ICD Coding Method and Device Based on Multi-Task Learning and Graph Attention Network

Abstract

The present invention provides an ICD coding method and apparatus based on multi-task learning and graph attention network. The method includes: obtaining medical texts corresponding to at least one coding prediction task; the at least one coding prediction task includes an ICD coding task; constructing corresponding text graphs according to the medical texts corresponding to the at least one coding prediction task; inputting the text graphs into a coding prediction model to obtain coding prediction results; the coding prediction results include ICD coding results, and the coding prediction model is obtained by training the graph attention network based on training data. The ICD coding method and apparatus based on multi-task learning and graph attention network provided by the present invention can alleviate the problem of data imbalance because the medical texts are converted into text graph form before coding prediction, effectively improve the automatic coding effect of medical texts, and the coding results are more accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of medical text processing, and in particular, to an ICD coding method and device based on multi-task learning and graph attention network. Background Art

[0002] The International Classification of Diseases (ICD) is an internationally unified disease classification method and an important part of the health information standard system. In the field of medical text processing, the ICD coding of medical texts is a very important task, usually completed by professional ICD coders. The coder needs to refer to detailed and heavy medical record information before giving the correct coding result, which is time-consuming, laborious and inefficient.

[0003] Therefore, the automatic ICD coding technology has emerged. However, the existing automatic ICD coding methods ignore the problem of data imbalance in medical texts, and the obtained ICD coding results are not accurate enough.

[0004] Therefore, there is an urgent need for an ICD coding method that can alleviate the problem of data imbalance in medical texts and obtain more accurate coding results. Summary of the Invention

[0005] The present invention provides an ICD coding method and device based on multi-task learning and graph attention network to solve the defect that the existing automatic ICD coding method ignores the problem of data imbalance in medical texts and the obtained ICD coding results are not accurate enough.

[0006] In a first aspect, the present invention provides an ICD coding method based on multi-task learning and graph attention network, the method comprising:

[0007] Obtaining medical texts corresponding to at least one coding prediction task; wherein, the at least one coding prediction task includes an ICD coding task;

[0008] Constructing a corresponding text graph according to the medical texts corresponding to the at least one coding prediction task;

[0009] Inputting the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model;

[0010] wherein, the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data.

[0011] An ICD coding method based on multi-task learning and graph attention network provided by the present invention constructs a corresponding text graph according to the medical text corresponding to at least one coding prediction task, including:

[0012] Obtain concepts related to the medical text according to the medical text corresponding to the coding prediction task;

[0013] Identify the relationships between each concept and the medical text respectively, and determine the relationships between each concept;

[0014] Construct a corresponding text graph by taking the medical text and the concepts as nodes, and taking the relationships between each concept and the medical text and the relationships between each concept as edges.

[0015] An ICD coding method based on multi-task learning and graph attention network provided by the present invention inputs the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model, including:

[0016] Input the text graph corresponding to at least one coding prediction task into the graph attention network coding layer; wherein, the graph attention network coding layer includes multiple sub-coding layers for processing different levels of coding prediction tasks;

[0017] Process the text graph corresponding to the coding prediction task through the corresponding sub-coding layer according to the processing level corresponding to the coding prediction task to obtain a coding prediction result.

[0018] An ICD coding method based on multi-task learning and graph attention network provided by the present invention inputs the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model, including:

[0019] Input the text graph corresponding to at least one coding prediction task into the corresponding specific task coding layer for processing to obtain at least one specific task processing result;

[0020] Input the text graph corresponding to at least one coding prediction task into the global shared coding layer for processing to obtain a global shared processing result;

[0021] Perform weighted summation of the at least one specific task processing result and the global shared processing result respectively to obtain a coding prediction result.

[0022] According to the ICD coding method based on multi-task learning and graph attention network provided by the present invention, the at least one coding prediction task further includes a treatment plan recommendation task.

[0023] An ICD coding method based on multi-task learning and graph attention network provided by the present invention, wherein the at least one coding prediction task further includes a death prediction task.

[0024] In a second aspect, the present invention further provides an ICD coding device based on multi-task learning and graph attention network, the device comprising:

[0025] An acquisition module, configured to acquire medical texts corresponding to at least one coding prediction task; wherein, the at least one coding prediction task includes an ICD coding task;

[0026] A first processing module, configured to construct a corresponding text graph according to the medical texts corresponding to the at least one coding prediction task;

[0027] A second processing module, configured to input the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model;

[0028] Wherein, the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data.

[0029] In a third aspect, the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps of the ICD coding method based on multi-task learning and graph attention network as described in any one of the above.

[0030] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the ICD coding method based on multi-task learning and graph attention network as described in any one of the above.

[0031] In a fifth aspect, the present invention further provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the steps of the ICD coding method based on multi-task learning and graph attention network as described in any one of the above.

[0032] The ICD coding method and device based on multi-task learning and graph attention network provided by the present invention, by constructing a corresponding text graph based on the medical texts corresponding to the coding prediction task, inputting the text graph into the coding prediction model, and obtaining the coding prediction result output by the coding prediction model. Since the medical text data is converted into the form of a text graph before coding prediction, the problem of data imbalance can be alleviated, the automatic ICD coding effect of medical texts can be effectively improved, and the obtained coding results are more accurate and reliable. Brief Description of the Drawings

[0033] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 is a schematic flow chart of the ICD coding method based on multi-task learning and graph attention network provided by the present invention;

[0035] Figure 2 is a schematic diagram of the construction principle of the text graph;

[0036] Figure 3 is a schematic diagram of the data processing principle based on the general shared hierarchical multi-task learning method;

[0037] Figure 4 is a schematic diagram of the data processing principle based on the task-specific shared multi-task learning method;

[0038] Figure 5 is a schematic structural diagram of the ICD coding device based on multi-task learning and graph attention network provided by the present invention;

[0039] Figure 6 is a schematic structural diagram of the electronic device provided by the present invention. Detailed Embodiments

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0041] The following will describe in conjunction with Figures 1 to 4 the implementation process of the ICD coding method based on multi-task learning and graph attention network provided by the embodiments of the present invention.

[0042] Figure 1 The embodiments of the present invention provide an ICD coding method based on multi-task learning and graph attention network, and the method includes:

[0043] Step 110: Obtain medical texts corresponding to at least one coding prediction task; wherein, the at least one coding prediction task includes an ICD coding task.

[0044] In this embodiment, the coding prediction task may include an ICD coding task, a treatment plan recommendation task, and a death prediction task, and the medical texts corresponding to different tasks are different.

[0045] Step 120: Construct a corresponding text graph according to the medical texts corresponding to at least one coding prediction task.

[0046] Since the coding prediction task in this embodiment may include multiple tasks and the medical texts corresponding to different tasks are different, corresponding text graphs need to be established for different tasks.

[0047] Specifically, referring to the appendix Figure 2 , the process of constructing the corresponding text graph in this embodiment includes:

[0048] First, obtain concepts related to the medical text according to the medical text corresponding to the coding prediction task.

[0049] In this embodiment, the medical text of a certain task is input into MetaMap to obtain medical-related concepts, that is, UMLS medical domain knowledge. MetaMap is a program that matches biomedical texts with concepts in the UMLS (Unified Medical Language System) metathesaurus. This program can set many parameters, which are used to control the output of MetaMap and its internal operation (such as the degree of word deformation, whether to ignore strings containing common words in the metathesaurus, whether to consider the order of letters, etc.).

[0050] Then, identify the relationships between each concept and the medical text respectively, and determine the relationships between each concept.

[0051] When determining the relationship between a concept and a text, this embodiment uses the TF-IDF method to identify the edge between the concept and the text node. TF-IDF is a statistical method used to evaluate the importance of a word for a document set or a single document in a corpus. The importance of a word increases in direct proportion to the number of times it appears in the document, but at the same time decreases in inverse proportion to the frequency of its appearance in the corpus.

[0052] When determining the relationship between concepts, first use UMLS to obtain the initial relationship between concepts, then count the co-occurrence times of concepts in the text, and then use Pointwise Mutual Information (PMI) to calculate the association between concepts. The calculation formula is as follows:

[0053]

[0054] Wherein, #W(i) represents the number of sliding windows containing concept i, #W(j) represents the number of sliding windows containing concept j, #W(i,j) represents the number of sliding windows containing both concept i and concept j, and #W represents the total number of sliding windows.

[0055] Finally, the medical text and concepts are used as nodes, and the relationships between each concept and the medical text as well as the relationships between each concept are used as edges to construct the corresponding text graph.

[0056] Since the text graph in this embodiment includes the associations between text and concepts as well as between concepts, compared with the original scattered medical text data, it can effectively alleviate the problem of data imbalance.

[0057] Step 130: Input the text graph into the encoding prediction model to obtain the encoding prediction result output by the encoding prediction model.

[0058] Among them, the encoding prediction result includes the ICD encoding result; the encoding prediction model is obtained by training the graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD encoding task and the corresponding ICD encoding result sample data.

[0059] It can be understood that the graph attention network used in this embodiment refers to a neural network that encodes using an attention mechanism on graph data.

[0060] In practical applications, there may be processing requirements for multiple encoding prediction tasks. For example, it may be necessary to obtain ICD encoding results, treatment plan recommendation results, and death prediction results. Therefore, the encoding prediction model in this embodiment can implement synchronous processing of multiple tasks. This model adopts a multi-task learning method. Specifically, this embodiment mainly adopts two methods: general shared hierarchical multi-task learning and task-specific shared multi-task learning. The main task in the learning process is automatic ICD encoding, that is, assigning an international disease code to a piece of text through calculation. The two auxiliary tasks are treatment plan recommendation and death prediction respectively.

[0061] It can be understood that general shared hierarchical multi-task learning means that different tasks share the same graph attention encoding method in text graph modeling, and task-specific shared multi-task learning means a method that retains the characteristics of each task while sharing knowledge among different tasks.

[0062] See Appendix Figure 3, when using the general shared hierarchical multi-task learning method, different tasks are processed through a shared graph attention network encoding layer. In this process, different encoding prediction tasks are hierarchically classified to enable low-level tasks to perform supervised learning at the lower levels of the graph attention network, while maintaining more complex interactions at deeper layers.

[0063] Each task makes a final prediction through its respective sub-encoding layer. The formula for convolutional processing is as follows:

[0064]

[0065] Among them, assuming that task k uses the t-th layer as the output layer, h represents the vector representation of nodes in the text graph, and Θ (s) represents the parameters of the graph attention network.

[0066] If the encoding prediction model is trained through the above general shared hierarchical multi-task learning method, correspondingly, input the text graph into the encoding prediction model to obtain the encoding prediction results output by the encoding prediction model, including:

[0067] Input the text graph corresponding to at least one encoding prediction task into the graph attention network encoding layer; among them, the graph attention network encoding layer includes multiple sub-encoding layers for processing different levels of encoding prediction tasks;

[0068] According to the processing level corresponding to the encoding prediction task, process the text graph corresponding to the encoding prediction task through the corresponding sub-encoding layer to obtain the encoding prediction result.

[0069] For example, for the three tasks of ICD coding, treatment plan recommendation, and death prediction, in the actual medical diagnosis process, it is necessary to first determine the cause of the disease according to the ICD coding result, and then give a treatment recommendation plan, that is, a medication plan, according to the treatment plan recommendation result. Finally, judge whether the patient will die or survive according to the death prediction result.

[0070] According to the execution order of each task in the above diagnosis process, it can be determined that ICD coding is a high-level task, treatment plan recommendation is a sub-level task, and death prediction is a low-level task. Therefore, in the graph attention network, the ICD coding task can be completed in the first n sub-encoding layers of the encoding layer, the treatment plan recommendation task can be completed in the first n + i sub-encoding layers, and the death prediction task can be completed in the first n + j sub-encoding layers, where i < j and both are positive numbers.

[0071] See Appendix Figure 3, in this embodiment, taking the ICD coding task and the death prediction task as examples, first, the text graphs corresponding to the ICD coding task and the text graphs corresponding to the death prediction task are both input into the shared graph attention network coding layer. After being processed by the corresponding sub-coding layers, the ICD coding results and the death prediction results are output at the corresponding layers.

[0072] See Appendix Figure 4 , when adopting the task-specific shared multi-task learning method, the processing process of the task data is as follows:

[0073] First, each task is processed through its respective task-specific coding layer. For the output of the task-specific task at layer t, the formula is as follows:

[0074]

[0075] Among them, α vu represents the attention coefficient from node u to v, and the specific expression is:

[0076]

[0077] In the above formula, W represents the weight matrix, represents the neighbor nodes of node v in the text graph, f represents the LeakyReLU function, and a T represents the weight vector.

[0078] Then, all tasks are processed through the globally shared global shared coding layer, and the processing results are respectively input into the task-specific coding layers of each task. For the globally shared output result at layer t, the formula is as follows:

[0079]

[0080] Among them, Θ (s) represents the parameters of the globally shared coding layer.

[0081] Finally, the final output of task k is the weighted sum of the output results of the task-specific coding layer and the globally shared coding layer, and its convolution processing formula is as follows:

[0082]

[0083] Among them, g i→k (i∈s,k) respectively control the part of the information flow transmitted from the globally shared coding layer and the task-specific coding layer to task k.

[0084] If the coding prediction model is trained through the above-mentioned task-specific shared multi-task learning method, correspondingly, input the text graph into the coding prediction model to obtain the coding prediction results output by the coding prediction model, including:

[0085] Input the text graph corresponding to at least one encoded prediction task into the corresponding specific task encoding layer for processing to obtain at least one specific task processing result;

[0086] Input the text graph corresponding to at least one encoded prediction task into the globally shared encoding layer for processing to obtain a globally shared processing result;

[0087] Perform weighted summation of at least one specific task processing result and the globally shared processing result respectively to obtain an encoded prediction result.

[0088] See the appendix Figure 4 In this embodiment, taking the ICD coding task and the death prediction task as examples for illustration, input the text graph corresponding to the ICD coding task and the text graph corresponding to the death prediction task into their respective corresponding specific task encoding layers for processing, and input the text graph corresponding to the ICD coding task and the text graph corresponding to the death prediction task into the globally shared encoding layer for processing together. The obtained globally shared processing results are respectively input into the specific task encoding layer corresponding to the ICD coding task and the specific task encoding layer corresponding to the death prediction task. Perform weighted summation of the specific task processing result corresponding to the ICD coding task and the globally shared processing result to output the ICD coding result; and perform weighted summation of the specific task processing result corresponding to the death prediction task and the globally shared processing result to obtain the death prediction result.

[0089] In the training of the encoded prediction model in this embodiment, mainly complete the process of defining the loss function and optimizing the parameters. The following loss function is adopted in this training process:

[0090]

[0091] Among them, N k represents the number of training samples, C represents the number of categories, represents the true value, represents the predicted probability, and Θ represents the set of all training parameters.

[0092] For the automatic ICD coding task and the treatment plan recommendation task, use sigmoid as the activation function. For the binary classification task such as death prediction, use softmax as the activation function. The final ICD coding result includes the ICD annotation result of the medical text, the treatment plan recommendation result includes the drugs required for treatment, that is, the medication plan, and the death prediction result is the death or survival information.

[0093] In this way, on the basis of the ICD coding task in the embodiment of the present invention, two auxiliary tasks are added, which can share information among different tasks. While realizing ICD coding, it can also perform treatment plan recommendation and death prediction synchronously, and the function is more perfect.

[0094] The ICD coding device based on multi-task learning and graph attention network provided by the present invention will be described below. The ICD coding device based on multi-task learning and graph attention network described below can be correspondingly referred to the ICD coding method based on multi-task learning and graph attention network described above.

[0095] Figure 5 The ICD coding device based on multi-task learning and graph attention network provided by an embodiment of the present invention is shown. The device includes:

[0096] An acquisition module 510, configured to acquire medical texts corresponding to at least one coding prediction task; wherein, the at least one coding prediction task includes an ICD coding task;

[0097] A first processing module 520, configured to construct a corresponding text graph according to the medical texts corresponding to at least one coding prediction task;

[0098] A second processing module 530, configured to input the text graph into a coding prediction model, and obtain a coding prediction result output by the coding prediction model;

[0099] Wherein, the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data.

[0100] It should be noted that the ICD coding device provided by the embodiment of the present invention mainly realizes the function of ICD annotation for clinical record texts, that is, medical texts. In addition, the coding prediction task may further include two auxiliary tasks, namely, a treatment plan recommendation task and a death prediction task.

[0101] The text graph mentioned in this embodiment refers to using knowledge to help construct a heterogeneous text graph at the entity level. Its essence is an undirected graph composed of texts and concepts in the text as nodes, and text-concept and concept-concept as edges.

[0102] Specifically, in this embodiment, the first processing module 520 includes:

[0103] A concept acquisition unit, configured to acquire concepts related to the medical text according to the medical text corresponding to the coding prediction task;

[0104] A relationship recognition unit, configured to respectively recognize the relationships between each concept and the medical text, and determine the relationships between each concept;

[0105] A text graph construction unit for constructing a corresponding text graph by using medical texts and concepts as nodes and using the relationships between each concept and the medical text and the relationships between each concept as edges.

[0106] In a specific embodiment, the second processing module 530 includes:

[0107] A first input unit for inputting the text graph corresponding to at least one coding prediction task into the graph attention network coding layer; wherein, the graph attention network coding layer includes a plurality of sub-coding layers for processing different levels of coding prediction tasks;

[0108] A first prediction unit for processing the text graph corresponding to the coding prediction task through the corresponding sub-coding layer according to the processing level corresponding to the coding prediction task to obtain a coding prediction result.

[0109] In another specific embodiment, the second processing module 530 includes:

[0110] A second input unit for inputting the text graph corresponding to at least one coding prediction task into the corresponding specific task coding layer for processing to obtain at least one specific task processing result;

[0111] A third input unit for inputting the text graph corresponding to at least one coding prediction task into the global shared coding layer for processing to obtain a global shared processing result;

[0112] A second prediction unit for performing weighted summation on at least one specific task processing result and the global shared processing result respectively to obtain a coding prediction result.

[0113] It can be seen that the ICD coding device based on multi-task learning and graph attention network provided by the embodiments of the present invention, on the one hand, uses the text graph to obtain the associations between different levels of semantic concepts and between the text and the concepts, thereby alleviating the problem of data imbalance. On the other hand, two auxiliary tasks, namely the treatment plan recommendation task and the death prediction task, are also added, so that information can be shared among different tasks. The improvements in the above two aspects greatly improve the effect of automatic ICD coding of medical texts.

[0114] Figure 6 Illustrates a schematic diagram of the physical structure of an electronic device, such as Figure 6As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communications interface 620, and the memory 630 complete communication with each other through the communication bus 640. The processor 610 may call the logical instructions in the memory 630 to execute an ICD coding method based on multi-task learning and a graph attention network. The method includes: obtaining medical texts corresponding to at least one coding prediction task; where at least one coding prediction task includes an ICD coding task; constructing a corresponding text graph according to the medical texts corresponding to at least one coding prediction task; inputting the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model; where the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data.

[0115] In addition, when the logical instructions in the above-mentioned memory 630 are implemented in the form of software functional units and sold or used as independent products, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0116] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the ICD coding method based on multi-task learning and graph attention network provided by the above-mentioned various methods. The method includes: obtaining medical texts corresponding to at least one coding prediction task; wherein, at least one coding prediction task includes an ICD coding task; constructing a corresponding text graph according to the medical texts corresponding to at least one coding prediction task; inputting the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model; wherein, the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on the medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data.

[0117] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it realizes the ICD coding method based on multi-task learning and graph attention network provided by the above-mentioned various methods. The method includes: obtaining medical texts corresponding to at least one coding prediction task; wherein, at least one coding prediction task includes an ICD coding task; constructing a corresponding text graph according to the medical texts corresponding to at least one coding prediction task; inputting the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model; wherein, the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on the medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data.

[0118] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0119] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An ICD coding method based on multi-task learning and graph attention network, characterized in that, it includes: Obtain medical texts corresponding to at least one coding prediction task; wherein, the at least one coding prediction task includes an ICD coding task; Construct corresponding text graphs according to the medical texts corresponding to the at least one coding prediction task; Input the text graphs into a coding prediction model to obtain coding prediction results output by the coding prediction model; wherein, the coding prediction results include ICD coding results; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data; wherein, the constructing corresponding text graphs according to the medical texts corresponding to the at least one coding prediction task includes: Obtain concepts related to the medical text according to the medical text corresponding to the coding prediction task; Identify the relationships between each concept and the medical text respectively, and determine the relationships between each concept; Construct corresponding text graphs by using the medical text and the concepts as nodes, and using the relationships between each concept and the medical text and the relationships between each concept as edges.

2. The ICD coding method based on multi-task learning and graph attention network according to claim 1, characterized in that, Inputting the text graphs into a coding prediction model to obtain coding prediction results output by the coding prediction model includes: Input the text graphs corresponding to at least one coding prediction task into a graph attention network coding layer; wherein, the graph attention network coding layer includes multiple sub-coding layers for processing different levels of coding prediction tasks; Process the text graphs corresponding to the coding prediction tasks through corresponding sub-coding layers according to the processing levels corresponding to the coding prediction tasks to obtain coding prediction results.

3. The ICD coding method based on multi-task learning and graph attention network according to claim 1, characterized in that, Inputting the text graphs into a coding prediction model to obtain coding prediction results output by the coding prediction model includes: Input the text graphs corresponding to at least one coding prediction task into corresponding specific task coding layers for processing to obtain at least one specific task processing result; Input the text graphs corresponding to at least one coding prediction task into a global shared coding layer for processing to obtain a global shared processing result; Perform weighted summation of the at least one specific task processing result and the global shared processing result respectively to obtain coding prediction results.

4. The ICD coding method based on multi-task learning and graph attention network according to claim 1, characterized in that, The at least one coding prediction task further includes a treatment plan recommendation task.

5. The ICD coding method based on multi-task learning and graph attention network according to claim 1 or 4, characterized in that, The at least one coding prediction task further includes a death prediction task.

6. An ICD coding device based on multi-task learning and graph attention network, It is characterized in that it includes an acquisition module, configured to acquire medical texts corresponding to at least one coding prediction task; wherein, the at least one coding prediction task includes an ICD coding task; a first processing module, configured to construct a corresponding text graph according to the medical texts corresponding to the at least one coding prediction task; a second processing module, configured to input the text graph into a coding prediction model to obtain a coding prediction result output by the coding prediction model; wherein, the coding prediction result includes an ICD coding result; the coding prediction model is obtained by training a graph attention network based on training data, and the training data includes text graph sample data constructed based on medical texts corresponding to the ICD coding task and corresponding ICD coding result sample data; wherein, the constructing a corresponding text graph according to the medical texts corresponding to the at least one coding prediction task includes acquiring concepts related to the medical text according to the medical text corresponding to the coding prediction task; respectively identifying the relationships between each concept and the medical text, and determining the relationships between each concept; using the medical text and the concepts as nodes, and using the relationships between each concept and the medical text and the relationships between each concept as edges to construct a corresponding text graph.

7. An electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, it is characterized in that when the processor executes the program, the steps of the ICD coding method based on multi-task learning and graph attention network according to any one of claims 1 to 5 are implemented.

8. A non-transitory computer-readable storage medium, on which a computer program is stored, it is characterized in that when the computer program is executed by a processor, the steps of the ICD coding method based on multi-task learning and graph attention network according to any one of claims 1 to 5 are implemented.

9. A computer program product, including a computer program, it is characterized in that when the computer program is executed by a processor, the steps of the ICD coding method based on multi-task learning and graph attention network according to any one of claims 1 to 5 are implemented.

Images