelihood function under constraints and reduces the regression coefficient of insignificant covariates to zero, thereby selecting key features to build an imaging model. Finally, LASSO regression selected 15 features from the original 58 features.

[0057] In a specific embodiment of the present invention, Figure 2 As shown, a Manhattan plot is shown showing the p-value of the association between symptomatic plaques and radiomic features in diabetic patients. The p-value is shown on the y-axis and 1197 radiomic parameters are listed on the x-axis. Figure 2 A in the figure represents the unadjusted linear regression analysis results. Figure 2 Middle B shows the results after adjusting for plaque burden and HbA1c values. The points above the red line represent the radiomic features that are significantly associated with symptomatic ICAS in diabetic patients.

[0058] In the specific linear regression analysis results, 165 (13.78%) of the 1197 radiomic features were significantly associated with symptomatic ICAS. Figure 2 Middle A; After adjusting for plaque burden, 66 (5.51%) features, including first-order and texture features (but not shape features), remained statistically different between the symptomatic ICAS group and the asymptomatic ICAS group. After adjusting for HbA1c values, 14.03% (168 / 1197) of the features showed significant differences between the symptomatic ICAS group and the asymptomatic ICAS group. After adjusting for both plaque burden and HbA1c levels, 58 (4.85%) radiomic features were associated with symptomatic ICAS. Figure 2 Middle B. Compared with shape features and first-order features, texture features (GLRLM and GLDM) were more different between the symptomatic ICAS group and the asymptomatic ICAS group. After screening by multivariate linear regression analysis, 58 radiomic features were selected for subsequent analysis. Then, 15 radiomic features with non-zero coefficients were identified using the LASSO regression model.

[0059] S4: Based on the XGBoost algorithm, the high-risk plaque features and the imaging genomics features are trained to obtain a prediction model.

[0060] Due to the high dimensionality of the dataset and the complex interrelationships between plaque features, this study used machine learning (ML) techniques in radiomics analysis. XGBoost is a cutting-edge ensemble boosting machine learning algorithm that iteratively trains a series of weak learners (single-layer decision trees) on a set of patient data or features and eventually combines them into a strong classifier to predict outcomes.

[0061] Wherein, step S4 further comprises: S41: Establish a parameter dictionary and create multiple XGBClassifier() classifiers.

[0062] In step S41, firstly, a XGBClassifier() classifier is created, and a parameter dictionary containing all parameters that need to be input into the classifier needs to be established, where the parameter combination includes 'min_child_weight', 'gamma', 'subsample', 'colsample_bytree' and 'max_depth'.

[0063] S42: Establishing a plurality of data sets based on the high-risk plaque characteristics and the imaging genomics characteristics.

[0064] S43: Perform 10-fold cross validation on multiple XGBClassifier() classifiers through multiple data sets respectively, and adjust the parameters of the training learner to converge the XGBClassifier() classifier to obtain multiple nested models; wherein each set of data sets corresponds to each XGBClassifier() classifier one by one.

[0065] S44: Combine multiple nested models to obtain a prediction model.

[0066] In steps S42 to S44, the GridSearchCV grid search and cross-validation method is used to adjust the parameters in the specified parameter range in sequence according to the step size, and the adjusted parameters are used to train the learner to find the parameters with the highest accuracy on the validation set from all the parameters. The model evaluation is labeled as 'roc_auc', the cross-validation parameter is 10, and the parallel number is -1; the best parameter combination is obtained, and the evaluation indicators corresponding to the training set and the test set are obtained using the best model.

[0067] The present invention trained and tested three machine learning models and adopted 10-fold cross validation with 10 repetitions. The 10-fold cross validation method can ensure that the model has robust performance when dealing with small sample sizes. The specific process is that the data is divided into 10 subsets of equal size to ensure that the proportion of cases in each subset is similar, of which 8 folds (80%) are used for training, 1 fold (10%) is used for adjusting model parameters, and 1 fold (10%) is used for testing.

[0068] Subsequently, stratified sampling is used to maintain the consistency of event distribution in all 10 folds. This process is repeated 10 times, and different data folds are used for model training, adjustment, and testing each time to obtain a more accurate estimate of the model generalization error. Finally, the machine learning scores of all 10 test folds are combined to evaluate the overall performance of the model on the entire data set, that is, to obtain the optimal parameters and obtain a predictive model.

[0069] The nested model in step S43 includes: The first nested model, the first nested model is trained with high-risk plaque characteristics as the data set; the second nested model, the second nested model is trained with high-risk plaque characteristics and the second screening factor as the data set; the third nested model, the first nested model is trained with high-risk plaque characteristics, the second screening factor, and imaging genomics characteristics as the data set.

[0070] In order to evaluate the added value of radiomics features in distinguishing symptomatic and asymptomatic plaques in addition to traditional plaque analysis features, the present invention developed three nested machine learning models based on HRMRI plaque parameters, each of which was used for a binary classification task: diabetic patients with symptomatic plaques and diabetic patients without symptomatic plaques. The specific models include: Model 1: including conventional vulnerable plaque features, including positive remodeling, intraplaque hemorrhage, and plaque surface irregularity; Model 2: adding plaque load and HbA1c value to Model 1; Model 3: adding 15 radiomics features selected by LASSO regression to Model 2.

[0071] In a specific embodiment, the present invention uses the XGBoost algorithm and adopts 10 repeated 10-fold cross validation to develop three nested machine learning models based on HRMRI plaque parameters. The 10-fold cross validation process divides the data into 10 folds: 8 folds (80%) are used for training, 1 fold (10%) is used to adjust model parameters, and 1 fold (10%) is used for testing. Stratified sampling is used to maintain the consistency of event distribution in all 10 folds. This process is repeated 10 times, and the machine learning scores of all 10 test folds are merged to evaluate the overall performance of the model on the entire data set.

[0072] like Figure 3 The figure shows the diagnostic performance of the three models in distinguishing symptomatic plaques in diabetic patients, which is displayed by the receiver operating characteristic curve (ROC curve). The horizontal axis is the false positive rate of the test, the vertical axis is the sensitivity of the model, the middle curve is Model 1, representing Model 1, the lower curve is Model 2, representing Model 2, and the upper curve is Model 3, representing Model 3. In the machine learning model, the plaque burden and HbA1c value are combined with the traditional vulnerable plaque features (surface irregularity, intraplaque hemorrhage and positive remodeling) (Model 2). Compared with the use of vulnerable plaque features (Model 1) alone, the ability to distinguish symptomatic ICAS patients is significantly improved (AUC 0.61, 95% confidence interval [CI] 0.49-0.73), while the AUC of using vulnerable plaque features (Model 1) alone is 0.54 (95% CI 0.44-0.63). The addition of imaging features (Model 3) further improves the differentiation ability, exceeding Model 1 and Model 2 (AUC 0.76, 95% CI 0.67–0.86; p<0.001).

[0073] Wherein, after step S4, the method further includes: A global explanation method is used to provide personalized explanations for each feature of the prediction model.

[0074] After obtaining the model, the present invention also performed feature importance interpretation, that is, the performance of the machine learning model in distinguishing symptomatic ICAS from asymptomatic ICAS was evaluated by the area under the receiver operating characteristic curve (AUC). In addition, the XGBoost algorithm also provided an importance score for each feature, which was determined by the gain of the feature and indicated its relative contribution to the model prediction in the decision tree. In order to enhance clinical interpretability, a summary chart was created to rank the feature importance based on gain to explain the output of the machine learning model.

[0075] Normality of continuous data was assessed using the Shapiro-Wilk test. Continuous data were presented as mean ± standard deviation (SD) or median with interquartile ranges according to normality of residuals. Symptomatic and asymptomatic subjects were compared using t-test or Mann-Whitney U test. Categorical variables were reported as frequency (%). Differences in proportions were assessed using Fisher's exact test or chi-square test. For clinical and traditional plaque parameters, a two-sided p value < 0.05 was considered statistically significant.

[0076] Receiver operating characteristic analysis was used to evaluate the ability of all models to discriminate symptomatic patients. Area under the curve (AUC) values ​​were compared using the method proposed by Delong et al. The accuracy, sensitivity, and specificity were calculated by optimizing the Youden index. All analyses were performed using Python (Python Software Foundation, Python Language Reference, Version 3.9) and R software (R Foundation for Statistical Computing, Vienna, Austria, Version 4.1.0).

[0077] The ranking of feature importance for identifying symptomatic ICAS patients in the final machine learning model is shown in Figure 4 , Figure 4 1 to 20 are the feature importance ranking numbers, specifically 1 is glycated hemoglobin; 2 is wavelet transform HLH, grayscale dimension matrix GLSZM, grayscale non-uniformity; 3 is 3D, , grayscale run matrix GLRLM, long-range emphasis; 4 is wavelet transform HLL, grayscale size matrix GLSZM, large area low grayscale emphasis; 5 is wavelet transform LLH, grayscale dependency matrix GLDM, dependent on non-uniformity; 6 is wavelet transform HLL, grayscale co-occurrence matrix GLCM, inverse difference matrix IDM; 7 is wavelet transform HHH, grayscale size matrix GLSZM, small area low grayscale emphasis; 8 is wavelet transform LHH, grayscale dependency matrix GLDM, dependent on variance; 9 is 3D, , grayscale run matrix GLRLM, short-range emphasis; 10 is wavelet transform LHL, first-order statistical characteristics, robust mean absolute deviation; 11 is positive reshaping; 12 is wavelet transform HLH, grayscale run matrix GLRLM, short-range emphasis; 13 is wavelet transform HHH, first-order statistical characteristics, skewness; 14 is plaque load; 15 is wavelet transform HLH, grayscale run matrix GLRLM, normalized non-uniform run length; 16 is 3D, , gray-level co-occurrence matrix GLCM, contrast inverse variance; 17 is intra-plaque hemorrhage; 18 is 3D, , grayscale dependence matrix GLDM, large area low grayscale emphasis; 19 is wavelet edge switching LHH, grayscale run length matrix GLRLM, long-range emphasis; 20 is the irregular surface of the patch. Figure 4 It can be seen that HbA1c value ranked highest in feature importance, followed by texture and first-order radiomics features, which occupied the top 10 parameters in the overall ranking, while traditional plaque features (such as plaque burden, surface irregularity, intraplaque hemorrhage, and positive remodeling) ranked low among all selected parameters.

[0078] S5: Predicting the onset of symptomatic ICAS in the patient to be predicted by using the prediction model to obtain a prediction result.

[0079] Figure 5 and Figure 6 Two examples of accurate predictions made by the machine learning model are shown, and individualized explanations are provided for the two subjects, with the contribution of each feature to the final prediction score and its specific value detailed in the cases of the two subjects.

[0080] Figure 5 Example of a correct prediction of symptomatic ICAS in a diabetic patient. High-resolution MRI showed symptomatic intracranial atherosclerotic plaques in a 45-year-old male patient with a history of diabetes (HbA1c=8.4%). Figure 5 A in the middle is TOF magnetic resonance angiography (MRA) showing severe stenosis of the left middle cerebral artery M1 segment (arrow, stenosis degree = 85.24%); Figure 5 Middle B is diffusion-weighted imaging (DWI) showing acute infarction in the left basal ganglia region; Figure 5 Middle C and Figure 5 In the middle (D), the T1-VISIA images were reconstructed at the lesion (C) and the reference site (D), respectively. The remodeling index = 1.12, indicating that the plaque was positively remodeled, and the plaque burden = 87.49%; Figure 5E in the middle is an irregular plaque surface shown on the coronal image; in addition, a personalized explanation is generated for this embodiment, specifically intra-plaque hemorrhage: no; irregular plaque surface: yes; positive remodeling: yes; actual patient status: symptomatic ICAS; optimal cutoff value: 0.50; model prediction value and classification result: 0.73, symptomatic ICAS; Figure 5 Figure F is the personalized explanation generated by the machine learning model using the Shapley Additive Explanation (SHAP) value, which details the individual contribution of each feature to the prediction likelihood of symptomatic ICAS, where the features are arranged from large to small in terms of effect size, with positive values ​​to the right of the bar representing a positive contribution and negative values ​​to the left of the bar representing a negative contribution.

[0081] Figure 6 Example of correct prediction of asymptomatic ICAS in a diabetic patient. High-resolution MRI shows asymptomatic intracranial atherosclerotic plaques in a 54-year-old diabetic male patient (HbA1c=5.8%). Figure 6 A in the middle is TOF magnetic resonance angiography (MRA) showing mild stenosis of the left middle cerebral artery M1 segment (arrow, stenosis degree = 79.74%); Figure 6 Middle B: DWI did not detect infarction lesions; Figure 6 Middle C and Figure 6 Middle D is a T1-VISIA image reconstructed at the lesion and reference site, with a remodeling index of 1.10, indicating positive remodeling, and plaque burden of 77.59%; Figure 6 E in the middle is a regular plaque surface displayed on the coronal image; in addition, a personalized explanation is generated for this embodiment, specifically intra-plaque hemorrhage: no; irregular plaque surface: no; positive remodeling: yes; actual patient status: asymptomatic ICAS; optimal cutoff value: 0.50; model prediction value and classification result: 0.45, asymptomatic ICAS; Figure 6 Figure F is a waterfall plot of personalized local explanations for asymptomatic patients. The bar chart shows the order of feature importance, which is measured by the mean absolute Shapley additive explanation (SHAP) value, where positive values ​​to the right of the bar represent positive contributions and negative values ​​to the left of the bar represent negative contributions. The sum of all SHAP values ​​for each local explanation corresponds to the predicted likelihood of an asymptomatic diagnosis.

[0082] The present invention also provides a system for predicting symptomatic ICAS in diabetic patients based on machine learning, comprising: Acquisition module: used for acquiring HRMRI image samples of target patients; wherein the target patients are diabetic patients with intracranial atherosclerotic diseases who are screened according to a preset first screening factor; Analysis module: used for performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features; Extraction module: used for extracting and screening radiomics features from the HRMRI image samples to obtain radiomics features; Training module: used to train the high-risk plaque features and the imaging genomics features based on the XGBoost algorithm to obtain a prediction model; Prediction module: used to predict the onset of symptomatic ICAS for the patient to be predicted by using the prediction model obtained through training in the training module to obtain a prediction result.

[0083] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0084] The present invention provides a method and system for predicting symptomatic ICAS in diabetic patients based on machine learning. A model for predicting new-onset symptomatic intracranial atherosclerotic diseases in diabetic patients is constructed through radiomics analysis and machine learning methods. The risk of symptomatic ICAS in diabetic patients can be individually and accurately assessed. On the basis of the original diagnostic process, intracranial high-resolution vascular wall magnetic resonance images are used as data basis for prediction, without increasing the original economic burden of patients, and is convenient for clinical promotion and applicability. Compared with the prior art, the present invention has more accurate prediction and is more conducive to early detection and treatment of symptomatic intracranial atherosclerotic diseases in diabetic patients.

[0085] The present invention provides a method and system for predicting symptomatic ICAS in diabetic patients based on machine learning. By combining high-risk plaque characteristics and imaging genomics characteristics and training with the XGBoost algorithm, a prediction model capable of predicting symptomatic ICAS (symptomatic intracranial atherosclerotic stenosis) in diabetic patients is obtained, and a variety of factors related to the onset of ICAS are comprehensively considered, thereby improving the accuracy and reliability of the prediction; secondly, before collecting HRMRI image samples, the present invention selects diabetic patients with intracranial atherosclerotic diseases as target patients by screening factors, and further obtains a comparison symptomatic group and a comparison asymptomatic group by propensity score matching, thereby ensuring the homogeneity and representativeness of the research samples and helping to improve the generalization ability of the prediction model; thirdly, in the feature extraction stage, the present invention not only focuses on high-risk plaque characteristics, but also deeply extracts imaging genomics features, and through multiple steps The screening process includes adjusting the screening factors, linear regression analysis, Z-score processing and LASSO operator screening, which effectively removes redundant and irrelevant features, retains the features most relevant to symptomatic plaques, and improves the prediction efficiency of the model; then in the model training stage, the present invention establishes multiple XGBClassifier() classifiers, and trains them through different data sets, including only high-risk plaque features, high-risk plaque features + second screening factors, high-risk plaque features + second screening factors + imaging genomics features, to form a multi-level nested model. This training strategy helps to evaluate the impact of different feature combinations on prediction performance, so as to select the optimal model structure; after the model is built, the present invention evaluates the prediction model through the AUC value corresponding to the target patient, which ensures the accuracy and reliability of the model, helps to verify the performance of the model in practical applications, and adjusts and optimizes it as needed.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting symptomatic ICAS in diabetic patients based on machine learning, characterized in that: include: S1: Collect HRMRI image samples of target patients; The target patients are diabetic patients with intracranial atherosclerotic diseases who are screened according to the preset first screening factor; S2: performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features; S3: extracting and screening radiomics features from the HRMRI image samples to obtain radiomics features; S4: Based on the XGBoost algorithm, the high-risk plaque features and the imaging genomics features are trained to obtain a prediction model; S5: Predicting the onset of symptomatic ICAS in the patient to be predicted by using the prediction model to obtain a prediction result.

2. A method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 1, characterized in that: The first screening factor includes: Patients with a history of stroke, cardioembolic stroke, responsible vessel occlusion, non-atherosclerotic intracranial artery stenosis, history of stent implantation or endarterectomy, substandard imaging quality, missing clinical information, or a history of cerebral ischemic events.

3. The method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 1, characterized in that: Step S1 also includes: S11: The target patients with ischemic events were classified into the basic symptomatic group, and the target patients without ischemic events were classified into the basic asymptomatic group; S12: performing propensity score matching on the basic symptomatic group and the basic asymptomatic group by using a greedy nearest neighbor algorithm to obtain a comparison symptomatic group and a comparison asymptomatic group.

4. The method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 1, characterized in that: The high-risk plaque features in step S2 include: The plaque load at the narrowest part of the blood vessel is expressed as: ； in, To calculate the plaque burden at the narrowest part of the vessel, is the area of ​​the blood vessel at its narrowest point, is the lumen area at the narrowest point; Vascular remodeling index, the expression of the vascular remodeling index is: ； in, To calculate the vascular remodeling index, is the vascular area at the site of stenosis, the vascular area of ​​the reference vessel; Intraplaque hemorrhage, which means a plaque showing a signal intensity 1.5 times higher than that of the adjacent muscle on T1-weighted images; The plaque surface is irregular, which means that there is a discontinuity in the surface of the plaque immediately adjacent to the lumen.

5. The method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 1, characterized in that: Step S3 further comprises: S31: performing plaque marking on the HRMRI image sample; S32: extracting the volume of interest based on the annotated plaque outline; S33: extracting radiomics features from the volume of interest to obtain a first image feature; S34: Screening the first image features to obtain radiomics features.

6. A method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 5, characterized in that: Step S34 further comprises: S341: adjusting the second screening factors including plaque load and glycated hemoglobin respectively, and screening the first image features by linear regression analysis to obtain second image features related to symptomatic plaques; S342: Processing the second image feature by using a Z score to obtain a standardized third image feature; S343: Filter the third image feature by using the LASSO operator to obtain radiomics features.

7. The method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 1, characterized in that: Step S4 further comprises: S41: Establish parameter dictionary and create multiple XGBClassifier() classifiers; S42: establishing multiple data sets based on the high-risk plaque features and the radiomics features; S43: performing 10-fold cross validation on multiple XGBClassifier() classifiers through multiple data sets respectively, and adjusting the parameters of the training learner to converge the XGBClassifier() classifiers to obtain multiple nested models; wherein each set of data sets corresponds to each XGBClassifier() classifier one by one; S44: Combine multiple nested models to obtain a prediction model.

8. The method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 7, characterized in that: The nested model in step S43 includes: A first nested model, wherein the first nested model is trained using high-risk plaque features as a data set; a second nested model, wherein the second nested model is trained using high-risk plaque characteristics and a second screening factor as a data set; The third nested model, the first nested model is trained with high-risk plaque characteristics, the second screening factor, and imaging omics characteristics as data sets.

9. The method for predicting symptomatic ICAS in diabetic patients based on machine learning according to claim 1, characterized in that: After step S4, the method further includes: A global explanation method is used to provide personalized explanations for each feature of the prediction model.

10. A machine learning-based prediction system for symptomatic ICAS in diabetic patients, characterized in that: include: Acquisition module: used to collect HRMRI image samples of target patients; The target patients are diabetic patients with intracranial atherosclerotic diseases who are screened according to the preset first screening factor; Analysis module: used for performing high-risk plaque feature analysis on the HRMRI image sample to obtain high-risk plaque features; Extraction module: used for extracting and screening radiomics features from the HRMRI image samples to obtain radiomics features; Training module: used to train the high-risk plaque features and the imaging genomics features based on the XGBoost algorithm to obtain a prediction model; Prediction module: used to predict the onset of symptomatic ICAS for the patient to be predicted by using the prediction model obtained through training in the training module to obtain a prediction result.

Images