CN110299194A - The similar case recommended method with the wide depth model of improvement is indicated based on comprehensive characteristics - Google Patents
The similar case recommended method with the wide depth model of improvement is indicated based on comprehensive characteristics Download PDFInfo
- Publication number
- CN110299194A CN110299194A CN201910490881.2A CN201910490881A CN110299194A CN 110299194 A CN110299194 A CN 110299194A CN 201910490881 A CN201910490881 A CN 201910490881A CN 110299194 A CN110299194 A CN 110299194A
- Authority
- CN
- China
- Prior art keywords
- model
- feature
- comprehensive characteristics
- similar case
- layer
- 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
- 238000000034 method Methods 0.000 title claims abstract description 39
- 230000006872 improvement Effects 0.000 title claims description 11
- 230000036541 health Effects 0.000 claims description 27
- 239000013598 vector Substances 0.000 claims description 24
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 claims description 18
- 201000010099 disease Diseases 0.000 claims description 15
- 239000003814 drug Substances 0.000 claims description 13
- 230000004913 activation Effects 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 4
- 241000208340 Araliaceae Species 0.000 claims description 3
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 3
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 3
- 238000012952 Resampling Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000000586 desensitisation Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 235000008434 ginseng Nutrition 0.000 claims description 3
- 208000019423 liver disease Diseases 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000007170 pathology Effects 0.000 claims description 3
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 238000003058 natural language processing Methods 0.000 abstract description 2
- 230000001351 cycling effect Effects 0.000 abstract 1
- 238000012216 screening Methods 0.000 abstract 1
- 230000006870 function Effects 0.000 description 21
- 238000003745 diagnosis Methods 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 208000010392 Bone Fractures Diseases 0.000 description 2
- 208000004044 Hypesthesia Diseases 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 229940079593 drug Drugs 0.000 description 2
- 208000034783 hypoesthesia Diseases 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 231100000862 numbness Toxicity 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000001902 propagating effect Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 208000008035 Back Pain Diseases 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 206010017711 Gangrene Diseases 0.000 description 1
- 208000008930 Low Back Pain Diseases 0.000 description 1
- 206010031252 Osteomyelitis Diseases 0.000 description 1
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000003759 clinical diagnosis Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 230000006378 damage Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002405 diagnostic procedure Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 208000014674 injury Diseases 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000000547 structure data Methods 0.000 description 1
- 210000001519 tissue Anatomy 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/80—Information retrieval; Database structures therefor; File system structures therefor of semi-structured data, e.g. markup language structured data such as SGML, XML or HTML
- G06F16/83—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/242—Dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Medical Informatics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Public Health (AREA)
- Primary Health Care (AREA)
- Epidemiology (AREA)
- Biomedical Technology (AREA)
- Bioethics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Security & Cryptography (AREA)
- Pathology (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computer Hardware Design (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The present invention relates to indicate to belong to Computer Natural Language Processing technical field with the similar case recommended method for improving wide depth model based on comprehensive characteristics.The present invention passes through first indicates that model obtains the comprehensive characteristics of case history description using comprehensive characteristics, carries out disorder in screening;Secondly, handling discrete features using cross feature method, it is input in linear segment, and the comprehensive characteristics of case history description and shallow Model feature is blended, is input to using door cycling element as in the recommendation sort sections of core;Finally, exporting dozens of case recommendation items on the basis of hundreds of candidate cases.The present invention realizes personalized case and recommends, and proposes a kind of recommendation sort algorithm model that traditional shallow-layer linear model is combined with depth network model, improves the accuracy of similar case recommendation.
Description
Technical field
It indicates to belong to the similar case recommended method for improving wide depth model based on comprehensive characteristics the present invention relates to a kind of
Computer Natural Language Processing technical field.
Background technique
With the fast development of intelligent medical treatment, artificial intelligence technology is gradually dissolved into medical industry, patient and medical worker,
The health medical treatment information platform of medical institutions' interaction comes into being.The research of clinical assisting in diagnosis and treatment at this stage, mainly by dividing
Magnanimity medical data and diagnostic data existing for analysis, processing in the multimedia form.It works as a result, the feature extraction of mass data
Essential, how to carry out personalized diagnosis and treatment application similarly has far-reaching research significance.
Compared with existing medical history recommended method, Traditional Chinese medical electronic case history medical treatment text is with semi-structured feature and continuously
Feature, so that traditional Text Representation method does not have universality and accuracy rate is often poor.Meanwhile it depending on unduly existing
Medical data preferable cannot must learn to hide feature.Also, usually there is plurality of target user in interrogation platform, and medical matters people
The interrogation habit of member is different.Meanwhile clinical diagnosis requires platform to have the features such as fast, quasi-.So that tradition is rule-based
Diagnosis and case recommended method fall flat.For two above problem, the invention proposes one kind based on comprehensive special
Sign indicates to realize personalized case with the similar case recommended method for improving wide depth model and recommend, propose a kind of tradition
The recommendation sort algorithm model that shallow-layer linear model is combined with depth network model improves the accurate of similar case recommendation
Property.
Summary of the invention
The present invention provides a kind of similar case recommended method indicated based on comprehensive characteristics with the wide depth model of improvement, needles
To Chinese medicine electronic health record text, preferable recommendation effect is achieved on the whole, also improves recommendation efficiency to a certain extent.
The technical scheme is that a kind of indicate to recommend with the similar case for improving wide depth model based on comprehensive characteristics
Method, specific step is as follows for the method:
Step1, medical text desensitization is carried out to text first, and medical text is segmented;By tcm characteristic vocabulary of terms,
The unique name of disease of Basic Theories of Chinese Medicine noun, Chinese medicine and disease name, hepatopathy noun and Chinese medicinal formulae noun, are added
THULAC dictionary, and segmented using THULAC, obtaining the corpus as unit of word indicates;
Step2, characteristic sub-area is carried out, discrete features is mapped to real-valued vectors;By continuous feature according to step in Step1
It is segmented, is indicated using the term vector that Word2Vec obtains corpus;
The comprehensive characteristics of Step3, building based on thresholding convolution variation self-encoding encoder indicate model;Firstly, according in Step2
Characteristic sub-area, two parts feature is carried out to the fusion of dimension, is compiled automatically wherein continuous feature is used based on thresholding convolution variation
Code device algorithm carries out character representation;Finally, the high-level semantics information obtained in Traditional Chinese medical electronic case history indicates;
Step4, building are based on the similar case recommended models for improving wide depth model;Each doctor is constructed respectively respectively
Similar case recommended models;It is indicated according to the case history comprehensive characteristics that Step3 is obtained, sequence exports dozens of case recommendation items to doctor
It is raw.
Further, the specific steps of the step Step1 are as follows:
Step1.1, privacy and Feature Selection operation are carried out to Chinese medicine electronic health record source data, remove in case history text
It is related to patient's individual privacy information, such as " name ", " admission number ", " home address ";It is screened out in conjunction with expert opinion and extraction is suffered from
Sick feature does not have contributive item, such as " occupation ", " wedding condition ", " nationality ", " physical examination ";
Step1.2, Chinese medicine associated disease pathology and medical terminology dictionary is added, work is segmented using THULAC Chinese text
Tool segments electronic health record;
Step1.3, due in electronic health record text missing values it is more, the illness feature of accurate description can not be showed, in
It is the electronic health record removed less than 150 words.
Further, the specific steps of the step Step2 are as follows:
Step2.1, " vital sign " in electronic health record data is extracted, " Gender ", the information such as " patient age ",
And its numerical value is mapped as one-dimensional vector, the discrete features as electronic health record;
Step2.2, content will be segmented obtained by Step1, term vector is mapped as using Word2vec method;And term vector is arranged
Table is arranged as matrix, the continuous feature as electronic health record.
Further, in the step Step3, continuous feature use based on thresholding convolution variation autocoder algorithm into
The specific steps of row character representation are as follows:
Step3.1, it is encoded using thresholding convolutional network, continuous feature obtained by Step2 is sent into pond layer and is compiled
Code result;And calculated using above-mentioned coding result and generate mean value and variance, generate Gaussian Profile and resampling;
Step3.2, building double stacked CNN model, the convolutional layer of nonlinear activation function is exported and be have passed through
The convolutional layer of sigmoid nonlinear activation function activation, which exports, to be multipliedWherein:
W and V indicates the weight of convolutional layer, and b and c indicate the bias term of convolutional layer, and * indicates convolution operation operation, and σ is thresholding convolution letter
Number;
Step3.3, loss function are logp (x)=DKL(qφ(z|x)||pθ(z | x)) so that final loss function reaches
To minimum;Wherein, θ is Optimal Parameters, and logp (x) indicates that model needs maximized log-likelihood function, DKLFor KL divergence, qφ
(z | x) it is encoder, pθ(x | z) it is decoder, z is hidden variable, and x is input variable;And if only if qφ(z | x)=pθ(z|x)
When, DKL(qφ(z|x)∥pθ(z | x))=0;
Step3.4, network model is trained come undated parameter using stochastic gradient rise method;First with prior distributionThe sample of one group of hidden variable z of stochastical sampling, is then input to decoder, finally exports a data point
The random sample of x;Consider different number hidden unit n, can be less than or higher than primitive character quantity.
Further, the specific steps of the step Step4 are as follows:
Step4.1, Logic Regression Models are definedHere x=[x1,x2,…,xd] indicate feature d's
One group of vector contains in characteristic set and is originally inputted feature and assemblage characteristic, w=[w1,w2,…,wd] indicate model ginseng
Number;
Step4.2, cross feature is definedHere cki∈ { 0,1 } is Boolean, and such as i-th special
Sign is k-th of conversion φkA part, ckiAs 1, it is otherwise 0, for binary features, for example, having and if only if assemblage characteristic
It all sets up and is only 1, be otherwise exactly 0;
Step4.3, GRU layers of core for defining depth module, and additional feedforward is added between the last layer and output
Layer, wherein using tanh function as the activation primitive of output layer;Connection is added between node layer hiding, and is followed with a door
Ring element controls the output of concealed nodes, effective variation of the Modelling feature on time-series dynamics;
It Step4.4, by the input feature vector of shallow-layer part include by structures such as regional feature, weather time and search keys
At cross feature as input, batch Stochastic Optimization Model parameter, while propagating backward to the shallow-layer and depth door of model again
Cyclic part;
Step4.5, conjunctive model anticipation function is definedIt uses
Joint output result takes the weighted sum of logarithm as predicted value, and the weighted sum is then fed to a common loss function,
It carries out joint training and optimizes;Final output is probability value;Wherein,For wide deep all model parameters,For GRU model
Parameter;It is finally ranked up from low to high using probability, takes preceding 5 to recommend as case.
In order to enable comprehensive characteristics to indicate that model fully represents input data set, the present invention enables each in data set
Data point x, there is one group or multiple groups hidden variable is corresponding to it.qφ(z | x) it is encoder, pθ(x | z) it is decoder.Probability is close
Degree function p (z) up-samples to obtain sample z in higher dimensional space Z, utilizes function f (z;Hidden variable z θ) is mapped to initial data sky
Between X;Make f (z by Optimal Parameters θ;It is θ) similar as far as possible to the truthful data in TO-EMR;Pass through pθ(x | z) replace f (z;
θ), the dependence between x and z can be clearly seen according to total probability formula.Maximize with lower probability:
P (x)=∫ pθ(x|z)p(z)dz
In the unbalanced data set of processing data distribution, tradition is using mean square error (MSE) although method can be simple
The error between network output and desired target value must be measured, but preferable effect can not be obtained, thus mean square error is not
It is a kind of effective error metrics method again.The present invention passes through building another set neural computing conditional probability distribution qφ(z
| x) it is used to approximate true posterior probability pθ(z | x), utilize KL divergence (Kullback-Leibler divergence) measurement two
Difference between a distribution calculates qφ(z | x) and pθSimilarity between (z | x):
Arrangement obtains:
Logp (x)=DKL(qφ(z|x)||pθ(z|x))+L(x;φ,θ)≥L(x;φ,θ)
Logp (x) indicates that model needs maximized log-likelihood function.Wherein, KL divergence is non-negative, and if only if
qφ(z | x)=pθ(z | x) when, σθ(z)) finally, by maximize objective function be converted into solve it is convex excellent
The problem of change.
The present invention is using stochastic gradient descent method come undated parameter training network model, loss function Loss=MSE+
DKLIn experiment, first with prior distributionThe sample of one group of hidden variable z of stochastical sampling, then inputs
To decoder, the random sample of a data point x is finally exported.The hidden unit n for considering different number, can be less than or be higher than
The quantity of primitive character.It can not only convert to the data compared with low dimensional, i.e., not exclusively indicate, can also indicate more high-dimensional
Data, i.e., excessive complete representation.
The beneficial effects of the present invention are:
1, different character representation algorithms is used to the data of different types of structure in medical text.For Traditional Chinese medical electronic disease
The different data of structure type in text are gone through, carry out character representation using different methods.For discrete features item, by half hitch
Structure data use structuring mapping method, obtain numerical value vector.For continuous characteristic item, using based on thresholding convolution variation from
Dynamic encoder algo carries out character representation, automatically extracts out the high-level semantics information in Traditional Chinese medical electronic case history.To different structure class
The case description vectorsization of type indicate after completing, and continuous characteristic vector is carried out merging for dimension with discrete features vector.Melt
Feature after conjunction can optimally indicate to be originally inputted.
2, reply medical data is distributed unbalanced status.It is distributed unbalanced problem to improve medical data, avoids counting
According to amount difference, prevents training set from having excessive deviation and ensure there are enough training datas, it can be more using convolution variation self-encoding encoder
Good learning data distribution solves medical data distribution imbalance problem.
3, personalized similar case recommended models are constructed.Similar case recommended models are constructed to each doctor respectively, are merged
Clinician user interrogation preference, solves the problems, such as individual difference.This method obtains the number of certain disease from a large amount of high-quality case library
After hundred candidate cases, sequence exports dozens of case recommendation items to doctor, by reference.
To sum up, the similar case recommended method proposed by the present invention indicated based on comprehensive characteristics with the wide depth model of improvement,
It realizes personalized case to recommend, proposes a kind of recommendation sequence that traditional shallow-layer linear model is combined with depth network model
Algorithm model improves the accuracy of similar case recommendation.
Detailed description of the invention
Fig. 1 is flow chart of the invention;
Fig. 2 is that comprehensive characteristics of the present invention indicate illustraton of model;
Fig. 3 is thresholding convolution variation self-encoding encoder illustraton of model in comprehensive characteristics of the present invention expression;
Fig. 4 is that the present invention improves wide depth model illustraton of model.
Specific embodiment
Embodiment 1: as shown in Figs 1-4, the similar case recommendation side with the wide depth model of improvement is indicated based on comprehensive characteristics
Method, specific step is as follows for the method:
Step1, medical text desensitization is carried out to text first, and medical text is segmented;By tcm characteristic vocabulary of terms,
The unique name of disease of Basic Theories of Chinese Medicine noun, Chinese medicine and disease name, hepatopathy noun and Chinese medicinal formulae noun, are added
THULAC dictionary, and segmented using THULAC, obtaining the corpus as unit of word indicates;
Step2, characteristic sub-area is carried out, discrete features is mapped to real-valued vectors;By continuous feature according to step in Step1
It is segmented, is indicated using the term vector that Word2Vec obtains corpus;
The comprehensive characteristics of Step3, building based on thresholding convolution variation self-encoding encoder indicate model;Firstly, according in Step2
Characteristic sub-area, two parts feature is carried out to the fusion of dimension, is compiled automatically wherein continuous feature is used based on thresholding convolution variation
Code device algorithm carries out character representation;Finally, the high-level semantics information obtained in Traditional Chinese medical electronic case history indicates;
Step4, building are based on the similar case recommended models for improving wide depth model;Each doctor is constructed respectively respectively
Similar case recommended models;It is indicated according to the case history comprehensive characteristics that Step3 is obtained, sequence exports dozens of case recommendation items to doctor
It is raw.
Further, the specific steps of the step Step1 are as follows:
Step1.1, privacy and Feature Selection operation are carried out to Chinese medicine electronic health record source data, remove in case history text
It is related to patient's individual privacy information, such as " name ", " admission number ", " home address ";It is screened out in conjunction with expert opinion and extraction is suffered from
Sick feature does not have contributive item, such as " occupation ", " wedding condition ", " nationality ", " physical examination ";
Step1.2, Chinese medicine associated disease pathology and medical terminology dictionary is added, work is segmented using THULAC Chinese text
Tool segments electronic health record;
Step1.3, due in electronic health record text missing values it is more, the illness feature of accurate description can not be showed, in
It is the electronic health record removed less than 150 words.
Further, the specific steps of the step Step2 are as follows:
Step2.1, " vital sign " in electronic health record data is extracted, " Gender ", the information such as " patient age ",
And its numerical value is mapped as one-dimensional vector, the discrete features as electronic health record;
Step2.2, content will be segmented obtained by Step1, term vector is mapped as using Word2vec method;And term vector is arranged
Table is arranged as matrix, the continuous feature as electronic health record.
Further, in the step Step3, continuous feature use based on thresholding convolution variation autocoder algorithm into
The specific steps of row character representation are as follows:
Step3.1, it is encoded using thresholding convolutional network, continuous feature obtained by Step2 is sent into pond layer and is compiled
Code result;And calculated using above-mentioned coding result and generate mean value and variance, generate Gaussian Profile and resampling;
Step3.2, building double stacked CNN model, the convolutional layer of nonlinear activation function is exported and be have passed through
The convolutional layer of sigmoid nonlinear activation function activation, which exports, to be multipliedWherein:
W and V indicates the weight of convolutional layer, and b and c indicate the bias term of convolutional layer, and * indicates convolution operation operation, and σ is thresholding convolution letter
Number;
Step3.3, loss function are logp (x)=DKL(qφ(z|x)||pθ(z | x)) so that final loss function reaches
To minimum;Wherein, θ is Optimal Parameters, and logp (x) indicates that model needs maximized log-likelihood function, DKLFor KL divergence, qφ
(z | x) it is encoder, pθ(x | z) it is decoder, z is hidden variable, and x is input variable;And if only if qφ(z | x)=pθ(z|x)
When, DKL(qφ(z|x)||pθ(z | x))=0;
Step3.4, network model is trained come undated parameter using stochastic gradient rise method;First with prior distributionThe sample of one group of hidden variable z of stochastical sampling, is then input to decoder, finally exports a data point
The random sample of x;Consider different number hidden unit n, can be less than or higher than primitive character quantity.
Further, the specific steps of the step Step4 are as follows:
Step4.1, Logic Regression Models are definedHere x=[x1,x2,…,xd] indicate feature d's
One group of vector contains in characteristic set and is originally inputted feature and assemblage characteristic, w=[w1,w2,…,wd] indicate model ginseng
Number;
Step4.2, cross feature is definedHere cki∈ { 0,1 } is Boolean, and such as i-th special
Sign is k-th of conversion φkA part, ckiAs 1, it is otherwise 0, for binary features, for example, having and if only if assemblage characteristic
It all sets up and is only 1, be otherwise exactly 0;
Step4.3, GRU layers of core for defining depth module, and additional feedforward is added between the last layer and output
Layer, wherein using tanh function as the activation primitive of output layer;Connection is added between node layer hiding, and is followed with a door
Ring element controls the output of concealed nodes, effective variation of the Modelling feature on time-series dynamics;
It Step4.4, by the input feature vector of shallow-layer part include by structures such as regional feature, weather time and search keys
At cross feature as input, batch Stochastic Optimization Model parameter, while propagating backward to the shallow-layer and depth door of model again
Cyclic part;
Step4.5, conjunctive model anticipation function is definedIt uses
Joint output result takes the weighted sum of logarithm as predicted value, and the weighted sum is then fed to a common loss function,
It carries out joint training and optimizes;Final output is probability value;Wherein,For wide deep all model parameters,For GRU model
Parameter;It is finally ranked up from low to high using probability, takes preceding 5 to recommend as case.
The wherein step Step5 are as follows: quality is recommended to use accuracy rate (Precision), recall rate (Recall), F1
It is worth (F-Measure) and is used as measurement index.Recommend efficiency by clinician user carry out personalized recommendation when, the training of model and
Predetermined speed is measured.
The present invention considers that the recommendation efficient design of conjunctive model is weighed using the generation unit time expense in recommendation process
Amount.I.e. respectively on data training set and test set, it is compared using the single case list average recommendation time, and and its
The advisory speed of his model compares experiment.Width depth recommended models proposed by the present invention, more due to its depth model part
Add simply, overcome the shortcomings that gradient disappears, and the problem of medical text data is unevenly distributed can be improved, has higher
Efficiency.
For clinician user u, R is enableduAs the case set that model is recommended, LuAs user u approve case set,
Recommend accuracy rate, recall rate and F1-Score value are as follows:
The true electronic health record data set that data used in this case study are collected from Traditional Chinese Medical Hospital, Yunnan Prov..Research choosing altogether
Taken wherein 2090 authentic and valid electronic health records, comprising the most common illness in department, department, the hospital of traditional Chinese hospital, have bone fracture bone fracture disease,
8 kinds of diseases such as numbness disease, injury of the tissues disease, lumbago diseases, dislocation disease, gangrene of finger or toe disease, hot numbness, suppurative osteomyelitis, the electronic health record data set being configured to.
After arranging data sample, 70% is taken as training set, 10% is used as cross validation collection, and 20% is used as test set.
Experiment: for verify auxiliary diagnosis quality, using based on thresholding convolution improvement width depth model (CFI) with
Logistic Regression, SVM traditional classification method and the character representation based on two step model of DBN+SVM and automatic
Diagnostic method compares experiment.Comparative experimental data utilizes identical ready-portioned TO-EMR data set.As a result as shown in Table X.
The comprehensive index value contrast table of the different models of table 1
In table 1, different model and method are compared, the CFI model is superior to other methods in various indexs, than
The comprehensive index value of existing method improves 2.63 percentage points.The result shows that CFI model has stronger electricity than conventional model
Sub- case history text information extractability, high-accuracy indicate that model mistaken diagnosis probability is very low, and higher recall rate indicates model leakage
Examine that probability is lower, comprehensive evaluation index F value indicates that the auxiliary diagnosis effect of Integrated Evaluation Model is prominent;CFI model with it is existing
Depth model is compared, and performance is promoted, and the method for indicating characteristic sub-area effectively raises the accuracy rate of character representation, is become
Point self-encoding encoder can preferably learn the distribution of electronic health record data, improve the ability of depth characteristic expression.In summary right
Than experimental result, CFI model is effectively demonstrated, there is significant practical value, realizes that comprehensive characteristics are indicated and carried out clinical auxiliary
It is feasible and effective for helping diagnosis.
Above in conjunction with attached drawing, the embodiment of the present invention is explained in detail, but the present invention is not limited to above-mentioned
Embodiment within the knowledge of a person skilled in the art can also be before not departing from present inventive concept
Put that various changes can be made.
Claims (5)
1. indicating the similar case recommended method with the wide depth model of improvement based on comprehensive characteristics, it is characterised in that: the method
Specific step is as follows:
Step1, medical text desensitization is carried out to text first, and medical text is segmented;By tcm characteristic vocabulary of terms, Chinese medicine
The unique name of disease of basic theory noun, Chinese medicine and disease name, hepatopathy noun and Chinese medicinal formulae noun, are added THULAC word
Library, and segmented using THULAC, obtaining the corpus as unit of word indicates;
Step2, characteristic sub-area is carried out, discrete features is mapped to real-valued vectors;Continuous feature is carried out according to step in Step1
Participle is indicated using the term vector that Word2Vec obtains corpus;
The comprehensive characteristics of Step3, building based on thresholding convolution variation self-encoding encoder indicate model;Firstly, according to the spy in Step2
Subregion is levied, two parts feature is carried out to the fusion of dimension, wherein continuous feature, which uses, is based on thresholding convolution variation autocoder
Algorithm carries out character representation;Finally, the high-level semantics information obtained in Traditional Chinese medical electronic case history indicates;
Step4, building are based on the similar case recommended models for improving wide depth model;Each doctor is constructed respectively respectively similar
Case recommended models;Indicate that sequence exports dozens of case recommendation items to doctor according to the case history comprehensive characteristics that Step3 is obtained.
2. the similar case recommended method according to claim 1 indicated based on comprehensive characteristics with the wide depth model of improvement,
It is characterized by: the specific steps of the step Step1 are as follows:
Step1.1, privacy and Feature Selection operation are carried out to Chinese medicine electronic health record source data, remove involved in case history text
Patient's individual privacy information, such as " name ", " admission number ", " home address ";It is screened out in conjunction with expert opinion special to illness is extracted
Sign does not have contributive item, such as " occupation ", " wedding condition ", " nationality ", " physical examination ";
Step1.2, Chinese medicine associated disease pathology and medical terminology dictionary is added, tool pair is segmented using THULAC Chinese text
Electronic health record is segmented;
Step1.3, due in electronic health record text missing values it is more, the illness feature of accurate description can not be showed, then gone
Fall to be less than the electronic health record of 150 words.
3. the similar case recommended method according to claim 1 indicated based on comprehensive characteristics with the wide depth model of improvement,
It is characterized by: the specific steps of the step Step2 are as follows:
Step2.1, " vital sign " in electronic health record data is extracted, " Gender ", the information such as " patient age ", and will
Its numerical value is mapped as one-dimensional vector, the discrete features as electronic health record;
Step2.2, content will be segmented obtained by Step1, term vector is mapped as using Word2vec method;And term vector list is arranged
It is classified as matrix, the continuous feature as electronic health record.
4. the similar case recommended method according to claim 1 indicated based on comprehensive characteristics with the wide depth model of improvement,
It is characterized by: continuous feature, which is used, carries out feature based on thresholding convolution variation autocoder algorithm in the step Step3
The specific steps of expression are as follows:
Step3.1, it is encoded using thresholding convolutional network, continuous feature obtained by Step2 is sent into pond layer and obtains coding knot
Fruit;And calculated using above-mentioned coding result and generate mean value and variance, generate Gaussian Profile and resampling;
Step3.2, building double stacked CNN model, by the output of the convolutional layer of nonlinear activation function with to have passed through sigmoid non-
The convolutional layer of linear activation primitive activation, which exports, to be multipliedWherein: W and V is indicated
The weight of convolutional layer, b and c indicate the bias term of convolutional layer, and * indicates convolution operation operation, and σ is thresholding convolution function;
Step3.3, loss function are logp (x)=DKL(qφ(z|x)||pθ(z | x)) so that final loss function reaches most
It is small;Wherein, θ is Optimal Parameters, and logp (x) indicates that model needs maximized log-likelihood function, DKLFor KL divergence, qφ(z|
It x) is encoder, pθ(x | z) it is decoder, z is hidden variable, and x is input variable;And if only if qφ(z | x)=pθ(z | x) when,
DKL(qφ(z|x)||pθ(z | x))=0;
Step3.4, network model is trained come undated parameter using stochastic gradient rise method;First with prior distributionThe sample of one group of hidden variable z of stochastical sampling, is then input to decoder, finally exports a data point
The random sample of x;Consider different number hidden unit n, can be less than or higher than primitive character quantity.
5. the similar case recommended method according to claim 1 indicated based on comprehensive characteristics with the wide depth model of improvement,
It is characterized by: the specific steps of the step Step4 are as follows:
Step4.1, Logic Regression Models are definedHere x=[x1,x2,…,xd] indicate one group of feature d
Vector contains in characteristic set and is originally inputted feature and assemblage characteristic, w=[w1,w2,…,wd] indicate model parameter;
Step4.2, cross feature is definedHere cki∈ { 0,1 } is Boolean, as ith feature is
K-th of conversion φkA part, ckiAs 1, it is otherwise 0, for binary features, for example, having and if only if assemblage characteristic is whole
Establishment is only 1, is otherwise exactly 0;
Step4.3, GRU layers of core for defining depth module, and additional feedforward layer is added between the last layer and output,
The middle activation primitive for using tanh function as output layer;Connection is added between node layer hiding, and single with a door circulation
Member controls the output of concealed nodes, effective variation of the Modelling feature on time-series dynamics;
It Step4.4, by the input feature vector of shallow-layer part include being made of regional feature, weather time and search key etc.
Cross feature propagates backward to the shallow-layer and depth door circulation of model as input, batch Stochastic Optimization Model parameter again
Part;
Step4.5, conjunctive model anticipation function is definedUse joint
Output result takes the weighted sum of logarithm as predicted value, and the weighted sum is then fed to a common loss function, carries out
Joint training simultaneously optimizes;Final output is probability value;Wherein,For wide deep all model parameters,For GRU model ginseng
Number;It is finally ranked up from low to high using probability, takes preceding 5 to recommend as case.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910490881.2A CN110299194B (en) | 2019-06-06 | 2019-06-06 | Similar case recommendation method based on comprehensive feature representation and improved wide-depth model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910490881.2A CN110299194B (en) | 2019-06-06 | 2019-06-06 | Similar case recommendation method based on comprehensive feature representation and improved wide-depth model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110299194A true CN110299194A (en) | 2019-10-01 |
CN110299194B CN110299194B (en) | 2022-11-08 |
Family
ID=68027589
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910490881.2A Active CN110299194B (en) | 2019-06-06 | 2019-06-06 | Similar case recommendation method based on comprehensive feature representation and improved wide-depth model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110299194B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111613339A (en) * | 2020-05-15 | 2020-09-01 | 山东大学 | Similar medical record searching method and system based on deep learning |
CN112699408A (en) * | 2020-12-31 | 2021-04-23 | 重庆大学 | Wearable device data privacy protection method based on self-encoder |
CN116189843A (en) * | 2023-04-23 | 2023-05-30 | 索思(苏州)医疗科技有限公司 | Treatment scheme recommendation method, device, system and storage medium |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140201126A1 (en) * | 2012-09-15 | 2014-07-17 | Lotfi A. Zadeh | Methods and Systems for Applications for Z-numbers |
CN105653840A (en) * | 2015-12-21 | 2016-06-08 | 青岛中科慧康科技有限公司 | Similar case recommendation system based on word and phrase distributed representation, and corresponding method |
US20170213000A1 (en) * | 2016-01-25 | 2017-07-27 | Shenzhen University | Metabolic mass spectrometry screening method for diseases based on deep learning and the system thereof |
US20170300814A1 (en) * | 2016-04-13 | 2017-10-19 | Google Inc. | Wide and deep machine learning models |
CN108647251A (en) * | 2018-04-20 | 2018-10-12 | 昆明理工大学 | The recommendation sort method of conjunctive model is recycled based on wide depth door |
CN108897834A (en) * | 2018-06-22 | 2018-11-27 | 招商信诺人寿保险有限公司 | Data processing and method for digging |
CN109447244A (en) * | 2018-10-11 | 2019-03-08 | 中山大学 | A kind of advertisement recommended method of combination gating cycle unit neural network |
WO2019047790A1 (en) * | 2017-09-08 | 2019-03-14 | 第四范式(北京)技术有限公司 | Method and system for generating combined features of machine learning samples |
-
2019
- 2019-06-06 CN CN201910490881.2A patent/CN110299194B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140201126A1 (en) * | 2012-09-15 | 2014-07-17 | Lotfi A. Zadeh | Methods and Systems for Applications for Z-numbers |
CN105653840A (en) * | 2015-12-21 | 2016-06-08 | 青岛中科慧康科技有限公司 | Similar case recommendation system based on word and phrase distributed representation, and corresponding method |
US20170213000A1 (en) * | 2016-01-25 | 2017-07-27 | Shenzhen University | Metabolic mass spectrometry screening method for diseases based on deep learning and the system thereof |
US20170300814A1 (en) * | 2016-04-13 | 2017-10-19 | Google Inc. | Wide and deep machine learning models |
WO2019047790A1 (en) * | 2017-09-08 | 2019-03-14 | 第四范式(北京)技术有限公司 | Method and system for generating combined features of machine learning samples |
CN108647251A (en) * | 2018-04-20 | 2018-10-12 | 昆明理工大学 | The recommendation sort method of conjunctive model is recycled based on wide depth door |
CN108897834A (en) * | 2018-06-22 | 2018-11-27 | 招商信诺人寿保险有限公司 | Data processing and method for digging |
CN109447244A (en) * | 2018-10-11 | 2019-03-08 | 中山大学 | A kind of advertisement recommended method of combination gating cycle unit neural network |
Non-Patent Citations (5)
Title |
---|
倪壮: "基于距离度量学习的医疗数据挖掘研究与应用", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
刘利军: "基于改进的宽深度模型的推荐方法研究", 《计算机应用与软件》 * |
王艺平: "面向中医骨科问诊的相似病例推荐方法研究", 《中国优秀博硕士学位论文全文数据库(硕士) 医药卫生科技辑》 * |
王静: "在线问诊平台相似病例推荐", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
陈培新: "文本语义的向量表示与建模方法研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111613339A (en) * | 2020-05-15 | 2020-09-01 | 山东大学 | Similar medical record searching method and system based on deep learning |
CN111613339B (en) * | 2020-05-15 | 2021-07-09 | 山东大学 | Similar medical record searching method and system based on deep learning |
CN112699408A (en) * | 2020-12-31 | 2021-04-23 | 重庆大学 | Wearable device data privacy protection method based on self-encoder |
CN116189843A (en) * | 2023-04-23 | 2023-05-30 | 索思(苏州)医疗科技有限公司 | Treatment scheme recommendation method, device, system and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110299194B (en) | 2022-11-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Sha et al. | Interpretable predictions of clinical outcomes with an attention-based recurrent neural network | |
Chattopadhyay et al. | A comparative study of fuzzy c-means algorithm and entropy-based fuzzy clustering algorithms | |
Hernández-Julio et al. | Framework for the development of data-driven Mamdani-type fuzzy clinical decision support systems | |
CN110299194A (en) | The similar case recommended method with the wide depth model of improvement is indicated based on comprehensive characteristics | |
Woodman et al. | A comprehensive review of machine learning algorithms and their application in geriatric medicine: present and future | |
Klüver | Steering clustering of medical data in a Self-Enforcing Network (SEN) with a cue validity factor | |
Kumar et al. | Gene expression data clustering using variance-based harmony search algorithm | |
Coban | A new modification and application of item response theory‐based feature selection for different machine learning tasks | |
Chattopadhyay et al. | Some studies on fuzzy clustering of psychosis data | |
Dadgar et al. | A hybrid method of feature selection and neural network with genetic algorithm to predict diabetes | |
Gulhane et al. | A Machine Learning based Model for Disease Prediction | |
Wibowo et al. | A K-Nearest Algorithm Based Application To Predict Snmptn Acceptance For High School Students In Indonesia | |
Nugraha et al. | Classification of Depression Expressions on Twitter Using Ensemble Learning with Word2Vec | |
Mandala et al. | A Study on the Development of Machine Learning in Health Analysis. | |
Azizan et al. | Hybridised Network of Fuzzy Logic and a Genetic Algorithm in Solving 3-Satisfiability Hopfield Neural Networks | |
Rendeiro et al. | Taxonomical associative memory | |
Reddy et al. | Diabetes Prediction using Extreme Learning Machine: Application of Health Systems | |
Kuang et al. | LSTM based classification model and its application for doctor-patient relationship evaluation | |
Lu et al. | Combining transformer-based model and GCN to predict ICD codes from clinical records | |
Mu et al. | Diagnosis prediction via recurrent neural networks | |
Chen et al. | Clinical knowledge graph embeddings with hierarchical structure for thyroid treatment recommendation | |
Akhila et al. | A review on sentiment analysis of Twitter data for diabetes classification and prediction | |
Szeląg | Application of the dominance-based rough set approach to ranking and similarity-based classification problems | |
Tebbe et al. | Is natural language processing the cheap charlie of analyzing cheap talk? A horse race between classifiers on experimental communication data | |
Behpour et al. | Understanding Machine Learning Through Data-Oriented and Human Learning Approaches |
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 |