Medical information processing method, system and medium, edge computing device

Abstract

The application provides a medical information processing method, system and medium, and an edge computing device. The medical information processing method comprises the following steps: analyzing an input question as a structured query; searching the structured query in a database by using a dual-mode retrieval engine to obtain a candidate evidence literature set; analyzing each candidate evidence literature in the candidate evidence literature set to generate an evidence quality matrix; and using a large language model to perform long-chain reasoning and self-check iteration processing on the structured query, the candidate evidence literature set, the evidence quality matrix and patient individual characteristics to generate a clinical decision support report. The medical information processing method of the application uses multi-agent cooperation and self-checking strategies to improve the accuracy and reliability of medical information processing.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and medical informatics technology, and relates to a medical information processing method, and in particular to a medical information processing method, system and medium, and edge computing device. Background Technology

[0002] With the in-depth development of artificial intelligence technology, the application scope of large language models in the medical field is constantly expanding, gradually covering multiple clinical support aspects such as intelligent consultation, automatic medical record summarization, and auxiliary generation of treatment suggestions. In clinical practice, traditional technical models for assisting medical decision-making have long faced core challenges in accuracy and reliability. These models heavily rely on the experience summaries of individual physicians, non-systematic literature reviews, and hierarchical expert consensus. These methods are inherently limited by the contingency of individual cognition, the limitations of local information, and the lag in knowledge updates. The fundamental problem is that, faced with massive and rapidly changing medical evidence, traditional methods struggle to achieve efficient and accurate screening and integration. The resulting decision support information often suffers from defects such as mixed evidence quality, contradictory conclusions, or detachment from the specific patient context. This directly leads to large fluctuations in the accuracy and low reproducibility of auxiliary conclusions, and the lack of transparent and objective evaluation standards in their formation process, ultimately weakening the credibility and consistency of clinical decisions.

[0003] The above issues restrict the reliable application of technology in real-world medical environments and highlight the urgent need to build reliable, explainable, and accurate intelligent assistance systems at the current stage. Summary of the Invention

[0004] This application provides a medical information processing method, system, medium, and edge computing device to address the limitations in the timeliness and objectivity of processing massive amounts of medical information in existing technologies.

[0005] Firstly, this application provides a medical information processing method. The medical information processing method includes: parsing an input question into a structured query; using a dual-mode retrieval engine to search the structured query in a database to obtain a set of candidate evidence documents; analyzing each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix; and using a large language model to perform long-chain reasoning and self-checking iterative processing on the structured query, the set of candidate evidence documents, the evidence quality matrix, and patient individual characteristics to generate a clinical decision support report.

[0006] In one implementation of the first aspect, using a large language model to perform long-chain reasoning and self-checking iteration on the structured query, the candidate evidence document set, the evidence quality matrix, and patient individual characteristics to generate a clinical decision support report includes: using a multi-class proxy mechanism of the large language model to perform multiple rounds of reasoning on the structured query, the candidate evidence document set, the evidence quality matrix, and patient individual characteristics to generate reasoning results; verifying the reasoning results based on a self-checking mechanism to determine whether the reasoning results meet a preset threshold; if so, generating the clinical decision support report; if not, rolling back the long-chain reasoning until the preset threshold or the number of iterations threshold is met.

[0007] In one implementation of the first aspect, the multiple agents include a problem decomposition agent, a retrieval agent, an evidence evaluation agent, a reasoning agent, and an auditing agent; the problem decomposition agent is used to break down a complex problem into multiple sub-problems, the retrieval agent is used to call a retrieval engine to extract relevant medical evidence, the evidence evaluation agent is used to evaluate the quality of evidence according to preset rules, the reasoning agent is used to generate the reasoning result by integrating the relevant medical evidence and the individual characteristics of the patient, and the auditing agent is used to verify the reasoning result.

[0008] In one implementation of the first aspect, verifying the inference result based on a self-checking mechanism and determining whether the inference result meets a preset threshold includes: verifying the inference result based on a self-checking mechanism and determining whether the inference result meets a preset consistency threshold; if yes, generating the clinical decision support report; if no, rolling back to the corresponding step of the long-chain inference until the preset consistency threshold or iteration number threshold is met; wherein, the corresponding step of the long-chain inference includes an inference step, an evidence processing step, and a preceding problem processing step.

[0009] In one implementation of the first aspect, using a dual-mode retrieval engine to search the structured query in a database to obtain a set of candidate evidence documents includes: using a keyword retrieval sub-engine and a vector retrieval sub-engine to form a dual-mode retrieval engine to search the structured query in an external medical literature database, a local clinical guideline database, and / or a real-world database to obtain the set of candidate evidence documents.

[0010] In one implementation of the first aspect, the medical information processing method further includes: pre-acquiring randomized controlled trial literature within a preset time range as candidate evidence literature; if the number of randomized controlled trial literature does not meet the conditions, then sequentially introducing systematic review literature, comparative study literature, and / or case-control study literature as the candidate evidence literature; and sorting the candidate evidence literature according to the relevance score and evidence level of the query to obtain the candidate evidence literature set.

[0011] In one implementation of the first aspect, analyzing each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix includes: analyzing each candidate evidence document in the candidate evidence document set from five dimensions: risk bias, consistency, directness, precision, and publication bias, to generate the evidence quality matrix. In the second aspect, this application provides a medical information processing system. The medical information processing system includes: a question parsing module for parsing an input question into a structured query; a retrieval query module for using a dual-mode retrieval engine to retrieve the structured query from a database to obtain a candidate evidence document set; a quality analysis module for analyzing each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix; and a report generation module for using a large language model to perform long-chain reasoning and self-checking iterative processing on the structured query, the candidate evidence document set, the evidence quality matrix, and patient individual characteristics to generate a clinical decision support report.

[0012] Thirdly, this application provides an edge computing device. The edge computing device includes: a device housing; a memory for storing computer programs; and a processor configured for the medical information processing method described in any one of the first aspects in an offline or low-bandwidth network environment.

[0013] Fourthly, this application provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the medical information processing method described in any one of the first aspects.

[0014] As described above, the medical information processing method, system, medium, and edge computing device described in this application have the following beneficial effects:

[0015] The medical information processing method in this application embeds the five-step evidence-based medicine approach into the reasoning process of a large language model, achieving standardization, structuring, and traceability of medical information processing. Medical information processing suggestions include evidence levels and cited literature, significantly improving the credibility and interpretability of clinical decision support reports. Multi-agent collaboration and self-checking fallback strategies also reduce the illusion rate of model output and improve decision accuracy.

[0016] In addition, the introduction of individual patient characteristics and real-world data improves the personalization accuracy of clinical decision support reports and reduces the probability of errors in the generated report content. Attached Figure Description

[0017] Figure 1 This diagram illustrates an application scenario of the medical information processing method described in this application.

[0018] Figure 2 The diagram shown is a structural schematic of the end-to-cloud interaction scenario described in the embodiments of this application.

[0019] Figure 3 The diagram shown is a flowchart illustrating the medical information processing method described in an embodiment of this application.

[0020] Figure 4 The diagram shown is a schematic representation of the rollback inference described in an embodiment of this application.

[0021] Figure 5 The diagram shown is a flowchart illustrating the medical information processing method described in the embodiments of this application.

[0022] Figure 6 The diagram shown is a schematic diagram of generating an evidence quality matrix as described in an embodiment of this application.

[0023] Figure 7 The diagram shown is a structural schematic of the medical information processing system described in an embodiment of this application.

[0024] Figure 8 The diagram shown is a schematic representation of the medical information processing system described in an embodiment of this application.

[0025] Figure 9 The diagram shown is a structural schematic of the edge computing device described in an embodiment of this application.

[0026] Component designation explanation

[0027] 1 Medical information processing device 11 User terminal 12 Local processor 13 Display terminal 2 End-to-cloud interaction system 20 terminal 21 cloud server 100 Medical Information Processing System 110 Problem Analysis Module 120 Search and query module 130 Quality Analysis Module 140 Report generation module 900 Edge computing devices 910 memory 920 processor 930 Device casing 940 monitor S11~S14 step S121~S123 step Detailed Implementation

[0028] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0029] It should be noted that in the embodiments of this application, the words "optionally" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "optionally" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "optionally" or "for example" is intended to present the relevant concepts in a specific manner.

[0030] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0031] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0032] To clearly describe the technical solutions of the embodiments of this application, the terms involved in this application are first defined as follows:

[0033] Dual-mode retrieval mechanism: an information retrieval method that combines two different retrieval modes, aiming to improve the comprehensiveness and accuracy of retrieval.

[0034] Long-chain reasoning: A complex reasoning process that derives conclusions step by step from initial information through multiple logical deductions or causal relationships, commonly found in artificial intelligence and decision analysis.

[0035] Audit agent: A monitoring program deployed in a system or network that is responsible for automatically recording and reviewing operational behavior to ensure compliance and security.

[0036] Real-world data: Routine health data derived from real-world medical environments (such as electronic medical records and health insurance databases), reflecting actual clinical practice, and distinct from rigorous clinical trial data.

[0037] Key clinical outcome measures: Core assessment indicators pre-defined in clinical trials that directly reflect the effectiveness of interventions, such as survival rate and disease remission, and are used to determine the value of treatment.

[0038] With the rapid development of artificial intelligence technology, the application of large language models in the medical field has gradually expanded, and can now support a number of auxiliary medical tasks, including intelligent consultation, case summary generation, and treatment suggestion generation. However, existing technologies still have the following main problems in practical applications: (1) the reasoning process lacks transparency, the reasoning chain is difficult to trace, leading to insufficient interpretability of decisions; (2) the output content has a high illusion rate, easily generating inaccurate or fictitious medical information; (3) the source and quality of evidence lack effective control, making it difficult to guarantee the scientific validity of medical suggestions; (4) the ability to integrate individualized information is insufficient, treatment suggestions are severely generalized, and lack specificity. Evidence-based medicine, as an important methodology for modern clinical decision-making, emphasizes a five-step closed-loop process of "proposing questions—retrieving evidence—evaluating evidence—combining with patients—assessing effects." However, there is currently no effective technical solution that can natively embed this standardized process into the long-chain reasoning process of large language models, thereby embedding traceable and quantifiable credible medical evidence in the reasoning chain to support highly reliable clinical decision-making. Therefore, there is an urgent need for a clinical decision support technology that integrates the reasoning capabilities of large language models with the five-step process of evidence-based medicine to improve the accuracy, interpretability, and personalization of intelligent diagnostic and treatment recommendations.

[0039] To address at least the aforementioned issues, this application provides a medical information processing method. The method includes: parsing an input question into a structured query; retrieving the structured query from a database using a dual-mode retrieval engine to obtain a set of candidate evidence documents; analyzing each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix; and using a large language model to perform long-chain reasoning and self-checking iterative processing on the structured query, the candidate evidence document set, the evidence quality matrix, and patient individual characteristics to generate a clinical decision support report.

[0040] In this embodiment, the five-step evidence-based medicine methodology is embedded into the reasoning process of a large language model, achieving standardization, structuring, and traceability of medical information processing. Medical information processing suggestions include evidence levels and cited literature, significantly improving the credibility and interpretability of clinical decision support reports. Multi-agent collaboration and self-checking fallback strategies also reduce the illusion rate of model output and improve decision accuracy.

[0041] Figure 1 This diagram illustrates an application scenario of the medical information processing method described in this application. The medical information processing device 1 can be used to implement the medical information processing method provided in the embodiments of this application, but the application scenarios of the medical information processing method provided in the embodiments of this application are not limited to this. Figure 1 The medical information processing device 1 shown is as follows. Figure 1As shown, the medical information processing device 1 includes a user terminal 11, a local processor 12, and a display terminal 13. The medical information processing method provided in this embodiment can be applied to the local processor 12.

[0042] in, Figure 1 The local processor 12 can be a single local processor, a cluster of multiple local processors, or a cloud computing center, etc., and is not specifically limited here. Although Figure 1 Only one user terminal 11, one local processor 12, and one display terminal 13 are shown in the diagram, but it should be understood that... Figure 1 The examples in this paper are only for understanding this solution. The specific number of local processors 12 and display terminals 13 should be flexibly determined based on the actual situation.

[0043] In some other implementations, the medical information processing device 1 may not include a display terminal 13, but only a local processor 12 with display and interaction functions. The medical information processing method provided in this application embodiment can be applied to the local processor 12. The local processor 12 with display functions may include tablet computers, laptops, PDAs, mobile phones, personal computers, and voice interaction devices, etc., and is not limited here.

[0044] In some other implementations, the medical information processing method described in this application can be applied to edge-cloud interaction scenarios. Figure 2 The diagram shown is a structural schematic of the end-to-cloud interaction scenario described in an embodiment of this application. For example... Figure 2 As shown, the terminal-cloud interaction system 2 includes a terminal 20 and a cloud server 21. The terminal 20 and the cloud server 21 can communicate with each other, and the communication method is not limited to wired or wireless.

[0045] The terminal 20 can be mobile or fixed. For example, it can be a wireless terminal or a wired terminal. A wireless terminal can refer to a device with wireless transceiver capabilities, which can be deployed indoors, outdoors, and in industrial workshops. The terminal 20 can be a mobile phone, tablet computer, laptop computer, etc., and is not limited thereto. The cloud server 21 can include one or more servers, or one or more processing nodes, or one or more virtual machines running on the server. The cloud server 21 can also be referred to as a server cluster, management platform, data processing center, etc., and is not limited thereto in this embodiment.

[0046] The technical solutions in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0047] The following embodiments of this application provide a medical information processing method, which, for example, can be... Figure 1 The local processor 12 shown Figure 2 The cloud server 21 shown is used to implement this. Figure 3 The diagram shown is a flowchart illustrating the medical information processing method described in an embodiment of this application. Figure 3 As shown, the medical information processing method includes steps S11 to S14.

[0048] Step S11: Parse the input question into a structured query.

[0049] Step S12: Use the dual-mode retrieval engine to search the database for structured queries in order to obtain a set of candidate evidence documents.

[0050] Step S13: Analyze each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix.

[0051] Step S14: Use a large language model to perform long-chain reasoning and self-checking iteration on structured queries, candidate evidence document sets, evidence quality matrix and patient individual characteristics to generate a clinical decision support report.

[0052] In some possible implementations, an input question is received from user input, the source of which may be, for example, the clinical oral statements of patients seen by a doctor, or the diagnostic records or images of patients seen by a doctor at the hospital. The input question is parsed into a structured query Q1 using a PICO (Patient-Intervention-Comparison-Outcome) or PECO (Patient-Exposure-Comparison-Outcome) model. PICO, an acronym for Patient / Population, Intervention, Comparison, and Outcome, is primarily used in interventional studies (e.g., randomized controlled trials) to answer questions such as "Which treatment is better?". PECO, an acronym for Patient / Population, Exposure, Comparison, and Outcome, evolved from PICO and is primarily used in observational studies (e.g., cohort studies, case-control studies) to answer questions such as "Is a certain exposure factor associated with the outcome?". The structured query Q1 is used to invoke a dual-mode retrieval engine to search multiple databases and obtain a candidate evidence literature set R2. Each candidate evidence literature in the set is analyzed according to the GRADE framework to generate an evidence quality matrix M3. The GRADE framework is a transparent, systematic, and universal framework used to assess evidence quality and determine the strength of clinical recommendations. A large language model is used to perform long-chain inference and self-checking iteration on the structured query Q1, the candidate evidence literature set R2, the evidence quality matrix M3, and the patient individual characteristics F4 to generate a clinical decision support report.

[0053] In some other possible implementations, a candidate evidence literature set The database was searched using a dual-mode search engine to retrieve structured queries, including randomized controlled trials, cohort studies, case-control studies, systematic reviews / meta-analyses, and / or guideline consensus. Meta-analysis is a statistical method that quantitatively combines the results of multiple independent studies on the same scientific question (usually a clinical question).

[0054] It should be noted that the clinical decision support report output by the medical information processing method of this application is only for reference by medical personnel and does not directly generate automatic diagnosis or automatic treatment instructions.

[0055] In this embodiment, the five-step evidence-based medicine methodology is embedded into the reasoning process of a large language model, achieving standardization, structuring, and traceability of medical information processing. Medical information processing suggestions include evidence levels and cited literature, significantly improving the credibility and interpretability of clinical decision support reports. Multi-agent collaboration and self-checking fallback strategies also reduce the illusion rate of model output and improve decision accuracy.

[0056] In one embodiment of this application, the use of a large language model to perform long-chain reasoning and self-checking iteration on structured queries, candidate evidence document sets, evidence quality matrices, and patient individual characteristics to generate a clinical decision support report includes: using a multi-class proxy mechanism of the large language model to perform multiple rounds of reasoning on structured queries, candidate evidence document sets, evidence quality matrices, and patient individual characteristics to generate reasoning results; verifying the reasoning results based on a self-checking mechanism to determine whether the reasoning results meet a preset threshold; if so, generating a clinical decision support report; if not, rolling back to perform long-chain reasoning until the preset threshold or iteration number threshold is met.

[0057] In some possible implementations, the structured query Q1, candidate evidence document set R2, evidence quality matrix M3, and patient individual characteristics F4 are input into a large language model. A multi-class surrogate mechanism is used to generate inference results through multiple rounds of reasoning. A self-checking mechanism is introduced during the inference process to verify the inference results and determine whether they meet a preset consistency threshold. If yes, a clinical decision support report is generated; otherwise, the process automatically rolls back and re-infers until the preset consistency threshold is met or the maximum number of iterations is reached. Optionally, the self-checking temperature parameter is set to 0.2, and the maximum number of iterations is set to 3 to balance the accuracy and response speed of the clinical decision support report and improve the consistency and stability of the inference output.

[0058] In one embodiment of this application, the multiple agents include a problem decomposition agent, a retrieval agent, an evidence evaluation agent, a reasoning agent, and an auditing agent; the problem decomposition agent is used to break down a complex problem into multiple sub-problems, the retrieval agent is used to call a retrieval engine to extract relevant medical evidence, the evidence evaluation agent is used to evaluate the quality of evidence according to preset rules, the reasoning agent is used to generate reasoning results by integrating relevant medical evidence and individual patient characteristics, and the auditing agent is used to verify the reasoning results.

[0059] In some possible implementations, long-chain inference is supported by a multi-agency collaborative mechanism, including problem decomposition agents, retrieval agents, evidence evaluation agents, inference agents, and auditing agents. Problem decomposition agents break down complex problems into multiple sub-problems; retrieval agents invoke retrieval engines to extract relevant medical evidence; evidence evaluation agents evaluate the quality of evidence according to preset rules; inference agents synthesize relevant medical evidence and individual patient characteristics to generate inference results; and auditing agents verify the consistency and logicality of the inference results.

[0060] In this embodiment, the illusion rate of the model output is reduced and the accuracy of the suggestions is improved by using multi-agent collaboration and self-checking strategies.

[0061] In one embodiment of this application, the reasoning result is verified based on a self-checking mechanism to determine whether the reasoning result meets a preset threshold. This includes: verifying the reasoning result based on the self-checking mechanism to determine whether the reasoning result meets a preset consistency threshold; if yes, a clinical decision support report is generated; if no, the process rolls back to the corresponding step of the long-chain reasoning until the preset consistency threshold or iteration number threshold is met; wherein, the corresponding step of the long-chain reasoning includes a reasoning step, an evidence processing step, and a preceding problem processing step.

[0062] Among some possible implementations, Figure 4 The diagram shown is a schematic representation of the rollback inference described in an embodiment of this application. For example... Figure 4 As shown, the preset consistency thresholds include three categories: the first category includes complete inference results, sufficient citations, and local logical consistency; the second category includes sufficient evidence quantity, acceptable evidence quality, and no conflicts between pieces of evidence; and the third category includes correct clinical problem decomposition, complete query conditions, and correct patient feature identification. When the preset consistency threshold is not met due to incomplete inference results, insufficient citations, or local logical inconsistencies, the system rolls back to the inference step and regenerates the inference result based on existing evidence. When the preset consistency threshold is not met due to insufficient evidence quantity, low evidence quality, or conflicts between pieces of evidence, the system rolls back to the evidence processing step, supplements, re-screens, or re-evaluates the evidence, and then executes the inference again. When the preset consistency threshold is not met due to incorrect clinical problem decomposition, missing query conditions, or incorrect patient feature identification, the system rolls back to the preceding problem processing step, reprocesses the input information, and then executes subsequent steps. When the preset iteration limit is reached, the system stops automatic inference and outputs a review prompt. In some other possible implementations, a summary of conflict causes, prompts for supplementary information, and / or suggestions for manual confirmation can also be output.

[0063] In other possible implementations, for example, generating medication recommendations for a patient with chronic heart failure and chronic kidney disease, the initial inference might suggest "spironolactone could be considered." Validation reveals that this inference does not significantly conflict with existing evidence, but the output fails to explicitly mention the risk of hyperkalemia, renal function monitoring requirements, and follow-up conditions. In this case, the basic evidence is complete; the main problem lies in the incomplete expression of the inference result. Therefore, the original evidence is retained, and the process is rolled back to the inference step, requiring a more complete inference result to be generated again. The second output, supplemented with information such as "the decision on whether to activate and continue should be based on renal function and serum potassium monitoring," meets the preset consistency threshold, generating a clinical decision support report.

[0064] For example, consider the reasoning process regarding whether to initiate anticoagulation therapy for a patient with atrial fibrillation. Initially, the system, based on partial medical record information, suggests "anticoagulation." However, verification reveals that the current evidence chain lacks complete extraction of stroke risk score-related elements and sufficient inclusion of bleeding risk-related information, leading to insufficient evidence for the conclusion. The problem here is not merely incomplete description, but rather insufficient extraction and integration of risk factors during the evidence processing stage. Therefore, the system reverts to the evidence processing step, supplementing the extraction of patient age, stroke history, hypertension, renal dysfunction, and bleeding history. After re-screening and evaluating the evidence, the reasoning is executed again.

[0065] For example, the original question was: "How should an adult patient with community-acquired pneumonia who is allergic to penicillin be selected as the initial antibiotic?" However, in the initial question processing stage, the system only extracted "adult community-acquired pneumonia," omitting the crucial constraint "penicillin allergy." Consequently, subsequent retrieval and inference might yield conventional β-lactam inference results, which would not match the patient's individual condition. Verification revealed that the root cause lay in the bias in the initial question representation, rather than in the inference or evidence processing steps. Therefore, the system rolled back to the initial question processing steps, re-extracted and added the crucial condition "penicillin allergy," and then re-executed the evidence retrieval, evidence processing, and inference processes.

[0066] For example, consider assessing the need to initiate anticoagulation therapy for an elderly patient with atrial fibrillation, a history of gastrointestinal bleeding, and fluctuating renal function. After multiple rounds of rollback and re-reasoning, it is still found that: on the one hand, the patient has a high risk of stroke; on the other hand, the risk of bleeding is also significantly increased, and some key laboratory indicators are missing, making it impossible to reach a stable conclusion that meets the preset consistency threshold. In this case, if automatic iteration continues, it may only repeatedly output conflicting conclusions, making it difficult to further improve the consistency of the reasoning results. Therefore, after reaching the preset iteration limit, automatic reasoning stops, and instead of outputting definitive recommendations, a review prompt message is output. In other possible implementations, the review prompt message may also include a summary of the conflict reasons, key missing information prompts, and suggestions for manual confirmation.

[0067] In one embodiment of this application, using a dual-mode retrieval engine to search for structured queries in a database to obtain a set of candidate evidence documents includes: using a keyword retrieval sub-engine and a vector retrieval sub-engine to form a dual-mode retrieval engine to search for structured queries in external medical literature databases, local clinical guideline databases, and / or real-world databases to obtain a set of candidate evidence documents.

[0068] In some possible implementations, the bimodal retrieval engine consists of a keyword retrieval sub-engine and a vector retrieval sub-engine. The bimodal retrieval engine is used to search for the structured query Q1 in external medical literature databases, local clinical guideline databases, and / or real-world databases to obtain a set of candidate evidence documents R2.

[0069] Figure 5 The diagram shown is a flowchart illustrating the medical information processing method described in an embodiment of this application. Figure 5 As shown, step S12 may further include steps S121 to S123.

[0070] Step S121: Obtain randomized controlled trial literature within a preset time range as candidate evidence literature.

[0071] Step S122: If the number of randomized controlled trials does not meet the requirements, then systematic reviews, comparative studies, and / or case-control studies are introduced as candidate evidence in sequence.

[0072] Step S123: Sort the candidate evidence documents according to the relevance score and evidence level of the query to obtain a set of candidate evidence documents.

[0073] In some possible implementations, priority is given to randomized controlled trials (RCTs) published within the last three years. If the number of RCTs is insufficient, systematic reviews, comparative studies, and / or case-control studies are introduced as candidate evidence in that order. The candidate evidence is then comprehensively ranked based on the relevance score and level of evidence to obtain a set of candidate evidence.

[0074] In one embodiment of this application, analyzing each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix includes: analyzing each candidate evidence document in the candidate evidence document set from five dimensions: risk bias, consistency, directness, accuracy, and publication bias, to generate an evidence quality matrix. In some possible implementations, Figure 6 This is a schematic diagram illustrating the generation of the evidence quality matrix as described in an embodiment of this application. Figure 6 As shown, a candidate evidence literature set R2 is received. Each candidate evidence literature in R2 is analyzed from five dimensions: risk bias, consistency, directness, precision, and publication bias. The scores obtained from the five dimensions are combined to generate an evidence quality matrix. For each literature in the candidate evidence literature set R2... Generate a five-dimensional rating vector To generate an evidence quality matrix Evidence quality matrix The expression is:

[0075] ,

[0076] in, For the risk bias dimension, the scoring vector is... For the consistency dimension of the score vector, For the directness dimension of the rating vector, For the accuracy dimension of the score vector, This is a rating vector for the bias dimension.

[0077] Evidence Quality Matrix Each row corresponds to a candidate evidence document, and each column corresponds to a dimension. The matrix elements are computable discrete downgrade values ​​or continuous scores. For example, the score vector for each dimension can be set to discrete values, such as 0, 1, and 2, which correspond to no problem, serious problem, and very serious problem, respectively. Alternatively, the score vector for each dimension can be set to values ​​of 0, -1, and -2 to represent the downgrade magnitude. Simultaneously, a traceable mapping table E of candidate evidence documents, dimensions, and fields is output to provide interpretability of the long-chain reasoning process, supporting subsequent auditing and rollback operations by the self-inspection agent.

[0078] In one embodiment of this application, the medical information processing method further includes: predicting key clinical outcome indicators based on clinical decision support reports and real-world data to output medical information processing recommendations, which include evidence level labels, risk-benefit assessment results, and / or literature citation links.

[0079] In some possible implementations, the key clinical outcome indicator OC5 is predicted based on clinical decision support reports and real-world data to output medical information processing recommendations. These recommendations include evidence level labeling, risk-benefit assessment results, and / or literature citation links. The key clinical outcome indicator OC5 may be, for example, the 30-day hospitalization rate, the one-year survival rate, or the predicted value of blood pressure reduction in hypertensive patients, and may be adjusted according to the patient's specific condition or possible circumstances. This application is not limited to this.

[0080] In this embodiment, individual patient characteristics and real-world data are incorporated to predict key clinical outcome indicators, improving the personalized accuracy of medical information processing recommendations and reducing the 30-day readmission rate. The medical information processing recommendations include evidence level annotations and literature citation links, enhancing the credibility and interpretability of the results.

[0081] Figure 7 The diagram shown is a structural schematic of the medical information processing system described in an embodiment of this application. Figure 7 As shown, the medical information processing system 100 includes a problem analysis module 110, a retrieval and query module 120, a quality analysis module 130, and a report generation module 140.

[0082] The question parsing module 110 is used to parse the input question into a structured query.

[0083] The retrieval module 120 is used to retrieve structured queries in the database using a dual-mode retrieval engine to obtain a set of candidate evidence documents.

[0084] The quality analysis module 130 is used to analyze each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix.

[0085] The report generation module 140 is used to perform long-chain reasoning and self-checking iteration on structured queries, candidate evidence literature sets, evidence quality matrices and individual patient characteristics using a large language model to generate clinical decision support reports.

[0086] It should be noted that the medical information processing system 100 includes modules 110 to 140 as described above. Figure 2 Steps S11 to S14 in the medical information processing method shown correspond one-to-one and will not be elaborated here.

[0087] Among some possible implementations, Figure 8 The diagram shown is a schematic representation of the medical information processing system described in an embodiment of this application. Figure 8The patient's information is as follows: Mr. Wang, 72 years old, diagnosed with heart failure with preserved ejection fraction (HFpEF), and also has chronic kidney disease (CKD stage 3b) with a risk of hyperkalemia. The doctor considered adding spironolactone (an MRA drug) to improve the patient's prognosis, but was concerned about the patient's renal function and serum potassium level, requiring a comprehensive risk-benefit assessment. The doctor entered the question "Will this patient benefit from adding spironolactone? What are the risks?" into the medical information processing system 100's dialog box. The medical information processing system 100 identified spironolactone as the core intervention. The medical information processing system 100 automatically scanned the patient's electronic medical record laboratory information system in the hospital through the API interface, extracting the patient's individual characteristics F4. The patient's individual characteristics F4 are, for example: Past history: hypertension, CKD stage 3b; Recent test values: serum creatinine 160 μmol / L, serum potassium 5.1 mmol / L (borderline high value), eGFR 38 ml / min; Recent examination: echocardiography showed LVEF 55% (HFpEF characteristics). The question parsing module 110 parses the input question into a structured query Q1, which may be, for example: {P: HFpEF + CKD3b + risk of hyperkalemia; I: spironolactone; C: standard treatment; O: readmission rate / incidence of hyperkalemia}. The retrieval module 120 uses a dual-mode retrieval engine to search the database for the structured query Q1 to obtain a candidate evidence literature set R2. External retrieval: Retrieves RCT literature from the past three years on "MRA treatment in HFpEF with CKD" (e.g., subgroup analysis of the TOPCAT study). Internal retrieval: Utilizes a real-world data warehouse (RWD) to retrieve adverse event records from the past 5 years in hospitals for patients with similar "eGFR < 45 and serum potassium > 5.0" after spironolactone use. The quality analysis module 130 analyzes each candidate evidence literature in the candidate evidence literature set R2 to generate an evidence quality matrix M3. The report generation module 140 utilizes a large language model to perform long-chain inference on the structured query Q1, candidate evidence literature set R2, evidence quality matrix M3, and patient individual characteristics F4, generating an inference result of "Recommended use, can reduce hospitalization risk". During the self-checking iteration phase, the audit agent detected a patient's serum potassium level of 5.1 mmol / L, and the evidence quality matrix M3 showed that guidelines strongly warn against the use of MRA in patients with serum potassium >5.0 mmol / L. Determining a high risk in the initial recommendation, a rollback mechanism was triggered, requiring the inference agent to regenerate the plan, which must include "contraindication assessment" and "monitoring plan". In the second iteration, the large language model corrected the plan, generating an inference result of "Deferred use" or "Start with very low dose and close monitoring". The inference result was deemed to meet a preset threshold, generating a clinical decision support report.The protocol evaluation module, based on clinical decision support reports and real-world data (i.e., the patient's eGFR trend and RWD data), predicts the key clinical outcome indicator OC5 to output medical information processing recommendations. Benefit prediction: If medication is used, the risk of 30-day heart failure readmission is relatively reduced by 12%. Risk prediction: However, the probability of developing severe hyperkalemia (>6.0 mmol / L) is as high as 18% (based on data from similar patients in our hospital), with evidence level annotations and literature citation links. The protocol presentation module displays a structured decision report on its screen, with the medical information processing recommendation: postpone initiation of spironolactone therapy (warning: high risk of hyperkalemia). The risk-benefit assessment result is: Although literature supports MRA improving HFpEF prognosis (evidence level: Moderate), based on this patient's **serum potassium (5.1 mmol / L) and renal function (eGFR38)** data, the model predicts an 18% risk of severe hyperkalemia (OC5 indicator). Alternative or adjusted protocols are: 1. It is recommended to prioritize the use of SGLT2 inhibitors (stronger evidence for renal protection benefits). 2. If necessary, it is recommended to recheck serum potassium levels one week later. If the potassium level drops below 5.0, the dosage can be increased to 12.5 mg every other day. Reference links include patient data and literature data. Patient data is linked to LIS report number #20231012-K (serum potassium trend chart). Literature data is cited from the chapter on potassium management in CKD patients in the 202x Heart Failure Guidelines.

[0088] In this embodiment, the system not only addresses textual issues but also deeply integrates objective test data from the hospital system. Through a self-checking and rollback mechanism, it successfully blocks general recommendations that could lead to medical accidents by utilizing individual abnormal test values ​​of patients. Simultaneously, it provides specific risk probabilities using real-world data. Furthermore, the medical information processing system 100 supports a dual-deployment architecture of cloud and edge computing, is adaptable to low-bandwidth (<64kbps) environments, and possesses strong usability and scalability.

[0089] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0090] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of this application, depending on actual needs. For example, the functional modules / units in the various embodiments of this application may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0091] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] This application also provides an edge computing device. Figure 9 The diagram shown is a structural schematic of the edge computing device described in an embodiment of this application. Figure 9 As shown, the edge computing device 900 in this embodiment includes a memory 910, a processor 920, and a device housing 930.

[0093] The memory 910 includes no less than 9GB of LPDDR4 dynamic random access memory and no less than 64GB of solid-state memory; preferably, the memory 910 includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0094] Specifically, memory 910 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Edge computing device 900 may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 910 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application. It is understood that memory 910 may be volatile memory or non-volatile memory, or both. Non-volatile memory may be read-only memory (ROM) or programmable read-only memory (PROM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memories.

[0095] The processor 920 is connected to the memory 910 and is configured to execute the medical information processing method described in any embodiment of this application in an offline or low-bandwidth network environment. The processor 920 uses a system-on-a-chip (SoC) based on the ARM Cortex-A76 architecture and integrates a neural network processing unit (NPU) with 4 TOPS of computing power. Specifically, in terms of software configuration, a lightweight language model with approximately 3B parameters is deployed to adapt to edge resource constraints, and a local medical knowledge base containing clinical guidelines and standard pathway information, with a data capacity of approximately 2GB, is pre-installed. During execution, the cloud completes the structured retrieval and GRADE evidence evaluation process, while the edge performs inference and outcome prediction tasks. When the network bandwidth is below 64kbps, the system automatically switches to a local pure offline mode, calling a pre-cached high-quality evidence library to maintain system availability.

[0096] Optionally, the processor 920 may also be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0097] Optionally, the edge computing device 900 in this embodiment may further include a display 940. The display 940 is communicatively connected to the memory 910 and the processor 920, and is used to display the graphical user interface (GUI) of the medical information processing method described in this application embodiment. The display 940 is, for example, a 10.1-inch capacitive touch screen.

[0098] In this embodiment, the edge computing device 900 hardware is limited to ARM architecture and NPU chip, and the software supports cloud-edge collaboration. When the bandwidth is less than 64kbps, it automatically switches to the offline lightweight model and local database, which increases the flexibility of the clinical decision support report of this application in primary healthcare scenarios.

[0099] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the program implements the medical information processing method described in any embodiment of this application. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0100] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0101] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0102] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A medical information processing method, characterized in that, include: The input question is parsed into a structured query; The structured query is retrieved from the database using a dual-mode search engine to obtain a set of candidate evidence documents; Each candidate evidence document in the candidate evidence document set is analyzed to generate an evidence quality matrix; The structured query, the candidate evidence document set, the evidence quality matrix, and the patient's individual characteristics are used to perform long-chain reasoning and self-checking iterative processing to generate a clinical decision support report.

2. The medical information processing method according to claim 1, characterized in that, The large language model is used to perform long-chain inference and self-checking iteration on the structured query, the candidate evidence document set, the evidence quality matrix, and individual patient characteristics to generate a clinical decision support report, including: The structured query, the candidate evidence document set, the evidence quality matrix, and the patient's individual characteristics are used to perform multiple rounds of reasoning through the multi-class proxy mechanism of the large language model to generate reasoning results; The reasoning results are verified based on a self-checking mechanism to determine whether they meet a preset threshold. If they do, a clinical decision support report is generated; otherwise, the long-chain reasoning is rolled back until the preset threshold or the number of iterations is met.

3. The medical information processing method according to claim 2, characterized in that, The various types of agents include problem decomposition agents, retrieval agents, evidence evaluation agents, reasoning agents, and audit agents; The problem decomposition agent is used to break down a complex problem into multiple sub-problems; the retrieval agent is used to call a retrieval engine to extract relevant medical evidence; the evidence evaluation agent is used to evaluate the quality of evidence according to preset rules; the reasoning agent is used to generate the reasoning result by integrating the relevant medical evidence and the individual characteristics of the patient; and the auditing agent is used to verify the reasoning result.

4. The medical information processing method according to claim 2, characterized in that, The inference result is verified based on a self-checking mechanism, and the determination of whether the inference result meets a preset threshold includes: The reasoning result is verified based on a self-checking mechanism to determine whether the reasoning result meets a preset consistency threshold. If it does, the clinical decision support report is generated. If not, the process rolls back to the corresponding step of the long-chain reasoning until the preset consistency threshold or iteration number threshold is met. The corresponding steps of the long-chain reasoning include reasoning steps, evidence processing steps, and preceding problem processing steps.

5. The medical information processing method according to claim 1, characterized in that, The structured query is retrieved from the database using a dual-mode search engine to obtain a set of candidate evidence documents, including: A dual-mode retrieval engine, consisting of a keyword retrieval sub-engine and a vector retrieval sub-engine, is used to search the structured query in external medical literature databases, local clinical guideline databases, and / or real-world databases to obtain the candidate evidence literature set.

6. The medical information processing method according to claim 1, characterized in that, Also includes: Randomized controlled trial literature within a predetermined time frame is obtained in advance as candidate evidence literature; If the number of randomized controlled trials does not meet the requirements, then systematic reviews, comparative studies, and / or case-control studies will be introduced in sequence as candidate evidence. as well as The candidate evidence documents are sorted according to the relevance score and evidence level of the query to obtain the candidate evidence document set.

7. The medical information processing method according to claim 1, characterized in that, Analyzing each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix includes: The candidate evidence documents in the candidate evidence document set are analyzed from five dimensions: risk bias, consistency, directness, accuracy, and publication bias, in order to generate the evidence quality matrix.

8. A medical information processing system, characterized in that, include: The question parsing module is used to parse the input question into a structured query; The retrieval module is used to retrieve the structured query in the database using a dual-mode retrieval engine to obtain a set of candidate evidence documents; The quality analysis module is used to analyze each candidate evidence document in the candidate evidence document set to generate an evidence quality matrix. The report generation module is used to perform long-chain reasoning and self-checking iteration on the structured query, the candidate evidence document set, the evidence quality matrix, and patient individual characteristics using a large language model to generate a clinical decision support report.

9. An edge computing device, characterized in that, include: Device casing; Memory, used to store computer programs; A processor configured to perform the medical information processing method of any one of claims 1 to 7 in an offline or low-bandwidth network environment.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the medical information processing method according to any one of claims 1 to 7.

Images