CN113658688B - Clinical decision support method based on word segmentation-free deep learning - Google Patents
Clinical decision support method based on word segmentation-free deep learning Download PDFInfo
- Publication number
- CN113658688B CN113658688B CN202110965560.0A CN202110965560A CN113658688B CN 113658688 B CN113658688 B CN 113658688B CN 202110965560 A CN202110965560 A CN 202110965560A CN 113658688 B CN113658688 B CN 113658688B
- Authority
- CN
- China
- Prior art keywords
- phrase
- character
- medical record
- decision support
- clinical decision
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 238000012549 training Methods 0.000 claims abstract description 49
- 239000013598 vector Substances 0.000 claims abstract description 45
- 238000003745 diagnosis Methods 0.000 claims abstract description 41
- 238000013528 artificial neural network Methods 0.000 claims abstract description 14
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 claims description 29
- 201000010099 disease Diseases 0.000 claims description 28
- 230000008569 process Effects 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 15
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000001351 cycling effect Effects 0.000 claims description 3
- 230000002194 synthesizing effect Effects 0.000 claims 1
- 230000011218 segmentation Effects 0.000 abstract description 8
- 238000005516 engineering process Methods 0.000 abstract description 7
- 230000006978 adaptation Effects 0.000 abstract description 4
- 238000003759 clinical diagnosis Methods 0.000 abstract description 2
- 230000008676 import Effects 0.000 abstract 1
- 238000013527 convolutional neural network Methods 0.000 description 6
- 208000024891 symptom Diseases 0.000 description 6
- 230000003993 interaction Effects 0.000 description 3
- 238000003058 natural language processing Methods 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 238000003491 array Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013523 data management Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000003203 everyday effect Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000474 nursing effect Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000013439 planning Methods 0.000 description 1
- 238000004393 prognosis Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- 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
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- 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
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Epidemiology (AREA)
- Pathology (AREA)
- Primary Health Care (AREA)
- Databases & Information Systems (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The invention belongs to the technical field of medical informatization, and particularly relates to a clinical decision support method based on deep learning. The invention adopts a Chinese semantic adaptation technology which is not based on word segmentation, namely, the invention can adapt to Chinese without word segmentation on the premise of unchanged frame of the original word-based deep learning technology, and can also import known word vectors to realize more ideal Chinese semantic adaptation; the method specifically comprises the following steps: importing diagnosed medical record data from a historical medical record library; extracting features of the medical record data, and extracting all k-character phrases to generate phrase packages serving as training samples; inputting training samples into a multi-layer feedforward neural network for end-to-end training to generate a clinical decision support model; medical record data to be diagnosed is imported, and a diagnosis suggestion result is output. The invention has high clinical diagnosis accuracy, strong robustness, strong availability and generalization capability, and can rapidly realize the multi-type and multi-level clinical decision support requirements of comprehensive, special, general and emergency.
Description
Technical Field
The invention belongs to the technical field of medical informatization, and particularly relates to a clinical decision support method based on deep learning.
Background
Along with the promotion of large data planning of health medical treatment in China, medical record systems of hospitals have entered an information age, massive electronic medical record data are accumulated in the hospitals, and detailed records of patients in hospital diagnosis are included, including symptom expression, diagnosis results, treatment measures, prognosis conditions and the like, so that the medical record system has high reference value for other doctors to diagnose and treat similar patients. The number of patients in the current three-in-one hospital is large, doctors need to visit a large number of patients with similar symptoms every day, and when the number of patients is too large, the problems of low working efficiency of the doctors, even high misdiagnosis rate and the like are easy to occur.
The existing commercial software product for clinical decision support mainly carries out semantic recognition and content structuring on medical records and inspection reports through an NLP technology, then carries out comparison through a rule-based expert system, and outputs auxiliary diagnosis results according to consistency of comparison results. The drawbacks of this type of product are mainly twofold: (1) The existing NLP technology-based semantic recognition accuracy is generally, especially under the scene that a great number of professional terms and symbols exist in medical records and the expression habit is inconsistent with the conventional text, the semantic recognition accuracy is lower, and the accuracy of the finally output auxiliary diagnosis result is further affected; (2) The ultimate basis for rule-based expert systems for reasoning is guidelines, specifications, expert consensus for various diseases (> 3000 common diseases in comprehensive hospitals), these guideline documents being in continual updating. Under the actual application scene, a large number of research personnel and medical professionals are required to be continuously input to update and upgrade the system, and the cost is relatively higher.
In recent years, deep learning has been rapidly developed, and great achievement has been achieved in the fields of speech recognition, image recognition, natural language processing, and the like. The medical record data is subjected to operations such as mining, classifying, regression and the like of association rules by using a deep learning method, and an auxiliary diagnosis application is generated, so that inexperienced doctors can be helped to make diagnosis or better decisions in the disease diagnosis and treatment process.
Related researches on the development of a clinical decision support system by utilizing deep learning on Chinese electronic medical records are still in a starting stage, and particularly for multi-feature multi-category medical data sets, a plurality of technical problems still remain to be solved. Specifically, the following points are:
(1) When deep learning is used for machine training, the conventional method is to firstly perform word segmentation operation on a text to be processed based on semantics, and then enter a subsequent model for processing as an input item; the word segmentation operation has the problems that raw words are difficult to process, a word segmentation method is not unique, non-Chinese character mixed conditions are difficult to process, the subsequent model operation is influenced, and the conditions of low data fitting degree, poor generalization and the like are caused;
(2) The method for extracting word senses based on the fixed dictionary has high requirements on the data standardization degree, can not deal with the processing of non-standardized data with multiple sources and different quality in a real scene, and has poor adaptability;
(3) The cases of different departments are various and have large differences, and in the face of the reality of multiple diseases of multiple departments, various related characteristics have complex interactions, and if only some local characteristics in the whole medical record are analyzed, the suspicion of being in a million is avoided.
Disclosure of Invention
The invention aims to provide a clinical decision support method with high diagnosis accuracy, strong generalization capability and good robustness.
The clinical decision support method provided by the invention mainly realizes the following two targets:
(1) Inputting medical records and inspection reports, and outputting a plurality of diagnosis prompts (each diagnosis prompt contains a disease type and a disease probability value corresponding to the disease type) through model reasoning to provide clinical decision support for doctors;
(2) And quantitatively evaluating the diagnosis quality and the diagnosis capability of doctors, departments and medical institutions through comparison of the diagnosis results of the doctors and the model diagnosis prompts.
The clinical decision support method provided by the invention is based on a word segmentation-free deep learning technology, and particularly provides a Chinese semantic adaptation technology not based on word segmentation, which can adapt to Chinese without word segmentation on the premise that the original word-based deep learning technology framework is basically unchanged, and can also introduce known word vectors under the possible condition, so that ideal Chinese semantic adaptation is realized.
The invention provides a clinical decision support method based on word segmentation-free deep learning, which comprises the following specific steps:
s1, importing diagnosed medical record data from a history medical record library;
s2, extracting features of the diagnosed medical record data, and extracting all k-character phrases to generate phrase packages serving as training samples;
or, further performing word vector conversion on the generated phrase package to serve as a training sample;
s3, inputting training samples into a multi-layer feedforward neural network for end-to-end training to generate a clinical decision support model;
s4, importing medical record data to be diagnosed, performing feature extraction processing same as that of the step S2 on the medical record data to be diagnosed, inputting the generated phrase package into a clinical decision support model generated in the step S3, and outputting a diagnosis suggestion result.
The k-character phrase described in step S2 is determined by the following rule:
the text contained in the diagnosed medical record data is set as T, n characters are contained in the T, and w is set as i Represents the ith character in T, wherein i is equal to or more than 1 and n is equal to or less than n from w i The initial k-character phrase is from w i The first k consecutive characters, namely: w (w) i w i+ 1 ...w i+k-1 Wherein k is a positive integer of n or less, if w i The tail part close to T causes that a k-character phrase with a certain length cannot be extracted, and the k-character phrase is abandoned;
the phrase package contains a population of k-character phrases, the population of k-character phrases comprising: from w 1 Initial global k-character phrase, from w 2 Initial global k-character phrase, from w 3 The initial whole k-character phrase, and so on, until w n An initial global k-character phrase;
the characters contained in the text may be Chinese characters or non-Chinese characters.
Preferably, k is 10 or less, or 9, or 8, or 7, or 6, or 5.
More preferably, k is equal to or less than 4, and when k is equal to or less than 4, the extracted phrase includes 1-character phrase, 2-character phrase, 3-character phrase, and 4-character phrase.
Further, the diagnosed medical record data in step S1 includes extracting a diagnosed electronic medical record from a medical information management system of a hospital; unstructured is carried out on the diagnosed electronic medical record, personal sensitive information and other information are deleted, the plain text content of the electronic medical record is obtained, and the original data related to each diagnosis conclusion is synthesized into an independent text set.
Further, the word vector conversion method in step S2 specifically includes:
for each k-character phrase contained in the phrase package, querying a word vector database, if the word is found, setting the word vector of the k-character phrase as a known word vector, otherwise, randomly initializing the word vector of the k-character phrase to be a normal distribution with the mean value of 0 and the standard deviation of 0.1.
Further, in step S3, the multi-layer feedforward neural network is divided into an input layer, a hidden layer, and an output layer, where:
the input layer is used for inputting a phrase package generated in the previous step;
the number of layers of the hidden layer and the hidden units of each layer can be determined according to actual conditions;
the hidden layer is composed of at least 1 connected hidden layers, and each layer comprises at least 64 hidden units;
preferably, it consists of at least 2 connected hidden layers; more preferably, it is composed of at least 3 connected hidden layers;
preferably, each layer contains at least 128 hidden units; preferably, at least 256 hidden units are included; more preferably, at least 512 hidden units are included;
the hidden layer gives different weights to carry out weighted calculation through the strip-by-strip vectors in the training sample phrase package, adds a bias term and outputs a feature h through an activation function i I.e. h i ＝f(w×X i:j +b i ) Where w is a weight, f is a nonlinear function, b i Is biased;
and the output layer classifies the extracted features in the medical records by using a classifier to obtain the disease probability of each disease of the diagnosed medical records or medical records to be diagnosed.
The step S3 specifically includes the following substeps:
s31, inputting training samples into a multilayer feedforward neural network to perform end-to-end learning;
s32, inputting the result of the S31 into a multi-classification output layer, and outputting a diagnosis suggestion result;
s33, comparing the result output by the S32 with an original diagnosis result of the medical record, and correcting the wrong S32 result through a back propagation algorithm;
and S34, during training, 5% of training samples are reserved to monitor the training process, the sub-step S33 is cycled, and when the loss on the reserved samples stops improving, the cycling is stopped, so that the clinical decision support model training is completed.
Further, the diagnosis proposal result in step S4 includes a disease type and a disease probability value corresponding to the disease type.
The deep learning model adopts a multi-layer feedforward neural network instead of the convolutional neural network which is more commonly used at present, and can better reflect complex interaction of multi-factor and multi-symptom, unlike the convolutional neural network which can only extract local symptoms.
The invention has the technical advantages that the full-mechanized model training is realized, and the accuracy, the robustness and the generalization capability of the product are stronger because the conventional technical paths of semantic word segmentation and feature extraction are not adopted, the research and development cost is lower, and the invention has the following specific points:
(1) The clinical diagnosis accuracy is high, and the absolute accuracy of the diagnosis result is more than 95%;
(2) The robustness is strong, the requirements on the software and hardware environment and the data management level are low, and the method can be quickly adapted to the changes of guideline specifications and diagnosis and treatment paths;
(3) The availability is strong, the implementation period is short, the investment is small, the application threshold is low, and the method is suitable for basic medical structures and Internet medical scenes;
(4) The generalization capability is strong, and the requirements of multiple types and multiple layers of clinical decision support of comprehensive, specialized, general and emergency treatment can be rapidly met.
The invention is especially suitable for the auxiliary diagnosis of diseases such as pediatrics and the like. The proportion rate of pediatricians in China is low, so that the consultation task of the pediatricians is heavy, meanwhile, due to the special constitution of children, the diagnosis needs to be carried out with cautions, and the working pressure of the pediatricians is extremely high. The invention has high diagnosis rate and accurate matching degree, can provide effective auxiliary diagnosis for doctors, avoids misdiagnosis to a great extent and reduces the working pressure of the doctors.
Drawings
The foregoing and/or additional aspects and advantages of the present invention will become apparent and may be better understood from the following description of embodiments, taken in conjunction with the accompanying drawings.
Fig. 1 is a flowchart of a clinical decision support method based on word segmentation-free deep learning according to an embodiment of the invention.
Fig. 2 is a flow chart of a deep learning training process of a clinical decision support method based on word segmentation-free deep learning according to an embodiment of the invention.
Fig. 3 is a flowchart of a clinical decision support method based on word segmentation-free deep learning according to another embodiment of the present invention.
Detailed Description
The invention is further described by the following specific examples, which are not to be construed as limiting the invention.
Those skilled in the art can combine and combine the features of the different embodiments described in this specification without contradiction.
Furthermore, the terms "first," "second," and the like, are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In the description of the present invention, the meaning of "plurality" means at least two, for example, two, three, etc., unless specifically defined otherwise.
Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps of the process, and additional implementations are included within the scope of the preferred embodiment of the present invention in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order from that shown or discussed, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the embodiments of the present invention.
Logic and/or steps represented in the flowcharts or otherwise described herein, e.g., a ordered listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium may even be paper or other suitable medium upon which the program is printed, as the program may be electronically captured, via, for instance, optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.
It is to be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above-described embodiments, the various steps or methods may be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. As with the other embodiments, if implemented in hardware, may be implemented using any one or combination of the following techniques, as is well known in the art: discrete logic circuits having logic gates for implementing logic functions on data signals, application specific integrated circuits having suitable combinational logic gates, programmable Gate Arrays (PGAs), field Programmable Gate Arrays (FPGAs), and the like.
Those of ordinary skill in the art will appreciate that all or part of the steps carried out in the method of the above-described embodiments are programs that can be implemented by hardware associated with the instructions of the programs, and that when executed, comprise one or a combination of the steps of the method embodiments.
In addition, each functional unit in the embodiments of the present invention may be integrated in one processing module, or each unit may exist alone physically, or two or more units may be integrated in one module. The integrated modules may be implemented in hardware or in software functional modules. The integrated modules may also be stored in a computer readable storage medium if implemented as software functional modules and sold or used as a stand-alone product.
The above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, or the like. While embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and not to be construed as limiting the invention, and that variations, modifications, alternatives and variations may be made to the above embodiments by one of ordinary skill in the art within the scope of the invention.
The medical record in the invention is the original record of the whole process of the diagnosis and treatment of the patient in the hospital, and comprises a department, the age, the sex, the complaint, the medical history, the course record, the diagnosis result, the examination result, the doctor's advice, the operation record, the nursing record and the like. Electronic medical records refer not only to static medical record information, but also to related services provided, which are information about the lifetime health status and the healthcare behavior of individuals managed in an electronic manner, and relate to all process information of collection, storage, transmission, processing and utilization of patient information.
Embodiment 1 provides a clinical decision support method, as shown in fig. 1, comprising the following specific steps:
s1, importing diagnosed medical record data from a historical medical record library.
For example, the diagnosed medical record data may be an electronic medical record that extracts the major medical conditions of each department over the past few years from the medical information management system of the hospital. After a large amount of diagnosed electronic medical records are obtained, the diagnosed electronic medical records are unstructured, and personal sensitive information and other information are deleted to obtain the plain text content of the electronic medical records.
After medical record data diagnosed in the process is imported, the original data related to each diagnosis conclusion is synthesized into an independent text set.
S2, extracting features of the diagnosed medical record data, and extracting all k-character phrases to generate phrase packages serving as training samples.
Wherein the k-character phrase is determined by the following rule:
the text contained in the diagnosed medical record data is set as T, n characters are contained in the T, and w is set as i Represents the ith character in T, wherein i is equal to or more than 1 and n is equal to or less than n from w i The initial k-character phrase is from w i The first k consecutive characters, namely: w (w) i w i+ 1 ...w i+k-1 Wherein k is a positive integer of n or less, if w i The tail near T results in the inability to extract a certain length of k-character phrase, which is discarded.
The phrase package contains a population of k-character phrases, the population of k-character phrases comprising: from w 1 Initial global k-character phrase, from w 2 Initial global k-character phrase, from w 3 The initial whole k-character phrase, and so on, until w n The initial ensemble of k-character phrases.
The characters contained in the text may be Chinese characters or non-Chinese characters.
In this embodiment, a phrase package is generated using k.ltoreq.4, that is, all 1-character phrases, all 2-character phrases, all 3-character phrases, and all 4-character phrases, as the training samples, for which feature extraction is performed in this step.
S3, inputting the training samples into a multi-layer feedforward neural network to perform end-to-end training, and generating a clinical decision support model.
The deep learning model adopts a multi-layer feedforward neural network instead of the conventional convolutional neural network, so that complex interaction of multiple factors and symptoms can be reflected better, and unlike the convolutional neural network, only local symptoms can be extracted.
The multi-layer feedforward neural network is divided into an input layer, a hidden layer and an output layer.
Wherein the input object of the input layer is a phrase package generated in the previous step.
The number of layers of the hidden layer and the hidden unit of each layer can be determined according to practical situations. The hidden layer gives different weights to carry out weighted calculation through each phrase or vector in the training sample phrase package, adds a bias term and outputs a feature h through an activation function i I.e. h i ＝f(w×X i:j +b i ) Where w is a weight and f is a nonlinear function, such as: hyperbolic tangent function (tanh), sigmoid function, reLU function, etc., b i Is biased.
The output layer classifies the extracted features in the medical records by using a classifier to obtain the disease probability of each disease of the diagnosed medical records or medical records to be diagnosed.
Specifically, taking a classifier as an example of a softmax classifier, a full connection layer of the feedforward multilayer neural network model can be connected with the softmax classifier, and the softmax classifier classifies feature vectors of the feedforward multilayer neural network model to obtain the disease probability of each disease:
；
wherein p is i Representing the probability of the occurrence of the ith condition of the diagnosed electronic medical record, y i The i element of y, i.e., the score of the i-th condition in the electronic medical record.
The deep learning training process is shown in fig. 2, and the specific training process is as follows:
(1) Inputting training samples after feature extraction and word vector conversion into a multi-layer feedforward neural network, wherein the neural network consists of 3 connected hidden layers, each layer comprises 512 hidden units, and entering a full-connection layer to execute a deep learning training process;
(2) Inputting the result output by the hidden layer into a multi-classification output layer (softmax layer) and outputting a diagnosis suggestion result;
(3) Comparing the diagnosis proposal result output in the previous step with the original diagnosis result of the medical record, and correcting the error result through a back propagation algorithm;
(4) During training, 5% of training samples are reserved to monitor the training process, the steps are cycled, and when the loss on the reserved samples stops improving, the cycling is stopped, so that the clinical decision support model training is completed.
Random gradient descent is used to train the neural network. The network is normalized at 50% rate to reduce the input.
And analyzing the disease probability of each disease of the diagnosed electronic medical record and the doctor diagnosis result corresponding to the diagnosed electronic medical record by using a preset algorithm, and correcting the parameters of the deep convolutional neural network model and the parameters of the classifier according to the analysis result.
Specifically, in the training process, parameters of the deep convolutional neural network model and parameters of the classifier are updated by a preset algorithm such as a Back Propagation (BP) algorithm.
For example, the loss function of the output of the entire network is expressed as:
it should be noted that the entire network may be understood as a network composed of a feedforward multi-layer neural network model and a classifier. Wherein P is T Representing the output of the classifier, each P T The elements represent the probability of developing the disorder of the training outcome. The training process minimizes the LOSS function through the BP algorithm until the network converges, the LOSS is not lowered any more, the training is completed at this time, and all parameters in the whole network are reserved.
S4, importing medical record data to be diagnosed, performing feature extraction processing the same as that of the step S2 on the medical record data to be diagnosed, inputting the generated phrase package into a clinical decision support model generated in the step S3, and outputting a diagnosis suggestion result, wherein the diagnosis suggestion result comprises a disease type and a disease probability value corresponding to the disease type.
Example 2, another clinical decision support method is provided, as shown in figure 3. The training samples are optimized by adding the word vector database, so that the efficiency can be further improved.
The word vector database stores word vectors corresponding to massive medical vocabularies, and simultaneously stores the corresponding relations between the medical vocabularies and the word vectors. After each medical vocabulary of the medical record to be diagnosed is obtained, the word vectors corresponding to the medical vocabulary of the medical record to be diagnosed can be efficiently queried in the preset word vector database based on the corresponding relation between the medical vocabulary and the word vectors.
There are many chinese word vector databases available from web publishing, such as: chinese Word vector databases published by Tencerting AI Lab or on Github (https:// Github. Com/editing/Chinese-Word-Vectors) can be used in the present invention.
The word vector database used in the present invention can also be constructed by the following method.
Acquiring each medical vocabulary in a medical vocabulary bank; inputting medical vocabulary in a medical vocabulary bank into a pre-established Word2Vec model, and obtaining Word vectors corresponding to the medical vocabulary; and forming word vector samples from word vectors corresponding to the medical vocabulary, and storing the word vector samples in a preset word vector database. Because the medical word stock records a large number of medical words, correspondingly, word vectors corresponding to the large number of medical words are stored in the established preset word vector database, and the high-efficiency operation of the whole system is facilitated. In addition, word vectors corresponding to medical vocabulary can be obtained efficiently by using the Word2Vec model.
The invention is not limited in any way to the word vector database that can be employed.
Specifically, after step S2 of the previous embodiment, the following steps are added:
s21, performing word vector conversion on the obtained phrase package of the medical record data diagnosed in the step S2, and taking the phrase package as a training sample.
The word vector conversion method used in the present embodiment is specifically as follows:
for each k-character phrase contained in the phrase package, querying a word vector database, if the word is found, setting the word vector of the k-character phrase as a known word vector, otherwise, randomly initializing the word vector of the k-character phrase to be a normal distribution with the mean value of 0 and the standard deviation of 0.1.
Other steps are the same as those of embodiment 1, and will not be described again.
Example 3 the method of the above example was used to construct a clinical decision support system and model training based on 726 ten thousand outpatient medical records provided by a university at Shanghai affiliated pediatric hospital (national pediatric medical center) as a training set. After model training is completed, 13 ten thousand medical records (covering 7213 disease types) of all outpatients in a certain month of all departments of the institution are taken as medical record sets to be diagnosed, the model after training is tested, the consistency rate of the final diagnosis result reaches 93.6%, and the accuracy of the algorithm found after single-blind comparison with doctor diagnosis is superior to the average level of the three-medical-department hospitals.
The present invention is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present invention are intended to be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims (8)
1. A clinical decision support method based on word segmentation-free deep learning is characterized by comprising the following specific steps:
s1, importing diagnosed medical record data from a history medical record library;
s2, extracting features of the diagnosed medical record data, and extracting all k-character phrases to generate phrase packages serving as training samples;
or, further performing word vector conversion on the generated phrase package to serve as a training sample;
wherein the k-character phrase is determined by the following rule:
the text contained in the diagnosed medical record data is set as T, n characters are contained in the T, and w is set as i Represents the ith character in T, wherein i is equal to or more than 1 and n is equal to or less than n from w i The initial k-character phrase is from w i The first k consecutive characters, namely: w (w) i w i+1 ...w i+k-1 Wherein k is a positive integer of n or less, if w i The tail part close to T causes that a k-character phrase with a certain length cannot be extracted, and the k-character phrase is abandoned;
the phrase package contains a population of k-character phrases, the population of k-character phrases comprising: from w 1 Initial global k-character phrase, from w 2 Initial global k-character phrase, from w 3 The initial whole k-character phrase, and so on, until w n An initial global k-character phrase;
s3, inputting training samples into a multi-layer feedforward neural network for end-to-end training to generate a clinical decision support model;
the multilayer feedforward neural network is divided into an input layer, a hidden layer and an output layer, wherein:
the input layer is used for inputting a phrase package generated in the previous step;
the hidden layer is composed of at least 1 connected hidden layers, and each layer comprises at least 64 hidden units;
the output layer classifies the extracted features in the medical records by using a classifier to obtain the disease probability of each disease of the diagnosed medical records or medical records to be diagnosed;
s4, importing medical record data to be diagnosed, performing feature extraction processing same as that of the step S2 on the medical record data to be diagnosed, inputting the generated phrase package into a clinical decision support model generated in the step S3, and outputting a diagnosis suggestion result.
2. The clinical decision support method according to claim 1, wherein the character contained in the text in step S2 is a chinese character or a non-chinese character;
wherein k is 10 or less.
3. The clinical decision support method according to claim 2, wherein the diagnosed medical record data in step S1 includes extracting a diagnosed electronic medical record from a medical information management system of a hospital; unstructured is carried out on the diagnosed electronic medical record, personal sensitive information and other information are deleted, and plain text content of the electronic medical record is obtained; and synthesizing the raw data associated with each of the diagnostic conclusions into a separate text set.
4. The clinical decision support method according to claim 2, wherein the word vector conversion method in step S2 is specifically:
for each k-character phrase contained in the phrase package, querying a word vector database, if the word is found, setting the word vector of the k-character phrase as a known word vector, otherwise, randomly initializing the word vector of the k-character phrase to be a normal distribution with the mean value of 0 and the standard deviation of 0.1.
5. The clinical decision support method according to claim 1, wherein the hidden layer in step S3 is composed of at least 2 connected hidden layers; or at least 3 connected hidden layers;
each layer contains at least 128 hidden units; or at least 256 hidden units; or at least 512 hidden units.
6. The clinical decision support method according to claim 5, wherein the hidden layer in step S3 performs weighted calculation by giving different weights to each of the phrases or vectors in the training sample phrase package, and adds a bias term to output a feature h by the activation function i I.e. h i ＝f(w×X i:j +b i ) Where w is a weight, f is a nonlinear function, b i Is biased.
7. The clinical decision support method according to claim 6, wherein step S3 specifically comprises the sub-steps of:
s31, inputting training samples into a multilayer feedforward neural network to perform end-to-end learning;
s32, inputting the result of the S31 into a multi-classification output layer, and outputting a diagnosis suggestion result;
s33, comparing the result output by the S32 with an original diagnosis result of the medical record, and correcting the wrong S32 result through a back propagation algorithm;
and S34, during training, 5% of training samples are reserved to monitor the training process, the sub-step S33 is cycled, and when the loss on the reserved samples stops improving, the cycling is stopped, so that the clinical decision support model training is completed.
8. The method according to any one of claims 1 to 7, wherein the diagnosis proposal result in step S4 includes a disease type and a disease probability value corresponding to the disease type.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110965560.0A CN113658688B (en) | 2021-08-23 | 2021-08-23 | Clinical decision support method based on word segmentation-free deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110965560.0A CN113658688B (en) | 2021-08-23 | 2021-08-23 | Clinical decision support method based on word segmentation-free deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113658688A CN113658688A (en) | 2021-11-16 |
| CN113658688B true CN113658688B (en) | 2024-02-13 |
Family
ID=78491930
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110965560.0A Active CN113658688B (en) | 2021-08-23 | 2021-08-23 | Clinical decision support method based on word segmentation-free deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113658688B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116646078B (en) * | 2023-07-19 | 2023-11-24 | 中国人民解放军总医院 | Cardiovascular critical clinical decision support system and device based on artificial intelligence |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2010003097A (en) * | 2008-06-20 | 2010-01-07 | Fuji Xerox Co Ltd | Document generation support device and document creation support program |
| CN107863147A (en) * | 2017-10-24 | 2018-03-30 | 清华大学 | The method of medical diagnosis based on depth convolutional neural networks |
| CN110263323A (en) * | 2019-05-08 | 2019-09-20 | 清华大学 | Keyword abstraction method and system based on the long Memory Neural Networks in short-term of fence type |
-
2021
- 2021-08-23 CN CN202110965560.0A patent/CN113658688B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2010003097A (en) * | 2008-06-20 | 2010-01-07 | Fuji Xerox Co Ltd | Document generation support device and document creation support program |
| CN107863147A (en) * | 2017-10-24 | 2018-03-30 | 清华大学 | The method of medical diagnosis based on depth convolutional neural networks |
| CN110263323A (en) * | 2019-05-08 | 2019-09-20 | 清华大学 | Keyword abstraction method and system based on the long Memory Neural Networks in short-term of fence type |
Non-Patent Citations (1)
| Title |
|---|
| 倪晓华 ; .利用GATE的XML配置文件实现病历短语抽取的机器学习方法.中国医疗设备.2017,(第07期),第124-126页. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113658688A (en) | 2021-11-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11810671B2 (en) | System and method for providing health information | |
| CN107863147B (en) | Medical diagnosis method based on deep convolutional neural network | |
| US11610678B2 (en) | Medical diagnostic aid and method | |
| US11631497B2 (en) | Personalized device recommendations for proactive health monitoring and management | |
| EP3734604A1 (en) | Method and system for supporting medical decision making | |
| Agarwal et al. | A natural language processing framework for assessing hospital readmissions for patients with COPD | |
| US7840511B2 (en) | Learning or inferring medical concepts from medical transcripts using probabilistic models with words or phrases identification | |
| US11915127B2 (en) | Prediction of healthcare outcomes and recommendation of interventions using deep learning | |
| US20180121618A1 (en) | System and method for extracting oncological information of prognostic significance from natural language | |
| KR20190102399A (en) | System and method for interpreting medical images through the generation of refined artificial intelligence reinforcement learning data | |
| Khan et al. | Chronic disease prediction using administrative data and graph theory: The case of type 2 diabetes | |
| Fang et al. | Feature Selection Method Based on Class Discriminative Degree for Intelligent Medical Diagnosis. | |
| CN111316281A (en) | Semantic classification of numerical data in natural language context based on machine learning | |
| US20230343426A1 (en) | Method and system for automatically generating a section in a radiology report | |
| US20160110502A1 (en) | Human and Machine Assisted Data Curation for Producing High Quality Data Sets from Medical Records | |
| US20210027889A1 (en) | System and Methods for Predicting Identifiers Using Machine-Learned Techniques | |
| Gupta et al. | Clinical decision support system to assess the risk of sepsis using tree augmented Bayesian networks and electronic medical record data | |
| Faris et al. | Classification of Arabic healthcare questions based on word embeddings learned from massive consultations: a deep learning approach | |
| US11281855B1 (en) | Reinforcement learning approach to decode sentence ambiguity | |
| CN113658688B (en) | Clinical decision support method based on word segmentation-free deep learning | |
| US11809826B2 (en) | Assertion detection in multi-labelled clinical text using scope localization | |
| US11727685B2 (en) | System and method for generation of process graphs from multi-media narratives | |
| US20230298589A1 (en) | Ai platform for processing speech and video information collected during a medical procedure | |
| WO2024006572A1 (en) | Apparatus and a method for detecting associations among datasets of different types | |
| JP2024061599A (en) | A system for identifying abnormalities in the course of medical treatment based on a hierarchical 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 |