CN108053885A - A kind of hemorrhagic conversion forecasting system - Google Patents

Abstract

The invention discloses a kind of hemorrhagic conversion forecasting systems, belong to field of medical technology；System includes：Acquiring unit, for obtaining a plurality of training patient data；Model generation unit is used to be used for the prediction model for predicting hemorrhagic conversion with patient data generation one according to the training of a plurality of acquisition, and model generation unit further comprises：Feature selection module, for being made choice to training with the training in patient data with state of an illness feature；Tagsort module, for carrying out tagsort with state of an illness feature to selected training；Model training module, for training to form prediction model with state of an illness feature according to the training by classification；Collecting unit, for collecting actual patient data；Predicting unit, for actual patient data to be sent into the prediction model that training is formed, to export corresponding prediction result.The advantageous effect of above-mentioned technical proposal is：Reduce hemorrhagic conversion the occurrence of probability, so as to reduce clinical risk and corresponding medical expense.

Description

A Prediction System for Hemorrhagic Transformation

technical field

The invention relates to the field of medical technology, in particular to a hemorrhagic transformation prediction system.

Background technique

Cerebral infarction is one of the major public health problems in the world, with high morbidity, mortality, disability and recurrence, and limited clinical treatment methods. Among them, intravenous thrombolytic therapy is a major breakthrough in the clinical treatment of ischemic stroke in the past 20 years, which can effectively reduce the death and disability rate. However, intravenous thrombolytic therapy is a high-risk treatment, and it may be accompanied by symptoms of hemorrhagic transformation after thrombolysis. Some bleeding can aggravate neurological damage, and even endanger life. One of the important reasons for further promotion. Specifically, in the clinical treatment process, intravenous thrombolysis has a strict time limit. Only patients with acute cerebral infarction who meet the conditions for intravenous thrombolysis within 4.5 hours of onset can receive this treatment. The so-called "qualified for intravenous thrombolysis" means The risk of hemorrhagic transformation is low or almost no risk of hemorrhagic transformation. Therefore, the prediction of the risk of hemorrhagic transformation is directly related to the communication between doctors and patients and the decision-making of treatment.

In the prior art, the prediction of hemorrhagic transformation is usually determined based on some relevant clinical studies, such as age, blood sugar, clinical neurological function at the time of onset, and the like. At the same time, there will be related clinical scoring tables to assist clinicians in predicting the possibility of hemorrhagic transformation, such as hemorrhage score after thrombolysis, multicenter stroke survey prediction score, SITS score, GRASPS score, and SEDAN score. However, these scoring tables are basically obtained based on the doctor's diagnosis and treatment experience or using simple logistic regression, and the use value and corresponding accuracy are yet to be tested. You can refer to the article "Comparison of five prediction models in the application of bleeding prediction after thrombolysis in Chinese population" to compare the performance of different scoring models. Among them, the GRASPS score has the best performance, but the AUC value corresponding to its receiver operating characteristic curve (ROC) is only 0.7056, and its working performance still needs to be improved. Moreover, the above-mentioned clinical prompts are relatively independent, and the scoring table is relatively simple. Basically, they only cover a few clinical factors that may have a greater impact on the prognosis, and do not comprehensively evaluate the clinically relevant factors of the patient, and lack individualized evaluation. .

Contents of the invention

According to the problems existing in the prior art, a technical solution of a hemorrhagic transformation prediction system is now provided, which aims to reduce the occurrence probability of hemorrhagic transformation during intravenous thrombolytic therapy, thereby reducing clinical risks and corresponding medical expenses.

The above-mentioned technical solutions specifically include:

A hemorrhagic transformation prediction system, including:

An acquisition unit, configured to acquire multiple pieces of patient data for training, each piece of patient data for training includes multiple disease characteristics for training;

A model generation unit, connected to the acquisition unit, is used to generate a prediction model for predicting hemorrhagic transformation according to the plurality of obtained training patient data, and the model generation unit further includes:

A feature selection module, configured to select the disease features for training in the patient data for training;

A feature classification module, connected to the feature selection module, for feature classification of the selected disease features for training;

A model training module, connected to the feature classification module, for training and forming the prediction model according to the classified training disease features;

an acquisition unit, configured to acquire actual patient data;

The prediction unit is connected to the acquisition unit and the model generation unit respectively, and is used to send the actual patient data into the prediction model formed by training, so as to output a corresponding prediction result.

Preferably, the hemorrhagic transformation prediction system, wherein the feature selection module further includes:

The first feature selection component is used to select the disease features for training by using the CM feature selection method;

The second feature selection component is used to select the disease features for training by adopting the packaging model feature selection method;

The third feature selection component is used to select the disease features for training by adopting a filter model feature selection method;

A selection control part is connected to the first feature selection part, the second feature selection part and the third feature selection part, and is used to select and enable the first feature selection part according to the correspondence between the training disease features. A feature selection component or said second feature selection component or said third feature selection component.

Preferably, in the hemorrhagic transformation prediction system, the feature classification module uses a random forest model to classify the disease features for training.

Preferably, in the hemorrhagic transformation prediction system, the feature classification module uses a support vector machine to classify the disease features for training.

Preferably, in the hemorrhagic transformation prediction system, the feature classification module uses Logistic regression or perceptron to classify the disease features for training.

Preferably, in the hemorrhagic transformation prediction system, the feature classification module uses the AdaBoost algorithm to classify the disease features for training.

Preferably, the hemorrhagic transformation prediction system further includes:

A data processing unit, connected between the acquisition unit and the model generation unit, for performing preset processing on the patient data for training, so as to achieve data balance of the patient data for training;

The preset processing is: processing the patient data for training by means of oversampling and/or multivariate support vector machine algorithm.

Preferably, the hemorrhagic transformation prediction system further includes:

A data processing unit, connected between the acquisition unit and the model generation unit, for performing preset processing on the patient data for training, so as to achieve data balance of the patient data for training;

The preset processing is: processing the training patient data in an over-sampling manner; and/or

processing said training patient data with a cost-sensitive loss function; and/or

The training patient data is processed using a cost sensitive learning rate.

Preferably, the hemorrhagic transformation prediction system, wherein the model generation unit further includes:

The risk rating module is connected to the model training module, and the risk rating module performs risk level discretization processing according to the obtained patient data for training to form a set of reference discrete points for risk rating as the model training module training Reference data when forming the predictive model.

The beneficial effect of the above technical solution is to provide a predictive system for hemorrhagic transformation, which can reduce the occurrence probability of hemorrhagic transformation during intravenous thrombolytic therapy, thereby reducing clinical risks and corresponding medical expenses.

Description of drawings

Figure 1 is a schematic diagram of the overall structure of a hemorrhagic transformation prediction system in a preferred embodiment of the present invention;

Fig. 2 is in the preferred embodiment of the present invention, the specific structural representation of feature selection module;

Fig. 3 is a schematic diagram of the specific working principle of the feature selection module in a preferred embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

It should be noted that, in the case of no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but not as a limitation of the present invention.

With the continuous development of science and technology, some machine learning algorithms have also been applied to the process of predicting hemorrhagic transformation, that is, using some machine learning algorithms to realize the prediction of hemorrhagic transformation, identifying Patterns that are difficult for humans to recognize. The existing risk prediction of thrombolytic hemorrhage only input the database corresponding to thrombolytic therapy and hemorrhagic transformation into a fixed machine learning model, without improving the algorithm according to the characteristics of medical data, without considering the imbalance of data, data The possible correlation between features and the impact of the doctor's prior diagnosis on the data set, so the prediction level of the model is relatively general. For example, in the research of H.Asadi et al. % But its recall rate is close to zero, and the predicted AUC value is only about 0.6. The design systems of these prediction models are all fixed, and the parameters are all based on the training of the data in the previous database, and there is no self-renewal function designed for the newly entered patient data. Over time, the accuracy of these systems' predictions degrades significantly.

Based on the above-mentioned problems in the prior art, a hemorrhagic transformation prediction system is now provided, which is applied to predict the hemorrhagic transformation that may occur during intravenous thrombolytic therapy.

Specifically, the above-mentioned hemorrhagic transformation prediction system is specifically as described in Figure 1, including:

An acquisition unit 1, configured to acquire multiple pieces of patient data for training, each piece of patient data for training includes multiple disease characteristics for training;

The model generation unit 2 is connected to the acquisition unit 1, and is used to generate a prediction model for predicting hemorrhagic transformation according to multiple obtained training patient data, and the model generation unit 2 further includes:

The feature selection module 21 is used to select the disease features for training in the patient data for training;

The feature classification module 22 is connected to the feature selection module 21, and is used for feature classification of the selected training disease features;

The model training module 23 is connected to the feature classification module 22, which is used to form a predictive model according to the classified training with disease feature training;

The acquisition unit 3 is used to acquire actual patient data;

The prediction unit 4 is respectively connected to the acquisition unit 3 and the model generation unit 2, and is used to send actual patient data into the prediction model formed by training, so as to output corresponding prediction results.

Specifically, in this embodiment, the acquisition unit 1 may be connected to an external database, or remotely connected to a server, and obtain pre-prepared patient data for training through the database or the remote server. Of course, the above patient data for training can also be directly input into the acquisition unit 1 by way of setting by the user.

In this embodiment, the patient data for training includes a plurality of disease features for training, and these disease features for training are features with clear physical meanings, so there is no need for the system to perform feature extraction from the patient data for training. The so-called features with clear physical meaning can be based on some basic information in the medical records of some patients in historical records, such as age and gender, etc., some inspection information for preoperative physical examination, and whether there is bleeding after intravenous thrombolytic therapy Information about the transformation status (no hemorrhagic transformation, slight hemorrhage or severe hemorrhage), etc.

In this embodiment, after the acquisition unit 1 acquires the data, it needs to preprocess the data. Specifically, it needs to filter out missing or obviously erroneous data items in the training patient data, and standardize the continuous data. Further, the above-mentioned data preprocessing process can be performed manually by the user, or can be automatically performed by the system according to some preset screening rules, such as filling templates with preset training patient data data, and according to the templates, training patients Match the data to determine whether there is data missing in the training patient data, and match it according to the data format of different filling bits in the template to determine whether there is an obvious data error in the training patient data.

In this embodiment, in the above-mentioned model generation unit 2, before performing training and generation of the prediction model, it is first necessary to select features. The reason why feature selection is needed is that in the case of different data set sizes, feature dimensions, and feature attributes, different training disease feature selection frameworks will have different performance and their respective suitable environments, so it is necessary to train the model The features are selected before to place different features into the most suitable feature selection framework to reflect their best test results. The specific feature selection method will be described in detail below.

In this embodiment, after feature selection, the feature classification module 22 needs to be used to classify the selected disease features for training. The above-mentioned feature classification module 22 can be realized by a classifier, and the disease features for training after feature classification can be applied to the process of model training.

In this embodiment, a method similar to that in the prior art is adopted to obtain a corresponding prediction model according to feature training, which will not be repeated here.

In this embodiment, the trained prediction model can be applied to the actual process of predicting hemorrhagic transformation. Specifically, according to the input requirements of the prediction model, the actual data of the patient are collected and sent into the prediction model. After the prediction of the prediction model, the probability of hemorrhagic transformation of the patient after intravenous thrombolytic therapy can be obtained. forecast result. Doctors can use the predicted results as reference information to communicate with patients and formulate relevant diagnosis and treatment plans, thereby reducing clinical risks and saving medical expenses.

In a preferred embodiment of the present invention, as shown in Figure 2, the feature selection module 21 further includes:

The first feature selection component 211 is used to select the disease feature for training using the CM feature selection method;

The second feature selection component 212 is used to select the disease features for training by adopting the package model feature selection method;

The third feature selection component 213 is used to select the disease features for training by adopting the filter model feature selection method;

The selection control part 214 is connected with the first feature selection part 211, the second feature selection part 212 and the third feature selection part 213 respectively, for selecting and enabling the first feature selection part 211 or the third feature selection part 211 according to the corresponding relationship between the training disease features. The second feature selection component 212 or the third feature selection component 213 .

In this embodiment, the first feature selection component 211, the second feature selection component 212, and the third feature selection component 213 respectively represent three different feature selection frameworks of the system, which can be automatically trained and tested by a computer system .

Specifically, the first feature selection component 211 uses a CM feature selection method (Conservative Mean Feature Selection) to select disease features for training. The CM feature selection method mainly selects a single feature, which provides a solution for sampling to improve the stability of feature selection. Specifically, the CM feature selection method utilizes the characteristic that the AUC value does not change in the case of monotonic function mapping, and uses K-fold validation to calculate the AUC value for a specific feature and classification result. Then, for these K AUC values, calculate their mean value μ and standard deviation α. Finally, by comparing the value of (μ-α) to select the best multiple training disease features to form a feature subset.

The above-mentioned second feature selection component 212 selects the disease features for training by adopting the wrapper model feature selection method (Wrapper). The third feature selection component 213 uses a filter model feature selection method (Filter) to select disease features for training. In these two feature selection methods, what needs to be considered is their evaluation function and search algorithm. In this technical solution, for these two feature selection methods, multiple search algorithms including forward search, reverse search, genetic algorithm and exhaustive search are provided. And for the feature selection method of the package model, the AUC value output by the CFS framework (Correlation-based feature Selection) learning algorithm is used as its evaluation function. For the filter model feature selection method, symmetrical uncertainty, RELIEF and minimum description length are used as its evaluation function.

In this embodiment, a selection control component 214 is used to control the operation of the above three feature selection frameworks. Specifically, as shown in Figure 3, the selection control component 214 first judges according to the association relationship between the training disease features:

1) If the correlation is relatively simple, then the selection control part 214 directly selects and enables the first feature selection part 211, that is, adopts the CM feature selection method to select the disease feature for training;

2) If the association relationship is relatively complicated, the selection control component 214 selects the other two traditional feature selection frameworks. Further, if enough data has been obtained, the selection control component 214 selects and enables the second feature selection component 212, that is, adopts the encapsulation model feature selection method to select the disease features for training;

3) If the amount of acquired data is small, the selection control unit 214 selects and activates the third feature selection unit 213, that is, selects the disease features for training by adopting a filtering model feature selection method.

In a preferred embodiment of the present invention, the feature classification module 22 uses a random forest model to classify the disease features for training.

Specifically, for the random forest model, the RandomForestClassifier in the Scikit-learn learning library can be directly used in the system. Because the decision tree of each lesson of the random forest is equivalent to feature selection during the training process, there is no need to consider the feature selection algorithm in the process of feature classification using the random forest model. In the process of cross validation (cross validation), directly perform grid search on the parameters that need to be adjusted, and finally determine the value of each parameter under the condition that the AUC value is optimal.

In a preferred embodiment of the present invention, the feature classification module 22 uses a support vector machine to classify the disease features for training.

Specifically, for a traditional support vector machine (Support Vector Machine, SVM), the libsvm library under Python 3.6 is used. Among them, the AUC value used to measure performance considers the relationship between the distance value of the data point to the optimal hyperplane and the final classification.

For Multivariate SVM, the AUC value can be used as the loss function between the predicted label and the real label. The kernels used are all linear kernels, and the processing of oversampling is not considered, and the svm-perf library written in C language is used.

In the actual processing process, different types of support vector machines, corresponding feature selection algorithms, and data equalization processing methods (described in detail below) can be selected according to the actual situation, and the current best SVM model can be output.

In a preferred embodiment of the present invention, the feature classification module 22 uses Logistic regression or a perceptron to classify the disease features for training.

Specifically, when using the Logistic regression method, different feature selection schemes and data equalization processing methods need to be compared during the training process (details will be described below). In this technical solution, the Theano framework under Python3.6 is used for Logistic regression.

When using the perceptron method, in terms of perceptron design, feature selection and proper nonlinear design of single hidden layer perceptron can make the perceptron model have a better fitting effect on the current data set. And considering the limited size of the current patient data set for training, the number of hidden layers of the perceptron should not be too large, otherwise it will lead to more errors.

For some hyperparameters of the perceptron, cross-validation can be used to determine. Furthermore, an appropriate feature selection and data equalization processing method can be selected according to the actual situation (details will be described below). Similarly, in this technical solution, the Theano framework under Python 3.6 is also used for the perceptron.

In a preferred embodiment of the present invention, the feature classification module 22 uses the AdaBoost algorithm to classify the disease features for training.

Specifically, in the AdaBoost algorithm, each weak classifier is designed as a relatively simple perceptron model. The number of weak classifiers can be determined by cross-validation. The above-mentioned AdaBoost algorithm is also mainly realized through the Theano framework under Python3.6.

In a preferred embodiment of the present invention, still as shown in Figure 1, the above-mentioned hemorrhagic transformation prediction system specifically further includes:

The data processing unit 5 is connected between the acquisition unit 1 and the model generation unit 2, and is used for performing preset processing on the patient data for training, so as to accomplish the following goals: screening patient data with obvious errors, filling missing data, realizing Data equalization of patient data for training.

Specifically, in terms of missing input features, for the existing multi-data center datasets, there is a certain amount of data missing because the features recorded in each data center are emphasized. The missing processing scheme adopted in the present invention is a missing indicate scheme, so as to avoid negative impacts on the accuracy of the data set caused by the filling of the mean or the filling of the median. In terms of target classification balance, for the existing patient data, the proportion of the training samples with hemorrhagic transformation and the training samples without hemorrhagic transformation is very unbalanced, and the samples with symptomatic bleeding may only account for the unbalanced proportion. About 1/20 of the bleeding samples occurred, resulting in a problem of data imbalance in the data set of the entire training sample. Therefore, the system needs to perform some data equalization processing on the data set to avoid the problem of inaccurate output of the final prediction model caused by using an unbalanced data set to train the model.

Further, for the different feature classification methods (i.e. different classifiers) adopted by the feature classification module 22, the above-mentioned preset processing (i.e. the data equalization processing method adopted by the system) is also different, specifically:

1) When the feature classification module 22 uses a support vector machine to classify features, the preset processing adopted in the above-mentioned data processing unit 5 is as follows:

① Oversampling, using the oversampling method to perform data equalization processing on the training patient data. Specifically, a series of few-class samples (ie, patient data for training with symptomatic bleeding, which will not be described in detail below) are randomly sampled, so that the number of samples corresponding to different types is similar.

②Multivariate support vector machine algorithm was used for classification, and the AUC value was used instead of the error rate to update the support vector machine.

2) When the feature classification module 22 uses Logistic regression or a perceptron to classify features, the preset processing adopted in the above-mentioned data processing unit 5 is as follows:

①Using oversampling to perform data equalization processing on the training patient data. Specifically, a series of few-class samples are randomly sampled so that the number of samples corresponding to different types is similar.

②A cost-sensitive loss function is used to replace the general loss function used in the system to increase the punishment for misjudgment of few-class samples.

③Use the cost-sensitive learning rate to replace the general learning rate used in the system, so that the learning rate for few-class samples is higher, and for multi-class samples (that is, patient data for training without symptomatic bleeding, which will not be described in detail below) The learning rate is lower. In this case, the parameters of the model are updated with a larger step size for few-class samples than for multi-class samples.

In a preferred embodiment of the present invention, still as shown in Figure 1, the above-mentioned model generation unit 2 also includes:

The risk rating module 24 is connected to the model training module 23, and the risk rating module 24 performs risk level discretization processing according to the obtained training patient data to form a set of reference discrete points for risk rating, which are used as the model training module training to form the prediction model. reference data.

Specifically, the system provides two solutions for the discretization of risk levels:

When the data in the system database is relatively scarce and the data quality is not very ideal, the system uses an unsupervised equal-frequency discretization scheme. The discretization of equal frequency numbers acts on the training set, and a set of discrete points of continuous data can be obtained. This group of discrete points is the reference point for subsequent risk rating for patients;

When the data in the system database is relatively sufficient and the data quality can be guaranteed to a certain extent, the system uses a supervised discretization scheme to minimize information entropy. This scheme acts on the training set, and through the sum of the information entropy of different groups in the training set, the difference in the ratio of bleeding samples to the total samples between different groups is as large as possible, and then a set of risk factors for new patients is generated. Discrete points for grading references.

The effectiveness of the classification of the above two wind direction grade discretization schemes can be verified by the Wilcoxon signed rank test (Wilcoxon rank sum test). The system selects the discrete point generated by the scheme whose Z value deviates more from the origin during the inspection process as the reference point for the prediction part.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the implementation and protection scope of the present invention. For those skilled in the art, they should be able to realize that all equivalents made by using the description and illustrations of the present invention The solutions obtained by replacement and obvious changes shall all be included in the protection scope of the present invention.

Claims (9)

1. A hemorrhagic transformation prediction system, comprising: An acquisition unit, configured to acquire multiple pieces of patient data for training, each piece of patient data for training includes multiple disease characteristics for training; A model generation unit, connected to the acquisition unit, is used to generate a prediction model for predicting hemorrhagic transformation according to the plurality of obtained training patient data, and the model generation unit further includes: A feature selection module, configured to select the disease features for training in the patient data for training; A feature classification module, connected to the feature selection module, for feature classification of the selected disease features for training; A model training module, connected to the feature classification module, for training and forming the prediction model according to the classified training disease features; an acquisition unit, configured to acquire actual patient data; The prediction unit is connected to the acquisition unit and the model generation unit respectively, and is used to send the actual patient data into the prediction model formed by training, so as to output a corresponding prediction result. 2. hemorrhagic transformation prediction system as claimed in claim 1, is characterized in that, further comprises in the described feature selection module: The first feature selection component is used to select the disease features for training by using the CM feature selection method; The second feature selection component is used to select the disease features for training by adopting the packaging model feature selection method; The third feature selection component is used to select the disease features for training by adopting a filter model feature selection method; A selection control part is connected to the first feature selection part, the second feature selection part and the third feature selection part, and is used to select and enable the first feature selection part according to the correspondence between the training disease features. A feature selection component or said second feature selection component or said third feature selection component. 3. The hemorrhagic transformation prediction system according to claim 1, wherein the feature classification module classifies the disease features for training using a random forest model. 4. The hemorrhagic transformation prediction system according to claim 1, wherein the feature classification module classifies the disease features for training by means of a support vector machine. 5. The hemorrhagic transformation prediction system according to claim 1, wherein the feature classification module uses Logistic regression or a perceptron to classify the disease features for training. 6. The hemorrhagic transformation prediction system according to claim 1, wherein the feature classification module uses the AdaBoost algorithm to classify the disease features for training. 7. The hemorrhagic transformation prediction system as claimed in claim 4, further comprising: A data processing unit, connected between the acquisition unit and the model generation unit, for performing preset processing on the patient data for training, so as to achieve data balance of the patient data for training; The preset processing is: processing the patient data for training by means of oversampling and/or multivariate support vector machine algorithm. 8. The hemorrhagic transformation prediction system as claimed in claim 5, further comprising: A data processing unit, connected between the acquisition unit and the model generation unit, for performing preset processing on the patient data for training, so as to achieve data balance of the patient data for training; The preset processing is: processing the training patient data in an over-sampling manner; and/or processing said training patient data with a cost-sensitive loss function; and/or The training patient data is processed using a cost sensitive learning rate. 9. hemorrhagic transformation prediction system as claimed in claim 1, is characterized in that, also comprises in the described model generating unit: The risk rating module is connected to the model training module, and the risk rating module performs risk level discretization processing according to the obtained patient data for training to form a set of reference discrete points for risk rating as the model training module training Reference data when forming the predictive model.