The specific embodiment
With reference to the accompanying drawings the present invention is described more fully, exemplary embodiment of the present invention is wherein described.In the accompanying drawings, identical label represents identical or similar assembly or element.
In this article, vector is the various index sets that certain patient provides, model f be mapping function a: x → 0,1,2 ...., n}, n is value 3,4 or other integers for example.That is, given patient's indicator vector x, this patient's of model prediction liver cirrhosis pathology is f (x) by stages, the value of f (x) for set 0,1,2 ...., in n}n kind centrifugal pump one.Specific index and disaggregated model are the important contents of this technology.Below by index and the disaggregated model introducing this patent and adopt.
Fig. 1 illustrates according to the structure chart of the first embodiment of hepatic fibrosis detection equipment of the present invention.As shown in Figure 1, in this embodiment, hepatic fibrosis detection equipment comprises input equipment 11, grader 12 and output device 13.Wherein, input equipment 11 is for receiving age and Biochemical Indices In Serum, and this Biochemical Indices In Serum at least comprises platelet, hyaluronic acid (HA), serum direct bilirubin (DBIL), prothrombin time (PT), serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT) and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT).Age and Biochemical Indices In Serum that grader 12 receives according to input equipment 11 carry out Liver Fibrosis Stages, and Liver Fibrosis Stages result is sent to output device 13; Grader 12 is according to the platelet, serum paddy (propylhomoserin) third (keto acid) transaminase (ALT that receive; GPT) and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT) three indexs obtain the ratio that two experts introduce: serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT)/platelet and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT)/serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT), blood substitute diarrhea with indigested food (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT), serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT) as the input parameter of grader.The Liver Fibrosis Stages result of output device 13 output category devices 12.For grader 12, can be grader, logistic regression (logistic regression analysis) grader, Adaboost integrated classifier or PCA (the Principal Component Analysis of support vector machine classifier, the grader based on decision-making device model, support vector regression model, principal component analysis)+KNN (K Nearest Neighbor, k nearest neighbor) category of model device.Grader 12 can be realized by software on computing equipment, or realizes by specialized hardware, circuit or equipment.
In above-described embodiment, grader carries out the detection of hepatic fibrosis by the Biochemical Indices In Serum of age and selection, can obtain than the detection method of prior art and detect more accurately effect, due to Biochemical Indices In Serum platelet, hyaluronic acid (HA), serum direct bilirubin (DBIL), prothrombin time (PT), serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT) and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT) detection is more universal, and general hospital can realize, and therefore, can expand the application of this scheme and popularize, and the cost and the difficulty that reduce whole detection.In addition, can select according to actual needs different graders, thereby improve the accuracy rate of the grader in reality.
Hepatic fibrosis detection equipment in the present invention, can, according to clinical demand, adopt multiple way of realization.According to one embodiment of present invention, this input equipment, grader and output device are positioned on a computer, and input equipment and output device are corresponding to outut devices such as the input equipments such as the keyboard of computer, touch screen, mouse, equipment interface and display screen, audio output apparatus, output interfaces; Grader can be realized by software, or realizes by the special-purpose grader circuit being connected with mainboard.By this checkout equipment of computer realization, can make full use of the characteristic that the penetration of computer use is high, reduce the cost of realizing of this checkout equipment.According to another embodiment of the invention, input equipment, grader and output device are positioned on same portable handheld device, and this handheld device can be general purpose hand-held computer, or for the special equipment of liver fibrosis diagnosis.Mode by checkout equipment with handheld device realizes, and has improved convenience and motility that this equipment is used.According to one embodiment of present invention, this hepatic fibrosis detection equipment also can be realized by the mode of in-circuit diagnostic system, introduces a kind of specific implementation of in-circuit diagnostic system below in conjunction with Fig. 2.
According to one embodiment of present invention, Biochemical Indices In Serum also comprises serum alkaline phosphatase (ALP; AKP), serum cholinesterase (ChE), mobility (PTA) three, or any one or two in above-mentioned three indexs.
According to one embodiment of present invention, Biochemical Indices In Serum also comprises transforminggrowthfactor-β1 (TGF-β 1) and alpha2 Macroglobulin (AMG); Grader carries out Liver Fibrosis Stages for age and the Biochemical Indices In Serum receiving according to input equipment.
In above-described embodiment, grader carries out the detection of hepatic fibrosis by the Biochemical Indices In Serum of age and selection, can obtain than the detection method of prior art and detect more accurately effect.
Fig. 2 illustrates according to the structure chart of the second embodiment of hepatic fibrosis detection equipment of the present invention.As shown in Figure 2, input equipment 21 can be the equipment such as computer, panel computer, PDA, as the equipment of input equipment, can be connected with grader 22 by the mode such as wired, wireless, and grader 22 can be server, computer or special equipment; The Liver Fibrosis Stages result of grader 22 outputs can be exported by output device 23, also can export to user by input equipment 21.Mode by in-circuit diagnostic system realizes this checkout equipment, only needs a grader that is positioned at backstage, can comprise a plurality of inputs and outfan, thereby can realize the detection support to more diagnosis department, thereby has reduced unit testing cost.
According to one embodiment of present invention, age and Biochemical Indices In Serum that the input of grader is mentioned except above-described embodiment, can also comprise the Transient elastography data of hepatic tissue, the hardness number of the hepatic tissue obtaining by Transient elastography equipment.
Fig. 3 illustrates according to the structure chart of the 3rd embodiment of hepatic fibrosis detection system of the present invention.As shown in Figure 3, in this embodiment, hepatic fibrosis detection system comprises input equipment 31, grader 32, output device 33 and Transient elastography equipment 34.Input equipment 31 and output device 33 can be referring to the descriptions of above-described embodiment, for being not described in detail at this for purpose of brevity.Transient elastography equipment 34 can obtain the Transient elastography data of hepatic tissue; Grader 33 receives the Transient elastography data of hepatic tissue from Transient elastography equipment 34, according to the Transient elastography data of age, Biochemical Indices In Serum and hepatic tissue, carry out Liver Fibrosis Stages.Transient elastography equipment 34 is for example FibroScan, obtains the FibroScan hardness number of hepatic tissue.
In above-described embodiment, in conjunction with the Transient elastography data of age, Biochemical Indices In Serum and hepatic tissue, carry out Liver Fibrosis Stages, can make full use of various testing results, make the result of Liver Fibrosis Stages more accurate.
According to one embodiment of present invention, this system also comprises Biochemical Indices In Serum checkout equipment, and Biochemical Indices In Serum checkout equipment detects the sample in test kit, obtains Biochemical Indices In Serum data, by input equipment, sends to grader.
Fig. 4 illustrates the schematic diagram of an embodiment of Transient elastography equipment and probe thereof.As shown in Figure 4, this elastogram equipment 44 comprises probe interface (Probe Socket) 441, for connecting ultrasonic probe 45, also comprises data transmission interface 442, for connecting computer or network with transmission data.Ultrasonic probe 45 comprises sonac (Ultrasound transducer) 443, shift knob (Button) 444, electrodynamic transducer (electrodynamics transducer) 445, transmission line (Connection cable) 446 and socket (Jack) 447.
If the data sample number of training classifier parameter is less, rely on the problem that over-fitting probably appears in single model, even can obtain good accuracy on existing sample, the popularization performance that may not have, is difficult to correctly predict unknown sample.
In order to address this problem, can adopt the method for Bagging: train a plurality of independently graders, last classification results obtains according to the result ballot of a plurality of graders.Can solve so to a certain extent the not robustness while only predicting with an independent model.Incomplete same with traditional Bagging method, use the method for similar cross validation, by sample random division, be n decile each time, with n-1 part training classifier (parameter is also during this time determined by grid search method) wherein, remaining portion is predicted.Like this, according to the result of prediction, carry out the screening of model.By the such random division of repeated several times, random division can be selected certain model each time.Finally, all models that obtain are merged, by the mode of ballot, carry out determining of last classification results.
Fig. 5 illustrates according to the structure chart of the 4th embodiment of hepatic fibrosis detection system of the present invention.As shown in Figure 5, in this embodiment, hepatic fibrosis detection system comprises input equipment 31, grader 52, output device 33 and Transient elastography equipment 34.Wherein, input equipment 31, output device 33 and Transient elastography equipment 34 can be referring to the descriptions of above-described embodiment, for being not described in detail at this for purpose of brevity.Grader 52 comprise voting machine 523, two or more sub-classifiers as the first sub-classifier 521, the second sub-classifier 522 ... etc.Each sub-classifier 521,522 etc. carries out Liver Fibrosis Stages acquisition Liver Fibrosis Stages result separately according to the Transient elastography data of age, Biochemical Indices In Serum and hepatic tissue, and its Liver Fibrosis Stages result is outputed to voting machine 523, voting machine 523 is determined the Liver Fibrosis Stages result of output by the mode of for example voting according to the Liver Fibrosis Stages result of each sub-classifier.
Fig. 6 illustrates according to the structure chart of the 5th embodiment of hepatic fibrosis detection equipment of the present invention.As shown in Figure 6, in this embodiment, hepatic fibrosis detection equipment comprises input equipment 31, grader 32, output device 33 and Transient elastography equipment 34 and parameter training device 65.Parameter training device 65 receives training sample data, determines the parameter of grader according to training sample data; Wherein, training sample data can comprise age, Biochemical Indices In Serum data and corresponding Liver Fibrosis Stages; Or training sample data can comprise Transient elastography data and the corresponding Liver Fibrosis Stages of age, Biochemical Indices In Serum data, hepatic tissue.According to existing sample, can train and obtain hepatic fibrosis disaggregated model.Consider that sample may enrich constantly, therefore, designed a kind of self-learning strategy of model.Write the learning strategy of above model as automatic training process completely, input interface is exactly sample set; Output interface is the final anticipation function using.Therefore,, once after having upgraded sample set, only need the automatic training function of caller, just can complete the self study process of model.Meanwhile, old model also can correspondingly be backed up preservation, with the model of tackling under emergency case, resumes work.
Object lesson below in conjunction with support vector machine is introduced disaggregated model.Below we will describe the training strategy of this disaggregated model in detail; While relating to characteristic vector, unification represents by vector x.
1. model " decomposes-merges " strategy
Decompose
Former problem need to predict which kind of in 5 class hardness numbers a sample belong to, and directly compares complexity, first classification problem is decomposed into four subproblems:
Each subproblem is a two-value classification problem.Such as subproblem 1, the meaning is exactly a given sample, and the hardness number that judge it is to be greater than to equal 1, is still less than 1.Remaining in like manner.
To each subproblem (two class classification problems), adopt support vector machine (Support Vector Machine, SVM) disaggregated model training.Finally learn to four submodel f altogether
i(x), i=1,2,3,4.F
i(x) be output as 0 or 1.
About support vector machine, next trifle is shown in detailed introduction.
Merge
After completing above-mentioned four sub-problem models, just subproblem can be merged into final decision rules.The result of four sub-model predictions is a sequence (f
1, f
2, f
3, f
4), each value value of sequence is 0 or 1, therefore has 16 kinds of possible sequence values.Finally predicting the outcome that each value is corresponding carried out decision-making according to the rule of table 2.
Four submodels finally predict the outcome
Table 1: submodel merges rule
2. support vector machine (Support Vector Machine, SVM) disaggregated model
Mention above, each model is decomposed into four submodels, and each submodel is two class classification problems.What we adopted is that support vector machine is as basic classification device.
Support vector machine is a kind of outstanding disaggregated model, it is classified the sample in sample space according to maximizing class interval (margin maximization) principle, when obtaining less training error rate, guarantee to promote preferably performance (ability that unknown sample is predicted).
Fig. 7 has provided a schematic diagram of svm classifier device in linear separability situation.
Fig. 7: largest interval svm classifier hyperplane schematic diagram.Solid and hollow 2 heap points represent two class sample points.The classifying face of middle solid line has larger class interval than all the other all dotted line classifying faces, thereby has better popularization performance.
Linear SVM
Simply, SVM is a kind of linear classifier.For two class classification problems, given training dataset
here x
i∈ R
d, i=1,2 ..n is characteristic vector, y
i∈+1, and-1}, i=1,2 ..n is the label of sample.The rule of classification is:
x is new samples to be sorted,
it is the result of svm classifier device category of model.Sign (x) is-symbol function, when x >=0, sign (x)=1; As x < 0, sign (x)=-1.
Here two variablees that determine grader are that they need to train out from data.By following equation, solve and obtain specifically:
Wherein C is a balance training error rate and the parameter of promoting performance, generally by cross validation, determines.
In fact, the definite optimization problem of equation 1 can be converted into following dual problem:
s.t.:0≤α≤C,α
Ty=0 (3)
Here α=[α
1..., α
n]
t, y=[y
1..., y
n]
t, D=(D
ij),
after solving the value that dual problem obtains dual variable α, the solution (w, b) of former problem can directly obtain:
Therefore final grader can be write as:
Non-linear SVM
SVM also can learn out nonlinear model.It utilize kernel method (kernel method) use specific nonlinear mapping by sample from former spatial mappings to feature space of higher-dimension more, making can be at higher dimensional space linear separability in the data of luv space linearly inseparable.Like this, at a linear model of higher dimensional space design, be just equivalent to the nonlinear model of designing at luv space.Fig. 8 has provided one, by polynomial kernel function, two-dimentional sample has been risen to dimension to three dimensions, thereby makes original existing inseparable sample at the schematic diagram of higher dimensional space linear separability.
Fig. 8: non-linear SVM algorithm schematic diagram.Original sample linearly inseparable.By following formula, change:
Φ:R
2→R
3
Use kernel method that original method is risen to dimension to higher dimensional space, at higher dimensional space linear separability, this is equivalent at luv space Nonlinear separability.
From Linear SVM above, can see, no matter be the dual form of SVM or last grader solution, can be write as the form of sample inner product
therefore, use kernel method to carry out the nonlinear mapping Φ of sample: x → Φ (x).Higher dimensional space after mapping like this, the inner product between sample can be calculated at an easy rate: Φ (x
i)
tΦ (x
j)=K (x
i, x
j).K is exactly kernel function, such as gaussian kernel function:
Therefore, the grader of non-linear SVM can be write as
wherein, dual variable can obtain by solving dual problem 6:
s.t.:0≤α≤C,α
Ty=0 (6)
Here α=[α
1..., α
n]
t, e=[1 ..., 1]
t, y=[y
1..., y
n]
t,
Kernel function (Kernel function) in general needs to meet Mercer condition, and common kernel function has three kinds:
1) polynomial kernel function: K (x
i, x
j)=(x
ty+c)
p, c ∈ R
2) gaussian kernel function: K (x
i, x
j)=exp ((x
i-x
j)
2/ (2 σ
2))
3) Sigmoid kernel function: K (x
i, x
j)=tanh (kx
ty-δ)
Can adopt non-linear SVM as the most basic grader, kernel selects gaussian kernel.
The biochemical indicator of above-mentioned employing and model obtain preferred parameter.In fact, adopt above-mentioned strategy, also extend and obtained other several groups of parameters:
1. index parameter
Table 2
The sequence number of the feature of below above-mentioned model being used is carried out the mark of medical science implication:
Table 3
Above-mentioned feature 5 and 8 is two ratio features introducing according to expert advice, and they relate to three features 2,14,15, and feature 2 occurs in upper table, and feature 14,15 is as follows:
Table 4
Table 5
Table 6
Table 7
Note, feature is order in no particular order.Various models relate to 13 kinds of different features, are divided into three types: age, Biochemical Indices In Serum and FibroScan hardness number; Biochemical Indices In Serum comprises platelet, hyaluronic acid (HA), serum direct bilirubin (DBIL), prothrombin time (PT), serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT) and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT), transforminggrowthfactor-β1 (TGF-β 1), alpha2 Macroglobulin (AMG), it is different obtaining their required costs.
Adopt different features, can train and obtain different models.
1) feature of model one use is minimum, use be all detect more universal, the biochemical indicator feature that general hospital can realize, thereby model is the simplest;
2) model four has been used whole features, model accuracy is relatively the highest, but because the feature of using relates to FibroScan hardness number and Biochemical Indices In Serum transforminggrowthfactor-β1 (TGF-β 1), alpha2 Macroglobulin (AMG), thereby in the collection of feature, cost is larger;
3) model two and model three have been taked balance strategy accordingly, have considered the precision of model and the required cost of collection apparatus, are two kinds of compromise schemes.
Following table 8 shows the result of the test data of above-mentioned 4 models:
Table 8
Embodiments of the invention have designed the disaggregated model based on medical science indexs such as Biochemical Indices In Serum and FibroScan indexs according to " goldstandard ", nondestructively Liver Fibrosis Stages are predicted.By patient being carried out to the chemical examination of specific biochemical indicator, obtain the characteristic vector of patient's patient's condition.Model, according to characteristic vector, is predicted the present pathological staging S0-S4 (or F0-F4) (rank is higher, represents that hepatic fibrosis is more serious) of patient.
Technical scheme of the present invention is comformed in a plurality of parameters and is screened, and key data has: sex, age, hepatitis B virus DNA carrying capacity, various liver enzyme index, cholesterol associated and nearly all biochemical indicator, hepatic fibrosis special detection index, FibroScan hardness number etc.By above all parameters are carried out to analyzing and processing and calculation, finally determine clinical n the Biochemical Indices In Serum best with hepatic fibrosis dependency, and draw hepatic fibrosis liver cirrhosis diagnosis model in conjunction with FS testing result.
Consider each hospital equipment and the skimble-scamble situation of biochemistry detection level, the detection for convenience of Different hospital, is also divided into two versions.
1) FS+ biochemistry detection model): for hepatopathy section hospital/outpatient service, index contains FS hardness number and Biochemical Indices In Serum platelet, hyaluronic acid (HA), serum direct bilirubin (DBIL), prothrombin time (PT), serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT) and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT).
2) detection model of all related biochemical indicator of FS+): for there being checkout equipment, hepatopathy section hospital/outpatient service that scientific research level is higher, the degree of depth solves diagnosis problem.Index contains Biochemical Indices In Serum platelet, hyaluronic acid (HA), serum direct bilirubin (DBIL), prothrombin time (PT), serum paddy (propylhomoserin) third (keto acid) transaminase (ALT; GPT) and serum paddy (propylhomoserin) grass (ethyl acetoacetic acid) transaminase (AST; GOT), FS hardness number and Biochemical Indices In Serum transforminggrowthfactor-β1 (TGF-β 1), alpha2 Macroglobulin (AMG).
The advantage such as the hepatic fibrosis detection system of the embodiment of the present invention has noinvasive, practical, method is easy, price, safety are good:
(1) devoid of risk.According to FibroScan noinvasive diagnostic equipment and relevant blood chemical result, diagnostic system can be by model analysis, and judgement hepatopath's degree of hepatic fibrosis, exists any risk hardly, also can not have any wound to patient.
(2) comprehensive cost is low.Owing to doing liver puncture except needs blood testing, also need to do puncture operation and later stage trauma care, so comprehensive cost will be higher than the method for non-invasive diagnosis.
(3) method is simple, clinical application range is wide.It is shorter that the operator of Fibroscan obtains certificate required time, and simple and easy to operate; Biochemical indicator detects without carrying out Special Training again, and itself satisfies the requirements hospital; The noinvasive model that both combine, clinical application range is extensive.
Description of the invention provides for example with for the purpose of describing, and is not exhaustively or limit the invention to disclosed form.Many modifications and variations are obvious for the ordinary skill in the art.Selecting and describing embodiment is for better explanation principle of the present invention and practical application, thereby and makes those of ordinary skill in the art can understand the various embodiment with various modifications that the present invention's design is suitable for special-purpose.