Diagnosis icd automatic coding method and system based on medical record semantic understanding

Abstract

The application discloses a diagnosis ICD automatic coding method and system based on medical record semantic understanding, which comprises the following steps: extracting medical record information from original medical records of a hospital through an information extraction program, wherein the extracted medical record information comprises original word data of discharge diagnosis and diagnosis and treatment data in a hospitalization process; performing diagnosis term splitting and medical record content cleaning on the extracted medical record information, wherein the diagnosis term splitting utilizes a named entity recognition technology to split the original word data of discharge diagnosis; constructing an evidence extraction model, extracting three-element text evidence for assisting in coding of diagnosis terms through the evidence extraction model; constructing a coding model based on three-element text implication, performing text implication scoring of diagnosis original words and candidate coding labels through the coding model based on three-element text implication; and obtaining the highest scoring candidate label for the current diagnosis original word, and obtaining corresponding coding from a standard ICD10 library. The application improves coding accuracy and can complete automatic coding according to complete medical record information.

Description

Technical Field

[0001] This invention belongs to the field of medical record data processing technology, specifically relating to a diagnostic ICD automatic coding method and system based on medical record semantic understanding. Background Technology

[0002] ICD is an abbreviation for International Classification of Diseases, a disease classification method compiled and published under the auspices of the World Health Organization. It is an important component of the health information standards system and is used by clinical research, medical monitoring, and health administration departments worldwide.

[0003] The International Classification of Diseases (ICD) is the core knowledge base of the WHO International Classification family and a standard in many medical fields. The first edition of the ICD was created in 1893, more than 120 years ago. It has been continuously updated to reflect developments in health and medicine, and the current version is the 10th edition. The complete statistical scope of the ICD covers causes of death, diseases, injuries, symptoms, reasons for seeking medical attention, and external causes of disease, and is widely used in clinical research and medical surveillance.

[0004] Currently, discharge diagnoses written by doctors often contain missing information, which affects the accuracy of the coding. The current mainstream ICD coding technology mainly relies on shallow text information such as keyword matching to complete the diagnostic coding, ignoring semantic matching, resulting in low coding reliability. Summary of the Invention

[0005] To address this issue, the present invention provides a diagnostic ICD automatic coding method and system based on medical record semantic understanding, which solves the problems of poor accuracy and reliability of traditional ICD coding.

[0006] To achieve the above objectives, a first aspect of the present invention provides the following technical solution: a diagnostic ICD automatic coding method based on medical record semantic understanding, comprising: Medical record information is extracted from the hospital's original medical records through an information extraction program. The extracted medical record information includes original discharge diagnosis data and treatment data during hospitalization. The extracted medical record information is subjected to diagnostic terminology segmentation and medical record content cleaning. The diagnostic terminology segmentation uses named entity recognition technology to segment the original discharge diagnosis data. An evidence extraction model is constructed, and diagnostic ternary text evidence with supporting information, encoded with auxiliary diagnostic terms, is extracted through the evidence extraction model. A coding model based on ternary textual implication is constructed. The diagnostic ternary textual evidence with supporting information is input into the coding model based on ternary textual implication. The textual implication scores of the diagnostic original words and candidate coding labels are calculated through the coding model based on ternary textual implication. For the current diagnostic term, obtain the candidate encoding label with the highest score and retrieve the corresponding encoding from the standard ICD library.

[0007] As a preferred solution for the diagnostic ICD automatic encoding method based on medical record semantic understanding, medical record content is cleaned using regular expressions. The medical record content cleaning includes abbreviation replacement, full-width / half-width conversion, and specified character conversion.

[0008] As a preferred scheme for the diagnostic ICD automatic coding method based on medical record semantic understanding, the ternary textual evidence for auxiliary diagnostic terminology coding extracted by the evidence extraction model includes: Using a pre-trained model, generate character representations of questions and several sentences with discourse information; The generated character representations are input into a bidirectional recurrent neural network to obtain a contextual sentence representation. The context sentence representation is mapped to two dimensions using MLP, and the score for whether the current sentence is a supporting sentence is calculated using the sigmoid function; For a given segment, each segment is processed by an answer prediction network to determine whether it contains a supporting sentence; Determine whether the score of a sentence in the segment is greater than a preset threshold; if it is greater than the preset threshold, it is determined to be a supporting sentence. Obtain a set of supporting sentences and concatenate diagnostic original words to form diagnostic ternary text evidence with supporting information.

[0009] As a preferred scheme for the diagnostic ICD automatic coding method based on medical record semantic understanding, the textual implication scores of diagnostic original words and candidate coding labels obtained through the coding model based on ternary textual implication include: Based on the classification model, candidate coding probabilities are obtained from the diagnostic ternary text evidence, and the top k with the highest probability values ​​are selected as the candidate coding label set.

[0010] As a preferred scheme for the diagnostic ICD automatic coding method based on medical record semantic understanding, single-label reasoning is adopted to perform textual entailment on the original diagnostic words and the descriptive text of each candidate coding label to obtain the prediction score of the candidate coding label.

[0011] As a preferred scheme for the diagnostic ICD automatic coding method based on medical record semantic understanding, it adopts dual-label reasoning, which uses the diagnostic original word and two similar candidate coding labels to perform textual entailment and obtain the pairwise comparison score within the candidate coding label set.

[0012] As a preferred scheme for the diagnostic ICD automatic coding method based on medical record semantic understanding, the absolute score of a single label and the pairwise comparison scores within the candidate coding label set are weighted to obtain the final score of each candidate coding label, and the highest score is selected as the final answer.

[0013] A second aspect of the present invention also provides a diagnostic ICD automatic coding system based on medical record semantic understanding, comprising: The medical record information acquisition module is used to extract medical record information from the hospital's original medical records through an information extraction program. The extracted medical record information includes original discharge diagnosis data and treatment data during hospitalization. The medical record information splitting and cleaning module is used to split diagnostic terms and clean medical record content from the extracted medical record information. The diagnostic term splitting uses named entity recognition technology to split the original discharge diagnosis data. The supporting information extraction module is used to construct an evidence extraction model and extract ternary text evidence of auxiliary diagnostic terminology encoding through the evidence extraction model. The ternary text implication scoring module is used to construct an encoding model based on ternary text implication, and to perform diagnostic original word and candidate encoding label text implication scores through the encoding model based on ternary text implication. The diagnostic coding module is used to obtain the candidate coding label with the highest score for the current diagnostic word and retrieve the corresponding code from the standard ICD10 library.

[0014] As a preferred solution for a diagnostic ICD automatic coding system based on medical record semantic understanding, the medical record information splitting and cleaning module uses regular expressions to clean the medical record content, which includes abbreviation replacement, full-width / half-width conversion, and specified character conversion. The supporting information extraction module includes: The character representation submodule is used to generate character representations of questions and several sentences with discourse information using a pre-trained model; The sentence representation submodule is used to input the generated character representation into a bidirectional recurrent neural network to obtain a contextual sentence representation; The sentence quantification module is used to map the context sentence representation to two dimensions using MLP, and then calculates the score of whether the current sentence is a supporting sentence using the sigmoid function; The supporting sentence judgment submodule is used to determine whether a supporting sentence exists for each specified segment through the answer prediction network; it also determines whether the score of a sentence in the segment is greater than a preset threshold, and if it is greater than the preset threshold, it is judged as a supporting sentence.

[0015] The diagnostic ternary text splicing submodule is used to obtain a set of supporting sentences and diagnostic original words and splice them into diagnostic ternary text evidence with supporting information.

[0016] As a preferred embodiment of a diagnostic ICD automatic coding system based on medical record semantic understanding, the ternary text implication scoring module includes: The candidate coding label set submodule is used to obtain candidate coding probabilities based on the classification model for diagnostic ternary text evidence, and take the top k with the highest probability values ​​as the candidate coding label set. The single-label inference submodule is used to perform textual entailment on the diagnostic original word and the description text of each candidate coding label using single-label inference, and obtain the prediction score of the candidate coding label. The dual-label reasoning submodule is used to perform textual entailment by diagnosing the original word and two similar candidate coding labels, and obtain the pairwise comparison score within the candidate coding label set; The sorting and integration submodule is used to weight the absolute score of a single label and the pairwise comparison scores within the candidate coding label set to obtain the final score of each candidate coding label, and select the one with the highest score as the final answer.

[0017] A third aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the diagnostic ICD auto-coding method based on medical record semantic understanding, as described in the first aspect or any possible implementation thereof.

[0018] A fourth aspect of the present invention provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute a diagnostic ICD autocoding method based on medical record semantic understanding, which implements the first aspect or any possible implementation thereof.

[0019] This invention has the following advantages: It extracts medical record information from original hospital medical records through an information extraction program. The extracted information includes discharge diagnosis terminology and treatment data during hospitalization. The extracted information undergoes diagnostic terminology segmentation and medical record content cleaning. Diagnostic terminology segmentation utilizes named entity recognition technology to segment the discharge diagnosis terminology. An evidence extraction model is constructed to extract ternary text evidence to assist in the encoding of diagnostic terms. A coding model based on ternary text implication is constructed, and the text implication scores of diagnostic terminology and candidate coding labels are calculated. For the current diagnostic terminology, the candidate coding label with the highest score is obtained, and the corresponding code is retrieved from the standard ICD10 library. This invention, through the evidence extraction model, can supplement the diagnostic text filled in by doctors, thereby improving coding accuracy. Through the ternary text implication coding model, it overcomes the shortcomings of traditional models that only perform shallow text matching and ignore semantic matching, enabling automated coding based on complete medical record information. Attached Figure Description

[0020] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the diagnostic ICD automatic coding method based on medical record semantic understanding provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the evidence extraction process in the diagnostic ICD automatic coding method based on medical record semantic understanding provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the text implication scoring process in the diagnostic ICD automatic coding method based on medical record semantic understanding provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the diagnostic ICD automatic coding system architecture based on medical record semantic understanding provided in Embodiment 2 of the present invention. Detailed Implementation

[0022] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0023] See Figure 1 Embodiment 1 of the present invention provides an automatic coding method for diagnostic ICDs based on medical record semantic understanding, comprising the following steps: S1. Extract medical record information from the hospital's original medical records through an information extraction program. The extracted medical record information includes original discharge diagnosis data and treatment data during hospitalization. S2. Perform diagnostic terminology segmentation and medical record content cleaning on the extracted medical record information. Diagnostic terminology segmentation uses named entity recognition technology to segment the original discharge diagnosis data. S3. Construct an evidence extraction model and extract diagnostic ternary text evidence with supporting information by encoding auxiliary diagnostic terms through the evidence extraction model. S4. Construct a coding model based on ternary textual implication, input the diagnostic ternary textual evidence with supporting information into the coding model based on ternary textual implication, and calculate the textual implication scores of the diagnostic original words and candidate coding labels through the coding model based on ternary textual implication. S5. For the current diagnostic term, obtain the candidate encoding label with the highest score and retrieve the corresponding encoding from the standard ICD library.

[0024] In this embodiment, the information extraction procedure involved in step S1 is itself a prior art. It is a basic natural language processing that can transform unstructured or semi-structured natural language text into structured features. It typically includes three types of subtasks: extracting entities of a specified type from the text (entity extraction / named entity recognition, NER); extracting semantic relationships between entities (relation extraction, RE); and extracting events from the text.

[0025] In this embodiment, the named entity recognition technology involved in step S2 is a subtask of information extraction. It can locate and classify named entities in text into predefined categories, such as people, organizations, locations, and time expressions. From the model level, it can be divided into rule-based methods, unsupervised learning methods, and supervised learning methods. From the input level, it can be divided into character-level methods, word-level methods, and character-word combination methods.

[0026] For example, before the separation: 1. Lung cancer; 2. Benign prostatic hyperplasia; 3. Urinary tract infection; 4. Bladder stones.

[0027] After breakdown: Lung cancer, benign prostatic hyperplasia, urinary tract infection, bladder stones.

[0028] In this embodiment, step S2 cleans the medical record content using regular expressions. The medical record content cleaning includes abbreviation replacement, full-width / half-width conversion, and specified character conversion.

[0029] Specifically, regular expressions are used to formulate regular expression matching rules based on the known characteristics of named entities, and then entity recognition is performed by matching regular expressions during prediction.

[0030] Suppose that the single clinical diagnostic term obtained through step S2 is problem q, and the relevant medical record information is as follows: This invention requires the extraction of relevant evidence that can assist in the coding of diagnostic terms.

[0031] For example: "q = lung cancer; C = After admission, relevant examinations were completed. ... Enhanced CT scan of the chest, abdomen, and pelvis showed: 1. A mass in the right upper lung. ... Bronchoscopy was performed on November 23, 2019. Postoperative pathology (right upper lobe, transbronchial lung biopsy specimen): lung adenocarcinoma... Among them, "mass in the upper right lung" and "postoperative pathology: (upper right lung lobe, bronchoscopic lung biopsy specimen): lung adenocarcinoma" are the corresponding evidence.

[0032] See Figure 2 In this embodiment, the extraction of ternary textual evidence encoding auxiliary diagnostic terms through the evidence extraction model includes: S31. Using a pre-trained model, generate character representations of questions and several sentences with discourse information; S32. Input the generated character representation into a bidirectional recurrent neural network to obtain the context sentence representation; S33. Map the context sentence representation to two dimensions using MLP, and calculate the score of whether the current sentence is a supporting sentence using the sigmoid function; S34. For a given segment, each segment is processed by an answer prediction network to determine whether there is a supporting sentence; S35. Determine whether the score of a sentence in the segment is greater than a preset threshold. If it is greater than the preset threshold, it is determined to be a supporting sentence.

[0033] S36. Obtain the set of supporting sentences and the diagnostic original words to form diagnostic ternary text evidence with supporting information.

[0034] Specifically, in step S31, a pre-trained model M is first used to generate word representations of question q and k sentences with discourse information. in, This is a commonly used model based on large-scale unsupervised text pre-training, used to generate text encodings with contextual information. For the generated representation, h is the dimension of a single character, and n is the length of the question plus the length of k sentences plus 2. Special classification identifiers that contain full-text information. Special separators used to distinguish between questions and sentences in the text.

[0035] Specifically, in step S32, the generated character representation is input into a bidirectional recurrent neural network to obtain a more detailed contextual representation. : Based on the context representation, obtain the sentence representations of k sentences. Then, using all representations of the start and end positions of the sentences, obtain the maximum amount of information through max pooling. in, Let k be the starting position of the k-th sentence in the model input. This is the end position.

[0036] Specifically, in step S33, the sentence representation is mapped to two dimensions using MLP, and the score for whether the current sentence is a supporting sentence is calculated using the sigmoid function: Specifically, in step S34, the sentence representation and the special classification identifier representation are combined as the task input to determine whether the segment has a supporting sentence. Similarly, the score is calculated through MLP mapping and the sigmoid function. Specifically, in step S35, for each of the m segments, the answer prediction network first determines whether the segment contains supporting sentences. in, The threshold is adjustable.

[0037] Specifically, in step S36, for segments with supporting sentences, it is further determined whether the scores of the sentences within them are greater than a threshold. greater than the threshold This is the supporting sentence: .

[0038] Finally, the obtained set of supporting sentences s and the original diagnostic text q are concatenated to form a diagnostic text d with supporting information, as shown in the following example: d= Lung cancer ## Right upper lung mass ## Postoperative pathology: (Right upper lobe, bronchoscopic lung biopsy specimen): Lung adenocarcinoma.

[0039] See Figure 3 In this embodiment, the text implication scores for diagnostic original words and candidate encoded labels obtained through the ternary text implication-based encoding model include: S41. Based on the classification model, obtain candidate coding probabilities for diagnostic ternary text evidence, and take the top k with the highest probability values ​​as the candidate coding label set. S42. Using single-label reasoning, textual entailment is performed on the diagnostic original word and the description text of each candidate coding label to obtain the prediction score of the candidate coding label. S43. Using dual-label reasoning, textual implication is performed by diagnosing the original word and two similar candidate coding labels to obtain pairwise comparison scores within the candidate coding label set; S44. The absolute score of a single label is weighted by the pairwise comparison scores within the candidate coding label set to obtain the final score of each candidate coding label, and the highest score is selected as the final answer.

[0040] Specifically, in step S41, candidate coding probabilities are obtained for a diagnosis d based on the classification model, and the top k with the highest probability values ​​are selected as the candidate coding label set. : ) MLP is a two-layer fully connected layer, and softmax is a normalization function.

[0041] Here, we assume that the diagnostic word plus supporting information is d, and the description text of the candidate encoded label is: ; Let's first assume a general implication model Inf: Inf(textA,textB)=MLP(M(<[cls],textA,[sep],textB>)) MLP is a two-layer fully connected layer that outputs an implication score.

[0042] Specifically, in step S42, the text entailment of the diagnostic original word and the description text of each candidate encoded label is used to obtain the prediction score of that label: .

[0043] In step S43, since some candidate encoded labels describe texts with high similarity, it is necessary to perform text entailment by diagnosing the original word and two similar candidate encoded labels to obtain a relative score representing the priority of the label: .

[0044] In step S44, the absolute score of a single label and the pairwise comparison scores within the candidate encoded label set are weighted to obtain the final score for each label, and the highest score is selected as the final answer. ,in, This represents the weighting coefficient.

[0045] After processing steps S41 to S44, the final result is shown in the following example: d = Lung cancer ## Right upper lung mass ## Postoperative pathology: (Right upper lobe, bronchoscopic lung biopsy specimen): Lung adenocarcinoma; o_best = Malignant tumors of the upper lobe, bronchus, or lung.

[0046] In step S5, for the current diagnostic term, the label with the highest score is obtained, and the corresponding code is retrieved from the standard ICD10 library, thus completing the entire diagnostic coding process, as shown in the following example: name = malignant tumor of the upper lobe, bronchus, or lung code = C34.1.

[0047] In summary, this invention extracts medical record information from original hospital medical records through an information extraction program. The extracted information includes discharge diagnosis terminology and treatment data during hospitalization. The extracted information undergoes diagnostic terminology segmentation and medical record content cleaning. Diagnostic terminology segmentation utilizes named entity recognition technology to segment the discharge diagnosis terminology. An evidence extraction model is constructed to extract ternary text evidence to assist in the encoding of diagnostic terms. A coding model based on ternary text implication is constructed to score the text implication of diagnostic terms and candidate coding labels. For the current diagnostic term, the candidate coding label with the highest score is obtained, and the corresponding code is retrieved from the standard ICD10 library. This invention, through its evidence extraction model, can supplement the diagnostic text filled in by doctors, thereby improving coding accuracy. Through the ternary text implication coding model, it overcomes the shortcomings of traditional models that only perform shallow text matching and ignore semantic matching, enabling automated coding based on complete medical record information.

[0048] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0049] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. Example 2

[0050] See Figure 4 Embodiment 2 of the present invention also provides an automatic ICD coding system for diagnosis based on medical record semantic understanding, comprising: The medical record information acquisition module 1 is used to extract medical record information from the original hospital medical records through an information extraction program. The extracted medical record information includes the original discharge diagnosis data and the diagnosis and treatment data during hospitalization. The medical record information splitting and cleaning module 2 is used to split diagnostic terms and clean medical record content from the extracted medical record information. The diagnostic term splitting uses named entity recognition technology to split the original discharge diagnosis data. The supporting information extraction module 3 is used to construct an evidence extraction model and extract diagnostic ternary text evidence with supporting information encoded in the auxiliary diagnostic terminology through the evidence extraction model. The ternary text implication scoring module 4 is used to construct a coding model based on ternary text implication. The diagnostic ternary text evidence with supporting information is input into the coding model based on ternary text implication, and the text implication scores of the diagnostic original word and candidate coding label are calculated through the coding model based on ternary text implication. Diagnostic coding module 5 is used to obtain the candidate coding label with the highest score for the current diagnostic word and retrieve the corresponding code from the standard ICD library.

[0051] In this embodiment, the medical record information splitting and cleaning module 2 uses regular expressions to clean the medical record content, which includes abbreviation replacement, full-width / half-width conversion, and specified character conversion. The supporting information extraction module 3 includes: The character representation submodule 31 is used to generate character representations of questions and several sentences with discourse information using a pre-trained model; The sentence representation submodule 32 is used to input the generated character representation into the bidirectional recurrent neural network to obtain the context sentence representation; The sentence numerator module 33 is used to map the context sentence representation to two dimensions using MLP, and calculate the score of whether the current sentence is a supporting sentence using the sigmoid function; The supporting sentence judgment submodule 34 is used to determine whether a supporting sentence exists for each specified segment through the answer prediction network; and to determine whether the score of the sentence in the segment is greater than a preset threshold. If it is greater than the preset threshold, it is determined to be a supporting sentence.

[0052] The diagnostic ternary text splicing submodule 35 is used to obtain a set of supporting sentences and diagnostic original words and splice them into diagnostic ternary text evidence with supporting information.

[0053] In this embodiment, the ternary text implication scoring module 4 includes: The candidate coding label set submodule 41 is used to obtain candidate coding probabilities based on the classification model for diagnostic ternary text evidence, and take the top k with the highest probability values ​​as the candidate coding label set. The single-label inference submodule 42 is used to perform textual entailment on the diagnostic original word and the description text of each candidate coding label using single-label inference, and obtain the prediction score of the candidate coding label. The dual-label reasoning submodule 43 is used to perform dual-label reasoning, which involves diagnosing the original word and two similar candidate coding labels to perform textual entailment, and obtaining pairwise comparison scores within the candidate coding label set. The sorting and integration submodule 44 is used to weight the absolute score of a single label and the pairwise comparison scores within the candidate coding label set to obtain the final score of each candidate coding label, and select the one with the highest score as the final answer.

[0054] It should be noted that the information interaction and execution process between the modules / submodules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here. Example 3

[0055] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for an automatic diagnostic ICD encoding method based on medical record semantic understanding. The program code includes instructions for executing the automatic diagnostic ICD encoding method based on medical record semantic understanding of Embodiment 1 or any possible implementation thereof.

[0056] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)). Example 4

[0057] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor; The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the diagnostic ICD automatic coding method based on medical record semantic understanding according to Embodiment 1 or any possible implementation thereof.

[0058] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0059] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0060] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0061] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A diagnostic ICD automatic coding method based on medical record semantic understanding, characterized in that, include: Medical record information is extracted from the hospital's original medical records through an information extraction program. The extracted medical record information includes original discharge diagnosis data and treatment data during hospitalization. The extracted medical record information is subjected to diagnostic terminology segmentation and medical record content cleaning. The diagnostic terminology segmentation uses named entity recognition technology to segment the original discharge diagnosis data. An evidence extraction model is constructed, and diagnostic ternary text evidence with supporting information, encoded with auxiliary diagnostic terms, is extracted through the evidence extraction model. A coding model based on ternary textual implication is constructed. The diagnostic ternary textual evidence with supporting information is input into the coding model based on ternary textual implication. The textual implication scores of the diagnostic original words and candidate coding labels are calculated through the coding model based on ternary textual implication. For the current diagnostic term, obtain the candidate encoding label with the highest score and retrieve the corresponding encoding from the standard ICD library.

2. The diagnostic ICD automatic coding method based on medical record semantic understanding according to claim 1, characterized in that, Medical record content cleaning is performed using regular expressions. The cleaning process includes abbreviation replacement, full-width / half-width character conversion, and specified character conversion.

3. The diagnostic ICD automatic coding method based on medical record semantic understanding according to claim 1, characterized in that, The diagnostic ternary text evidence with supporting information, encoded using the evidence extraction model, includes: Using a pre-trained model, generate character representations of questions and several sentences with discourse information; The generated character representations are input into a bidirectional recurrent neural network to obtain a contextual sentence representation. The context sentence representation is mapped to two dimensions using MLP, and the score for whether the current sentence is a supporting sentence is calculated using the sigmoid function; For a given segment, each segment is processed by an answer prediction network to determine whether it contains a supporting sentence; Determine whether the score of a sentence in the segment is greater than a preset threshold; if it is greater than the preset threshold, it is determined to be a supporting sentence. Obtain a set of supporting sentences and concatenate diagnostic original words to form diagnostic ternary text evidence with supporting information.

4. The diagnostic ICD automatic coding method based on medical record semantic understanding according to claim 3, characterized in that, The diagnostic text implication scores for original words and candidate encoded labels obtained through the ternary text implication-based coding model include: Based on the classification model, candidate coding probabilities are obtained from the diagnostic ternary text evidence, and the top k with the highest probability values ​​are selected as the candidate coding label set.

5. The diagnostic ICD automatic coding method based on medical record semantic understanding according to claim 4, characterized in that, Single-label inference is used to perform textual entailment on the diagnostic original word and the descriptive text of each candidate coding label to obtain the prediction score of the candidate coding label.

6. The diagnostic ICD automatic coding method based on medical record semantic understanding according to claim 5, characterized in that, A dual-label inference approach is adopted, which performs textual entailment by diagnosing the original word and two similar candidate coding labels, and obtains the pairwise comparison score within the candidate coding label set.

7. The diagnostic ICD automatic coding method based on medical record semantic understanding according to claim 6, characterized in that, The final score for each candidate coding label is obtained by weighting the absolute score of the single label with the pairwise comparison scores within the candidate coding label set, and the highest score is selected as the final answer.

8. A diagnostic ICD automatic coding system based on medical record semantic understanding, characterized in that, include: The medical record information acquisition module is used to extract medical record information from the hospital's original medical records through an information extraction program. The extracted medical record information includes original discharge diagnosis data and treatment data during hospitalization. The medical record information splitting and cleaning module is used to split diagnostic terms and clean medical record content from the extracted medical record information. The diagnostic term splitting uses named entity recognition technology to split the original discharge diagnosis data. The supporting information extraction module is used to construct an evidence extraction model and extract diagnostic ternary text evidence with supporting information encoded in the auxiliary diagnostic terminology through the evidence extraction model. The ternary text implication scoring module is used to construct a coding model based on ternary text implication. The diagnostic ternary text evidence with supporting information is input into the coding model based on ternary text implication, and the text implication scores of the diagnostic original words and candidate coding labels are calculated through the coding model based on ternary text implication. The diagnostic coding module is used to obtain the candidate coding label with the highest score for the current diagnostic word and retrieve the corresponding code from the standard ICD library.

9. The diagnostic ICD automatic coding system based on medical record semantic understanding according to claim 8, characterized in that, In the medical record information splitting and cleaning module, regular expressions are used to clean the medical record content, which includes abbreviation replacement, full-width / half-width conversion, and specified character conversion. The supporting information extraction module includes: The character representation submodule is used to generate character representations of questions and several sentences with discourse information using a pre-trained model; The sentence representation submodule is used to input the generated character representation into a bidirectional recurrent neural network to obtain a contextual sentence representation; The sentence quantification module is used to map the context sentence representation to two dimensions using MLP, and then calculates the score of whether the current sentence is a supporting sentence using the sigmoid function; The supporting sentence judgment submodule is used to determine whether a supporting sentence exists for each specified segment after passing through the answer prediction network; it also determines whether the score of a sentence in the segment is greater than a preset threshold, and if it is greater than the preset threshold, it is judged as a supporting sentence; The diagnostic ternary text splicing submodule is used to obtain a set of supporting sentences and diagnostic original words and splice them into diagnostic ternary text evidence with supporting information.

10. The diagnostic ICD automatic coding system based on medical record semantic understanding according to claim 9, characterized in that, The ternary text implication scoring module includes: The candidate coding label set submodule is used to obtain candidate coding probabilities based on the classification model for diagnostic ternary text evidence, and take the top k with the highest probability values ​​as the candidate coding label set. The single-label inference submodule is used to perform textual entailment on the diagnostic original word and the description text of each candidate coding label using single-label inference, and obtain the prediction score of the candidate coding label. The dual-label reasoning submodule is used to perform textual entailment by diagnosing the original word and two similar candidate coding labels, and obtain the pairwise comparison score within the candidate coding label set; The sorting and integration submodule is used to weight the absolute score of a single label and the pairwise comparison scores within the candidate coding label set to obtain the final score of each candidate coding label, and select the one with the highest score as the final answer.

Images