CN108053885A - A kind of hemorrhagic conversion forecasting system - Google Patents
A kind of hemorrhagic conversion forecasting system Download PDFInfo
- Publication number
- CN108053885A CN108053885A CN201711209120.2A CN201711209120A CN108053885A CN 108053885 A CN108053885 A CN 108053885A CN 201711209120 A CN201711209120 A CN 201711209120A CN 108053885 A CN108053885 A CN 108053885A
- Authority
- CN
- China
- Prior art keywords
- training
- feature
- model
- state
- patient data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Measuring And Recording Apparatus For Diagnosis (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
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
Technical field
The present invention relates to field of medical technology more particularly to a kind of hemorrhagic conversion forecasting systems.
Background technology
Cerebral infarction is one of global major public health problem, incidence, lethality, disability rate and high recurrence rate, and
Clinical treatment means are limited.Wherein intravenous thrombolysis therapy is a nearly 20 years weights in cerebral arterial thrombosis clinical treatment means
Quantum jump can effectively reduce dead and disability rate.But intravenous thrombolysis therapy is a kind for the treatment of means of excessive risk, in treatment
Simultaneously nerve function lesion or even threat to life may can be aggravated, is gone out with the symptom of hemorrhagic conversion after thrombolysis, some bleedings
Blood conversion be also intravenous thrombolysis therapy means can not further genralrlization one of the major reasons.Specifically, in clinical economics process
In, intravenous thrombolysis has a stringent time restriction, only fall ill 4.5 it is small when within meet the acute cerebral infarction of intravenous thrombolysis condition and suffer from
Person can just carry out this treatment, and the risk that so-called " meeting intravenous thrombolysis condition " refers to generation hemorrhagic conversion is relatively low or basic
There is no hemorrhagic conversion danger.Therefore, the communication of doctor and patient are directly related to and to controlling for the anticipation of hemorrhagic conversion risk
The decision-making for the treatment of is held.
It is determined in the prior art for the prediction of hemorrhagic conversion usually according to some relevant clinical researches, such as example
Clinical nervous function when age, blood glucose, morbidity etc..Also having relevant clinical score table carrys out adjuvant clinical doctor to going out simultaneously
The possibility of blood conversion is prejudged, such as hemorrhage score, multicenter palsy the poll projected score after thrombolysis, SITS scores,
GRASPS scores and SEDAN scores etc..But these grade forms are substantially the diagnosis and treatment experience based on doctor and either use
Obtained from simple logistic regression, use value and corresponding accuracy have to be tested.It is referred to《Five kinds of prediction moulds
Comparison of the type after Chinese population thrombolysis in bleeding prediction application》In one text, the performance of different Rating Models is compared
Compared with.Wherein GRASPS scoring best performances, but its Receiver operating curve (receiver operating
Characteristiccurve, ROC) corresponding to AUC value only have 0.7056, working performance still has to be hoisted.Also,
Above-mentioned clinical prompting is relatively independent, and grade form is also relatively easy, and all only covering several substantially clinically may be to influence prognosis
Larger factor carries out Comprehensive Evaluation without the Relevant Clinical Factors to patient, also lacks individuation and judge.
The content of the invention
According to problems of the prior art, a kind of technical solution of hemorrhagic conversion forecasting system is now provided, it is intended to drop
Probability the occurrence of hemorrhagic conversion during low intravenous thrombolysis therapy, so as to reduce clinical risk and corresponding medical expense.
Above-mentioned technical proposal specifically includes：
A kind of hemorrhagic conversion forecasting system, wherein, including：
Acquiring unit, for obtaining a plurality of training patient data, every training patient data includes multiple
Training state of an illness feature；
Model generation unit connects the acquiring unit, is given birth to for the training according to a plurality of acquisition with patient data
Further comprise into one for the prediction model predicted hemorrhagic conversion, the model generation unit：
Feature selection module, for being made choice to the training with the training in patient data with state of an illness feature；
Tagsort module connects the feature selection module, for it is selected it is described training with state of an illness feature into
Row tagsort；
Model training module connects the tagsort module, for according to special with the state of an illness by the training of classification
Sign training forms the prediction model；
Collecting unit, for collecting actual patient data；
Predicting unit connects the collecting unit and the model generation unit respectively, for by the actual patient number
According to being sent into the prediction model that training is formed, to export corresponding prediction result.
Preferably, the hemorrhagic conversion forecasting system, wherein, the feature selection module further comprises：
Fisrt feature alternative pack, for being made choice using CM feature selectings mode to the training with state of an illness feature；
Second feature alternative pack, for being carried out using packaging model feature selecting mode to the training with state of an illness feature
Selection；
Third feature alternative pack, for being carried out using filtering model feature selecting mode to the training with state of an illness feature
Selection；
Control unit is selected, connects the fisrt feature alternative pack, the second feature alternative pack and described respectively
Third feature alternative pack, for enabling the fisrt feature according to the correspondence selection between the training state of an illness feature
Alternative pack either the second feature alternative pack or the third feature alternative pack.
Preferably, the hemorrhagic conversion forecasting system, wherein, the tagsort module is by the way of Random Forest model
Classified to the training with state of an illness feature.
Preferably, the hemorrhagic conversion forecasting system, wherein, the tagsort module is by the way of support vector machines pair
The training is classified with state of an illness feature.
Preferably, the hemorrhagic conversion forecasting system, wherein, the tagsort module is returned or felt using Logistic
Know that the mode of device classifies to the training with state of an illness feature.
Preferably, the hemorrhagic conversion forecasting system, wherein, the tagsort module is using AdaBoost algorithms to described
Training is classified with state of an illness feature.
Preferably, the hemorrhagic conversion forecasting system, wherein, it further includes：
Data processing unit is connected between the acquiring unit and the model generation unit, for the training
Default processing is carried out with patient data, to realize the data balancing of the training patient data；
It is described it is default processing be：The training is suffered from by the way of over-sampling and/or Multi-SVM algorithm
Person's data are handled.
Preferably, the hemorrhagic conversion forecasting system, wherein, it further includes：
Data processing unit is connected between the collecting unit and the model generation unit, for the training
Default processing is carried out with patient data, to realize the data balancing of the training patient data；
It is described it is default processing be：The training is handled with patient data by the way of over-sampling；And/or
The training is handled with patient data using cost-sensitive loss function；And/or
The training is handled with patient data using cost sensitive learning rate.
Preferably, the hemorrhagic conversion forecasting system, wherein, it is further included in the model generation unit：
Risk rating module, connects the model training module, and the risk rating module is according to the training of acquisition
Risk class sliding-model control is carried out with patient data, to form the reference discrete point of one group of risk rating, as the model
Reference data when training module trains to form the prediction model.
The advantageous effect of above-mentioned technical proposal is：A kind of hemorrhagic conversion forecasting system is provided, intravenous thrombolysis can be reduced and controlled
Probability the occurrence of hemorrhagic conversion during treatment, so as to reduce clinical risk and corresponding medical expense.
Description of the drawings
Fig. 1 is a kind of general structure schematic diagram of hemorrhagic conversion forecasting system in the preferred embodiment of the present invention；
Fig. 2 is the concrete structure schematic diagram of feature selection module in the preferred embodiment of the present invention；
Fig. 3 is the concrete operating principle schematic diagram of feature selection module in the preferred embodiment of the present invention.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present invention, the technical solution in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is only part of the embodiment of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art obtained on the premise of creative work is not made it is all its
His embodiment, belongs to the scope of protection of the invention.
It should be noted that in the case where there is no conflict, the feature in embodiment and embodiment in the present invention can phase
Mutually combination.
The invention will be further described in the following with reference to the drawings and specific embodiments, but not as limiting to the invention.
With the continuous development of science and technology, some machine learning algorithms, which have been also applied to, predicts hemorrhagic conversion
During, i.e., the prediction to hemorrhagic conversion is realized using some machine learning algorithms, from substantial amounts of non-structured data
In identify the pattern that the mankind are difficult to.Existing is only merely to turn thromboembolism treatment and bleeding to the prediction of thrombolysis bleeding risk
Change corresponding database and inputed to a fixed machine learning model, algorithm is not carried out for the feature of medical data
Improve, do not account for the priori of incidence relation that may be present and doctor between the unbalanced of data, data characteristics diagnose to
The influence that data set is brought, therefore the prediction level of model is more general, such as in the research of 14 years H.Asadi et al., mould
Although type accuracy has reached 70% but its recall rate is close to zero, and the AUC value of prediction also only has 0.6 or so.These predictions
The design system of model is all fixed, and parameter is all based on what data in database before were trained, to newly entering
Patient data do not design self refresh function.Over time, under the accuracy of these system predictions has significantly
Drop.
Based on the above-mentioned problems in the prior art, a kind of hemorrhagic conversion forecasting system is now provided, which should
For predicting issuable hemorrhagic conversion situation during intravenous thrombolysis therapy.
Specifically, in above-mentioned hemorrhagic conversion forecasting system it is specific as described in Figure 1, including：
Acquiring unit 1, for obtaining a plurality of training patient data, every training patient data includes multiple training
With state of an illness feature；
Model generation unit 2 connects acquiring unit 1, is used for the patient data generation one of the training according to a plurality of acquisition
In the prediction model predicted hemorrhagic conversion, model generation unit 2 further comprises：
Feature selection module 21, for being made choice to training with the training in patient data with state of an illness feature；
Tagsort module 22, connection features selecting module 21, for carrying out spy with state of an illness feature to selected training
Sign classification；
Model training module 23, connection features sort module 22, for being instructed according to the state of an illness feature of the training by classification
White silk forms prediction model；
Collecting unit 3, for collecting actual patient data；
Predicting unit 4 connects collecting unit 3 and model generation unit 2 respectively, is trained for actual patient data to be sent into
In the prediction model of formation, to export corresponding prediction result.
Specifically, in the present embodiment, above-mentioned acquiring unit 1 can connect the database outside one or remotely connect one
Server-side, and pass through database or the preprepared training patient data of long-range server-side acquisition.Certainly, above-mentioned instruction
White silk can also be directly inputted in acquiring unit 1 with patient data by way of user setting.
In the present embodiment, above-mentioned training patient data includes multiple training state of an illness features, these training state of an illness
Feature is all the feature of some explicit physical meanings, therefore system is not required to be carried again from training with carrying out feature in patient data
It takes.The feature of so-called explicit physical meaning, some bases in the medical record information of some patients that can be according in historical record
This information is such as age and gender, after carrying out some inspection information of preoperative sign inspection and carrying out intravenous thrombolysis therapy
Whether relevant information (no hemorrhagic conversion, slight bleeding or severe haemorrhage) of hemorrhagic conversion situation etc. is had.
In the present embodiment, above-mentioned acquiring unit 1, it is necessary to pre-processed to data, specifically needs after data are obtained
It screens out in training patient data and there is missing or apparent vicious data entry, and to continuous data into rower
Quasi-ization processing.Further, the process of above-mentioned data prediction can be manually performed by user, can also be by system according to default
Some screening rule automated executions, such as the data filling template of default training patient data, and according to the template come pair
Training matched with patient data, in training of judgement patient data whether there is shortage of data situation and according to
The data format of different fillers matches it in template, in training of judgement patient data apparent with the presence or absence of data
The situation of mistake.
In the present embodiment, in above-mentioned model generation unit 2, before the training generation of prediction model is carried out, it is necessary first to
Feature is made choice.Why need to carry out feature selecting, be because in data set size, characteristic dimension and characteristic attribute
In the case of difference, the Selection Framework of different training state of an illness features can all have different performances and each suitable for using
Environment, it is therefore desirable to be made choice before training pattern to feature, different features is placed into most suitable feature selects
It selects in frame, embodies its optimal test effect.Specific feature selection approach can hereinafter be described in detail.
In the present embodiment, by feature selecting afterwards, it is necessary to using tagsort module 22 to being used by the training of selection
State of an illness feature is classified.Grader realization may be employed in features described above sort module 22, the training use after tagsort
During state of an illness feature can be applied to model training.
In the present embodiment, using with similar mode in the prior art, trained to obtain corresponding prediction model according to feature,
Details are not described herein.
In the present embodiment, during training obtained prediction model that can be applied to actual hemorrhagic conversion prediction.
Specifically, gather the real data of sufferer according to the input requirements of prediction model and be sent into prediction model, by prediction model
Prediction after can just obtain representing the prediction of the possibility that hemorrhagic conversion occurs after intravenous thrombolysis therapy in the sufferer
As a result.Doctor can be using the prediction result as with reference to relevant diagnosis and treatment plan is linked up and formulated between information and patient
Deng so as to reduce clinical risk, Economy type medicine expense.
In the preferred embodiment of the present invention, as shown in Figure 2, feature selection module 21 further comprises：
Fisrt feature alternative pack 211, for being made choice using CM feature selectings mode to training with state of an illness feature；
Second feature alternative pack 212, for being carried out using packaging model feature selecting mode to training with state of an illness feature
Selection；
Third feature alternative pack 213, for being carried out using filtering model feature selecting mode to training with state of an illness feature
Selection；
Control unit 214 is selected, connects fisrt feature alternative pack 211, second feature alternative pack 212 and the 3rd respectively
Feature selecting component 213, for enabling fisrt feature alternative pack according to the correspondence selection between training state of an illness feature
211 either second feature alternative pack 212 or third feature alternative packs 213.
In the present embodiment, above-mentioned fisrt feature alternative pack 211, second feature alternative pack 212 and third feature choosing
Three kinds of different feature selecting frames that component 213 represents system respectively are selected, can be realized by computer system automatic
Training and test.
Specifically, fisrt feature alternative pack 211 is using CM feature selectings mode (Conservative Mean
Feature Selection) training is made choice with state of an illness feature.CM feature selectings mode is carried out mainly for single feature
Selection provides a kind of scheme for sampling and improving feature selecting stability.Specifically, it make use of list in CM feature selectings mode
The constant characteristic of AUC value in the case of tune Function Mapping is calculated specific special for some using K-fold validation
It seeks peace the AUC value of classification results.Subsequently, for this K AUC value, then ask for its average value mu and standard deviation α.Finally by comparing
The value of (μ-α) selects optimal multiple training state of an illness features to form character subset.
Above-mentioned second feature alternative pack 212 uses training using packaging model feature selecting mode (Wrapper)
State of an illness feature makes choice.Third feature alternative pack 213 is using filtering model feature selecting mode (Filter) to instruction
White silk is made choice with state of an illness feature., it is necessary to which what is considered is that their valuation functions and search are calculated in both feature selecting modes
Method.In the technical program, for both feature selecting modes, provide including sweep forward, reverse search, genetic algorithm
And a variety of searching algorithms such as exhaustive search.And for packaging model feature selecting mode, using CFS frames
The AUC value of (Correlation-based feature Selection) learning algorithm output is as its valuation functions.For mistake
Aspect of model selection mode is filtered, is commented using symmetrical uncertainty, RELIEF and minimum description length as it
Estimate function.
In the present embodiment, the operation of above-mentioned three kinds of feature selecting frames is controlled using a selection control unit 214.Tool
Body, as shown in Figure 3, control unit 214 is selected to be judged first according to training with the incidence relation between state of an illness feature：
If 1) incidence relation is relatively simple, control unit 214 is selected to directly select and enables fisrt feature alternative pack
211, i.e., training is made choice with state of an illness feature using CM feature selectings mode；
If 2) incidence relation is complex, control unit 214 is selected to select more traditional other two feature selecting frame
Frame.Further, if having got enough data, the selection of control unit 214 is selected to enable second feature alternative pack
212, i.e., training is made choice with state of an illness feature using packaging model feature selecting mode；
If 3) data volume got is less, the selection of control unit 214 is selected to enable third feature alternative pack 213,
Training is made choice with state of an illness feature using filtering model feature selecting mode.
In the preferred embodiment of the present invention, features described above sort module 22 is by the way of Random Forest model to training
Classified with state of an illness feature.
It specifically, can be in systems directly using in Scikit-learn learning databases for Random Forest model
RandomForestClassifier.Because every class decision tree of random forest is equivalent to doing feature during training
It chooses, therefore additional examine need not be carried out to Feature Selection algorithm during carrying out tagsort using Random Forest model
Consider.Grid search directly is carried out to the parameter that needs adjust during cross validation (cross validation), finally
Determine its parameters value in the case where AUC value is optimal.
In the preferred embodiment of the present invention, features described above sort module 22 uses training by the way of support vector machines
State of an illness feature is classified.
Specifically, for traditional support vector machines (Support Vector Machine, SVM), use and be
Libsvm storehouses under Python 3.6.Wherein, as the AUC value for weighing performance consider be data point to optimal hyperlane away from
From the relation between value and last classification.
And for Multi-SVM (Multivariate SVM), can by the use of AUC value as prediction label and
Loss function between true tag.The core used is all linear kernel, and the processing without considering over-sampling and use C languages
Say the svm-perf storehouses write.
In actual process, different types of support vector machines, corresponding spy can be selected according to actual conditions
Levy selection algorithm and data balancing processing method (can hereinafter be described in detail), the current optimal SVM models of output.
In the preferred embodiment of the present invention, features described above sort module 22 is returned using Logistic or perceptron
Mode classifies to training with state of an illness feature.
Specifically, when by the way of Logistic recurrence, need to compare different feature selecting sides in training process
Case and data balancing processing method (can hereinafter be described in detail).In the technical program, for Logistic return using
Theano frames under Python3.6.
When by the way of perceptron, in the design aspect of perceptron, Feature Selection and single hidden layer perceptron are appropriate
Nonlinear Design enables to perceptron model to have preferable fitting effect to current data set.And in view of training at present
Limited with the data set size of patient data, the hidden layer number of perceptron is also unsuitable excessive, otherwise can attract more mistakes instead
Difference.
For some hyper parameters of perceptron, cross validation may be employed to determine.It further, can be according to reality
The suitable feature selecting used of situation selection and data equalization processing method (can hereinafter be described in detail).Similarly, the technical program
In, for perceptron also using the Theano frames under Python3.6.
In the preferred embodiment of the present invention, features described above sort module 22 is using AdaBoost algorithms to the training state of an illness
Feature is classified.
Specifically, in AdaBoost algorithms, each Weak Classifier is designed to relatively simple perceptron model.Its
The number of middle Weak Classifier can be determined by cross validation.Above-mentioned AdaBoost algorithms are also mainly by under Python3.6
Theano frames are realized.
In the preferred embodiment of the present invention, still as shown in fig. 1, above-mentioned hemorrhagic conversion forecasting system is specifically also wrapped
It includes：
Data processing unit 5 is connected between acquiring unit 1 and model generation unit 2, for training patient data
Default processing is carried out, to complete following target：Screening is clearly present the patient data of mistake, missing data is filled, is real
The now data balancing of training patient data.
Specifically, in terms of input feature vector missing, for existing multiple data centers data set, due to each data center
The feature of record is all given priority to, therefore there are a certain amount of shortage of data.The missing processing scheme that the present invention uses for
Missing indicate schemes are brought the accuracy of data set to avoid average filling or median filling negative
It influences.In terms of target classification equilibrium, for existing patient data, finally occur hemorrhagic conversion training sample and
The accounting not occurred between the training sample of hemorrhagic conversion is very unbalanced, the sample of symptomatic hemorrhagic occurs may only account for not sending out
1/20 or so of blood sample is born, the data set of entire training sample is caused the unbalanced problem of data occur.Therefore, system
It needs to carry out data set some data balancing processing, cause most to avoid being trained using unbalanced data the set pair analysis model
The problem of whole prediction model output is inaccurate.
Further, for tagsort module 22 used by different tagsort method (i.e. different classification
Device), above-mentioned default processing (the data balancing processing method that i.e. system uses) is also different, is specially：
1) when tagsort module 22 is when classifying to feature by the way of support vector machines, above-mentioned data processing list
The default processing used in member 5 is as follows：
1. over-sampling carries out data balancing processing to training by the way of over-sampling with patient data.Specifically, at random
A series of few class sample (having the training patient data of symptomatic hemorrhagic, be hereinafter no longer described in detail) is sampled, so that different
Sample size corresponding to type is close.
2. classifying using Multi-SVM algorithm, and error rate is substituted to update support vector machines with AUC value.
2) when tagsort module 22 using Logistic return or perceptron by the way of classify to feature when, on
It is as follows to state the default processing used in data processing unit 5：
1. data balancing processing is carried out with patient data to training by the way of over-sampling.Specifically, stochastical sampling one
Few class sample of series, so that the sample size corresponding to different type is close.
2. using the general loss function used in cost-sensitive loss function alternative system, to increase to few class sample
The punishment dynamics of misjudgement.
3. the general learning rate used in utilization cost sensitivity learning rate alternative system so that for few class sample
Habit rate higher, the study for multiclass sample (the training patient data of i.e. no symptomatic hemorrhagic is hereinafter no longer described in detail)
Rate is lower.In this case, the step-length that the parameter of model is updated for few class sample is greater than multiclass sample.
In the preferred embodiment of the present invention, still as shown in fig. 1, further included in above-mentioned model generation unit 2：
Risk rating module 24, link model training module 23, risk rating module 24 use patient according to the training of acquisition
Data carry out risk class sliding-model control, to form the reference discrete point of one group of risk rating, are instructed as model training module
Practice reference data when forming prediction model.
Specifically, system provides two schemes for risk class discretization：
In the case that the deficienter quality of data of data is also less desirable in system database, system use is without prison
The equal frequency discretization scheme superintended and directed.Equal frequency discretization is acted on training set, can obtain one group of continuous data discrete point.This
One group of discrete point is reference point when subsequently making risk rating for patient；
Data are more sufficient and in the case that the quality of data can centainly be ensured in system database, and system makes
With the minimum comentropy discretization scheme for having supervision.This scheme is acted on training set, passes through training set different grouping
The sum of comentropy so that go out between different grouping group blood sample account for total sample rate variance it is as big as possible, and then also generate one
The group patient risk that can be used for newly arriving is classified the discrete point of reference.
The validity of the classification of above two wind direction grade discretization scheme can pass through Wilcoxen signed rank test
(Wilcoxon rank sum tests) is verified.In Systematic selection checkout procedure the scheme of Z values deviation from origin bigger generated from
Reference point of the scatterplot as predicted portions.
The foregoing is merely preferred embodiments of the present invention, not thereby limit embodiments of the present invention and protection model
It encloses, to those skilled in the art, should can appreciate that all with made by description of the invention and diagramatic content
Equivalent substitution and obviously change obtained scheme, should all include within the scope of the present invention.
Claims (9)
1. a kind of hemorrhagic conversion forecasting system, which is characterized in that including：
Acquiring unit, for obtaining a plurality of training patient data, every training patient data includes multiple training
With state of an illness feature；
Model generation unit connects the acquiring unit, and one is generated for the patient data of the training according to a plurality of acquisition
For the prediction model predicted hemorrhagic conversion, the model generation unit further comprises：
Feature selection module, for being made choice to the training with the training in patient data with state of an illness feature；
Tagsort module connects the feature selection module, for carrying out spy with state of an illness feature to the selected training
Sign classification；
Model training module connects the tagsort module, for being instructed according to by the training state of an illness feature of classification
White silk forms the prediction model；
Collecting unit, for collecting actual patient data；
Predicting unit connects the collecting unit and the model generation unit respectively, for the actual patient data to be sent
Enter in the prediction model that training is formed, to export corresponding prediction result.
2. hemorrhagic conversion forecasting system as described in claim 1, which is characterized in that further wrapped in the feature selection module
It includes：
Fisrt feature alternative pack, for being made choice using CM feature selectings mode to the training with state of an illness feature；
Second feature alternative pack, for being selected with state of an illness feature the training using packaging model feature selecting mode
It selects；
Third feature alternative pack, for being selected with state of an illness feature the training using filtering model feature selecting mode
It selects；
Control unit is selected, connects the fisrt feature alternative pack, the second feature alternative pack and the described 3rd respectively
Feature selecting component selects for enabling the fisrt feature according to the correspondence selection between the training state of an illness feature
Component either the second feature alternative pack or the third feature alternative pack.
3. hemorrhagic conversion forecasting system as described in claim 1, which is characterized in that the tagsort module is using random gloomy
The mode of woods model classifies to the training with state of an illness feature.
4. hemorrhagic conversion forecasting system as described in claim 1, which is characterized in that the tagsort module using support to
The mode of amount machine classifies to the training with state of an illness feature.
5. hemorrhagic conversion forecasting system as described in claim 1, which is characterized in that the tagsort module uses
Logistic is returned or the mode of perceptron classifies to the training with state of an illness feature.
6. hemorrhagic conversion forecasting system as described in claim 1, which is characterized in that the tagsort module uses
AdaBoost algorithms classify to the training with state of an illness feature.
7. hemorrhagic conversion forecasting system as claimed in claim 4, which is characterized in that further include：
Data processing unit is connected between the acquiring unit and the model generation unit, for suffering to the training
Person's data carry out default processing, to realize the data balancing of the training patient data；
It is described it is default processing be：To training patient's number by the way of over-sampling and/or Multi-SVM algorithm
According to being handled.
8. hemorrhagic conversion forecasting system as claimed in claim 5, which is characterized in that further include：
Data processing unit is connected between the collecting unit and the model generation unit, for suffering to the training
Person's data carry out default processing, to realize the data balancing of the training patient data；
It is described it is default processing be：The training is handled with patient data by the way of over-sampling；And/or
The training is handled with patient data using cost-sensitive loss function；And/or
The training is handled with patient data using cost sensitive learning rate.
9. hemorrhagic conversion forecasting system as described in claim 1, which is characterized in that further included in the model generation unit：
Risk rating module, connects the model training module, and the risk rating module is suffered from according to the training of acquisition
Person's data carry out risk class sliding-model control, to form the reference discrete point of one group of risk rating, as the model training
Reference data when module trains to form the prediction model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711209120.2A CN108053885B (en) | 2017-11-27 | 2017-11-27 | Bleeding conversion prediction system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711209120.2A CN108053885B (en) | 2017-11-27 | 2017-11-27 | Bleeding conversion prediction system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108053885A true CN108053885A (en) | 2018-05-18 |
| CN108053885B CN108053885B (en) | 2023-03-28 |
Family
ID=62120664
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711209120.2A Active CN108053885B (en) | 2017-11-27 | 2017-11-27 | Bleeding conversion prediction system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108053885B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108766559A (en) * | 2018-05-22 | 2018-11-06 | 合肥工业大学 | Clinical decision support method and system for intelligent disorder in screening |
| CN109003679A (en) * | 2018-06-28 | 2018-12-14 | 众安信息技术服务有限公司 | A kind of cerebrovascular hemorrhage and ischemic prediction technique and device |
| CN112599250A (en) * | 2020-12-24 | 2021-04-02 | 中国人民解放军总医院第三医学中心 | Postoperative data analysis method and device based on deep neural network |
| CN115861662A (en) * | 2023-02-22 | 2023-03-28 | 脑玺(苏州)智能科技有限公司 | Prediction method, device, equipment and medium based on combined neural network model |
| CN116721771A (en) * | 2023-08-11 | 2023-09-08 | 首都医科大学附属北京朝阳医院 | Bleeding transformation risk judging method and device, storage medium and terminal |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2000079472A2 (en) * | 1999-06-17 | 2000-12-28 | Raytheon Company | A multi-resolution object classification method employing kinematic features and system therefor |
| CN105203749A (en) * | 2015-09-22 | 2015-12-30 | 深圳市第二人民医院 | Serum marker capable of evaluating cerebral hemorrhage risk before thrombolysis and application thereof |
-
2017
- 2017-11-27 CN CN201711209120.2A patent/CN108053885B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2000079472A2 (en) * | 1999-06-17 | 2000-12-28 | Raytheon Company | A multi-resolution object classification method employing kinematic features and system therefor |
| CN105203749A (en) * | 2015-09-22 | 2015-12-30 | 深圳市第二人民医院 | Serum marker capable of evaluating cerebral hemorrhage risk before thrombolysis and application thereof |
Non-Patent Citations (3)
| Title |
|---|
| ADITYA KHOSLA等: ""An integrated machine learning approach to stroke prediction"", 《PROCEEDINGS OF THE 16TH ACM SIGKDD INTERNATIONAL CONFERENCE ON KNOWLEDGE DISCOVERY AND DATA MINING》 * |
| PAUL BENTLEY等: ""Prediction of stroke thrombolysis outcome using CT brain"", 《NEUROIMAGE: CLINICAL》 * |
| 计智伟等: ""特征选择算法综述"" * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108766559A (en) * | 2018-05-22 | 2018-11-06 | 合肥工业大学 | Clinical decision support method and system for intelligent disorder in screening |
| CN108766559B (en) * | 2018-05-22 | 2020-12-11 | 合肥工业大学 | Clinical decision support method and system for intelligent disease screening |
| CN109003679A (en) * | 2018-06-28 | 2018-12-14 | 众安信息技术服务有限公司 | A kind of cerebrovascular hemorrhage and ischemic prediction technique and device |
| CN109003679B (en) * | 2018-06-28 | 2021-06-08 | 众安信息技术服务有限公司 | Method and device for predicting cerebral hemorrhage and ischemia |
| CN112599250A (en) * | 2020-12-24 | 2021-04-02 | 中国人民解放军总医院第三医学中心 | Postoperative data analysis method and device based on deep neural network |
| CN115861662A (en) * | 2023-02-22 | 2023-03-28 | 脑玺(苏州)智能科技有限公司 | Prediction method, device, equipment and medium based on combined neural network model |
| CN116721771A (en) * | 2023-08-11 | 2023-09-08 | 首都医科大学附属北京朝阳医院 | Bleeding transformation risk judging method and device, storage medium and terminal |
| CN116721771B (en) * | 2023-08-11 | 2023-12-19 | 首都医科大学附属北京朝阳医院 | Bleeding transformation risk judging method and device, storage medium and terminal |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108053885B (en) | 2023-03-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108053885A (en) | A kind of hemorrhagic conversion forecasting system | |
| Patil et al. | A comparative analysis on the evaluation of classification algorithms in the prediction of diabetes | |
| Maurya et al. | Chronic kidney disease prediction and recommendation of suitable diet plan by using machine learning | |
| Siristatidis et al. | Artificial intelligence in IVF: a need | |
| CN107403072A (en) | A kind of diabetes B prediction and warning method based on machine learning | |
| CN107358014A (en) | The clinical pre-treating method and system of a kind of physiological data | |
| CN109273096A (en) | A kind of risk management grading evaluation method based on machine learning | |
| WO2021179514A1 (en) | Novel coronavirus patient condition classification system based on artificial intelligence | |
| Aatila et al. | Keratoconus severity classification using features selection and machine learning algorithms | |
| CN113470816A (en) | Machine learning-based diabetic nephropathy prediction method, system and prediction device | |
| CN111145912A (en) | Machine learning-based prediction device for personalized ovulation promotion scheme | |
| CN109273093A (en) | A kind of construction method and building system of Kawasaki disease risk evaluation model | |
| CN113380413A (en) | Method and device for constructing invalid re-circulation (FR) prediction model | |
| Firke et al. | Convolutional neural network for diabetic retinopathy detection | |
| KR102421172B1 (en) | Smart Healthcare Monitoring System and Method for Heart Disease Prediction Based On Ensemble Deep Learning and Feature Fusion | |
| CN117116477A (en) | Construction method and system of prostate cancer disease risk prediction model based on random forest and XGBoost | |
| Revathi et al. | Particle Swarm Optimization based Detection of Diabetic Retinopathy using a Novel Deep CNN | |
| CN117612703A (en) | Diabetes retinopathy classification method based on medical test indexes | |
| CN114093448B (en) | Construction method of disease risk prediction model | |
| Kabir et al. | Multi-classification based Alzheimer's disease detection with comparative analysis from brain MRI scans using deep learning | |
| Banerjee et al. | Food recommendation using machine learning for chronic kidney disease patients | |
| Ebin et al. | A Variant Binary Classification Model for No-DR Mild-DR Detection using CLAHE Images With Transfer Learning | |
| Mahbub et al. | Machine learning approaches to identify significant features for the diagnosis and prognosis of chronic kidney disease | |
| CN115964504B (en) | Food safety risk prediction method and system | |
| KR102572317B1 (en) | Arrhythmia classification method using long short-term memory based on recurrent neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |