Multi-label disease diagnosis support system based on doctor decision making pattern recognition

Abstract

To provide a multi-label disease diagnosis support system based on doctor decision making pattern recognition.SOLUTION: A multi-label disease diagnosis support system based on doctor decision making pattern recognition is disclosed, which includes a doctor individual case list construction module, a doctor individual decision making pattern fitting module, a common component extraction and individual component separation module, and a disease diagnosis support module. The invention extracts a common component existing commonly in different decision making patterns from an individual decision making pattern by individual decision making pattern recognition, and overcomes adverse effects caused by subjective clinical experience to disease diagnosis support by separating more individual components dominated by the subjective clinical experience of doctors. The invention recognizes the individual decision making patterns of the doctors, overcomes field characteristics by dividing common components and individual components in terms of doctors, and can achieve more accurate disease diagnosis support by utilizing the common components.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to the field of medical and health information technology, and in particular to a multi-label disease diagnosis support system based on physician decision-making pattern recognition. [Background technology]

[0002] With the continuous improvement of medical information technology, a large amount of patient treatment and clinical decision-making data has been accumulated in medical information systems such as electronic medical records, and the research of clinical diagnosis support methods based on clinical big data and artificial intelligence to promote the realization of high-level disease prevention and control has become the focus of research in the field. Researchers often use models and methods such as machine learning and deep learning to analyze and mine electronic medical record data to build disease diagnosis decision-making support models, methods and systems.

[0003] Current disease diagnosis support methods and systems generally use the entire electronic medical record data directly to model, and the modeling effect is affected by the problem of heterogeneity of doctor decision-making in the clinical decision-making process. Electronic medical record data is a record of the clinical process, and the clinical disease diagnosis data stored in it is derived from the doctor's clinical decision-making, and is not gold standard data. In the clinical diagnosis process, the decision-making pattern of each doctor is not only related to objective medical knowledge, but also depends heavily on the doctor's subjective judgment, and is influenced by the doctor's specialty, past experience, favorite diseases, prudence, etc., so there are huge differences in both decision-making logic and data completeness.

[0004] Research has revealed that there are obvious differences and individual tendencies among doctors' clinical decision-making logic, that is, heterogeneity in doctor decision-making. For example, TIFF2024174836000002.tif792 shows that a study of clinical hyperactivity disorder showed that individual factors of the doctor caused a 28% difference in the diagnosis and management of hyperactivity disorder. TIFF2024174836000003.tif1192 shows that a retrospective analysis of electronic medical record data revealed that factors in physician decision-making, such as disease-department inconsistency and lack of knowledge of the corresponding disease, made it difficult for patients with chronic kidney disease with indicator abnormalities to receive a timely diagnosis, and the undiagnosed rate reached 61.4%.

[0005] As is clear from the above research results, if the model is built by directly integrating clinical data without considering the heterogeneity of doctor's decision-making in the clinical decision-making process, the disease detection effect will be greatly restricted. The data in the electronic medical record only shows the electronic record of the diagnosis given by the doctor who visited the patient, and is affected by the heterogeneity of doctor's decision-making in the clinical process. In the most intuitive expression, if a patient lacks a diagnosis record of a certain disease in the electronic medical record data, it does not mean that the patient does not actually suffer from the disease.

[0006] Technical solutions close to the present invention include: 1) In the weakly supervised learning method, positive and negative diagnostic samples are simultaneously corrected for errors using a data correction method, and a model is directly based on incomplete diagnostic data; 2) Semi-supervised learning methods use a small portion of validated medical data and a large portion of unlabeled samples to model a small number of samples with precision. 3) In positive sample-unlabeled learning, negative samples are regarded as unlabeled samples and positive samples are treated as refinement samples, and false negative samples are modeled using methods such as recognition or maximum likelihood estimation.

[0007] The above conventional techniques tentatively perform learning and modeling on clinical data using weakly supervised, semi-supervised, positive-unlabeled, etc. data, taking into account domain characteristics such as inconsistency and incompleteness of clinical data, and have better effects than direct supervised learning. However, the above techniques are tentatively deployed based on general-purpose data, and rarely consider the heterogeneity of doctor decision-making that generally exists.

[0008] The individual clinical decision-making patterns of doctors have field characteristics such as "mixture of subjective and objective factors" and "cannot be offset by each other in the training task by simple data integration", and it is difficult to eliminate the data deviation caused by the difference in doctor decision-making patterns by data integration training. Therefore, the above-mentioned assumptions based on generic data and the technology based on generic machine learning methods have difficulty in achieving sufficient analysis and mining of electronic medical record data in the heterogeneity of doctor decision-making, and the disease diagnosis support ability of the constructed system is limited.

[0009] The mixture of subjective and objective factors means that a doctor's individual clinical decision-making pattern is the result of a mixture of objective medical knowledge and the doctor's subjective clinical experience, of which the subjective clinical experience affects disease diagnosis support. The present invention uses individual decision-making pattern recognition to extract common components that exist in common among different decision-making patterns from the individual decision-making patterns, and separates more individual components that are dominated by the doctor's subjective clinical experience, thereby overcoming the adverse effects that may be caused by subjective clinical experience on disease diagnosis support.

[0010] The fact that simple data integration cannot cancel each other out in the training task means that there are differences between doctors' subjective clinical experiences, but these subjective clinical experiences are not randomly generated, so the differences cannot be offset by big data integration. The conventional similar technical solutions of the present invention recognize the inconsistency and incompleteness of data, and 1) satisfy random distribution in the global domain, or 2) satisfy random distribution in each local domain, thereby modeling disease diagnosis support in a simple data integration manner. The above similar technical solutions do not actually meet the actual generation logic of clinical data, so that the effect is greatly limited. The present invention overcomes the characteristics of the field by recognizing the individual decision-making patterns of doctors and dividing them into common components and individual components in the doctor's perspective, and can achieve more effective disease diagnosis support by utilizing common components.

[0011] In general, the present invention is based on physician decision-making pattern recognition, and by analyzing and modeling clinical data in accordance with clinical practice, it realizes fitting to the physician individual decision-making pattern, and obtains the individual components and common components in the physician individual decision-making pattern, thereby overcoming the problem of heterogeneity in physician decision-making in clinical data, and realizing more accurate disease diagnosis support through the common components. Summary of the Invention [Problem to be solved by the invention]

[0012] In response to the shortcomings of the prior art, the objective of the present invention is to provide a multi-label disease diagnosis support system based on physician decision-making pattern recognition, overcome the problem of heterogeneity in physician decision-making in clinical data, and realize more accurate disease diagnosis support. [Means for solving the problem]

[0013] The objective of the present invention is realized by the following technical solution: A multi-label disease diagnosis support system based on doctor decision-making pattern recognition, comprising: a doctor individual case collection construction module, a doctor individual decision-making pattern fitting module, a common component extraction and individual component separation module, and a disease diagnosis support module; The doctor individual case collection construction module is used for dividing the entire case collection in the electronic medical record into doctor individual case collections according to doctors; The doctor individual decision-making pattern fitting module is used for fitting an individual decision-making pattern in the clinical process of each doctor through machine learning modeling based on a doctor individual case collection; The common component extraction and individual component separation module is used to extract common components in each physician decision-making pattern through machine learning modeling and separate individual components; The physician individual decision-making pattern fitting module comprises: Construct a machine learning model, separate one common component branch and multiple individual component branches from the model, define mapping relationships between the inputs and outputs of all branches, train the model using a distribution similarity loss function, and minimize the product of the output of the common component branch and the output of each individual component branch and the distribution difference in each branch of the predicted result of the individual decision-making pattern; After completing the model training, the common component branch mapping relationship reflects the common components common to all doctors' decision patterns, and each individual component branch mapping relationship reflects the specific individual components of each doctor; The disease diagnosis support module is used to obtain common components and individual components corresponding to the input vectors of cases based on the trained model, and to realize disease diagnosis support.

[0014] Furthermore, in the doctor individual case collection construction module, case data from electronic medical records is acquired and matrixed into a total case collection including multiple cases, all cases of any one doctor are extracted from the total case collection, and a doctor individual case collection corresponding to that doctor is constructed, and the doctor individual case collection includes a feature matrix and a label matrix of diagnosis cases by that doctor.

[0015] Furthermore, in the doctor individual decision-making pattern fitting module, fitting is performed using a set of machine learning models, and each model learns the individual decision-making pattern of the corresponding doctor based on the corresponding doctor individual case collection, and after training is completed, a prediction matrix is ​​obtained consisting of prediction results in different doctor individual decision-making patterns of the input vectors of the cases. Alternatively, a single machine learning model is used for fitting, the model adopts a multi-branch network structure and includes a control assembly, and during the training process, the control assembly ensures that any one branch learns the individual decision-making pattern of the corresponding doctor based only on the corresponding individual doctor case collection, and after training is completed, a prediction matrix is ​​obtained, which consists of prediction results for different individual doctor decision-making patterns of the input vector of the case.

[0016] Furthermore, in the common component extraction and individual component separation module, a machine learning model is constructed, the input layer of the model is divided into two branches, one branch is a common component branch, and the other branch is further divided into multiple individual component branches, mapping relationships from the input layer of the model to the output layer corresponding to all branches are respectively defined, and model training is performed using a distribution similarity loss function.

[0017] Furthermore, the distribution similarity loss function adopts the KL divergence loss function or the JS divergence loss function.

[0018] Furthermore, based on the distribution similarity loss function, a multi-label loss function component for disease relationship learning is added, and model training is performed jointly.

[0019] Furthermore, in the common component extraction and individual component separation module, the constructed machine learning model is divided into a plurality of machine learning models for associative learning.

[0020] Furthermore, the disease diagnosis support module includes one fixed model whose structure and parameters are completely consistent with the machine learning model trained in the common component extraction and individual component separation module, and the input vector of the case is input to the fixed model, a common prediction result is obtained based on the mapping relationship of the common component branches, and the probability of an individual disease being found under the input conditions of each physician is obtained based on the mapping relationship of each individual component branch.

[0021] Furthermore, the disease diagnosis support module includes the following interface: A common prediction probability output interface is used to output the common prediction result and provide a reference probability of incidence of each disease in diagnostic support; A physician diagnosis probability output interface that outputs the probability of an individual disease being observed under each physician's input conditions; and a doctor recommendation level interface that outputs a sorted array of corresponding doctors sorted in descending order according to the probability of finding an individual disease under the input conditions of each doctor, and provides a suitable doctor recommendation service. Effect of the Invention

[0022] The beneficial effects of the present invention are as follows: 1. By constructing an innovative multi-label method for recognizing physician decision-making patterns from the perspective of individual physician decision-making patterns, the heterogeneity of physician decision-making can be overcome, common components in individual physician decision-making patterns can be extracted, individual components can be separated, and the subjective and objective factors in individual physician decision-making patterns can be separated and recognized. 2. Through decision-making pattern recognition, the shortcomings of traditional similar methods in data utilization through simple data integration are overcome, and more objective and effective extraction of common components in physician decision-making patterns is realized. A multi-label disease diagnosis support system is built around the extracted common components, thereby achieving more accurate disease diagnosis support. [Brief description of the drawings]

[0023] [Figure 1] FIG. 1 is a configuration diagram of a multi-label disease diagnosis support system based on physician decision-making pattern recognition shown in an exemplary embodiment. [Diagram 2] 1 is a framework for fitting individual physician decision-making patterns using a set of machine learning models as presented in one illustrative embodiment. [Diagram 3] 1 is a framework for fitting individual physician decision-making patterns using a single machine learning model as shown in one illustrative embodiment. [Figure 4] FIG. 2 is an architecture diagram of a machine learning model G shown in one illustrative embodiment. [Diagram 5] FIG. 11 is a comparison diagram of disease diagnosis support effects shown in an exemplary embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] In order to better understand the technical solution of the present application, the following describes in detail the embodiments of the present application with reference to the accompanying drawings.

[0025] It should be clear that the described embodiments are only some of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments that a person skilled in the art can obtain without inventive efforts fall within the scope of protection of the present application.

[0026] The terms used in the examples of this application are for the purpose of describing particular examples only and are not intended to limit the present application. The singular forms "a," "the," and "the" used in the examples of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0027] An embodiment of the present invention provides a multi-label disease diagnosis support system based on doctor decision-making pattern recognition, and as shown in Figure 1, the system includes a doctor individual case collection construction module, a doctor individual decision-making pattern fitting module, a common component extraction and individual component separation module, and a disease diagnosis support module. The doctor individual case collection construction module is used for dividing the entire case collection in the electronic medical record into doctor individual case collections according to the doctor, the doctor individual decision-making pattern fitting module is used for fitting the individual decision-making pattern of each doctor through machine learning modeling, the common component extraction and individual component separation module is used for extracting the common components in each doctor decision-making pattern and separating the individual components through machine learning modeling, and the disease diagnosis support module is used for obtaining the common components and individual components corresponding to the input vector based on the settlement model, and is used for the disease diagnosis support service. The specific functions of each module are as follows:

[0028] 1) Doctor individual case collection construction module Extract case data from electronic medical records stored in .csv files or databases such as Oracle, import them into the system, and create a whole case collection containing N cases. TIFF2024174836000004.tif729 and each case is The whole case collection D is written as TIFF2024174836000005.tif740. TIFF2024174836000006.tif843 and label matrix TIFF2024174836000007.tif843 and the feature matrix In TIFF2024174836000008.tif842, TIFF2024174836000009.tif622 is the input vector for each example, and the range of possible values ​​for each input vector component is real numbers. TIFF2024174836000010.tif66, features include age, name, BMI, blood normal index, urine normal index, etc., T denotes transposition, label matrix In TIFF2024174836000011.tif841, TIFF2024174836000012.tif522 is the output vector for each case, and the range of possible values ​​for each output vector component is TIFF2024174836000013.tif715, 0 indicates that the case does not have a diagnosis of the disease corresponding to the output vector component, and 1 indicates that the case has a diagnosis of the disease corresponding to the output vector component. The total number of doctors related to the entire case collection D is defined as K, and each doctor is Identified by TIFF2024174836000014.tif622. Any one doctor from the entire case collection D Extracting all the cases in TIFF2024174836000015.tif741 will result in a doctor individual case collection corresponding to the doctor in question. TIFF2024174836000016.tif736 can be constructed, TIFF2024174836000017.tif716 is a doctor Feature matrix and label matrix of a diagnosis case from TIFF2024174836000018.tif57.

[0029] 2) Physician individual decision-making pattern fitting module A set or a single machine learning model M is constructed, and a neural network architecture or a non-neural network architecture is used as the machine learning model, and the model M can fit the individual decision-making patterns of doctors in the clinical process by learning the doctor's decision-making patterns based on a collection of individual doctor cases.

[0030] When fitting with a set of machine learning models, as shown in Figure 2, TIFF2024174836000019.tif744, and each model TIFF2024174836000020.tif743 is the corresponding doctor individual case collection Based on TIFF2024174836000021.tif736, The doctor completed the training by following the process shown in TIFF2024174836000022.tif733. We simulated the individual decision-making patterns of TIFF2024174836000023.tif67. After training is completed, for any one of the input vectors x, the model M will select different doctors. Prediction of x in the individual decision-making pattern of TIFF2024174836000024.tif67 Prediction matrix consisting of TIFF2024174836000025.tif77 You can get TIFF2024174836000026.tif8124.

[0031] When fitting using a single machine learning model, M may select a multi-branch network structure, as shown in Figure 3, with K branches. Includes TIFF2024174836000027.tif825, control assembly gate, and full case study Based on TIFF2024174836000028.tif729, The training is completed by the process shown in TIFF2024174836000029.tif746. In the training process, the multiple selectable control assembly Gate selects one of the branches. TIFF2024174836000030.tif743 is a doctor After the training is completed, for any one input vector x, the model M can determine the decision patterns of different doctors. Branch corresponding to TIFF2024174836000032.tif67 Prediction result of x in TIFF2024174836000033.tif78 Prediction matrix consisting of TIFF2024174836000034.tif77 You can get TIFF2024174836000035.tif9125.

[0032] Regardless of the modeling and training method used above, TIFF2024174836000036.tif874 is the doctor predicted by the model TIFF2024174836000037.tif67 shows the probability of diagnosing the corresponding disease in the clinical process. The cross-entropy loss function is used to keep the output probability value of TIFF2024174836000038.tif77 within the interval (0,1). It is recommended to optimize using TIFF2024174836000039.tif711.

[0033] 3) Common component extraction and individual component separation module This module builds and trains a machine learning model G to The individual decision-making pattern of TIFF2024174836000040.tif67 is divided into a common component f that is common to all doctors' patterns and an individual component f that is specific to the doctor. Separate into two components: TIFF2024174836000041.tif67.

[0034] Theoretically, for any unknown input vector x, 1) a mapping of common components to x TIFF2024174836000042.tif852 shows the probability that a patient actually has the disease, and 2) the mapping of individual components to x TIFF2024174836000043.tif776 shows the medical history of the disease when a patient actually suffers from it. TIFF2024174836000044.tif67 indicates the probability of diagnosing the disease in question, and 3) TIFF2024174836000045.tif774 is the doctor predicted by the model TIFF2024174836000046.tif67 shows the probability of diagnosing the corresponding disease in the clinical process. At the same time, assuming that all diseases diagnosed by doctors in the clinical process are accurate, The result is TIFF2024174836000047.tif764.

[0035] Based on the above definitions and assumptions, Bayes' theorem TIFF2024174836000048.tif5123 It can be rewritten to TIFF2024174836000049.tif749.

[0036] The construction of a machine learning model G logically involves two stages of branching. As shown in Figure 4, in the case of the first stage of branching, the input layer of model G is TIFF2024174836000050.tif85 and TIFF2024174836000051.tif75 is split into two branches, and in the case of the second branch, TIFF2024174836000052.tif75 branch has K sub-branches The input layer of model G is divided into TIFF2024174836000053.tif823. The mapping relationship up to the output layer corresponding to TIFF2024174836000054.tif829 is We define it as TIFF2024174836000055.tif856. Then, we perform model training using a distribution similarity loss function L that has the ability to measure the distribution similarity. TIFF2024174836000056.tif834 and obtained via Model M The difference in distribution between each branch and TIFF2024174836000057.tif77 is minimized. In this embodiment, the KL divergence loss function is selected as the distribution similarity loss function, i.e., TIFF2024174836000058.tif1192, except TIFF2024174836000059.tif718 is the KL divergence loss function. After training is complete, the model will TIFF2024174836000060.tif813 reflects the common component f common to all physician patterns, and the mapping relationship Doctor by TIFF2024174836000061.tif716 TIFF2024174836000062.tif57 Unique individual components Reflected TIFF2024174836000063.tif67.

[0037] The feature of this module is that it uses the machine learning training process to generate mapping relationships. Using TIFF2024174836000064.tif813, we reflect the common component f and Using TIFF2024174836000065.tif716, each individual component A model structure reflecting TIFF2024174836000066.tif67 is constructed, and the model may use a neural network architecture or a non-neural network architecture. Based on a certain distribution similarity loss function L, modeling prediction results of individual decision-making patterns. TIFF2024174836000067.tif87 and the product of two components The advantage of this is that it is possible to minimize the distribution difference with TIFF2024174836000068.tif933. In actual operation, the above branching process of this module may be integrated into the same step, i.e., from model G to Directly separate K+1 branches containing TIFF2024174836000069.tif829. The machine learning model G can be divided into multiple machine learning models to perform federated learning, and the model structure of each branch is not limited, and selectable model structures include, but are not limited to, multi-layer neural network (MLP), radial basis function neural network (RBF), etc. The selection of the distribution similarity loss function L is not limited, and includes, but is not limited to, KL divergence (Kullback-Leibler divergence), JS divergence (Jensen-Shannon divergence), etc. In addition, it is possible to add a multi-label loss function component of disease relationship learning, such as BP-MLL loss function (Backpropagation for Multi-Label Learning).

[0038] 4) Disease diagnosis support module This module includes one established model whose structure and parameters are completely consistent with the machine learning model G, one common prediction probability output interface, one doctor diagnosis probability output interface, and one doctor recommendation level interface.

[0039] After any one of the input vectors x that calls the disease diagnosis support service is input to the module, the mapping relationship is first obtained using the settling model. Common prediction results from TIFF2024174836000070.tif813 Get TIFF2024174836000071.tif913 and extract each mapping relationship TIFF2024174836000072.tif717 for each doctor Probability of finding an individual disease under the input conditions of TIFF2024174836000073.tif67 Then, the common prediction probability output interface is used to directly provide the reference probability of each disease in diagnostic support. TIFF2024174836000075.tif913 is output as the common prediction result, and each doctor is Probability of finding an individual disease under the input conditions of TIFF2024174836000076.tif67 Output TIFF2024174836000077.tif716 and use the doctor recommendation interface A sorted array of corresponding doctors sorted in descending order according to TIFF2024174836000078.tif716 is output to provide a suitable doctor recommendation service.

[0040] In one embodiment, the Area Under Receiver Operating Characteristic Curve (AUC ROC ) to compare the disease diagnosis support effect of the invented system with that of a multilayer neural network (MLP) and a radial basis function neural network (RBF) that have not been processed by the common component extraction and individual component separation modules. As shown in Figure 5, the comparison results show that the disease diagnosis support ability of the present system is superior to the two general-purpose machine learning models used for comparison.

[0041] The present invention focuses on doctors, extracts patient data, constructs individual doctor case collections corresponding to different doctors, and constructs a multi-label disease diagnosis support system from the perspective of doctor decision-making patterns.The present invention realizes the clarification and utilization of common components and individual components through two processes: individual doctor decision-making pattern fitting, and common component extraction and individual component separation in the decision patterns.

[0042] The above describes certain embodiments of the present specification. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the examples and still achieve desirable results. Also, the processes depicted in the accompanying figures do not necessarily require the particular order shown or sequential order to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or advantageous.

[0043] The terms used in one or more embodiments herein are for the purpose of describing particular embodiments only and are not intended to limit one or more embodiments herein. The singular forms "a," "said," and "the" used in one or more embodiments herein and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. The term "and / or" used herein should be understood to mean any or all possible combinations of one or more associated listed items.

Claims

1. A multi-label disease diagnosis support system based on doctor decision-making pattern recognition, comprising: a doctor individual case collection construction module, a doctor individual decision-making pattern fitting module, a common component extraction and individual component separation module, and a disease diagnosis support module; The doctor individual case collection construction module is used for dividing the entire case collection in the electronic medical record into doctor individual case collections according to doctors, and the total number of doctors related to the entire case collection is K, and each doctor is Identify with The doctor individual decision-making pattern fitting module is used to build a set or a single machine learning model M based on the doctor individual case collection, and fit the individual decision-making pattern in the clinical process of each doctor through machine learning modeling; The common component extraction and individual component separation module is used to extract common components that indicate objective medical knowledge commonly present in each doctor's decision-making pattern through machine learning modeling, separate individual components that are dominated by the doctor's subjective clinical experience, and perform separation and recognition of subjective and objective factors in the doctor's individual decision-making pattern; The common component extraction and individual component separation module comprises: The construction of the machine learning model G involves two stages of branching. In the first stage of branching, the input layer of the machine learning model G is and It is divided into two branches: is the common component branch, and in the case of the second stage branch, A branch has K sub-branch The input layer of the machine learning model G is divided into K individual component branches. The mapping relationship to the output layer corresponding to where x is the input vector of cases, the mapping of the common component to x indicates the probability that the patient actually has the disease, and the mapping of the individual component to x indicates the probability that the patient actually has the disease, The model is trained using a distribution similarity loss function that has the ability to measure the distribution similarity and outputs the common component branch. and the output of each individual component branch Product of and each doctor obtained through the machine learning model M Prediction of individual decision-making patterns and minimizing the difference in distributions at each branch of After completing the model training, the common component branch mapping relationship reflects the common components common to all doctors' decision patterns, and each individual component branch mapping relationship reflects the specific individual components of each doctor; The disease diagnosis support module includes a fixed model whose structure and parameters are completely consistent with the trained machine learning model G in the common component extraction and individual component separation module, a common prediction probability output interface, a doctor diagnosis probability output interface, and a doctor recommendation level interface; After an input vector x of any one of the cases for which a disease diagnosis support service is called is input to the disease diagnosis support module, a mapping relationship of common component branches is first calculated using a settling model. The common prediction result and obtain the mapping relationship for each individual component branch. Each doctor Probability of individual disease presence under the input conditions Get Next, the common prediction probability output interface is used to directly provide the reference probability of each disease in diagnostic support. is output as a common prediction result, and each doctor is Probability of individual disease presence under the input conditions and the doctor recommendation interface , and provide a suitable doctor recommendation service by outputting a sorted array of corresponding doctors sorted in descending order according to A multi-label disease diagnosis support system based on physician decision-making pattern recognition.

2. In the doctor individual case collection construction module, case data from the electronic medical record is acquired and matrixed into a total case collection including a plurality of cases, all cases of any one doctor are extracted from the total case collection, and a doctor individual case collection corresponding to the doctor is constructed, and the doctor individual case collection includes a feature matrix and a label matrix of diagnosis cases by the doctor. The multi-label disease diagnosis support system based on physician decision-making pattern recognition according to claim 1.

3. In the doctor individual decision-making pattern fitting module, a set of machine learning models is used to perform fitting, and each model learns the individual decision-making pattern of the corresponding doctor based on the corresponding doctor individual case collection, and after training is completed, a prediction matrix is ​​obtained, which is composed of prediction results in different doctor individual decision-making patterns of the input vectors of the cases. The multi-label disease diagnosis support system based on physician decision-making pattern recognition according to claim 1.

4. In the doctor individual decision-making pattern fitting module, a single machine learning model is used for fitting, the model adopts a multi-branch network structure, and includes a control assembly. In the training process, the control assembly ensures that any one branch learns the corresponding doctor individual decision-making pattern only based on the corresponding doctor individual case collection. After the training is completed, a prediction matrix is ​​obtained, which is composed of prediction results in different doctor individual decision-making patterns of the input vector of the case. The multi-label disease diagnosis support system based on physician decision-making pattern recognition according to claim 1.

5. The distribution similarity loss function adopts the KL divergence loss function or the JS divergence loss function. The multi-label disease diagnosis support system based on physician decision-making pattern recognition according to claim 1.

6. Based on the distribution similarity loss function, a multi-label loss function component for disease relationship learning is added to jointly train the model. The multi-label disease diagnosis support system based on physician decision-making pattern recognition according to claim 1.

7. In the common component extraction and individual component separation module, the constructed machine learning model is divided into multiple machine learning models to perform federated learning. The multi-label disease diagnosis support system based on physician decision-making pattern recognition according to claim 1.

Images