ICD intelligent coding method based on multi-level knowledge driven large language model

Abstract

The application discloses an ICD intelligent coding method based on a multi-level knowledge-driven large language model, and belongs to the technical field of intelligent coding. The method comprises the following steps: S1, a sub-model matching knowledge base is constructed to select an adaptive sub-model for a diagnosis text; S2, a single diagnosis coding knowledge base is constructed, the adaptive sub-model is called, and a preliminary coding result is generated for a single diagnosis of the diagnosis text; S3, a global verification knowledge base is constructed to perform global verification on all preliminary coding results; and S4, the global verification result is input into a strong rule quality control engine to perform final verification and merging. The application optimizes the knowledge processing process, only filters and vectorizes effective characters that are strongly related to ICD coding when constructing the three-layer knowledge base, eliminates redundant and invalid characters, greatly improves the recall accuracy and efficiency of retrieval enhancement, and provides accurate knowledge support for model decision-making.

Description

Technical Field

[0001] This invention belongs to the field of intelligent coding technology, specifically relating to an ICD intelligent coding method based on a multi-level knowledge-driven large language model. Background Technology

[0002] Currently, there are two main technical solutions in the industry for addressing the issue of ICD intelligent encoding: The first approach involves constructing a vectorized mini-model. This method projects the semantics of the diagnostic text into a vector space and uses similarity calculations to obtain the ICD encoding results. However, medical semantics is highly complex, specialized, and diverse, containing a large number of technical terms, variations in disease descriptions, and associations with complications. The parameter scale and knowledge coverage of the vectorized mini-model are limited, making it impossible to fully encompass the complex medical semantic information, resulting in low encoding accuracy. Furthermore, the design logic of this approach can only handle the encoding task of a single diagnostic text and cannot accommodate the mutual influence between multiple diagnoses (such as encoding association rules in scenarios involving complications and comorbidities). This is incompatible with the actual medical practice scenario where "multiple discharge diagnoses for the same patient need to be co-coded," making it difficult to meet practical application needs.

[0003] The second approach uses a "large language model + naive RAG" method. This approach inputs ICD encoding rules, thought processes, and encoding methods into the large language model, relying on the model's semantic understanding capabilities for encoding. However, this approach has two major problems: First, large language models have an inherent "illusion" defect; without strict output constraints, they are prone to creating non-existent ICD codes, leading to invalid encoding results. Second, publicly available and accurate data related to ICD encoding is scarce. Large language models struggle to acquire sufficient and accurate encoding case knowledge during training, resulting in biases in the model's understanding and application of encoding rules, further reducing encoding accuracy. Summary of the Invention

[0004] To address the problems of low accuracy, high model illusion rate, and mismatch with actual business scenarios in existing ICD intelligent coding methods, this invention proposes an ICD intelligent coding method based on a multi-level knowledge-driven large language model.

[0005] The technical solution of this invention is: an ICD intelligent encoding method based on a multi-level knowledge-driven large language model, comprising the following steps: S1. Construct a sub-model matching knowledge base to select an appropriate sub-model for diagnostic text; S2. Construct a single-diagnosis coding knowledge base, call the adaptation sub-model, and generate preliminary coding results for a single diagnosis of the diagnostic text; S3. Construct a global verification knowledge base to perform global verification on all preliminary coding results; S4. Input the global verification results into the Zhiqiang rule quality control engine for final verification and merging.

[0006] Furthermore, in S1, the comprehensive fit score of several candidate sub-models is calculated, and the Top-K method is used to vote on each candidate sub-model to determine the fit sub-model for the diagnostic text. Overall fit score of candidate sub-models The expression is: ; in, Indicates the first One candidate sub-model Indicates the first Dynamic weights in each dimension This indicates the diagnostic text. The candidate sub-model's specialty domain and diagnostic text are represented in the first... Similarity across multiple dimensions.

[0007] Furthermore, S2 includes the following sub-steps: S21. Calculate the medical semantic contribution of each character in a single diagnosis; S22. Set a filtering threshold and filter characters based on their contribution to medical semantics; S23. Based on the adaptor sub-model, calculate the confidence level of the candidate codes of the filtered characters; S24. Set a dynamic confidence threshold and filter candidate codes based on their confidence levels; S25. Use the multi-source evidence weighted voting method to vote on the screened candidate codes, and take the candidate code with the largest vote value as the preliminary coding result.

[0008] Furthermore, in S21, the first Medical semantic contribution of each character The expression is: ; in, The first in the diagnostic text One character, Indicates the first Term frequency of each character - inverse document frequency, Indicates the first Weight of medical terms per character, Indicates the first The encoding correlation of each character The weighting coefficients represent the term frequency-inverse document frequency ratio. Weighting coefficients representing the weights of medical terms. Weighting coefficients representing the degree of correlation between codes; In S23, the confidence level of the candidate code The expression is: ; in, Indicates the first One candidate ICD code, This indicates the diagnostic text. Indicates the first Knowledge base matching degree between candidate ICD codes and diagnostic texts The sub-model output number The probability values ​​of each candidate ICD code. Indicates the weighting coefficient; In S25, the expression for the multi-source evidence weighted voting method is: ; in, Indicates the first The voting results for each candidate ICD code. Indicates the first Consistency of medical terminology in candidate ICD-coded diagnostic texts Indicates the first Consistency of medical terminology in candidate ICD-coded diagnostic texts The voting weights represent the degree of knowledge base matching. Voting weights representing historical successful coding rates Voting weights indicating consistency in medical terminology.

[0009] Furthermore, in S3, the global verification knowledge base is used for knowledge enhancement when verifying the preliminary coding results of the single diagnostic coding knowledge base.

[0010] Furthermore, S4 includes the following sub-steps: S41. Based on the global verification result, receive the registration request of the identifier factory, complete the identifier registration, and assign a unique processing factory for each type of identifier; S42. Extract the type and path information of the identifier and create an identifier instance; S43. Based on the identifier instance, complete the identifier resolution; S44. Perform a redirection operation based on the identifier resolution result.

[0011] Furthermore, in S41, the identifier types include business data identifiers, knowledge base identifiers, and knowledge point identifiers; Business data identifiers are used to access business data and to retrieve and reference it; business data includes user information, business process data, and transaction records. Knowledge base identifiers are used to complete knowledge base retrieval; Knowledge point identifiers are used to retrieve and reference knowledge point data; knowledge point data includes concept definitions and terminology explanations.

[0012] Furthermore, in S43, the identifier resolution process includes an initialization phase, a path resolution phase, a data retrieval phase, and a result processing phase; The initialization phase specifically involves: acquiring runtime input data as the parsing context, and obtaining the basic parsing data based on the identifier type; The path resolution phase specifically involves: breaking down the path information into several path units, each representing an operation type; path units include sub-objects, references, properties, static properties, group operations, and filtering operations; The data retrieval stage specifically involves: obtaining the retrieved data based on the type of the identifier; The result processing stage specifically involves: setting the retrieved data as a reference object of the identifier, updating the identifier's state, parsing it, and returning the parsing result to the requester.

[0013] Furthermore, in the result processing stage, either precise analysis or fuzzy analysis is used for parsing; Precise parsing specifically involves: starting from the root object of the identifier's data source, performing the positioning operation sequentially according to the path unit order; Fuzzy parsing specifically involves performing a fuzzy matching operation on the final level after performing the sub-final level positioning.

[0014] Furthermore, in S44, when the fuzzy identifier resolution result is empty, the redirection target is determined based on the identifier type and path, a redirection request is constructed, and identifier resolution is re-executed.

[0015] The beneficial effects of this invention are: (1) The present invention optimizes the knowledge processing process. When constructing a three-layer knowledge base, it only selects effective characters that are strongly related to ICD encoding for vectorization processing, eliminates redundant and invalid characters, greatly improves the recall accuracy and efficiency of enhanced retrieval, and provides accurate knowledge support for model decision-making.

[0016] (2) This invention innovatively designs a three-layer knowledge base structure (sub-model matching knowledge base, single diagnostic coding knowledge base and global verification knowledge base). Through layered knowledge empowerment, it not only comprehensively improves the accuracy of model coding, but also suppresses the illusion problem of large language models from two dimensions through the sub-model matching constraints of the first-layer knowledge base and the output range limitation of the sub-model itself, while being compatible with the actual business scenarios of multi-diagnostic collaborative coding. Attached Figure Description

[0017] Figure 1 This is a flowchart of the ICD intelligent encoding method based on a multi-level knowledge-driven large language model. Detailed Implementation

[0018] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0019] like Figure 1 As shown, this invention provides an ICD intelligent encoding method based on a multi-level knowledge-driven large language model, comprising the following steps: S1. Construct a sub-model matching knowledge base to select an appropriate sub-model for diagnostic text; S2. Construct a single-diagnosis coding knowledge base, call the adaptation sub-model, and generate preliminary coding results for a single diagnosis of the diagnostic text; S3. Construct a global verification knowledge base to perform global verification on all preliminary coding results; S4. Input the global verification results into the Zhiqiang rule quality control engine for final verification and merging.

[0020] Each sub-model (which includes a second-layer knowledge base, prompts from the main model, and output constraints) has strict pre-defined output range constraints (only valid codes and related descriptions conforming to the ICD coding standard are allowed to be output), fundamentally avoiding illusions such as the main language model creating its own codes and outputting invalid results. The division of sub-models is based on the classification logic of ICD coding (such as disease systems, coding levels, etc.), ensuring that each sub-model focuses on coding tasks in a specific domain, improving processing specificity.

[0021] In this embodiment of the invention, in S1, the comprehensive adaptation score of several candidate sub-models is calculated, and the Top-K method is used to vote on each candidate sub-model to determine the adaptation sub-model of the diagnostic text. Overall fit score of candidate sub-models The expression is: ; in, Indicates the first One candidate sub-model Indicates the first Dynamic weights in each dimension This indicates the diagnostic text. The candidate sub-model's specialty domain and diagnostic text are represented in the first... Similarity across multiple dimensions.

[0022] k=1,2,3 correspond to three matching dimensions: 1-semantic similarity, 2-scene adaptability, and 3-historical encoding accuracy. By using the entropy weight method for real-time calculation, the weights can be adaptively adjusted in different scenarios.

[0023] A first-layer knowledge base (sub-model matching knowledge base) is constructed to enhance the retrieval and recall of the large language model. This knowledge base stores the association data of "diagnostic text features - sub-models" and clearly annotates the association logic (such as "diagnostic text involving cardiovascular symptoms is preferentially matched with the cardiovascular disease sub-model" and "diagnostic text containing surgical descriptions is matched with the surgical coding sub-model"), providing accurate knowledge support for the large language model to select suitable sub-models and improving the accuracy of sub-model selection.

[0024] Existing technologies often employ "single-dimensional matching" (such as relying solely on semantic similarity) when selecting sub-models, which can easily lead to sub-model adaptation bias. This invention innovatively designs a multi-dimensional dynamic weighted voting algorithm for sub-models. It calculates the comprehensive adaptation score of candidate sub-models through multi-dimensional quantitative indicators, and selects the optimal set of sub-models based on the voting results, thereby improving the accuracy and robustness of sub-model matching.

[0025] All candidate sub-models are sorted in descending order, and a Top-K voting + threshold filtering strategy is used: [Selection / Selection] Sub-models with a score ≥0.7 are included in the final set. If the number of sub-models meeting the criteria is less than 3, the Top-3 are selected; if it is more than 5, the remaining sub-models are selected. The first 5 (to avoid computational redundancy due to too many sub-models).

[0026] In this embodiment of the invention, S2 includes the following sub-steps: S21. Calculate the medical semantic contribution of each character in a single diagnosis; S22. Set a filtering threshold and filter characters based on their contribution to medical semantics; S23. Based on the adaptor sub-model, calculate the confidence level of the candidate codes of the filtered characters; S24. Set a dynamic confidence threshold and filter candidate codes based on their confidence levels; S25. Use the multi-source evidence weighted voting method to vote on the screened candidate codes, and take the candidate code with the largest vote value as the preliminary coding result.

[0027] A second-layer knowledge base (single diagnostic code knowledge base) is constructed to enhance the knowledge of the large language model when processing a single diagnostic code. This knowledge base adopts a bidirectional knowledge structure that is compatible with both "many-to-one" (multiple diagnostic texts with different expressions correspond to the same ICD code) and "one-to-many" (a single diagnostic text corresponds to multiple candidate ICD codes due to different scenarios), covering diverse expression scenarios of medical diagnoses and exceptions to coding rules, thereby improving the adaptability and flexibility of business knowledge.

[0028] Existing technologies for character filtering in knowledge bases often employ qualitative methods such as "keyword extraction" or "part-of-speech filtering," lacking quantitative standards. This results in some low-value characters being mixed into the vectorization process, reducing retrieval accuracy. This invention innovatively designs a medical semantic contribution quantification model. Through a formula, it accurately calculates the contribution value of each character to ICD encoding matching, retaining only high-contribution characters for vectorization. This ensures retrieval accuracy while avoiding the exposure of the core knowledge base construction logic.

[0029] Set the filter threshold (Dynamically adaptive adjustment, initial value 0.6, can be iteratively optimized according to the coding scenario), only retain those that meet the requirements. The characters enter the vectorization process, and low-contribution redundant characters (such as interjections, general modifiers, etc.) are removed.

[0030] Existing technologies, when handling "one-to-many" code matching (a single diagnosis corresponding to multiple candidate ICD codes), often employ strategies such as "maximum model probability" or "random selection," which can easily lead to coding bias. This invention innovatively designs a two-stage conflict resolution mechanism: the first stage filters valid candidate codes through a confidence threshold, and the second stage determines the final code through weighted voting based on multi-source evidence, resolving the "one-to-many" conflict problem while improving coding accuracy.

[0031] In this embodiment of the invention, in S21, the first... Medical semantic contribution of each character The expression is: ; in, The first in the diagnostic text One character, Indicates the first Term frequency of each character - inverse document frequency, Indicates the first Weight of medical terms per character, Indicates the first The encoding correlation of each character The weighting coefficients represent the term frequency-inverse document frequency ratio. Weighting coefficients representing the weights of medical terms. Weighting coefficients representing the degree of correlation between codes; It reflects the general distinguishability of characters in the diagnostic text set. The value range is [0,1], which is jointly marked by the ICD official terminology dictionary and the clinical diagnosis and treatment guidelines dictionary (e.g., the MTW of "myocardial infarction" and "infarction" is 0.95, and the MTW of modifiers such as "mild" and "suspected" is 0.3). It reflects the probability of a direct association between a character and its ICD encoding.

[0032] α+β+γ=1, and optimized using gradient descent algorithm (the optimal values ​​after experimental optimization are α=0.3, β=0.5, γ=0.2).

[0033] In S23, the confidence level of the candidate code The expression is: ; in, Indicates the first One candidate ICD code, This indicates the diagnostic text. Indicates the first Knowledge base matching degree between candidate ICD codes and diagnostic texts The sub-model output number The probability values ​​of each candidate ICD code. Indicates the weighting coefficient; In S25, the expression for the multi-source evidence weighted voting method is: ; in, Indicates the first The voting results for each candidate ICD code. Indicates the first Consistency of medical terminology in candidate ICD-coded diagnostic texts Indicates the first Consistency of medical terminology in candidate ICD-coded diagnostic texts The voting weights represent the degree of knowledge base matching. Voting weights representing historical successful coding rates Voting weights indicating consistency in medical terminology.

[0034] In this embodiment of the invention, in S3, the global verification knowledge base is used for knowledge enhancement when verifying the preliminary coding results of the single diagnostic coding knowledge base.

[0035] A third-layer knowledge base (global validation knowledge base) is constructed as a global knowledge support to enhance the knowledge when the large language model validates the coding results of all diagnoses. This knowledge base stores global rules for ICD coding (such as coding exclusivity rules, comorbidity coding priority rules, multi-diagnosis collaborative coding specifications, etc.), providing a knowledge basis for cross-diagnosis coding consistency and rationality verification.

[0036] When the service starts, the discharge diagnosis text (which may contain one or more diagnostic messages) is passed as input parameters to the large language model.

[0037] The large language model is used for retrieval enhancement generation based on the first-layer knowledge base. By matching diagnostic text features with the association logic in the knowledge base, multiple suitable sub-models are selected for subsequent single diagnostic encoding processing.

[0038] The input discharge diagnosis text is split into independent individual diagnoses. For each independent diagnosis, the corresponding sub-model is called. Combined with the retrieval enhancement generation capability of the second-layer knowledge base, the individual diagnosis is encoded to obtain the preliminary encoding result of each independent diagnosis.

[0039] The preliminary coding results of all independent diagnoses are aggregated and fed back to the large language model. The large language model performs retrieval enhancement and generation based on the third-layer knowledge base, and performs global verification on all preliminary coding results (including coding consistency, priority adaptation, multi-diagnosis collaborative rule matching, etc.) to obtain the coding results to be verified.

[0040] The encoding results to be verified are input into the strong rule quality control engine. This engine has built-in hard constraint rules for ICD encoding (such as encoding format verification, required field verification, business logic conflict verification, etc.) to perform final verification and merging processing on the results, eliminating invalid encodings and correcting logical conflicts.

[0041] In this embodiment of the invention, S4 includes the following sub-steps: S41. Based on the global verification result, receive the registration request of the identifier factory, complete the identifier registration, and assign a unique processing factory for each type of identifier; S42. Extract the type and path information of the identifier and create an identifier instance; S43. Based on the identifier instance, complete the identifier resolution; S44. Perform a redirection operation based on the identifier resolution result.

[0042] In this embodiment of the invention, in S41, the identifier type includes business data identifier, knowledge base identifier, and knowledge point identifier; Business data identifiers are used to access business data and to retrieve and reference it; business data includes user information, business process data, and transaction records. Knowledge base identifiers are used to complete knowledge base retrieval; Knowledge point identifiers are used to retrieve and reference knowledge point data; knowledge point data includes concept definitions and terminology explanations.

[0043] In this embodiment of the invention, in S43, the identifier parsing process includes an initialization stage, a path parsing stage, a data retrieval stage, and a result processing stage; The initialization phase specifically involves: acquiring runtime input data as the parsing context, and obtaining the basic parsing data based on the identifier type; The path resolution phase specifically involves: breaking down the path information into several path units, each representing an operation type; path units include sub-objects, references, properties, static properties, group operations, and filtering operations; The data retrieval stage specifically involves: obtaining the retrieved data based on the type of the identifier; The result processing stage specifically involves: setting the retrieved data as a reference object of the identifier, updating the identifier's state, parsing it, and returning the parsing result to the requester.

[0044] Precise parsing specifically involves: starting from the data source of the identifier, performing the positioning operation sequentially according to the order of the path units; Fuzzy parsing specifically involves performing a fuzzy matching operation on the final level after performing the sub-final level positioning.

[0045] In this embodiment of the invention, when the fuzzy parsing result is empty, the redirection target is determined according to the type and path of the identifier, a redirection request is constructed, and the identifier parsing is re-executed.

[0046] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. An ICD intelligent encoding method based on a multi-level knowledge-driven large language model, characterized in that, Includes the following steps: S1. Construct a sub-model matching knowledge base to select an appropriate sub-model for diagnostic text; S2. Construct a single-diagnosis coding knowledge base, call the adaptation sub-model, and generate preliminary coding results for a single diagnosis of the diagnostic text; S3. Construct a global verification knowledge base to perform global verification on all preliminary coding results; S4. Input the global verification results into the Zhiqiang rule quality control engine for final verification and merging.

2. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 1, characterized in that, In step S1, the comprehensive adaptation score of several candidate sub-models is calculated, and the Top-K method is used to vote on each candidate sub-model to determine the adaptation sub-model for the diagnostic text. The comprehensive fit score of the candidate sub-model The expression is: ； in, Indicates the first One candidate sub-model Indicates the first Dynamic weights for each dimension This indicates diagnostic text. The candidate sub-model's specialty domain and diagnostic text are represented in the first... Similarity across multiple dimensions.

3. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Calculate the medical semantic contribution of each character in a single diagnosis; S22. Set a filtering threshold and filter characters based on their contribution to medical semantics; S23. Based on the adaptor sub-model, calculate the confidence level of the candidate codes of the filtered characters; S24. Set a dynamic confidence threshold and filter candidate codes based on their confidence levels; S25. Use the multi-source evidence weighted voting method to vote on the screened candidate codes, and take the candidate code with the largest vote value as the preliminary coding result.

4. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 3, characterized in that, In S21, the first Medical semantic contribution of each character The expression is: ； in, The first in the diagnostic text One character, Indicates the first Term frequency of each character - inverse document frequency, Indicates the first Weight of medical terms per character, Indicates the first The encoding correlation of each character The weighting coefficients represent the term frequency-inverse document frequency ratio. Weighting coefficients representing the weights of medical terms. Weighting coefficients representing the degree of correlation between codes; In S23, the candidate coding confidence level The expression is: ； in, Indicates the first One candidate ICD code, This indicates diagnostic text. Indicates the first Knowledge base matching degree between candidate ICD codes and diagnostic texts The sub-model output number The probability values ​​of each candidate ICD code. Indicates the weighting coefficient; In S25, the expression for the multi-source evidence weighted voting method is: ； in, Indicates the first The voting results for each candidate ICD code. Indicates the first Consistency of medical terminology in candidate ICD-coded diagnostic texts. Indicates the first Consistency of medical terminology in candidate ICD-coded diagnostic texts. The voting weights represent the degree of knowledge base matching. Voting weights representing historical successful coding rates Voting weights indicating consistency in medical terminology.

5. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 1, characterized in that, In S3, the global verification knowledge base is used for knowledge enhancement when verifying the preliminary coding results of the single diagnostic coding knowledge base.

6. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Based on the global verification result, receive the registration request of the identifier factory, complete the identifier registration, and assign a unique processing factory for each type of identifier; S42. Extract the type and path information of the identifier and create an identifier instance; S43. Based on the identifier instance, complete the identifier resolution; S44. Perform a redirection operation based on the identifier resolution result.

7. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 6, characterized in that, In S41, the identifier types include business data identifiers, knowledge base identifiers, and knowledge point identifiers; The business data identifier is used to access business data and to retrieve and reference business data; the business data includes user information, business process data, and transaction records. The knowledge base identifier is used to complete the knowledge base retrieval; The knowledge point identifier is used to complete the retrieval and citation of knowledge point data; the knowledge point data includes concept definitions and term explanations.

8. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 6, characterized in that, In S43, the identifier resolution process includes an initialization phase, a path resolution phase, a data retrieval phase, and a result processing phase. The initialization phase specifically involves: acquiring runtime input data as the parsing context, and acquiring the basic parsing data based on the identifier type; The path resolution stage specifically involves: breaking down the path information into several path units, each representing an operation type; the path units include sub-objects, references, attributes, static attributes, group operations, and filtering operations; The data retrieval stage specifically involves: obtaining retrieval data based on the type of the identifier; The result processing stage specifically involves: setting the retrieved data as a reference object of the identifier, updating the state of the identifier, parsing it, and returning the parsing result to the requester.

9. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 8, characterized in that, The result processing stage uses either precise parsing or fuzzy parsing for analysis. The precise parsing specifically involves: starting from the data source of the identifier, performing the positioning operation sequentially according to the order of the path units; The fuzzy parsing specifically involves performing a fuzzy matching operation on the final level after performing the sub-final level positioning.

10. The ICD intelligent encoding method based on a multi-level knowledge-driven large language model according to claim 9, characterized in that, When the fuzzy parsing result is empty, the redirection target is determined based on the identifier type and path, a redirection request is constructed, and the identifier parsing is re-executed.

Images