Method for evaluating prognosis of aneurysmal subarachnoid hemorrhage based on dynamic synergy

Abstract

This invention discloses a dynamic and collaborative prognostic assessment method for aneurysmal subarachnoid hemorrhage, comprising the following steps: training five sets of machine learning models based on common assessment features in the training set using a 10-fold cross-validation method to obtain five common prediction models for aneurysmal subarachnoid hemorrhage (FO); training five sets of machine learning models based on personalized assessment features in the training set using a 10-fold cross-validation method to obtain five individual prediction models for FO; dynamically combining the optimal FO common prediction model and the optimal FO individual prediction model using dynamic weights to obtain the optimal FO prediction model; and deploying a prediction tool based on the optimal FO prediction model in a web portal. This invention ensures a synergistic match between the model's generalization and accuracy, establishing a network-based prediction tool to provide clinicians with more convenient and objective assessment and to offer individualized prognostic functional status predictions for each patient.

Description

Technical Field

[0001] This invention relates to the field of prognostic functional assessment technology, specifically to a prognostic assessment method for aneurysmal subarachnoid hemorrhage based on dynamic synergy. Background Technology

[0002] Delayed ischemic attack (DCI) is a major cause of poor functional outcome (FO) after aneurysmal subarachnoid hemorrhage (aSAH). DCI is defined as neurological deterioration between 4 and 21 days after the initial hemorrhage, affecting approximately 30% of aSAH patients. FO in aSAH patients depends largely on the severity of the hemorrhage and complications occurring within 2 weeks. Early prediction of DCI and FO is crucial for the management of aSAH patients.

[0003] In most studies, subarachnoid hematoma volume (quantified using the modified Fisher Scale [mFS]3) and clinical severity at admission (quantified using the Hunt-Hess Scale [HH]4) have been used as effective predictors of DCI and FO, and are widely used to predict DCI. Furthermore, recent studies have found that some CTP parameters are important for predicting DCI and 3-month FO in patients with aSAH.

[0004] Machine learning (ML), a crucial component of artificial intelligence, places greater emphasis on the accuracy of predictions. Compared to traditional statistics, it may reveal previously undetected variables in the data. In recent years, an increasing number of studies have attempted to predict functional prognostic factors (DCI) using clinical data at admission, but few have successfully developed reliable ML prediction models. Current methods for assessing functional prognosis after aneurysmal subarachnoid hemorrhage (aSAH) routinely select characteristic parameters, resulting in rigid prediction models and low accuracy. Furthermore, these abstract prediction models are difficult for medical personnel, except for data processing specialists, to use, thus increasing the demands on healthcare professionals and creating a user-unfriendly experience. Summary of the Invention

[0005] The purpose of this invention is to provide a prognostic assessment method for aneurysmal subarachnoid hemorrhage based on dynamic collaboration, in order to solve the technical problems in the prior art, such as the routine selection of assessment feature parameters, the rigidity of prediction models, resulting in low prediction accuracy, and the increased requirements for medical staff and unfriendly user experience when abstracting into prediction models.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0007] A prognostic assessment method for aneurysmal subarachnoid hemorrhage based on dynamic synergy includes the following steps:

[0008] Step S1: Obtain clinical data and CTP data of patients suspected of having aSAH at the time of admission, as well as data on the probability of delayed cerebral ischemia and the evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage in patients suspected of having aSAH.

[0009] Step S2: Perform Z-score standardization on the clinical data, CTP data, incidence probability data, and evaluation data. Based on the principle that the two sets of datasets are indistinguishable, divide the clinical data, CTP data, incidence probability data, and evaluation data into training set and test set.

[0010] Step S3: Perform SMOTE oversampling technology on the training set, and select common features based on the oversampling data items to obtain common evaluation features, and use the probability of occurrence of delayed cerebral ischemia as a personalized evaluation feature.

[0011] Step S4: Using the 10-fold cross-validation method, a machine learning model is trained on the training set based on common evaluation features to obtain a DCI prediction model, so as to predict the probability of delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage.

[0012] Step S5: Using the 10-fold cross-validation method, train five sets of machine learning models based on common evaluation features in the training set to obtain five common prediction models for FO, so as to achieve common evaluation and prediction of functional prognosis of aneurysmal subarachnoid hemorrhage.

[0013] Step S6: Using the 10-fold cross-validation method, five sets of machine learning models are trained in the training set based on personalized evaluation features to obtain five FO personalized prediction models, so as to realize personalized evaluation and prediction of functional prognosis of aneurysmal subarachnoid hemorrhage.

[0014] Step S7: Precision, sensitivity, specificity, and area under the receiver operating characteristic curve (ROC) are used as model performance metrics. The performance of the five common prediction models and the five individual prediction models are evaluated in the test set. The DeLon test is performed in the test set to compare the model performance metrics and select the best common prediction model and the best individual prediction model for FO.

[0015] Step S8: Dynamically combine the optimal FO commonality prediction model and the optimal FO individuality prediction model using dynamic weights to obtain the optimal FO prediction model.

[0016] Step S9: Based on the optimal FO prediction model, deploy prediction tools in the web portal to enable clinicians to conveniently and objectively predict the probability of delayed cerebral ischemia and the functional prognosis of aneurysmal subarachnoid hemorrhage.

[0017] In a preferred embodiment of the present invention, the CTP data in step S1 is obtained by qualitative and quantitative analysis of whole-brain CTP scans, wherein...

[0018] The qualitative analysis involved radiologists classifying the CTP pseudo-color images in the whole-brain CTP scan as normal perfusion, localized hypoperfusion, or diffuse hypoperfusion.

[0019] The quantitative analysis delineated 32 regions of interest (ROIs) based on the brain's blood supply areas, and obtained the average values ​​of mCBF, mCBV, mMTT, mTTD, mTTS, mTmax, and mFEP of the ROIs from the whole-brain CTP scan.

[0020] As a preferred embodiment of the present invention, the method used to screen out the common evaluation features in step S3 is LASSO.

[0021] As a preferred embodiment of the present invention, in step S1, the evaluation data of functional prognosis of aneurysmal subarachnoid hemorrhage is divided into a common prediction model of good and poor and an optimal FO individual prediction model based on the mRS value of the modified Rankin scale. The evaluation data is obtained by neurosurgeons assessing clinically suspected aSAH patients who are admitted for follow-up examination 3 months later, and clinically suspected aSAH patients who are not followed up are assessed by telephone follow-up.

[0022] As a preferred embodiment of the present invention, the five sets of machine learning models include K-Nearest Neighbors (KNN), Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and CatBoost (CAT).

[0023] As a preferred embodiment of the present invention, the construction of the DCI prediction model includes:

[0024] Common evaluation features are used as input to the machine learning model, and the probability data of delayed cerebral ischemia are used as output to the machine learning model.

[0025] The DCI prediction model was obtained by training a machine learning model on a training set using common evaluation features and the probability data of delayed cerebral ischemia.

[0026] The model expression for the DCI prediction model is as follows:

[0027] P DIC =model DIC (S H );

[0028] In the formula, P DIC This data represents the probability of delayed cerebral ischemia. H To standardize evaluation features, the model DICFor machine learning models.

[0029] As a preferred embodiment of the present invention, the construction of the FO common prediction model includes:

[0030] Common assessment features were used as input to the machine learning model, and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage were used as output.

[0031] The FO common prediction model was obtained by training a machine learning model on the training set with common evaluation features and evaluation data on functional prognosis of aneurysmal subarachnoid hemorrhage.

[0032] The model expression for the FO commonality prediction model is as follows:

[0033] P FOH =model FOH (S H );

[0034] In the formula, model FOH ∈[KNN,LR,SVM,RF,CAT], P FOH S is a data point for evaluating the functional prognosis of aneurysmal subarachnoid hemorrhage. H To standardize evaluation features, the model FOH For machine learning models, KNN, LR, SVM, RF, and CAT represent K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost, respectively.

[0035] As a preferred embodiment of the present invention, the construction of the FO personality prediction model includes:

[0036] Personalized assessment features were used as input to the machine learning model, and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage were used as output.

[0037] The common prediction model of FO was obtained by training a machine learning model on the training set with personalized assessment features and evaluation data on functional prognosis of aneurysmal subarachnoid hemorrhage.

[0038] The model expression for the FO personality prediction model is as follows:

[0039] P FOh =model FOh (S h );

[0040] In the formula, model FOh ∈[KNN,LR,SVM,RF,CAT], P FOh S is a data point for evaluating the functional prognosis of aneurysmal subarachnoid hemorrhage.h For personalized evaluation features, the model FOh For machine learning models, KNN, LR, SVM, RF, and CAT represent K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost, respectively.

[0041] As a preferred embodiment of the present invention, the construction of the optimal FO prediction model includes:

[0042] The dynamic weights of the optimal FO commonality prediction model and the optimal FO individuality prediction model are defined, and the functional expression of the dynamic weights is as follows:

[0043]

[0044]

[0045] In the formula, H t h represents the dynamic weights of the optimal FO commonality prediction model at time t. t Let L be the dynamic weight of the optimal FO personality prediction model at time t, K be the total number of clinically suspected aSAH patients, and L be the dynamic weight of the model. t For clinically suspected aSAH patients whose functional prognostic data for aneurysmal subarachnoid hemorrhage at time t are consistent with the actual situation, N is the total number of time series predicted by the model, t is the model predicted time series representation value, and lenr t For clinically suspected aSAH patients whose functional prognosis data obtained from the optimal FO personalized prediction model at time t is consistent with the actual situation, lenR i For clinically suspected aSAH patients whose functional prognosis data obtained from the optimal FO common prediction model at time t is consistent with the actual situation, L t =lenr t +lenR t ;

[0046] The optimal FO prediction model is obtained by weighting the optimal common FO prediction model and the optimal individual FO prediction model with their dynamic weights. The functional expression of the optimal FO prediction model is as follows:

[0047] P FO,t =H t *best(P FOH,t )+h t *best(P FOh,t );

[0048] In the formula, P FO,tFor the evaluation data of functional prognosis of aneurysmal subarachnoid hemorrhage obtained at time t, best(P) FOH,t The best (P) value represents the evaluation data of functional prognosis for aneurysmal subarachnoid hemorrhage obtained at time t using the optimal FO common prediction model. FOh,t The optimal FO individual prediction model obtains the evaluation data of functional prognosis of aneurysmal subarachnoid hemorrhage at time t.

[0049] As a preferred embodiment of the present invention, the layout prediction tool for web portals based on the optimal FO prediction model includes:

[0050] In the web portal, set up a data input field for common evaluation features, and set up DCI prediction function options and FO prediction function options;

[0051] Link the DCI prediction function option to the DCI prediction model. Selecting the DCI prediction function option triggers the DCI prediction model to retrieve the data of the common assessment features cached in the data input field of the common assessment features. The model calculates the probability data of the patient's delayed cerebral ischemia.

[0052] Link the FO prediction function option to the optimal FO prediction model. Selecting the FO prediction function option triggers the optimal FO prediction model to retrieve the data of the common assessment features cached in the data input field of the common assessment features, as well as the probability data of delayed cerebral ischemia as a personalized assessment feature output by the DCI prediction model. The model calculates and obtains the evaluation data of the functional prognosis of the patient's aneurysmal subarachnoid hemorrhage.

[0053] Compared with the prior art, the present invention has the following advantages:

[0054] This invention significantly improves predictive efficacy compared to previous methods by incorporating CTP data into the predictive model. It can predict the probability of delayed cerebral ischemia in aSAH patients upon admission and the functional prognosis of aneurysmal subarachnoid hemorrhage at 3 months. Furthermore, it uses the probability of delayed cerebral ischemia as a personalized feature to construct the FO predictive model. The optimal FO common prediction model and the optimal FO personalized prediction model are dynamically combined using dynamic weights to obtain the optimal FO predictive model, ensuring a synergistic match between the model's generalization and accuracy. This establishes a network-based predictive tool, enabling clinicians to more conveniently and objectively assess and provide individualized prognostic functional status predictions for each patient. Attached Figure Description

[0055] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0056] Figure 1 A flowchart of the prognostic assessment method for aneurysmal subarachnoid hemorrhage provided in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of CTP qualitative analysis provided in an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of CTP quantitative analysis provided in an embodiment of the present invention;

[0059] Figure 4 This is a schematic diagram illustrating the feature selection process and importance ranking provided in an embodiment of the present invention;

[0060] Figure 5 Box plot of the model performance provided in the embodiments of the present invention;

[0061] Figure 6 A schematic diagram of a web portal for the prediction tool provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Machine learning (ML), a crucial component of artificial intelligence, places greater emphasis on the accuracy of predictions. Compared to traditional statistics, it may reveal previously undetected variables in the data. In recent years, an increasing number of studies have attempted to predict delayed cerebral ischemia (DCI) using admission clinical data, but few have established reliable ML prediction models. Previous studies have also demonstrated that admission CTP data, relative to clinical data, is effective in predicting the probability of delayed cerebral ischemia (DCI) in patients with aneurysmal subarachnoid hemorrhage (aSAH). Therefore, this invention provides a prognostic assessment method for aneurysmal subarachnoid hemorrhage using admission CT perfusion data. The combination of clinical and CTP data can further improve the prediction of DCI and functional prognostic evaluation (FO) data in aSAH patients. By evaluating the efficacy of five ML models, the optimal model is selected. Finally, a prediction tool based on the optimal model is developed and deployed on a webpage for easy and intuitive use by healthcare professionals.

[0064] like Figure 1 As shown, this invention provides a prognostic assessment method for aneurysmal subarachnoid hemorrhage based on dynamic synergy, comprising the following steps:

[0065] Step S1: Obtain clinical data and CTP data of patients suspected of having aSAH at the time of admission, as well as data on the probability of delayed cerebral ischemia and the evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage in patients suspected of having aSAH.

[0066] Step S2: Perform Z-score standardization on the clinical data, CTP data, incidence probability data, and evaluation data. Based on the principle that the two sets of datasets are indistinguishable, divide the clinical data, CTP data, incidence probability data, and evaluation data into training set and test set.

[0067] Step S3: Perform SMOTE oversampling technology on the training set, and select common features based on the oversampling data items to obtain common evaluation features, and use the probability of occurrence of delayed cerebral ischemia as a personalized evaluation feature.

[0068] Step S4: Using the 10-fold cross-validation method, a machine learning model is trained on the training set based on common evaluation features to obtain a DCI prediction model, so as to predict the probability of delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage.

[0069] Step S5: Using the 10-fold cross-validation method, train five sets of machine learning models based on common evaluation features in the training set to obtain five common prediction models for FO, so as to achieve common evaluation and prediction of functional prognosis of aneurysmal subarachnoid hemorrhage.

[0070] Step S6: Using the 10-fold cross-validation method, five sets of machine learning models are trained in the training set based on personalized evaluation features to obtain five FO personalized prediction models, so as to realize personalized evaluation and prediction of functional prognosis of aneurysmal subarachnoid hemorrhage.

[0071] Step S7: Precision, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve were used as model performance metrics. The performance of the five common prediction models and the five individual prediction models were evaluated in the test set. The DeLon test was performed in the test set to compare the model performance metrics and select the optimal common prediction model and the optimal individual prediction model for FO.

[0072] Step S8: Dynamically combine the optimal FO commonality prediction model and the optimal FO individuality prediction model using dynamic weights to obtain the optimal FO prediction model.

[0073] Step S9: Based on the optimal FO prediction model, deploy prediction tools in the web portal to enable clinicians to conveniently and objectively predict the probability of delayed cerebral ischemia and the functional prognosis of aneurysmal subarachnoid hemorrhage.

[0074] Compared to using only traditional hospital admission data for model prediction, this invention incorporates CTP data to increase model accuracy. The steps for obtaining CTP data are as follows:

[0075] like Figure 2 and Figure 3 As shown, the CTP data in step S1 is obtained by qualitative and quantitative analysis of whole-brain CTP scans, wherein...

[0076] The qualitative analysis involved radiologists classifying the CTP pseudo-color images in the whole-brain CTP scan as normal perfusion, localized hypoperfusion, or diffuse hypoperfusion.

[0077] The quantitative analysis delineated 32 regions of interest (ROIs) based on the brain's blood supply areas, and obtained the average values ​​of mCBF, mCBV, mMTT, mTTD, mTTS, mTmax, and mFEP of the ROIs from the whole-brain CTP scan.

[0078] All patients clinically suspected of having aSAH underwent a one-stop CT scan within 24 hours of admission, including plain CT scan and whole brain CTP scan. The clinical and CTP data of the patients at the time of admission were jointly evaluated and recorded by one neurosurgeon and one radiologist.

[0079] The criterion for determining no difference between the two groups in step S2 is: in the inter-group difference test, the p-values ​​for all variables between the two groups are greater than 0.05.

[0080] Considering the slight imbalance in the experimental data, we adopted the SMOTE oversampling technique, and then used LASSO to select features. LASSO is suitable for simplifying model variables, can compensate for the shortcomings of least squares and stepwise regression in local optima estimation, and can effectively select features and solve the multicollinearity problem among features. Therefore, this invention uses LASSO to select common evaluation features. The feature selection process and importance ranking are as follows: Figure 4 As shown.

[0081] In step S1, the evaluation data for the functional prognosis of aneurysmal subarachnoid hemorrhage were categorized into a common prediction model (good and poor) and an optimal FO individual prediction model based on the mRS value in the modified Rankin scale. The evaluation data were obtained by neurosurgeons from clinically suspected aSAH patients who were admitted for follow-up examination 3 months later. Clinically suspected aSAH patients who were not followed up were evaluated by telephone follow-up.

[0082] The five sets of machine learning models include K-Nearest Neighbors (KNN), Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and CatBoost (CAT).

[0083] Since previous studies have shown that CTP at admission relative to clinical data can effectively predict the probability of delayed cerebral ischemia (DCI) in aSAH patients, this invention uses squared assessment features to construct a DCI prediction model, which can predict the probability of delayed cerebral ischemia in patients at admission, allowing for early medical and nursing prevention and improving the timeliness of treatment.

[0084] Specifically, the construction of the DCI prediction model includes:

[0085] Common evaluation features are used as input to the machine learning model, and the probability data of delayed cerebral ischemia are used as output to the machine learning model.

[0086] The DCI prediction model was obtained by training a machine learning model on a training set using common evaluation features and the probability data of delayed cerebral ischemia.

[0087] The model expression for the DCI prediction model is as follows:

[0088] P DIC =model DIC (S H );

[0089] In the formula, P DIC This data represents the probability of delayed cerebral ischemia. H To standardize evaluation features, the model DIC For machine learning models.

[0090] This invention utilizes admission data and CTP data to construct a common prediction model for functional outcomes (FO). By combining common assessment features of CTP data, it can predict the evaluation data of functional prognosis for aneurysmal subarachnoid hemorrhage. This allows for the prediction of whether a patient's functional prognosis for aneurysmal subarachnoid hemorrhage is good or bad upon admission, enabling early assessment and providing patients with treatment expectations. This avoids medical disputes caused by patients having high psychological expectations. Therefore, this invention constructs a common prediction model for FO.

[0091] The FO common prediction model utilizes common features, thus exhibiting high generalization performance and the ability to be applied to a large number of patients. However, it has some shortcomings in individual accuracy. The construction of the FO common prediction model includes:

[0092] Common assessment features were used as input to the machine learning model, and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage were used as output.

[0093] The FO common prediction model was obtained by training a machine learning model on the training set with common evaluation features and evaluation data on functional prognosis of aneurysmal subarachnoid hemorrhage.

[0094] The model expression for the FO commonality prediction model is as follows:

[0095] P FOH =model FOH (S H );

[0096] In the formula, model FOH ∈[KNN,LR,SVM,RF,CAT], P FOH S is a data point for evaluating the functional prognosis of aneurysmal subarachnoid hemorrhage. H To standardize evaluation features, the model FOH For machine learning models, KNN, LR, SVM, RF, and CAT represent K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost, respectively.

[0097] Delayed cerebral ischemia (DCI) is a major cause of poor functional prognosis (FO) after aneurysmal subarachnoid hemorrhage (aSAH). Therefore, there is a correlation between the probability of DCI and the state of functional prognosis. This invention uses the probability of DCI as a personalized assessment feature to construct a common prediction model for FO. By predicting the state of functional prognosis through the personalized assessment feature model, this invention enables the prediction of evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage. This allows for the prediction of whether the patient's functional prognosis of aneurysmal subarachnoid hemorrhage is good or bad at the time of admission, providing patients with treatment expectations in advance and avoiding medical disputes caused by patients having high psychological expectations. Therefore, this invention constructs a personalized prediction model for FO.

[0098] The FO Personal Prediction Model utilizes personalized features, namely, the probability of delayed cerebral ischemia is the feature among all individual characteristics that can be directly used to predict functional prognosis. Therefore, the FO Personal Prediction Model has high accuracy and can be matched to individual patients, but it has some shortcomings in generalization.

[0099] The construction of the FO personality prediction model includes:

[0100] Personalized assessment features were used as input to the machine learning model, and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage were used as output.

[0101] The common prediction model of FO was obtained by training a machine learning model on the training set with personalized assessment features and evaluation data on functional prognosis of aneurysmal subarachnoid hemorrhage.

[0102] The model expression for the FO personality prediction model is as follows:

[0103] P FOh =model FOh (S h );

[0104] In the formula, model FOh ∈[KNN,LR,SVM,RF,CAT], P FOh S is a data point for evaluating the functional prognosis of aneurysmal subarachnoid hemorrhage. h For personalized evaluation features, the model FOh For machine learning models, KNN, LR, SVM, RF, and CAT represent K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost, respectively.

[0105] Due to the insufficient accuracy but strong generalization performance of the common prediction model for FO and the strong accuracy but insufficient generalization performance of the individual prediction model for FO, this invention proposes an optimal combination of the two types of models to achieve a balance between accuracy and generalization performance. This is achieved by setting dynamic weights for the optimal common prediction model and the optimal individual prediction model for FO, and by combining these weights to achieve a balance between accuracy and generalization performance. In the early stages of the prediction model's use, high generalization performance is maintained, while in the later stages, high accuracy performance is maintained.

[0106] In the early stages of model use, considering the significant increase in patient usage and training data (meaning the model encounters more diverse real-world patient data types), strong predictive inclusiveness is required, meaning it should be applicable to a larger number or type of patients. Therefore, the common prediction model of FO (Front-Offset Infarction) is given a high weight to ensure high generalization of the combined optimal FO prediction model. Later in the model's lifespan, as the model has adapted to a wider variety of real-world patient data types, it already possesses high generalization ability. Furthermore, the accuracy of the DCI (Digital Cognitive Impairment) prediction model and the accuracy of deriving functional prognosis from the probability of delayed cerebral ischemia also improve. To further match individualized specificity, the individual prediction model of FO is given a high weight to ensure high individual accuracy of the combined optimal FO prediction model. Therefore, this invention sets the combined weight of the optimal common FO prediction model and the optimal individual FO prediction model as a dynamic weight. In the early stages of model use... t H is close to 0 t It is close to 1, while the later h t Gradually increase to 1, H t The model is gradually reduced to 0, and the matching is adaptively adjusted according to the model's usage cycle to achieve a reasonable balance between generalization and individual accuracy throughout the entire model's usage cycle, thus meeting the needs of real-world scenarios.

[0107] The construction of the optimal FO prediction model includes:

[0108] The dynamic weights of the optimal FO commonality prediction model and the optimal FO individuality prediction model are defined, and the functional expression of the dynamic weights is as follows:

[0109]

[0110]

[0111] In the formula, H t h represents the dynamic weights of the optimal FO commonality prediction model at time t. tLet L be the dynamic weight of the optimal FO personality prediction model at time t, K be the total number of clinically suspected aSAH patients, and L be the dynamic weight of the model. t For clinically suspected aSAH patients whose functional prognostic data for aneurysmal subarachnoid hemorrhage at time t are consistent with the actual situation, N is the total number of time series predicted by the model, t is the model predicted time series representation value, and lenr t For clinically suspected aSAH patients whose functional prognosis data obtained from the optimal FO personalized prediction model at time t is consistent with the actual situation, lenR i For clinically suspected aSAH patients whose functional prognosis data obtained from the optimal FO common prediction model at time t is consistent with the actual situation, L t =lenr t +lenR t ;

[0112] The optimal FO prediction model is obtained by weighting the optimal common FO prediction model and the optimal individual FO prediction model with their dynamic weights. The functional expression of the optimal FO prediction model is as follows:

[0113] P FO,t =H t *best(P FOH,t )+h t *best(P FOh,t );

[0114] In the formula, P FO,t For the evaluation data of functional prognosis of aneurysmal subarachnoid hemorrhage obtained at time t, best(P) FOH,t The best (P) value represents the evaluation data of functional prognosis for aneurysmal subarachnoid hemorrhage obtained at time t using the optimal FO common prediction model. FOh,t The optimal FO individual prediction model obtains the evaluation data of functional prognosis of aneurysmal subarachnoid hemorrhage at time t.

[0115] like Figure 5 As shown, step S7 uses box plots to visually compare the model performance of KNN, LR, SVM, RF, and CAT.

[0116] like Figure 6 As shown, the layout prediction tool for web portals based on the optimal FO prediction model includes:

[0117] In the web portal, set up a data input field for common evaluation features, and set up DCI prediction function options and FO prediction function options;

[0118] Link the DCI prediction function option to the DCI prediction model. Selecting the DCI prediction function option triggers the DCI prediction model to retrieve the data of the common assessment features cached in the data input field of the common assessment features. The model calculates the probability data of the patient's delayed cerebral ischemia.

[0119] Link the FO prediction function option to the optimal FO prediction model. Selecting the FO prediction function option triggers the optimal FO prediction model to retrieve the data of the common assessment features cached in the data input field of the common assessment features, as well as the probability data of delayed cerebral ischemia as a personalized assessment feature output by the DCI prediction model. The model calculates and obtains the evaluation data of the functional prognosis of the patient's aneurysmal subarachnoid hemorrhage.

[0120] This invention provides a predictive example, which continuously collected data from 561 suspected aSAH patients who met the inclusion criteria between July 2020 and August 2022.

[0121] The inclusion criteria for experimental data were as follows: (1) age > 18 years; (2) whole brain CTP performed within 24 hours after symptom onset; (3) digital subtraction angiography confirmed aSAH.

[0122] The exclusion criteria for experimental data are as follows: (1) 215 cases of intracranial hemorrhage caused by other reasons; (2) 44 cases with a history of cerebrovascular disease; (3) 20 patients with poor image quality or incomplete images; (4) 40 patients with incomplete clinical data or who could not be assessed for 3 months of FO.

[0123] All patients received CTP upon admission prior to treatment and then underwent treatment according to current guidelines. Aneurysms were treated within 24 hours of symptom onset. Nimodipine was administered intravenously via a microinfusion pump during hospitalization.

[0124] A total of 242 patients were eventually included. The training set and the test set were randomly assigned in a 4:1 ratio, ensuring that the difference between the groups passed the intergroup test. The training set contained 194 patients, and the test set contained 48 patients.

[0125] The baseline variables used in the model at admission included age, sex, history of hypertension, Glass Coma Scale (GCS) score, World Federation of Neurosurgical Societies (WFNS) classification, Hunt-Hess classification (HH), modified Fisher score (mFS), early cerebral edema score (SEBES), treatment mode [conservative treatment, aneurysm embolization, aneurysm clipping], aneurysm location [anterior circulation, posterior circulation], qualitative CTP analysis [normal perfusion, localized hypoperfusion, diffuse hypoperfusion], and quantitative CTP analysis [mCBF, mCBV, mMTT, mTTD, mTTS, mTmax, mFEP], which were used to build a machine learning model to predict FO and DCI.

[0126] In the selection of common assessment features, the two most important variables for predicting DCI are, in order, mFEP, a parameter reflecting blood-brain barrier permeability in qualitative and quantitative CTP analysis. Among the most important variables for predicting FO, age is first, CTP qualitative analysis is second, and mFEP is fourth. Previous studies have shown that the disruption of blood-brain barrier permeability is reversible, so the importance of mFEP decreases for predicting FO at 3 months. However, overall, the importance of CTP data is self-evident, significantly improving model performance.

[0127] For DCI prediction, the CAT model performed best among the five ML models, showing a 10x CV AUC of 0.935 ± 0.05 on the training set and the highest AUC (0.890) on the test set among the five models.

[0128] For predicting commonalities of FO, the CAT model still performs best, with a 10-fold CV AUC of 0.953±0.05 on the training set and the highest AUC (0.897) on the test set among the five models. When the incidence of DCI is used as an evaluation feature, the CAT model still performs best among the five models for predicting FO personality, with a 10-fold CV AUC of 0.958±0.03 on the training set and the highest AUC (0.911) on the test set, indicating higher individual accuracy in predicting FO personality.

[0129] Ultimately, a predictive tool was established based on common predictions and individual predictions (optimal FO common prediction model and optimal FO individual prediction model) based on the CAT model. This tool can be used to predict the DCI of aSAH patients upon admission and their prognostic functional status 3 months later.

[0130] This invention significantly improves predictive efficacy compared to previous methods by incorporating CTP data into the predictive model. It can predict the probability of delayed cerebral ischemia in aSAH patients upon admission and the functional prognosis of aneurysmal subarachnoid hemorrhage at 3 months. Furthermore, it uses the probability of delayed cerebral ischemia as a personalized feature to construct the FO predictive model. The optimal FO common prediction model and the optimal FO personalized prediction model are dynamically combined using dynamic weights to obtain the optimal FO predictive model, ensuring a synergistic match between the model's generalization and accuracy. This establishes a network-based predictive tool, enabling clinicians to more conveniently and objectively assess and provide individualized prognostic functional status predictions for each patient.

[0131] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A prognostic assessment method for aneurysmal subarachnoid hemorrhage based on dynamic synergy, characterized in that: Includes the following steps: Step S1: Obtain clinical data and CTP data of patients suspected of having aSAH at the time of admission, as well as data on the probability of delayed cerebral ischemia and the evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage in patients suspected of having aSAH. Step S2: Perform Z-score standardization on the clinical data, CTP data, incidence probability data, and evaluation data. Based on the principle that the two sets of datasets are indistinguishable, divide the clinical data, CTP data, incidence probability data, and evaluation data into training set and test set. Step S3: Perform SMOTE oversampling technology on the training set, and select common features based on the oversampling data items to obtain common evaluation features, and use the probability of occurrence of delayed cerebral ischemia as a personalized evaluation feature. Step S4: Using the 10-fold cross-validation method, a machine learning model is trained on the training set based on common evaluation features to obtain a DCI prediction model, so as to predict the probability of delayed cerebral ischemia after aneurysmal subarachnoid hemorrhage. Step S5: Using the 10x cross-validation method, train five sets of machine learning models on the training set based on common evaluation features to obtain five common prediction models P for FO. FOH To achieve a common evaluation and prediction of functional prognosis in aneurysmal subarachnoid hemorrhage; Step S6: Using the 10x cross-validation method, train five sets of machine learning models based on personalized evaluation features in the training set to obtain five FO personalized prediction models P. FOh To achieve personalized evaluation and prediction of functional prognosis in aneurysmal subarachnoid hemorrhage; Step S7, precision, sensitivity, specificity, and the area under the receiver operating characteristic (ROC) curve were used as model performance metrics to evaluate the five common prediction models P for FO on the test set. FOH Performance and five FO personality prediction models P FOh To assess the performance of the models, a DeLong test was performed on the test set to compare the model performance metrics and select the optimal FO common prediction model and the optimal FO individual prediction model. Step S8: Dynamically combine the optimal FO commonality prediction model and the optimal FO individuality prediction model using dynamic weights to obtain the optimal FO prediction model. Step S9: Based on the optimal FO prediction model, a prediction tool is placed in the web portal to provide clinicians with convenient and objective prediction of the probability of delayed cerebral ischemia and the functional prognosis of aneurysmal subarachnoid hemorrhage. The common prediction model P of FO FOH The model was trained on the training set using common evaluation features and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage. The FO personality prediction model P FOh The model was trained on the training set using personalized assessment features and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage. The construction of the optimal FO prediction model includes: The dynamic weights of the optimal FO commonality prediction model and the optimal FO individuality prediction model are defined, and the functional expression of the dynamic weights is as follows: ； ； In the formula, H t h represents the dynamic weights of the optimal FO commonality prediction model at time t. t Let L be the dynamic weight of the optimal FO personality prediction model at time t, K be the total number of clinically suspected aSAH patients, and L be the dynamic weight of the model. t For clinically suspected aSAH patients whose functional prognostic data for aneurysmal subarachnoid hemorrhage at time t are consistent with the actual situation, N is the total number of time series predicted by the model, t is the model-predicted time series representation value, and lenr t For clinically suspected aSAH patients whose functional prognosis data obtained from the optimal FO personalized prediction model at time t is consistent with the actual situation, lenR t For clinically suspected aSAH patients whose functional prognosis data obtained from the optimal FO common prediction model at time t is consistent with the actual situation, L t =lenr t +lenR t ; The optimal FO prediction model is obtained by weighting the optimal common FO prediction model and the optimal individual FO prediction model with their dynamic weights. The functional expression of the optimal FO prediction model is as follows: P FO,t =H t * best(P FOH,t )+h t * best(P FOh,t ); In the formula, P FO,t For the evaluation data of functional prognosis of aneurysmal subarachnoid hemorrhage obtained at time t, best(P) FOH,t The optimal FO common prediction model provides evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage at time t. FOh,t The optimal FO individual prediction model provides evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage at time t.

2. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 1, characterized in that: In step S1, the CTP data is obtained through qualitative and quantitative analysis of whole-brain CTP scans. The qualitative analysis involved radiologists classifying the CTP pseudo-color images in the whole-brain CTP scan as normal perfusion, localized hypoperfusion, or diffuse hypoperfusion. The quantitative analysis delineated 32 regions of interest (ROIs) based on the brain's blood supply areas, and obtained the average values ​​of mCBF, mCBV, mMTT, mTTD, mTTS, mTmax, and mFEP of the ROIs from the whole-brain CTP scan.

3. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 1, characterized in that: The method used to screen out the common evaluation features in step S3 is LASSO.

4. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 1, characterized in that: In step S1, the evaluation data for the functional prognosis of aneurysmal subarachnoid hemorrhage were categorized into good and poor optimal FO common prediction models and optimal FO individual prediction models based on the mRS value in the modified Rankin scale. The evaluation data were obtained by neurosurgeons from clinically suspected aSAH patients who were admitted for follow-up examination 3 months later, and clinically suspected aSAH patients who were not followed up were evaluated by telephone follow-up.

5. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 1, characterized in that: The five sets of machine learning models include K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost.

6. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 1, characterized in that: The construction of the DCI prediction model includes: Common evaluation features are used as input to the machine learning model, and the probability data of delayed cerebral ischemia are used as output to the machine learning model. The DCI prediction model was obtained by training a machine learning model on a training set using common evaluation features and the probability data of delayed cerebral ischemia. The model expression for the DCI prediction model is as follows: P DIC =model DIC (S H ); In the formula, P DIC This data represents the probability of delayed cerebral ischemia. H To standardize evaluation features, the model DIC For machine learning models.

7. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 5, characterized in that, The common prediction model P of FO FOH The construction includes: Common assessment features were used as input to the machine learning model, and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage were used as output. The FO common prediction model P was obtained by training a machine learning model on the training set with common assessment features and evaluation data on functional prognosis of aneurysmal subarachnoid hemorrhage. FOH ; The common prediction model P of FO FOH The model expression is: P FOH =model FOH (S H ); In the formula, model FOH ∈[KNN, LR, SVM, RF, CAT], P FOH S is a data point for evaluating the functional prognosis of aneurysmal subarachnoid hemorrhage. H To standardize evaluation features, the model FOH For machine learning models, KNN, LR, SVM, RF, and CAT represent K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost, respectively.

8. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 7, characterized in that, The FO personality prediction model P FOh The construction includes: Personalized assessment features were used as input to the machine learning model, and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage were used as output. The FO personalized prediction model P was obtained by training a machine learning model on the training set using personalized assessment features and evaluation data on the functional prognosis of aneurysmal subarachnoid hemorrhage. FOh ; The FO personality prediction model P FOh The model expression is: P FOh =model FOh (S h ); In the formula, model FOh ∈[KNN, LR, SVM, RF, CAT], P FOh S is a data point for evaluating the functional prognosis of aneurysmal subarachnoid hemorrhage. h For personalized evaluation features, the model FOh For machine learning models, KNN, LR, SVM, RF, and CAT represent K-nearest neighbors, logistic regression, support vector machine, random forest, and CatBoost, respectively.

9. The method for prognostic assessment of aneurysmal subarachnoid hemorrhage based on dynamic synergy according to claim 1, characterized in that, The layout prediction tool for web portals based on the optimal FO prediction model includes: In the web portal, set up a data input field for common evaluation features, and set up DCI prediction function options and FO prediction function options; Link the DCI prediction function option to the DCI prediction model. Selecting the DCI prediction function option triggers the DCI prediction model to retrieve the data of the common assessment features cached in the data input field of the common assessment features. The model calculates the probability data of the patient's delayed cerebral ischemia. Link the FO prediction function option to the optimal FO prediction model. Selecting the FO prediction function option triggers the optimal FO prediction model to retrieve the data of the common assessment features cached in the data input field of the common assessment features, as well as the probability data of delayed cerebral ischemia as a personalized assessment feature output by the DCI prediction model. The model calculates and obtains the evaluation data of the functional prognosis of the patient's aneurysmal subarachnoid hemorrhage.

Images