Intelligent auditing method and system for clinical test accommodating and arranging rule
By using intelligent slicing and multi-level retrieval and sorting methods for medical record texts, the problems of low efficiency, poor accuracy, and insufficient interpretability in the review of clinical trial inclusion and exclusion rules are solved. This achieves efficient and accurate automated review, applicable to cross-visit medical records of any length, and meets the high concurrency requirements of clinical trials.
Patent Information
- Application Number
- CN202610543687.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-23
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies suffer from low efficiency, poor accuracy, limited applicability, and insufficient interpretability in the review of clinical trial inclusion and exclusion rules. In particular, they are prone to truncation or excessive consumption of computational resources when processing long text medical records, and the model output lacks traceability.
By intelligently slicing medical record texts into standardized short text fragments, and combining a multi-level retrieval and ranking mechanism, only the most relevant fragments are input into the large language model for inference. Combined with terminology expansion and rule decomposition, a three-level retrieval strategy of recall, coarse ranking, and fine ranking is adopted to filter redundant information, ensure that the text length is within the model's context window, and enhance expressive power and accuracy.
It significantly improves review efficiency and accuracy, reduces computational overhead and response time, meets the high concurrency requirements of large-scale clinical trials, provides interpretable judgment criteria, supports result traceability, and has good versatility and scalability.
Smart Images
Figure CN122067685A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical information technology, and in particular to an intelligent review method and system for clinical trial inclusion and exclusion rules. Background Technology
[0002] Clinical trials are a crucial step in drug development and medical technology evaluation, and the accuracy and efficiency of patient recruitment and screening directly impact the trial's progress and quality. Clinical trials typically establish detailed inclusion and exclusion criteria (hereinafter referred to as "inclusion and exclusion rules") to screen eligible patients. Currently, the review of inclusion and exclusion rules mainly relies on manual review or large language models directly processing complete medical records.
[0003] The manual review process involves medical professionals meticulously checking patient medical records against inclusion and exclusion rules, extracting key information, and manually determining compliance. While this method ensures interpretability to some extent, it suffers from low efficiency, high subjectivity, high error rates, and high labor costs, making it difficult to meet the demands of large-scale clinical trials for efficient and standardized screening.
[0004] The direct review approach using a large language model concatenates the patient's complete medical record text with the inclusion and exclusion rules before inputting it into the large language model, which then directly outputs the judgment result. However, this method has the following problems: The length of medical record texts is likely to exceed the model's token limit, leading to input truncation or inference failure; The medical records contain a large amount of redundant information that is irrelevant to the inclusion and exclusion rules, which interferes with the accuracy of the model's judgment. Full-text inference consumes significant computational resources and has a high response latency. The model output lacks interpretability and fails to meet the traceability requirements of clinical trials for the basis of judgment.
[0005] In summary, existing technologies suffer from low efficiency, poor accuracy, limited applicability, and insufficient interpretability in the review of clinical trial inclusion and exclusion rules. There is an urgent need for an automated review method and system that is efficient, accurate, and interpretable. Summary of the Invention
[0006] Based on this, it is necessary to provide an intelligent review method and system for clinical trial inclusion and exclusion rules to address the aforementioned technical issues. By intelligently slicing medical record text, long texts are divided into standardized short text fragments. Combined with a multi-level retrieval and sorting mechanism, only the most relevant few fragments are input into a large language model for inference, ensuring that the length of the input text is always within the model context window. This avoids truncation or error problems caused by excessively long texts and is applicable to cross-visit medical records of any length.
[0007] This invention provides an intelligent review method for clinical trial inclusion and exclusion rules, the method comprising: Obtain the original inclusion and exclusion rules of clinical trials, and parse and enhance the original inclusion and exclusion rules to generate an enhanced rule query set; The complete medical record text of the patient is obtained, and the complete medical record text is intelligently sliced to obtain multiple standardized short medical record text fragments with metadata tags. An index library is built based on all the short medical record text fragments. The enhanced rule query set and the index are sorted and retrieved at multiple levels. A preset number of short medical record text fragments that meet the preset requirements in relevance to the enhanced rule query set are retrieved from the index. The preset number of short medical record text fragments are concatenated with the original inclusion and exclusion rules to construct inference prompt words, which are then input into the large language model for inference to obtain the judgment result and judgment basis of a single inclusion and exclusion rule. Based on the preset rule aggregation logic, the judgment results of all inclusion and exclusion rules are integrated to output the final inclusion or exclusion recommendations for patients and the comprehensive judgment basis.
[0008] In one embodiment, the step of parsing and enhancing the original inclusion and sorting rules to generate an enhanced rule query set includes: Based on a medical thesaurus, the core medical terms in the original inclusion and exclusion rules are expanded with synonyms, near-synonyms, and medical abbreviations. The complex Na-Square rule containing multiple logical conditions is broken down into multiple atomic sub-rules; The original declarative expression of the sorting rules is rewritten into a combination of keywords or natural language queries suitable for retrieval, generating an enhanced rule query set that includes extended terms, atomic sub-rules, and retrieval expressions.
[0009] In one embodiment, the intelligent slicing process on the complete medical record text yields multiple standardized short medical record text fragments with metadata tags, including: Based on the type of medical record, one or more strategies are used to segment long text medical records into short text segments of a preset length, including fixed-size sliding window slicing, semantic segmentation based on a medical domain semantic segmentation model, or structure preservation segmentation based on inherent fields of medical records. Metadata tags including medical record type, consultation time, and field name are added to each short text fragment to obtain standardized medical record short text fragments.
[0010] In one embodiment, building an index based on all short text fragments of medical records includes: Based on the keywords and medical terms in the short medical record text fragments, an inverted index for fast keyword matching is constructed; The Embedding service is invoked to perform vector encoding on each of the short medical record text fragments, generating semantic vectors, and a vector index for similarity retrieval is constructed based on the semantic vectors.
[0011] In one embodiment, the step of performing multi-level retrieval and sorting of the enhanced rule query set and the index, and retrieving a preset number of short medical record text fragments from the index that meet preset relevance requirements to the enhanced rule query set, includes: Based on the inverted index, the BM25 algorithm is used to recall a first preset number of candidate medical record short text fragments related to the enhanced rule query from the index database; Calculate the vector semantic similarity and TF-IDF term frequency similarity between the enhanced rule query and the first preset number of candidate segments, perform weighted fusion of the two similarities, and filter out the second preset number of candidate segments after sorting by the comprehensive score; The enhanced rule query and the second preset number of candidate fragments are input into the cross-encoder model for deep interactive calculation. Based on the output relevance score, the third preset number of short medical record text fragments that are most relevant to the inclusion and exclusion rules are selected.
[0012] In one embodiment, the step of concatenating a preset number of recalled short medical record text fragments with the original exclusion rule to construct inference prompts is input into a large language model for inference, obtaining the judgment result and judgment basis for a single exclusion rule, including: The predetermined number of recalled medical record short text fragments are concatenated according to their relevance scores from high to low, and the metadata tags of each fragment are retained. Construct standardized reasoning prompts that include the instruction part, the original arrangement rules part, and the spliced medical record basis part; The inference prompts are input into a large language model for inference, and the structured single rule judgment results and corresponding judgment criteria are extracted from the model output.
[0013] In one embodiment, the step of synthesizing the judgment results of all inclusion and exclusion rules according to a preset rule aggregation logic, and outputting the patient's final inclusion or exclusion recommendation and comprehensive judgment basis, includes: Obtain the judgment results of all inclusion and exclusion rules; Based on the preset aggregation logic, a comprehensive judgment is made. If all inclusion rules are judged to be compliant and all exclusion rules are judged to be ineligible, then a patient inclusion suggestion is output; otherwise, a patient exclusion suggestion is output. Integrate the judgment criteria of all individual rules to generate a comprehensive judgment criterion for source tracing.
[0014] The present invention also provides an intelligent review system for clinical trial inclusion and exclusion rules, applied to the intelligent review method for clinical trial inclusion and exclusion rules described in any of the above embodiments, the system comprising: The rule processing module is used to obtain the original inclusion and exclusion rules of clinical trials, and to parse and enhance the original inclusion and exclusion rules to generate an enhanced rule query set; The case processing module is used to obtain the patient's complete medical record text, and to perform intelligent slicing processing on the complete medical record text to obtain multiple standardized short medical record text fragments with metadata tags, and to build an index library based on all the short medical record text fragments; The three-level retrieval and sorting module is used to perform multi-level retrieval and sorting of the enhanced rule query set and the index library, and to recall a preset number of short medical record text fragments from the index library that meet the preset requirements for relevance to the enhanced rule query set. The judgment module is used to concatenate the recalled preset number of short medical record text fragments with the original inclusion and exclusion rules, construct inference prompt words, and input them into the large language model for inference to obtain the judgment result and judgment basis of a single inclusion and exclusion rule; The results aggregation module is used to synthesize the judgment results of all inclusion and exclusion rules according to the preset rule aggregation logic, and output the final inclusion or exclusion suggestions and comprehensive judgment basis for the patient.
[0015] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the intelligent review method for clinical trial inclusion and exclusion rules as described above.
[0016] The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the intelligent review method for clinical trial inclusion and exclusion rules as described above.
[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent review method for clinical trial inclusion and exclusion rules as described above.
[0018] The aforementioned intelligent review method and system for clinical trial inclusion and exclusion rules intelligently segments medical record texts into standardized short text fragments. Combined with a multi-level retrieval and ranking mechanism, only the most relevant fragments are input into a large language model for inference, ensuring that the length of the input text remains within the model's context window. This avoids truncation or error issues caused by excessively long texts, making it suitable for cross-visit medical records of any length. The system enhances the expressive power of the original inclusion and exclusion rules through terminology expansion, rule decomposition, and retrieval-based rewriting. Simultaneously, a three-level retrieval strategy (recall, coarse ranking, and fine ranking) accurately selects medical record fragments highly relevant to the rules, filtering redundant information. This allows the large language model to focus on key judgment criteria for inference, significantly improving the accuracy and consistency of the judgment results. By automating the entire process from medical record acquisition, rule processing, retrieval and ranking to model inference and result aggregation, no manual intervention is required, increasing review efficiency by more than 100 times. This meets the needs of large-scale, high-concurrency patient screening in clinical trials while eliminating subjective differences from manual review, improving the consistency and reproducibility of results. By inputting only a small number of key medical record fragments into the large model, the number of input tokens is significantly reduced, lowering the computational overhead and response time of model inference. This makes it suitable for high-concurrency deployment in production environments, reducing the cost of technology implementation. The large language model outputs judgment results along with clear judgment criteria and retains metadata tags for the medical record fragments. The result aggregation module integrates the judgment criteria of all rules to form a comprehensive judgment description, supporting manual review and result traceability, meeting the traceability and compliance requirements of clinical trials. The decoupled design of each functional module allows for flexible replacement of components such as embedding, sorting, and the large language model. Search parameters, slicing strategies, and aggregation logic support configurable adjustments, enabling rapid adaptation to different clinical trial inclusion and exclusion rules and data types, demonstrating good versatility and scalability. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0020] Figure 1 The flowchart of the intelligent review method for clinical trial inclusion and exclusion rules provided by this invention; Figure 2 A block diagram of the intelligent review system for clinical trial inclusion and exclusion rules provided by this invention; Figure 3 An internal structural diagram of the computer device provided by the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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.
[0022] The following is combined with Figures 1 to 3 This invention describes an intelligent review method and system for clinical trial inclusion and exclusion rules.
[0023] In one embodiment, a smart review method for clinical trial inclusion / exclusion rules includes the following steps: Step S100: Obtain the original inclusion and exclusion rules of the clinical trial, and parse and enhance the original inclusion and exclusion rules to generate an enhanced rule query set.
[0024] Step S200: Obtain the patient's complete medical record text, and perform intelligent slicing processing on the complete medical record text to obtain multiple standardized short medical record text fragments with metadata tags, and build an index library based on all the short medical record text fragments.
[0025] Step S300: Perform multi-level retrieval and sorting of the enhanced rule query set and the index library, and retrieve a preset number of short medical record text fragments from the index library that meet the preset requirements for relevance to the enhanced rule query set.
[0026] Step S400: The preset number of recalled medical record short text fragments are concatenated with the original inclusion and exclusion rules to construct inference prompt words, which are then input into the large language model for inference to obtain the judgment result and judgment basis of a single inclusion and exclusion rule.
[0027] Step S500: Based on the preset rule aggregation logic, the judgment results of all inclusion and exclusion rules are integrated, and the final inclusion or exclusion suggestions and comprehensive judgment basis for the patient are output.
[0028] The aforementioned intelligent review method for clinical trial inclusion and exclusion rules intelligently segments medical record texts into standardized short text fragments. Combined with a multi-level retrieval and ranking mechanism, only the most relevant fragments are input into the large language model for inference, ensuring that the input text length remains within the model's context window. This avoids truncation or error issues caused by excessively long texts and is applicable to cross-visit medical records of any length. The method enhances the expressive power of the original inclusion and exclusion rules by expanding terminology, decomposing rules, and rewriting retrieval-based rules. Simultaneously, a three-level retrieval strategy (recall, coarse ranking, and fine ranking) accurately selects medical record fragments highly relevant to the rules, filtering redundant information. This allows the large language model to focus on key judgment criteria for inference, significantly improving the accuracy and consistency of judgment results. By automating the entire process from medical record acquisition, rule processing, retrieval and ranking to model inference and result aggregation, no manual intervention is required, improving review efficiency by more than 100 times. This meets the needs of large-scale, high-concurrency patient screening in clinical trials while eliminating subjective differences from manual review, improving the consistency and reproducibility of results. By inputting only a small number of key medical record fragments into the large model, the number of input tokens is significantly reduced, lowering the computational overhead and response time of model inference. This makes it suitable for high-concurrency deployment in production environments, reducing the cost of technology implementation. The large language model outputs judgment results along with clear judgment criteria and retains metadata tags for the medical record fragments. The result aggregation module integrates the judgment criteria of all rules to form a comprehensive judgment description, supporting manual review and result traceability, meeting the traceability and compliance requirements of clinical trials. The decoupled design of each functional module allows for flexible replacement of components such as embedding, sorting, and the large language model. Search parameters, slicing strategies, and aggregation logic support configurable adjustments, enabling rapid adaptation to different clinical trial inclusion and exclusion rules and data types, demonstrating good versatility and scalability.
[0029] In one embodiment, the step of parsing and enhancing the original inclusion and sorting rules to generate an enhanced rule query set includes the following steps: Step S110: Based on a medical professional thesaurus, expand the core medical terms in the original inclusion rules with synonyms, near-synonyms, and medical abbreviations.
[0030] Specifically, based on a medical thesaurus, the core medical terms in the rules (such as disease names, treatment methods, examination indicators, drug names, etc.) are expanded with synonyms, near-synonyms, and medical abbreviations.
[0031] For example, expanding "hypertension" in the rule to "hypertension, hypertensive disorders, HTN" can improve the coverage of subsequent searches. The expanded term set is used as part of the enhanced rule query for subsequent searches.
[0032] Step S120: Decompose the composite array rule containing multiple logical conditions into multiple atomic sub-rules.
[0033] Specifically, the composite nano-row rule is broken down into multiple atomic sub-rules.
[0034] For example, the rule "excluding patients who have received chemotherapy in the past 3 months and have diabetes" can be broken down into two sub-rules: "excluding patients who have received chemotherapy in the past 3 months" and "excluding patients with diabetes". This makes it easier to search and verify them separately later.
[0035] Step S130: Rewrite the declarative expression of the original inclusion and exclusion rules into a keyword combination or natural language query suitable for retrieval, and generate an enhanced rule query set containing extended terms, atomic sub-rules and retrieval expressions.
[0036] Specifically, the declarative expressions of the inclusion and exclusion rules are converted into retrieval-based expressions.
[0037] For example, rewriting "patients diagnosed with lung cancer who have not received surgical treatment" as "patients diagnosed with lung cancer who have not received surgical treatment" can generate keyword combinations or natural language queries that are adapted to the search engine.
[0038] Finally, steps S110 to S130 generate an enhanced rule query set containing extended terms, atomic sub-rules, and retrieval expressions, providing rich query input for subsequent retrievals.
[0039] In one embodiment, the intelligent slicing process of the complete medical record text to obtain multiple standardized short medical record text fragments with metadata tags includes the following steps: Step S210: Based on the medical record type, one or more strategies are used to segment long text medical records into short text segments of a preset length, including fixed-size sliding window slicing, semantic segmentation based on a medical domain semantic segmentation model, or structure preservation segmentation based on inherent fields of the medical record.
[0040] Specifically, a strategy combining three slicing methods is employed. The appropriate slicing method is automatically selected based on the medical record type, and all sliced results are uniformly formatted as short texts of a preset length (e.g., 512 tokens or 1024 tokens). Among these: For plain text medical records without obvious structure (such as medical records), a fixed-size sliding window is used for slicing. The sliding window is segmented according to a preset number of tokens, and the window step size is set to 50% of the preset number of tokens to avoid key information being cut into two fragments. For medical records with strong semantic logic, a semantic segmentation model in the medical field is used to perform semantic segmentation based on the semantic logic and paragraph division of the medical record text, ensuring that each segment contains a complete semantic unit. For example, the "present illness" paragraph is treated as an independent segment to avoid semantic breaks. For structured or semi-structured medical records (such as laboratory reports and HL7 / CDA format medical records), a structure-preserving block-based approach is adopted, which segments the records according to their inherent fields (such as "diagnosis results", "medication records", and "examination indicators") to strictly preserve the structural information and metadata of the medical records.
[0041] Step S220: Add metadata tags including medical record type, consultation time and field name to each short text fragment to obtain standardized medical record short text fragments.
[0042] Specifically, metadata tags are added to the short text fragments generated for each slice, including medical record type (such as discharge summary, progress notes, and test reports), consultation time, and field names (such as "chief complaint", "present illness", and "diagnosis result"), and stored in a standardized manner to facilitate subsequent retrieval and traceability.
[0043] In one embodiment, building an index based on all short text fragments of medical records includes the following steps: Step S230: Based on the keywords and medical terms in the short medical record text fragment, construct an inverted index for fast keyword matching.
[0044] Specifically, an inverted index is built based on keywords and medical terms in each short medical record text fragment, supporting fast keyword matching retrieval. The inverted index uses terms as keys to record a list of fragments containing those terms, which is used for fast retrieval in the subsequent recall phase.
[0045] Step S240: Call the Embedding service to perform vector encoding on each of the medical record short text fragments, generate semantic vectors, and construct a vector index for similarity retrieval based on the semantic vectors.
[0046] Specifically, an embedding service (such as the BGE-M3 model) is invoked to vectorize each short medical record text fragment, generating a corresponding semantic vector. All semantic vectors are then used to construct a vector index library, supporting subsequent retrieval based on vector similarity. The vector index and the inverted index together constitute the index library, providing the foundation for multi-level retrieval and ranking.
[0047] In one embodiment, the step of performing multi-level retrieval and sorting of the enhanced rule query set and the index database, and retrieving a preset number of short medical record text fragments from the index database that meet preset relevance requirements to the enhanced rule query set, includes the following steps: Step S310: Based on the inverted index, the BM25 algorithm is used to recall a first preset number of candidate medical record short text fragments related to the enhanced rule query from the index.
[0048] Specifically, during the recall phase, the BM25 keyword matching algorithm is used, based on an inverted index, to quickly retrieve medical record fragments containing keywords or medical terms from enhanced rule queries.
[0049] This stage aims for high recall, appropriately relaxing the screening criteria and outputting the number of candidate segments as Top N, such as Top 200, to ensure that no medical record segments potentially related to the rules are missed.
[0050] Step S320: Calculate the vector semantic similarity and TF-IDF word frequency similarity between the enhanced rule query and the first preset number of candidate segments, perform weighted fusion of the two similarities, and filter out the second preset number of candidate segments after sorting by the comprehensive score.
[0051] Specifically, in the coarse ranking stage, the Embedding service (e.g., the BGE-M3 model) is first invoked to perform real-time semantic vector encoding on the enhanced rule query and Top N candidate segments, generating rule query vectors and segment vectors. Then, the TF-IDF term frequency similarity and vector semantic similarity between the rule query and the candidate segments are calculated. The two similarities are then weighted and fused, and the weights can be optimized according to the actual scenario, for example, semantic similarity weight 0.7 and TF-IDF similarity weight 0.3, to obtain a comprehensive coarse ranking score. Finally, the comprehensive coarse ranking scores are sorted from high to low to select the Top M candidate segments, such as the Top 50, in order to reduce the computational load of subsequent fine ranking.
[0052] Step S330: Input the enhanced rule query and the second preset number of candidate fragments into the cross-encoder model for deep interactive calculation, and filter out the third preset number of short medical record text fragments that are most relevant to the inclusion and exclusion rules based on the output relevance score.
[0053] Specifically, in the fine-grained ranking stage, the cross-encoder model (e.g., BGE Reranker) in the ranking service is invoked to perform deep interactive calculations between the enhanced rule query and each Top M candidate fragment. The model outputs a precise relevance score between the two. The fragments are then sorted from highest to lowest relevance score, and the Top K medical record fragments most relevant to the ranking rules are selected, for example, the Top 10. The value of K can be flexibly adjusted according to the context length limitations of the large language model. Finally, a set of short medical record text fragments highly relevant to the ranking rules is obtained.
[0054] In one embodiment, the step of concatenating a preset number of recalled short medical record text fragments with the original exclusion rule to construct inference prompts, and inputting them into a large language model for inference to obtain the judgment result and judgment basis of a single exclusion rule, includes the following steps: Step S410: The pre-set number of recalled medical record short text fragments are concatenated according to their relevance scores from high to low, and the metadata tags of each fragment are retained.
[0055] Specifically, the Top K medical record short text fragments obtained after fine ranking are concatenated according to their relevance scores from high to low. Each fragment retains its metadata tags (such as medical record type, consultation time, and field name) to facilitate the large language model in identifying the source of information and improve the accuracy of inference.
[0056] Step S420: Construct standardized reasoning prompts that include the instruction part, the original inclusion and exclusion rules part, and the spliced medical record basis part.
[0057] Specifically, a standardized inference prompt is constructed, which consists of three parts: the instruction part, the inclusion and exclusion rules part, and the medical record basis part.
[0058] For example: "Based on the provided patient medical record excerpt, strictly determine whether the patient meets the following clinical trial inclusion and exclusion rules, output only 'compliant' / 'non-compliant', and briefly explain the basis for the judgment."
[0059] Arrangement rules: {Original arrangement rules}.
[0060] Medical record basis: "{Top K medical record fragments spliced together}". This prompt has a clear structure, guiding the model to focus on key information for judgment.
[0061] Step S430: Input the inference prompt words into the large language model for inference, and extract the structured single-rule judgment result and corresponding judgment basis from the model output.
[0062] Specifically, the constructed inference prompts are input into a large language model service (e.g., GPT-4) for online real-time inference. The model outputs a judgment result based on key medical record fragments and inclusion / exclusion rules in the prompts. Structured single inclusion / exclusion rule judgment results ("compliant" or "incompatible") and the corresponding judgment basis (usually key information from the medical record fragments referenced by the model) are extracted from the model output.
[0063] In one embodiment, the step of synthesizing the judgment results of all inclusion and exclusion rules according to a preset rule aggregation logic, and outputting the final inclusion or exclusion recommendation and comprehensive judgment basis for the patient, includes the following steps: Step S510: Obtain the judgment results of all inclusion and exclusion rules.
[0064] Specifically, the LLM judgment module obtains the judgment results and corresponding judgment basis of all inclusion and exclusion rules (including inclusion and exclusion rules) as input to the aggregation module.
[0065] Step S520: Perform a comprehensive judgment based on the preset aggregation logic. If all inclusion rules are judged to be compliant and all exclusion rules are judged to be ineligible, output a patient inclusion suggestion; otherwise, output a patient exclusion suggestion.
[0066] Specifically, a comprehensive judgment is made based on the aggregation logic according to preset standardized rules. This aggregation logic supports configurable modification to adapt to the needs of different clinical trials.
[0067] The core logic is as follows: if all inclusion rules are judged as "compliant" and all exclusion rules are judged as "incompatible", then output the patient's "inclusion" suggestion; otherwise, output the patient's "exclusion" suggestion.
[0068] Step S530: Integrate the judgment criteria of all individual rules to generate a comprehensive judgment criteria for tracing the source.
[0069] Specifically, the criteria for each inclusion / exclusion rule are integrated to form a comprehensive judgment basis, clearly marking the key medical record segments (including metadata information) on which the judgment is based, thus achieving traceability. Finally, the system outputs the patient's final clinical trial inclusion or exclusion recommendation and the comprehensive judgment basis.
[0070] The intelligent review system for clinical trial inclusion and exclusion rules provided by this invention is described below. The intelligent review system for clinical trial inclusion and exclusion rules described below can be referred to in correspondence with the intelligent review method for clinical trial inclusion and exclusion rules described above.
[0071] In one embodiment, a clinical trial inclusion and exclusion rule intelligent review system is applied to the clinical trial inclusion and exclusion rule intelligent review method described in any of the above embodiments, including a rule processing module, a case processing module, a three-level retrieval and sorting module, a judgment module, and a result aggregation module.
[0072] The rule processing module is used to obtain the original inclusion and exclusion rules of clinical trials, and to parse and enhance the original inclusion and exclusion rules to generate an enhanced rule query set.
[0073] The case processing module is used to obtain the patient's complete medical record text, and to perform intelligent slicing processing on the complete medical record text to obtain multiple standardized short medical record text fragments with metadata tags, and to build an index library based on all the short medical record text fragments.
[0074] The three-level retrieval and sorting module is used to perform multi-level retrieval and sorting of the enhanced rule query set and the index library, and to recall a preset number of short medical record text fragments from the index library that meet the preset requirements for relevance to the enhanced rule query set.
[0075] The judgment module is used to concatenate a preset number of short medical record text fragments with the original inclusion and exclusion rules, construct inference prompt words, and input them into the large language model for inference to obtain the judgment result and judgment basis of a single inclusion and exclusion rule.
[0076] The results aggregation module is used to aggregate the judgment results of all inclusion and exclusion rules according to the preset rule aggregation logic, and output the final inclusion or exclusion suggestions for patients and the comprehensive judgment basis.
[0077] Figure 3 This example illustrates a schematic diagram of the physical structure of an electronic device, which can be a smart terminal. Its internal structure diagram can be as follows: Figure 3 As shown, the electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an intelligent review method for clinical trial inclusion and exclusion rules.
[0078] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0079] On the other hand, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements an intelligent review method for clinical trial inclusion and exclusion rules.
[0080] On another front, a computer program product or computer program is provided, comprising computer instructions stored in a computer storage medium. A processor of an electronic device reads the computer instructions from the computer storage medium, and when the processor executes the computer instructions, it implements an intelligent review method for clinical trial inclusion and exclusion rules.
[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.
[0082] By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0084] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.
Claims
1. A method for intelligent review of clinical trial inclusion and exclusion rules, characterized in that, The method includes: Obtain the original inclusion and exclusion rules of clinical trials, and parse and enhance the original inclusion and exclusion rules to generate an enhanced rule query set; The complete medical record text of the patient is obtained, and the complete medical record text is intelligently sliced to obtain multiple standardized short medical record text fragments with metadata tags. An index library is built based on all the short medical record text fragments. The enhanced rule query set and the index are sorted and retrieved at multiple levels. A preset number of short medical record text fragments that meet the preset requirements in relevance to the enhanced rule query set are retrieved from the index. The preset number of short medical record text fragments are concatenated with the original inclusion and exclusion rules to construct inference prompt words, which are then input into the large language model for inference to obtain the judgment result and judgment basis of a single inclusion and exclusion rule. Based on the preset rule aggregation logic, the judgment results of all inclusion and exclusion rules are integrated to output the final inclusion or exclusion recommendations for patients and the comprehensive judgment basis.
2. The intelligent review method for clinical trial inclusion and exclusion rules according to claim 1, characterized in that, The process of parsing and enhancing the original sorting rules to generate an enhanced rule query set includes: Based on a medical thesaurus, the core medical terms in the original inclusion and exclusion rules are expanded with synonyms, near-synonyms, and medical abbreviations. The complex Na-Square rule containing multiple logical conditions is broken down into multiple atomic sub-rules; The original declarative expression of the sorting rules is rewritten into a combination of keywords or natural language queries suitable for retrieval, generating an enhanced rule query set that includes extended terms, atomic sub-rules, and retrieval expressions.
3. The intelligent review method for clinical trial inclusion and exclusion rules according to claim 2, characterized in that, The intelligent slicing process of the complete medical record text yields multiple standardized short medical record text fragments with metadata tags, including: Based on the type of medical record, one or more strategies are used to segment long text medical records into short text segments of a preset length, including fixed-size sliding window slicing, semantic segmentation based on a medical domain semantic segmentation model, or structure preservation segmentation based on inherent fields of medical records. Metadata tags including medical record type, consultation time, and field name are added to each short text fragment to obtain standardized medical record short text fragments.
4. The intelligent review method for clinical trial inclusion and exclusion rules according to claim 3, characterized in that, The index library built based on all short text fragments of medical records includes: Based on the keywords and medical terms in the short medical record text fragments, an inverted index for fast keyword matching is constructed; The Embedding service is invoked to perform vector encoding on each of the short medical record text fragments, generating semantic vectors, and a vector index for similarity retrieval is constructed based on the semantic vectors.
5. The intelligent review method for clinical trial inclusion and exclusion rules according to claim 4, characterized in that, The step of performing multi-level retrieval and sorting of the enhanced rule query set and the index, and retrieving a preset number of short medical record text fragments from the index that meet preset relevance requirements to the enhanced rule query set, includes: Based on the inverted index, the BM25 algorithm is used to recall a first preset number of candidate medical record short text fragments related to the enhanced rule query from the index database; Calculate the vector semantic similarity and TF-IDF term frequency similarity between the enhanced rule query and the first preset number of candidate segments, perform weighted fusion of the two similarities, and filter out the second preset number of candidate segments after sorting by the comprehensive score; The enhanced rule query and the second preset number of candidate fragments are input into the cross-encoder model for deep interactive calculation. Based on the output relevance score, the third preset number of short medical record text fragments that are most relevant to the inclusion and exclusion rules are selected.
6. The intelligent review method for clinical trial inclusion and exclusion rules according to claim 5, characterized in that, The process involves concatenating a predetermined number of recalled short medical record text fragments with the original exclusion rules to construct inference prompts, which are then input into a large language model for inference. This yields the judgment result and basis for each exclusion rule, including: The predetermined number of recalled medical record short text fragments are concatenated according to their relevance scores from high to low, and the metadata tags of each fragment are retained. Construct standardized reasoning prompts that include the instruction part, the original arrangement rules part, and the spliced medical record basis part; The inference prompts are input into a large language model for inference, and the structured single rule judgment results and corresponding judgment criteria are extracted from the model output.
7. The intelligent review method for clinical trial inclusion and exclusion rules according to claim 6, characterized in that, The system aggregates the results of all inclusion and exclusion rules according to a preset rule aggregation logic, and outputs the final inclusion or exclusion recommendation for the patient and the comprehensive judgment basis, including: Obtain the judgment results of all inclusion and exclusion rules; Based on the preset aggregation logic, a comprehensive judgment is made. If all inclusion rules are judged to be compliant and all exclusion rules are judged to be ineligible, then a patient inclusion suggestion is output; otherwise, a patient exclusion suggestion is output. Integrate the judgment criteria of all individual rules to generate a comprehensive judgment criterion for source tracing.
8. A smart review system for clinical trial inclusion and exclusion rules, applied to the smart review method for clinical trial inclusion and exclusion rules according to any one of claims 1 to 7, characterized in that, The system includes: The rule processing module is used to obtain the original inclusion and exclusion rules of clinical trials, and to parse and enhance the original inclusion and exclusion rules to generate an enhanced rule query set; The case processing module is used to obtain the patient's complete medical record text, and to perform intelligent slicing processing on the complete medical record text to obtain multiple standardized short medical record text fragments with metadata tags, and to build an index library based on all the short medical record text fragments; The three-level retrieval and sorting module is used to perform multi-level retrieval and sorting of the enhanced rule query set and the index library, and to recall a preset number of short medical record text fragments from the index library that meet the preset requirements for relevance to the enhanced rule query set. The judgment module is used to concatenate the recalled preset number of short medical record text fragments with the original inclusion and exclusion rules, construct inference prompt words, and input them into the large language model for inference to obtain the judgment result and judgment basis of a single inclusion and exclusion rule; The results aggregation module is used to synthesize the judgment results of all inclusion and exclusion rules according to the preset rule aggregation logic, and output the final inclusion or exclusion suggestions and comprehensive judgment basis for the patient.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.