Methods, apparatus, media, and program products for obtaining a physical examination analysis report

Abstract

The purpose of the present application is to provide a method, device, medium and program product for obtaining a physical examination analysis report, comprising: performing content recognition on a physical examination report file to obtain structured test index data; obtaining structured multi-modal observation data through enhanced interpretation, writing the test index data and the multi-modal observation data into an evidence mapping table; generating an abnormal evidence combination corresponding to at least one organ system based on the evidence mapping table; generating first query information according to the abnormal evidence combination, performing first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result, generating second query information based on the first retrieval result to perform second-stage medical evidence retrieval to obtain a second retrieval result, and obtaining a retrieval text block hit by the first retrieval result and the second retrieval result; constructing context information according to the abnormal evidence combination, the first retrieval result and the second retrieval result, and obtaining a physical examination analysis report output by a target large model based on the context information.

Description

[0001] This application claims priority to an earlier application (CN 202610769162.4, Method, apparatus, media and procedure for obtaining a medical examination analysis report, application date 2026-05-31), the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of computer technology, and in particular to a technology for obtaining a physical examination analysis report. Background Technology

[0003] Intelligent analysis of medical examination reports typically involves steps such as OCR (Optical Character Recognition) recognition, table structure restoration, medical indicator standardization, anomaly detection, medical evidence retrieval, and large language model report generation. Traditionally, the electronic PDF version of the medical examination report can be converted to text, and indicators can be extracted using templates, keywords, or regular expressions; alternatively, the entire report can be directly input into a large medical model to generate health recommendations. With the development of RAG (Retrieval-Augmented Generation) technology, some systems have begun to incorporate medical textbooks, expert guidelines, or paper excerpts into vector libraries, generating medical interpretations through retrieval-enhanced generation.

[0004] However, medical examination reports exhibit significant engineering complexities. Different institutions use different report formats, and a single page may simultaneously contain HTML tables, Markdown tables, multi-column side-by-side indicators, headerless indicator rows, detailed item blocks, text on imaging or examination opinions, abnormality markers, units, reference ranges, positive / negative qualitative results, and manual remarks. OCR intermediate results may also contain two sets of content: a PDF text layer and an OCR layout parsing layer, with varying quality. Furthermore, ultrasound, imaging, and electrocardiogram data often exist in the form of images, waveforms, measurement parameters, or report conclusions, and cannot be simply equated with examination indicator tables. If parsing is done using only fixed regular expressions or a single table structure, it is easy to miss indicators or incorrectly merge fields. If the data is directly given to the model for interpretation, the model will have difficulty distinguishing between true anomalies, OCR pseudo-anomalies, and multimodal edge detections. If the full text of the report is directly input into a large language model, problems such as OCR misalignment, table misalignment, positive misreading, missed readings, and edge detections from images or ECGs failing to enter the inference chain will be amplified, resulting in the output of medical advice lacking factual basis and source traceability. Summary of the Invention

[0005] One object of this application is to provide a method, apparatus, medium, and procedure for obtaining a medical examination analysis report.

[0006] According to one aspect of this application, a method for obtaining a medical examination analysis report is provided, the method comprising:

[0007] Content recognition is performed on the physical examination report document to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items;

[0008] Enhanced interpretation of the multimodal file corresponding to the physical examination report file is performed to obtain structured multimodal observation data, and the test index data and the multimodal observation data are written into the evidence mapping table;

[0009] Based on the evidence mapping table, at least one abnormal evidence combination corresponding to an organ system is generated, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data.

[0010] First query information is generated based on the combination of abnormal evidence. First search results are obtained by performing a first-stage medical evidence retrieval based on the first query information. Second query information is generated based on the first search results. Second search results are obtained by performing a second-stage medical evidence retrieval based on the second query information. The search text blocks matched by the first search results and the second search results are obtained. The first query information is used to discover candidate diseases and / or risk directions. The second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions.

[0011] Contextual information is constructed based on the abnormal evidence combination, the first search result, and the second search result to obtain a physical examination analysis report output by the target large model based on the contextual information. The physical examination analysis report includes a reference number for the searched text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

[0012] According to one aspect of this application, a computer device for obtaining a physical examination analysis report is provided, comprising a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of any of the methods described above.

[0013] According to one aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0014] According to one aspect of this application, a computer program product is provided, comprising a computer program, characterized in that, when executed by a processor, the computer program implements the steps of any of the methods described above.

[0015] According to one aspect of this application, a system for obtaining a medical examination analysis report is provided, the system comprising:

[0016] The first module is used to perform content recognition on the physical examination report file to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items.

[0017] The second module is used to enhance the interpretation of the multimodal file corresponding to the physical examination report file, obtain structured multimodal observation data, and write the test index data and the multimodal observation data into the evidence mapping table;

[0018] The third module is used to generate at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data.

[0019] The fourth module is used to generate first query information based on the abnormal evidence combination, perform a first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result, generate second query information based on the first retrieval result, perform a second-stage medical evidence retrieval based on the second query information to obtain a second retrieval result, and obtain the retrieval text block matched by the first retrieval result and the second retrieval result. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management suggestions and / or clinical evidence management suggestions and / or clinical evidence related to the candidate diseases and / or risk directions.

[0020] The fifth module is used to construct context information based on the abnormal evidence combination, the first search result and the second search result, and obtain a physical examination analysis report output by the target large model based on the context information. The physical examination analysis report includes a reference number for the search text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

[0021] According to one aspect of this application, a computer device for obtaining a medical examination analysis report is provided, the device comprising:

[0022] The module is used to perform content recognition on physical examination report files to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items.

[0023] The first and second modules are used to enhance the interpretation of the multimodal file corresponding to the physical examination report file, obtain structured multimodal observation data, and write the test index data and the multimodal observation data into the evidence mapping table;

[0024] The first and third modules are used to generate at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data.

[0025] The first four modules are used to generate first query information based on the abnormal evidence combination, perform a first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result, generate second query information based on the first retrieval result, perform a second-stage medical evidence retrieval based on the second query information to obtain a second retrieval result, and obtain the retrieval text block matched by the first retrieval result and the second retrieval result. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions.

[0026] The first and fifth modules are used to construct context information based on the abnormal evidence combination, the first search result and the second search result, and obtain a physical examination analysis report output by the target large model based on the context information. The physical examination analysis report includes a reference number for the search text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

[0027] Compared with existing technologies, this application focuses on a traceable chain of evidence, enabling doctors to obtain a physical examination analysis report with structured abnormal indicators, structured multimodal observation data, abnormal evidence combinations, candidate diseases and / or risk directions, management recommendations and / or clinical evidence and citation constraints in a shorter time. This ensures that the key facts, evidence sources and model citations from the original materials to the final report are traceable and verifiable. Attached Figure Description

[0028] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0029] Figure 1 This invention provides a flowchart illustrating a method for obtaining a physical examination analysis report according to an embodiment of the present application.

[0030] Figure 2 This diagram illustrates a system architecture for obtaining a physical examination analysis report according to an embodiment of this application.

[0031] Figure 3 This invention provides a flowchart illustrating a method for obtaining a physical examination analysis report according to an embodiment of the present application.

[0032] Figure 4 This diagram illustrates a system architecture for obtaining a physical examination analysis report according to an embodiment of this application.

[0033] Figure 5 This diagram illustrates a computer device structure for obtaining a physical examination analysis report according to an embodiment of this application.

[0034] Figure 6 Exemplary systems that can be used to implement the various embodiments described in this application are shown.

[0035] The same or similar reference numerals in the accompanying drawings represent the same or similar parts. Detailed Implementation

[0036] The present application will now be described in further detail with reference to the accompanying drawings.

[0037] In a typical configuration of this application, the terminal, the device of the service network, and the trusted party all include one or more processors (e.g., a central processing unit (CPU)), input / output interfaces, network interfaces, and memory.

[0038] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory. Memory is an example of computer-readable media.

[0039] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PCM), programmable random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0040] The devices referred to in this application include, but are not limited to, user equipment, network equipment, or devices composed of user equipment and network equipment integrated through a network. The user equipment includes, but is not limited to, any mobile electronic product capable of human-computer interaction (e.g., via a touchpad), such as smartphones and tablets. These mobile electronic products can use any operating system, such as Android or iOS. The network equipment includes an electronic device capable of automatically performing numerical calculations and information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), and embedded devices. The network equipment includes, but is not limited to, computers, network hosts, single network servers, multiple network server clusters, or clouds composed of multiple servers. Here, a cloud consists of a large number of computers or network servers based on cloud computing, where cloud computing is a type of distributed computing, consisting of a virtual supercomputer composed of a group of loosely coupled computer clusters. The network includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, VPN network, wireless ad hoc network, etc. Preferably, the device can also be a program running on the user equipment, network device, or a device formed by integrating user equipment and network device, network device, touch terminal, or network device and touch terminal through a network.

[0041] Of course, those skilled in the art should understand that the above-described devices are merely examples, and other existing or future devices that are applicable to this application should also be included within the scope of protection of this application, and are hereby incorporated by reference.

[0042] In the description of this application, "multiple" means two or more, unless otherwise expressly and specifically defined.

[0043] Figure 1This document illustrates a flowchart of a method for obtaining a physical examination analysis report according to an embodiment of this application. The method includes steps S11, S12, S13, S14, and S15. In step S11, a computer device performs content recognition on the physical examination report file to obtain structured test indicator data, wherein the test indicator data includes normal and abnormal indicator items. In step S12, the computer device performs enhanced interpretation on the multimodal file corresponding to the physical examination report file to obtain structured multimodal observation data, and writes the test indicator data and the multimodal observation data into an evidence mapping table. In step S13, the computer device generates at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data. In step S14, the computer device generates first query information based on the abnormal evidence combination, and performs a first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result. As a result, second query information is generated based on the first search result, and a second-stage medical evidence retrieval is performed based on the second query information to obtain a second search result. The search text blocks matched by the first and second search results are obtained. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions. In step S15, the computer device constructs context information based on the abnormal evidence combination, the first search result, and the second search result to obtain a physical examination analysis report output by the target large model based on the context information. The physical examination analysis report includes a reference number for the search text block, and the physical examination analysis report is associated with the test indicator data and the multimodal observation data through the evidence mapping table.

[0044] In step S11, the computer device performs content recognition on the physical examination report file to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items.

[0045] In some embodiments, the file format of the physical examination report file includes, but is not limited to, PDF, photographed images, etc., and this example embodiment does not impose any special limitations on this. In some embodiments, content recognition (e.g., OCR optical character recognition) is performed on the physical examination report file to obtain structured multiple test indicator data. Each test indicator data includes, but is not limited to, indicator name, indicator value, abnormality mark, qualitative result, unit, reference range, etc., and this example embodiment does not impose any special limitations on this. In some embodiments, the multiple test indicator data includes at least one normal indicator item (i.e., test indicator data determined to be normal) and at least one abnormal indicator item (i.e., test indicator data determined to be abnormal). In some embodiments, the physical examination report file uploaded by the user is received, the file type and size are verified, and a task identifier and a case identifier are generated. The task identifier can be in the form of "task_date_random fragment", and the case identifier can be in the form of "case_date_random fragment", and this example embodiment does not impose any special limitations on this. In some embodiments, a metadata JSON file (meta.json) is created in the task directory, and the user, clinic, invitation code, or de-identified patient identifier of the task is written into the metadata. In some embodiments, if user directory isolation is enabled, a user security directory is also generated and a task index is maintained to facilitate historical report queries by user.

[0046] In step S12, the computer device performs enhanced interpretation of the multimodal file corresponding to the physical examination report file to obtain structured multimodal observation data, and writes the test index data and the multimodal observation data into the evidence mapping table.

[0047] In some embodiments, the multimodal files corresponding to the physical examination report file include, but are not limited to, ultrasound images or ultrasound reports, CT / MRI / X-ray imaging images or reports, electrocardiogram waveforms / screenshots / parameter tables, and other examination images. This example embodiment does not impose any special limitations on these. In some embodiments, the images, waveforms, and examination conclusions in the multimodal files are enhanced and interpreted to obtain structured multimodal observation data. This multimodal observation data includes, but is not limited to, examination type, anatomical location, left / right side, measurement values, abnormal signs, severity, conclusion tendency, confidence level, source page / frame number, and original fragment. This example embodiment does not impose any special limitations on these. In some embodiments, enhanced interpretation can be achieved through medical multimodal models, visual language models, specialist imaging models, electrocardiogram analysis models, and structured interfaces of hospital information systems. This example embodiment does not impose any special limitations on these. In some embodiments, test index data and multimodal observation data are jointly written into an evidence mapping table to achieve the fusion of test index data and multimodal observation data. The role of the evidence mapping table is to map the structured conclusions in the subsequently generated physical examination analysis report back to the source of evidence (including test index data and multimodal observation data). For example, the evidence mapping table includes evidenceIds (evidence ID), observationIds (multimodal observation data ID), and findingIds (test index data ID) corresponding to each conclusion in the physical examination analysis report. That is, each piece of multimodal observation data is written into the evidence mapping table as a piece of evidence, and each piece of test index data is also written into the evidence mapping table as a piece of evidence. The front end can use the evidence mapping table for traceability display, and these IDs will also be used when performing medical evidence retrieval in the future to convert clinical indicators into searchable standardized indicator items.In some embodiments, after obtaining structured multimodal observation data, operations such as confidence gating, source localization, and structured field constraints are performed to upgrade it into a structured evidence stream that runs parallel to the test index data. This stream then enters subsequent RAG retrieval enhancement generation and large model inference. Confidence gating involves scoring structured items (structured multimodal observation data). Items below a first threshold are directly excluded; those between the first and second thresholds (where the second threshold is greater than the first) enter a review queue; and only those above the second threshold enter a high-confidence pool. Only multimodal observation data in the high-confidence pool are written into the evidence mapping table and then enter subsequent RAG retrieval enhancement generation and large model inference. Source localization involves binding each structured item to original OCR evidence, such as rawText (original evidence text), evidenceIds (a list of evidence IDs), and sourceMaterial (source material, i.e., the retained original or near-original input material, such as a structured file obtained by content recognition (e.g., OCR) of a medical examination report document, such as Markdown). (Document) Structured field constraints require structured items to have fixed fields, such as indicator name, value, unit, reference range, status, confidence level, and evidence ID. By enhancing the interpretation of multimodal files such as ultrasound, imaging, and electrocardiogram data, marginal cases that were previously only indicated in materials such as CT images and currently require deliberate manual review are upgraded into traceable, referable, and searchable parallel structured multimodal observation data. By fusing multimodal observation data with test indicator data, the large model can simultaneously consider numerical anomalies, imaging / ultrasound / ECG findings, literature evidence, and conflicting evidence when generating reports, which can reduce misjudgments caused by single textual evidence.

[0048] In step S13, the computer device generates at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data.

[0049] In some embodiments, the abnormal indicator items and multimodal observation data in the evidence mapping table are grouped according to organ systems. That is, abnormal indicator items and multimodal observation data under the same organ system are grouped to generate at least one abnormal evidence combination corresponding to an organ system. The abnormal evidence combination corresponding to each organ system is a combination of at least two abnormal indicator items under that organ system or a combination of at least one abnormal indicator item under that organ system and at least one multimodal observation data under that organ system, thus generating the clinical question "abnormal combination related to a certain organ system". In some embodiments, grouping according to organ systems can be achieved through organ system grouping, rule constraint search, graph algorithms (such as graph clustering, community detection, minimum connected subgraph algorithm, etc.), frequent itemset mining, knowledge graph (graph data representation) mining, graph neural networks (graph deep learning or graph machine learning), etc. In some embodiments, anomalous evidence combinations include, but are not limited to, combination identifiers, evidence identifier lists, multimodal observation data identifier lists, anomalous indicator item identifier lists, problem tags (such as "an indicator is elevated / decreased / positive"), pattern signatures (such as "indicator name.abnormal direction"), retrieval intent (fixed retrieval purpose, such as disease mapping, clinical significance, differential diagnosis, etc.), and query seeds (query seeds automatically generated from combinations of at least two anomalous indicator items, or query seeds automatically generated from combinations of at least one anomalous indicator item and at least one multimodal observation data item). This example embodiment does not specifically limit these. In some embodiments, an organ system can correspond to one or more anomalous evidence combinations. In some embodiments, the generation of anomalous evidence combinations can be achieved not only through organ system grouping but also through rule-constrained search, graph algorithms (such as graph clustering, community detection, minimum connected subgraph algorithms, etc.), frequent itemset mining, knowledge graph (graph data representation) mining, or graph neural networks (graph deep learning or graph machine learning). By combining anomalous evidence, the interpretation of a single indicator is extended to a multi-indicator combination interpretation, which can improve the recall capability when multiple mild anomalies jointly point to a certain risk direction. In some embodiments, test index data and multimodal observation data are merged in an evidence map table and enter the same inference chain through anomalous evidence combination, RAG retrieval enhancement generation, and large model context construction.

[0050] In some embodiments, the list of test indicator data, the list of multimodal observation data, and the evidence mapping table are converted into a data payload. The data payload refers to a structured JSON (structured data format) data packet transmitted between modules. Each anomaly indicator item retains the original indicator ID, the evidence identifier list, the multimodal observation data identifier list, the test indicator data identifier list, the original evidence text, the anomaly source, the anomaly priority, and the anomaly confidence level. Each multimodal observation data retains the modality type, anatomical location, discovery text, standardized discovery, measurement value, severity, confidence level, source frame, and source page. In some embodiments, based on the data payload, a set of anomalous evidence for enhanced RAG retrieval generation, a set of clinical questions (clinical questions generated from anomalous evidence), a set of RAG queries (a set of query seeds), system groupings (grouped by organ system), risk summary candidates, and syndrome retrieval packages (retrieval combinations by organ system tissues) are further generated. Then, based on this information, at least one combination of anomalous evidence corresponding to an organ system is generated, and enhanced RAG retrieval generation is performed based on the combination of anomalous evidence. This makes the subsequent retrieval independent of the original report format and does not leave CT, ultrasound, or ECG findings in an unstructured attachment state. This makes the subsequent retrieval chain independent of the original report format and image attachment format.

[0051] In step S14, the computer device generates first query information based on the combination of abnormal evidence, performs a first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result, generates second query information based on the first retrieval result, performs a second-stage medical evidence retrieval based on the second query information to obtain a second retrieval result, and obtains the retrieval text block matched by the first retrieval result and the second retrieval result. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions.

[0052] In some embodiments, a first query is generated based on anomalous evidence combinations. For example, based on the anomalous evidence combinations, the corresponding organ systems, and the patient's background, a first query is generated regarding clinical statements, disease hypotheses, and key points for subsequent assessment. This first query is used to search a knowledge base (including but not limited to medical textbooks, papers, guidelines, expert consensus, clinical pathways, patent literature, clinical trial registries, and institutionally built knowledge bases) to identify candidate diseases and / or risk directions. In some embodiments, a first-stage medical evidence retrieval (RAG, Retrieval Enhancement Generation) is performed based on the first query information to obtain a first search result. Then, a second query is generated based on the first search result regarding the candidate disease. This second query is used to search corpora such as guidelines, textbooks, literature, papers, or institutional knowledge bases around the candidate disease and / or risk direction to retrieve management recommendations and / or clinical evidence related to the candidate disease and / or risk direction. Finally, a second-stage medical evidence retrieval (RAG, Retrieval Enhancement Generation) is performed based on the second query information to obtain a second search result. In some embodiments, the retrieved text block refers to a searchable text block in the knowledge base, that is, the smallest searchable text block after the knowledge base document is segmented, including but not limited to the main text (text block text), parent context, source name, level, page number, PMID / URL (paper number / link), and other metadata. This example embodiment does not impose any special limitations on this. In some embodiments, the first stage of medical evidence retrieval may employ vector retrieval, and the first retrieval results may be structurally weighted according to the matching degree between indicator keywords and organ systems. The second stage may employ keyword retrieval, vector retrieval, or a hybrid retrieval, and the second retrieval results may be reordered according to the matching degree between disease keywords, indicator keywords, therapy keywords, and organ systems. In some embodiments, the same search text block may be hit by multiple search results (including the first search result and the second search result) corresponding to multiple query information (including the first query information and the second query information). When aggregating the results across query information (the combined set of search results returned by the first query information and the second query information), duplicates are removed by text block identifier (such as text block id). Search text blocks with duplicate hits are appended to the list of hit query information, the list of hit abnormal evidence combination identifiers, the list of hit evidence identifiers, and the number of hits. The score is increased according to the number of hits. Then, the sorting of search text blocks and / or the sorting of search results that hit search text blocks are adjusted according to the score, which can improve the relevance and stability of the search context.

[0053] In step S15, the computer device constructs context information based on the abnormal evidence combination, the first search result, and the second search result, and obtains a physical examination analysis report output by the target large model based on the context information. The physical examination analysis report includes a reference number for the searched text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

[0054] In some embodiments, contextual information of a large model is constructed based on anomalous evidence combinations, first search results, and second search results. Specifically, contextual information of the large model is constructed based on anomalous indicator items in the anomalous evidence combinations, multimodal observation data, candidate diseases and / or risk directions in the first search results, and management recommendations and / or clinical evidence in the second search results. In some embodiments, for the paper sources in the first or second search results, metadata such as title, year, PMID, URL, abstract, Medical Subject Headings (MeSH), keywords, therapy type, and indication are retained. For multimodal observation data, data such as examination type, image or waveform source, structured findings, confidence level, whether it supports the test indicator data, and whether manual review is required are retained. In some embodiments, the contextual information is input into the target large model to obtain a physical examination analysis report output by the target large model based on the contextual information. The target large model includes, but is not limited to, general large models, medical-specific large models, local models, and cloud models. This example embodiment does not impose special limitations on the model structure, model parameters, or model training method of the target large model. In some embodiments, the health checkup analysis report includes citation numbers (such as R1, R2, etc.) for the retrieved text blocks. During the generation of the health checkup analysis report, the target large model uses an evidence mapping table to associate the report with the test indicator data and multimodal observation data. This citation number allows for the rapid identification of the retrieved text blocks referenced by the report, as well as the associated test indicator data and multimodal observation data. In other words, the health checkup analysis report is subject to citation and association constraints, ensuring that the medical interpretations and recommendations within the report originate from medical evidence and the evidence mapping table, thus reducing unfounded interpretations or citation illusions. In some embodiments, the health checkup analysis report can be a Markdown file. The target large model first generates structured data (such as JSON) and then renders it as a Markdown file.

[0055] In some embodiments, the step of obtaining structured test indicator data by performing content recognition on the medical examination report file includes: performing content recognition on the medical examination report file to obtain a structured file; and parsing the structured file to obtain structured test indicator data. In some embodiments, content recognition (e.g., OCR) is performed on the medical examination report file to obtain a corresponding structured file (such as a Markdown file), and then the structured file is parsed to obtain structured test indicator data. In some embodiments, content recognition of the medical examination report file is performed by calling an OCR service. For development environments, a Hypertext Transfer Protocol (HTTP) OCR Worker (processing unit) can be used, and for production environments, a queue and a Graphics Processing Unit (GPU) Worker can be used. In some embodiments, if the OCR service is unavailable, the front end can prompt that OCR is not ready without losing the uploaded medical examination report file, maintaining the traceability of the task.

[0056] In some embodiments, obtaining a structured file by performing content recognition on the medical examination report file includes: performing content recognition on the medical examination report file, and obtaining a structured file according to a preset priority rule in response to a successful content recognition callback. In some embodiments, in response to a successful OCR service callback, the structured file (such as a Markdown file) produced by the OCR service is located according to a preset priority rule. The preset priority rule characterizes the priority used to locate the output of the OCR service. For example, the preset priority rule may be to locate the structured file (such as a Markdown file) according to the following priorities: 1. markdown_url or output_dir in the callback payload; 2. markdown / subdirectory under the task directory; 3. Task root directory; 4. Prioritize selecting a Markdown file named source.md from the candidate directories; 5. If multiple Markdown files are available, sort them by whether they are source.md and their path length, and select the most likely main file. In some embodiments, the preset priority rule can achieve compatibility with the output directory structures of different OCRWorkers and automatically use a higher-quality Markdown file when a new text layer fallback file is added to the OCR service.

[0057] In some embodiments, parsing the structured file to obtain structured inspection index data includes: parsing at least one piece of content in the structured file based on a parser corresponding to the content type of the content to obtain structured inspection index data. In some embodiments, for each piece of content in the structured file, parsing the content based on a parser corresponding to the content type of the content to obtain the structured inspection index data corresponding to the content. In some embodiments, the content type includes, but is not limited to, HTML tables, Markdown pipe tables, tables with headers, tables without headers, text layers, item detail blocks, etc., and the corresponding parsers include, but are not limited to, HTML table parsers, Markdown pipe table parsers, PDF text layer parsers, headerless row parsers, item detail block parsers, etc., which are not specifically limited in this example embodiment.

[0058] In some embodiments, the content type includes any of the following: a table, wherein the table includes any of the following: an HTML table, a Markdown pipeline table, a table with a header, and a table without a header; a text layer; and a project detail block. In some embodiments, for the text layer (such as a PDF text layer), based on the PDF text layer parser, context such as "inspection item" and "item name / result / reference" is identified, and data such as numerical values, anomaly markers, qualitative results, units, and reference ranges are extracted from the context, and consecutive rows are combined into inspection index data. In some embodiments, for the project detail block, based on the project detail block parser, fields such as project name, result / unit, and reference value are identified, and fields such as numerical values, anomaly markers, qualitative results, units, and reference ranges are extracted to obtain the inspection index data.

[0059] In some embodiments, the content type includes a table; wherein, for at least one piece of content in the structured file, parsing the content based on a parser corresponding to the content type of the content to obtain structured inspection index data includes: for a table in the structured file, parsing the table based on a table parser, identifying multiple columns in the table, and obtaining structured inspection index data corresponding to the multiple columns, wherein the multiple columns include at least two of the following: item column, numerical column, anomaly marker column, qualitative result column, unit column, and reference range column. In some embodiments, for tables in a structured file, the tables are parsed using a table parser (such as an HTML table parser, a Markdown pipe table parser, etc.) to identify multiple columns in the table. These multiple columns include, but are not limited to, at least two of the following: item column, numerical column, anomaly marker column, qualitative result column, unit column, and reference range column. For example, "item, item name, inspection item, test item, indicator, name" are identified as item columns; "result, test result, result value, current result" are identified as numerical columns; "unit, unit of measurement, result unit" are identified as unit columns; "prompt, marker, anomaly, anomaly prompt, flag" are identified as anomaly marker columns; "reference range, reference value, reference interval, normal range, normal value" are identified as reference range columns; and "negative / positive" are identified as qualitative result columns. In some embodiments, structured test indicator data corresponding to the multiple columns are obtained. For example, the test indicator data includes data corresponding to the numerical column "value = 6.76", data corresponding to the anomaly flag column "anomaly flag = ↑", data corresponding to the unit column "unit = mmol / L", data corresponding to the reference range column "reference = 4.11-5.89", and data corresponding to the qualitative result column "positive (+)".

[0060] In some embodiments, obtaining the structured inspection indicator data corresponding to the plurality of columns includes: if the plurality of columns include at least two item columns, dividing the table into multiple column groups based on the position of the item columns in the table, and obtaining the structured inspection indicator data corresponding to each column group. In some embodiments, if the plurality of columns include at least two item columns (for example, columns with headers of item / item name / indicator are called item columns), that is, when multiple groups of items are arranged horizontally side by side in the same table, it is necessary to perform multi-column indicator group splitting, splitting one row into multiple inspection indicator data. The specific process is: first scan the header, identify the position of each "item / item name / indicator" column (i.e., item column) in the table as the starting point of a group of items, and form a column group from the starting point to the starting point of the next group of items, thereby dividing the table into multiple column groups and obtaining the structured inspection indicator data corresponding to each column group.

[0061] In some embodiments, for a table in the structured file, parsing the table based on a table parser to identify multiple columns in the table and obtaining structured inspection index data corresponding to the multiple columns includes: for a table in the structured file, performing semantic recovery on the table based on a table parser to identify multiple columns in the recovered table and obtaining structured inspection index data corresponding to the multiple columns. In some embodiments, semantic recovery of tables in structured files is first performed using a table parser. Then, multiple columns in the recovered table are identified. For example, for HTML tables, the table (table), tr (table row), td (table data cell), and th (table header cell) are identified, and empty columns are filled in based on the colspan (column span). For tables with headers, column mappings are established based on header aliases such as "item," "result," "unit," "hint," and "reference range." For tables without headers, item rows are inferred based on item name appearance, numerical / positive / negative / reference range features. In some embodiments, semantic recovery can be assisted by rule parsing, Document Object Model (DOM) parsing, machine learning table structure recovery models, visual language models, or large language models. Semantic recovery can improve the stability of indicator extraction under different OCR intermediate formats and reduce erroneous or missed extractions caused by misaligned columns, mixed columns, merged columns, missing headers, and item detail blocks.

[0062] In some embodiments, the method further includes: determining whether the test indicator data is an abnormal indicator item based on priority rules corresponding to multiple anomaly sources, and identifying whether the test indicator data is an abnormal indicator item. The multiple anomaly sources include at least two of explicit anomaly markers, qualitative reference sources, reference range sources, and context-dependent reference range sources. In some embodiments, spaces, colons, and suffixes such as "measurement," "detection," "quantitative," and "qualitative" are removed from the indicator names in the test indicator data. In some embodiments, a preset indicator vocabulary is used to match standardized names, synonyms, abbreviations, and recommended merged words to the indicator names in the test indicator data to achieve indicator name standardization. The indicator vocabulary is a medical vocabulary used for the standardization of physical examination indicators and system attribution. Indicator name standardization can be achieved using methods such as manual vocabularies, medical terminology ontology, embedding similarity, LLM classifiers, and knowledge graphs. In some embodiments, based on priority rules corresponding to multiple anomaly sources, anomaly determination is performed on the test indicator data with respect to these multiple anomaly sources to determine whether the test indicator data is an anomalous indicator item. These multiple anomaly sources include, but are not limited to, at least two of the following: explicit anomaly markers, qualitative reference sources, reference range sources, and context-dependent reference range sources. For example, the priority rule could be "normal / negative protection takes priority to avoid misjudgment; explicit anomaly markers have the highest priority; qualitative positive results are secondary; and numerical reference range comparisons are secondary." In some embodiments, anomaly determination can be implemented by rule engines, medical knowledge graphs, statistical models, supervised learning classification models, or physician review result feedback models.

[0063] In some embodiments, the method further includes: performing anomaly consistency recalculation on the anomaly indicator item based on its actual indicator value, lower limit of the reference range, and upper limit of the reference range. In some embodiments, it is necessary to perform anomaly consistency recalculation on the anomaly indicator item based on its actual indicator value (actual), lower limit of the reference range (refMin), and upper limit of the reference range (refMax). If actual is greater than refMax, the status needs to be changed to "high"; if actual is less than refMin, the status needs to be changed to "low"; otherwise, the status needs to be changed to "normal". In some embodiments, if there is no available reference range for the explicit anomaly marker, the original explicit marker can be retained for the anomaly indicator item. In some embodiments, if the recalculation result conflicts with the original status, at least one of the anomaly indicator item, multimodal observation data, and evidence mapping table needs to be updated synchronously. By prioritizing the anomaly source and performing anomaly consistency recalculation based on the actual indicator value / lower limit of the reference range / upper limit of the reference range, erroneous retrieval, erroneous interpretation, and erroneous suggestions caused by misjudgment of anomaly direction can be reduced.

[0064] In some embodiments, the method further includes: if the source of the abnormality judgment corresponding to the abnormal indicator item is a context-dependent reference range source, determining that the abnormal indicator item does not need to undergo abnormality consistency recalculation. In some embodiments, if the source of the abnormality judgment corresponding to the abnormal indicator item (i.e., the abnormal source used as the basis for judging the test indicator data as an abnormal indicator item) is a context-dependent reference range source (a multi-stage reference range dependent on the patient's state, such as menstruation, pregnancy, follicular phase, luteal phase, etc.), abnormality consistency recalculation is not mandatory to avoid incorrectly applying multiple cycle ranges.

[0065] In some embodiments, the method further includes: performing fact-checking on the abnormal indicator items according to the structured file corresponding to the medical examination report file. In some embodiments, fact-checking on the abnormal indicator items is performed according to the structured file corresponding to the medical examination report file (such as a Markdown file obtained from OCR), comparing the indicator name, value, reference range, arrow / high / low direction, and final report status with the structured file, and outputting correction suggestions at the failure, warning, or uncertainty level. In some embodiments, if the verification result conflicts with the original status, at least one of the abnormal indicator items, multimodal observation data, and evidence mapping table needs to be updated synchronously.

[0066] In some embodiments, the method further includes: determining whether the abnormal indicator needs manual review based on the extraction confidence level and / or abnormal confidence level corresponding to the abnormal indicator. In some embodiments, determining whether the abnormal indicator needs manual review based on the extraction confidence level and / or abnormal confidence level corresponding to the abnormal indicator, wherein the extraction confidence level is used to measure "whether this indicator is reliably extracted", and the abnormal confidence level is used to measure "whether the basis for judging this indicator as abnormal is reliable". In some embodiments, a dashboard threshold and a review threshold can be set. Abnormal indicator items with extraction confidence level and / or abnormal confidence level lower than the review threshold are not entered into the dashboard, that is, they are not input into the target large model or do not appear in the physical examination analysis report. Abnormal indicator items between the review threshold and the dashboard threshold enter the queue for manual review, thereby realizing confidence level gating of abnormal indicator items. This can prevent abnormal indicator items with low confidence levels from directly entering the dashboard, RAG, and report generation.

[0067] In some embodiments, the method further includes: determining the extraction confidence level corresponding to the abnormal indicator item based on whether the indicator name of the abnormal indicator item matches a preset indicator term list. In some embodiments, the extraction confidence level corresponding to the abnormal indicator item is determined based on whether the indicator name of the abnormal indicator item matches a preset indicator term list. For example, if the indicator name of the abnormal indicator item can match at least one of the following in the preset indicator term list: a standard name, synonym, abbreviation, and recommended merged word, then the indicator name matches the indicator term list; otherwise, the indicator name does not match the indicator term list.

[0068] In some embodiments, the method further includes: determining the anomaly confidence level corresponding to the anomaly indicator item based on the plurality of anomaly sources. In some embodiments, the anomaly confidence level corresponding to the anomaly indicator item is determined based on the values ​​of the anomaly indicator item with respect to the plurality of anomaly sources, for example, based on whether the anomaly indicator item lacks a value with respect to at least one of the plurality of anomaly sources. For example, the anomaly confidence level is reduced when a reference range or anomaly marker is missing. In some embodiments, the anomaly confidence level corresponding to the anomaly indicator item is determined based on the anomaly determination source (i.e., the anomaly source among the plurality of anomaly sources that determines the test indicator data as an anomaly indicator item and is used as the determination basis) among the plurality of anomaly sources. The confidence level corresponding to each anomaly determination source can be preset, and the confidence level of the anomaly determination source corresponding to the anomaly indicator item can be used as the anomaly confidence level corresponding to the anomaly indicator item.

[0069] In some embodiments, writing the test index data and the multimodal observation data into an evidence mapping table includes: writing the test index data into the evidence mapping table; determining the multimodal confidence of the multimodal observation data based on the model confidence, source location confidence, cross-modal consistency, and clinical relevance corresponding to the multimodal observation data; and if the multimodal confidence meets a preset condition, writing the multimodal observation data into the evidence mapping table, wherein the cross-modal consistency is used to characterize whether the multimodal observation data and the test index data support each other, and the clinical relevance is used to characterize the correlation strength between the multimodal observation data and the abnormal indicator item. In some embodiments, model confidence is used to characterize the model's credibility with the multimodal observation data itself; source location confidence is used to characterize whether the multimodal observation data can be located to a specific page, frame, image region, or original text fragment; cross-modal consistency is used to characterize whether the multimodal observation data and the test indicator data support each other; and clinical relevance is used to characterize the strength of the association between the multimodal observation data and the abnormal indicator item. The multimodal confidence of the multimodal observation data (i.e., the comprehensive score or fusion score of the reliability of the multimodal observation data) can be calculated based on these four data. For example, the multimodal confidence of the multimodal observation data can be obtained by multiplying these four data by their respective weights and then weighting and summing them. The multimodal observation data will only be written into the evidence mapping table if the multimodal confidence meets the preset conditions (e.g., greater than or equal to the preset threshold).

[0070] In some embodiments, generating at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table includes: determining the organ system corresponding to the abnormal indicator item and the organ system to which the multimodal observation data belongs, based on the abnormal indicator items and multimodal observation data in the evidence mapping table; grouping the abnormal indicator items and multimodal observation data under the same organ system to generate at least one abnormal evidence combination corresponding to an organ system. In some embodiments, for abnormal indicator items in the evidence mapping table, the organ system corresponding to the abnormal indicator item is inferred first using the diagnostic-related organ system field in a preset indicator terminology. If the terminology is missing, the organ system corresponding to the abnormal indicator item is inferred using external organ system mapping. If it is still missing, rule keywords are used as a fallback, such as transaminase belonging to the liver and gallbladder, creatinine / urea belonging to the kidney, blood glucose belonging to glucose metabolism, and blood lipids belonging to the cardiovascular system. The rules cover organ systems such as the liver and gallbladder, kidney / urinary system, glucose metabolism, lipid metabolism / cardiovascular system, thyroid / endocrine system, blood, inflammation / immunity, coagulation, tumor markers, reproductive endocrine system, bone metabolism, digestive system, respiratory system, and nervous system. In some embodiments, for multimodal observation data in the evidence mapping table, the organ system to which the multimodal observation data belongs is inferred based on information such as anatomical location, examination type, and abnormal signs in the multimodal observation data. For example, fatty liver on liver ultrasound is an imaging finding and should be classified as a hepatobiliary system. In some embodiments, abnormal indicators and multimodal observation data under the same organ system are grouped to generate at least one abnormal evidence combination corresponding to an organ system. In some embodiments, the inference of organ systems can be achieved by means of vocabulary fields, external mapping tables, rule bases, graph models, or large language model classifiers.

[0071] In some embodiments, the anomalous evidence combination is a combination of at least one anomalous indicator and at least one multimodal observation data; wherein, the method further includes: determining the retrieval priority corresponding to the anomalous evidence combination based on the degree of support between the at least one anomalous indicator and the at least one multimodal observation data. In some embodiments, the retrieval priority corresponding to the anomalous evidence combination (retrieval priority in the first stage of medical evidence retrieval and / or retrieval priority in the second stage of medical evidence retrieval) is determined based on the degree of support between the anomalous indicator and the multimodal observation data in the anomalous evidence combination (e.g., whether they support each other), for example, if they support each other, the retrieval priority of the corresponding anomalous evidence combination is increased.

[0072] In some embodiments, the method further includes: generating target reference numbers for the test index data, the multimodal observation data, and the retrieval text block, and adding the target reference numbers to a reference map table; wherein, the method further includes: verifying whether the reference numbers in the physical examination analysis report belong to the reference map table, and if not, performing a first processing operation on the physical examination analysis report. In some embodiments, target reference numbers for the test index data, multimodal observation data, and retrieval text block are generated, and the target reference numbers are added to a reference map table. Subsequently, after the physical examination analysis report is output by the target large model, a reference verification is performed on the physical examination analysis report to verify whether the reference numbers contained in the physical examination analysis report belong to the reference map table. If not, a first processing operation is performed on the physical examination analysis report. The first processing operation includes, but is not limited to, direct failure, regeneration, and entering the correction process. This example embodiment does not specifically limit this. The reference map table is used to maintain the mapping relationship between the test index data, multimodal observation data, and retrieval text block and the reference numbers. Reference verification can be implemented through JSON schema verification, claim-level evidence verifiers, reference graphs, or manual review workbench, etc.

[0073] In some embodiments, the method further includes: determining conflict information between the test indicator data cited in the physical examination analysis report and the multimodal observation data cited in the physical examination analysis report based on the reference number; and performing a second processing operation on the physical examination analysis report if the conflict information meets preset conditions. In some embodiments, obtaining conflict information between the test indicator data cited in the physical examination analysis report and the multimodal observation data cited in the physical examination analysis report includes whether the two conflict or whether evidence of conflict between the two is disclosed. Whether the two conflict refers to whether the physical examination analysis report describes the conflict, and whether evidence of conflict is disclosed refers to whether the physical examination analysis report describes the conflict as a controlled expression such as "pending review," "conflicting evidence," or "cannot be confirmed," rather than directly drawing a definitive conclusion. In some embodiments, if the conflict information meets preset conditions, for example, if the two conflict or if evidence of conflict between the two is not disclosed, then performing a second processing operation on the physical examination analysis report. The second processing operation includes, but is not limited to, direct failure, regeneration, or entering a correction process. This example embodiment does not impose any special limitations on this.

[0074] In some embodiments, the method further includes: generating prompt words according to a preset clinical inference path; wherein, obtaining the physical examination analysis report output by the target large model based on the context information includes: obtaining the physical examination analysis report output by the target large model based on the context information and the prompt words. In some embodiments, prompt words are generated according to a preset clinical inference path (i.e., the clinical inference path for doctors reading physical examination reports), requiring the target large model to output in the following order: patient background, positive / abnormal findings, multimodal structured findings, abnormal combination interpretation, candidate risk directions, evidence citation, lifestyle and follow-up recommendations, frontier intervention direction boundaries, prohibitive expressions, and review prompts. The prompt words can stabilize the target large model's path for reproducible fact confirmation, abnormal combination analysis, risk assumptions, evidence citation, recommendation boundaries, and review prompts. In some embodiments, the preset clinical inference path is converted into a prompt word template, and then prompt words are generated based on the prompt word template. The prompt word template can be generated by a fixed template, a role-based template, a dynamic template selector, a fine-tuning model, or a prompt word parameter search, etc.

[0075] In some embodiments, the method further includes: responding to an editing operation performed by a target user on the physical examination analysis report, generating third query information based on the edited physical examination analysis report, performing a third-stage medical evidence retrieval based on the third query information to obtain a third search result, wherein the third query information is used to enhance evidence around the auxiliary analysis direction specified by the target user; and obtaining an updated physical examination analysis report output by the target large model based on the third search result. In some embodiments, after a target user (such as a doctor) edits the physical examination analysis report on the front end, they can trigger the generation of third query information based on the edited physical examination analysis report. The third query information is used to enhance evidence (i.e., in-depth exploration) around the auxiliary analysis direction specified by the target user. The auxiliary analysis direction includes, but is not limited to, auxiliary analysis directions related to the content of the physical examination analysis report, such as cell therapy, immune cell therapy, stem cell therapy, exosomes, gene therapy or gene editing, nutritional intervention, follow-up examination, and chronic disease management topics. That is, the third query information will contain keywords of these auxiliary analysis directions for in-depth exploration of cell / gene therapy evidence, and can be combined with external biomedical literature to enhance the search service. In other words, the third query information is an in-depth exploration path triggered by the target user. In some embodiments, a third-stage medical evidence retrieval (RAG, Retrieval Enhancement Generation) is performed based on the third query information to obtain third search results. These third search results filter out advanced therapeutic terms that are incompatible with or overly specialized in routine physical examination management, such as chimeric antigen receptor T-cell therapy (CAR-T), T-cell receptor engineered T-cell therapy (TCR-T), and CRISPR (Clustered Regularly Interval Short Palindromic Repeat) gene editing technology. In some embodiments, the third search results, after evidence compression, are incorporated into a target large model for updating the physical examination analysis report, such as rewriting sections like "Cell Application Protocol" and "Cell Type Definitions." In some embodiments, third query information is generated based on the edited physical examination analysis report, the first search results, the second search results, and a combination of abnormal evidence. In some embodiments, a health intervention document can be generated based on the edited report, using the report content and structured indicators as input to generate a separate intervention recommendation document.

[0076] In some embodiments, the front-end provides capabilities for uploading, progress display, viewing test indicator data, viewing multimodal observation data, risk analysis, displaying physical examination analysis reports, report editing, PDF export, historical analysis, in-depth exploration, and generating health intervention reports. After saving modifications, users can save the edited content, in-depth exploration results, and export records as history and audit records for subsequent manual (doctor) review and system optimization. This means that the doctor's editing and in-depth exploration of the physical examination analysis report are transformed into review and optimization data for subsequent rules, searches, and suggestion terms.

[0077] Figure 2This diagram illustrates a system architecture for obtaining a physical examination analysis report according to one embodiment of the present application.

[0078] like Figure 2As shown, in response to requests initiated by doctors / users, the authentication and quota module authenticates doctors / users and assigns quotas to requests. The report upload module receives the medical examination report files uploaded by doctors / users. The task status and event module generates task identifiers and case identifiers, and writes the user, clinic, invitation code, or anonymized patient identifier to the metadata. This metadata is then written to the task metadata module in the storage and audit layer. If user directory isolation is enabled, the system also generates a user security directory and maintains a task index for easy querying of historical reports by user. The OCR scheduling and result location module calls the OCR service to perform content recognition on the medical examination report files, locates structured files (such as Markdown files) in the recognition results, and stores the medical examination report files and structured files in the file / object storage module in the storage and audit layer. The structured extraction module extracts the test indicator data from the structured files based on the indicator vocabulary / organ system vocabulary in the medical knowledge layer. The consistency recalculation and fact-checking module performs consistency recalculation on abnormal indicator items in the test indicator data. In addition to fact-checking, the multimodal data access module receives multimodal files such as ultrasound, CT, and ECG. The physician multimodal interpretation module enhances the interpretation of images, waveforms, and report conclusions in the multimodal files. The multimodal structured evidence flow module generates multimodal observation data in parallel with the test index data based on the enhanced interpretation results. The evidence fusion module writes the ClinicalIndicator (test index data) and multimodal observation data into an evidence mapping table. The RAG (Retrieval Enhancement Generation) Payload module converts the evidence mapping table into a data payload, and further generates at least one abnormal evidence combination corresponding to an organ system based on the data payload. The phased retrieval and arrangement module performs the first and second phases of medical evidence retrieval based on the abnormal evidence combination in the textbook / guideline vector library, the paper / PubMed vector library, and external literature enhancement services at the medical knowledge layer. The weighted ranking and evidence aggregation module weights and rearranges the retrieval results and aggregates the retrieval text blocks that are hit by the retrieval. Prompt The Words (prompt words) report generation module generates prompt words based on search results. These prompt words are then input into a large model to obtain the output health checkup analysis report. The weighted ranking and evidence aggregation module, along with the Prompt Words report generation module, stores the corresponding operation structure in the Pipeline Artifacts module of the storage and audit layer. The citation verification module verifies the citation numbers contained in the health checkup analysis report. The front-end editing / deep exploration / health intervention module responds to doctors' editing operations on the health checkup analysis report, triggering deep exploration. Based on the edited report, a health intervention document is generated, using the report content and structured indicators as input to generate a separate intervention suggestion document. The front-end editing / deep exploration / health intervention module stores the corresponding operation records in the history and audit records module of the storage and audit layer.The feedback learning module generates feedback samples based on the operations performed by the front-end editing / deep exploration / health intervention modules. These samples are used to update anomaly detection corrections, search relevance, prompt word preferences, and report wording preferences, thus achieving a closed-loop mechanism for front-end editing feedback. This transforms the doctor's editing and deep exploration of the physical examination analysis report into data for subsequent rule, search, and prompt word optimization.

[0079] Figure 3 A flowchart illustrating a method for obtaining a physical examination analysis report according to an embodiment of this application is shown.

[0080] like Figure 3As shown, the system first receives the uploaded physical examination report file (PDF), generates a task_id (task identifier) / case_id (case identifier) ​​and a task directory, then saves source.pdf (source PDF file) and meta.json (metadata), and simultaneously accesses ultrasound / CT / ECG data. After completing the above saving operations, it publishes task_created (task created) / queued (task entered queue) events, then calls the OCR Worker to determine if the OCR was successful. If not, it publishes an error and retains the recoverable state; if successful, it locates the OCR... Markdown outputs are then parsed, followed by HTML / Markdown / PDF text layers / item details, splitting values, units, reference ranges, and anomaly markers. Next, indicator name cleaning and thesaurus matching are performed, followed by multi-level anomaly detection and confidence gating. High-confidence data is then checked for compliance with gating; if not, manual review or low-confidence alerts are issued. If yes, actual / refMin / refMax consistency recalculation is performed to generate clinicalIndicator test data. After the parallel access to ultrasound / CT / ECG data is completed, a medical multimodal model is used to enhance image / waveform interpretation, generating multimodal observation data. CT edge cases are then upgraded from review alerts to structured evidence nodes. Multi-source evidence fusion is then performed, fusing test data and multimodal observation data to generate an evidence map, which is then converted to RAG. The payload and bundle (anomalous evidence combination) are used to identify candidate diseases / risks based on query 1 (first query information). Query 2 (second query information) retrieves management suggestions and cutting-edge evidence. Then, it is determined whether the doctor has triggered in-depth exploration. If so, external literature enhancement search is performed for query 3 (third query information). After that, evidence aggregation, reordering, deduplication, and multi-hit scoring are performed. Otherwise, evidence aggregation, reordering, deduplication, and multi-hit scoring are performed directly. Then, an LLM (Large Language Model) context with citation numbers is constructed. A structured report is generated based on PromptWords. Citation verification and indicator fact-checking are then performed. Finally, the front-end is displayed, edited, exported, and feedback learning is provided.

[0081] Figure 4 This application illustrates a system for obtaining a physical examination analysis report (hereinafter referred to as the "physical examination analysis report system") according to one embodiment of the present application. The physical examination analysis report system includes a first module, a second module, a third module, a fourth module, and a fifth module.

[0082] The first module is used to perform content recognition on the physical examination report document to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items. Here, the relevant operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0083] The second module is used to enhance the interpretation of the multimodal file corresponding to the physical examination report file, obtain structured multimodal observation data, and write the test indicator data and the multimodal observation data into an evidence mapping table. Here, the relevant operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0084] The third module is used to generate at least one combination of anomalous evidence corresponding to an organ system based on the evidence mapping table. This combination of anomalous evidence is either a combination of at least two anomalous indicator items or a combination of at least one anomalous indicator item and at least one multimodal observation data point. The related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0085] The fourth module is used to generate first query information based on the combination of abnormal evidence, perform a first-stage medical evidence retrieval based on the first query information to obtain a first search result, generate second query information based on the first search result, perform a second-stage medical evidence retrieval based on the second query information to obtain a second search result, and obtain the search text block matched by the first and second search results. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0086] The fifth module is used to construct contextual information based on the abnormal evidence combination, the first search result, and the second search result, and obtain a physical examination analysis report output by the target large model based on the contextual information. The physical examination analysis report includes a citation number for the searched text block, and the physical examination analysis report is associated with the test indicator data and the multimodal observation data through the evidence mapping table. Here, related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0087] Figure 5The diagram shows a computer device structure for obtaining a physical examination analysis report according to an embodiment of the present application. The device includes module 11, module 12, module 13, module 14, and module 15.

[0088] Module 11 is used to perform content recognition on the physical examination report file to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items.

[0089] In some embodiments, the file format of the physical examination report file includes, but is not limited to, PDF, photographed images, etc., and this example embodiment does not impose any special limitations on this. In some embodiments, content recognition (e.g., OCR optical character recognition) is performed on the physical examination report file to obtain structured multiple test indicator data. Each test indicator data includes, but is not limited to, indicator name, indicator value, abnormality mark, qualitative result, unit, reference range, etc., and this example embodiment does not impose any special limitations on this. In some embodiments, the multiple test indicator data includes at least one normal indicator item (i.e., test indicator data determined to be normal) and at least one abnormal indicator item (i.e., test indicator data determined to be abnormal). In some embodiments, the physical examination report file uploaded by the user is received, the file type and size are verified, and a task identifier and a case identifier are generated. The task identifier can be in the form of "task_date_random fragment", and the case identifier can be in the form of "case_date_random fragment", and this example embodiment does not impose any special limitations on this. In some embodiments, a metadata JSON file (meta.json) is created in the task directory, and the user, clinic, invitation code, or de-identified patient identifier of the task is written into the metadata. In some embodiments, if user directory isolation is enabled, a user security directory is also generated and a task index is maintained to facilitate historical report queries by user.

[0090] Module 12 is used to enhance the interpretation of the multimodal file corresponding to the physical examination report file, obtain structured multimodal observation data, and write the test index data and the multimodal observation data into the evidence mapping table.

[0091] In some embodiments, the multimodal files corresponding to the physical examination report file include, but are not limited to, ultrasound images or ultrasound reports, CT / MRI / X-ray imaging images or reports, electrocardiogram waveforms / screenshots / parameter tables, and other examination images. This example embodiment does not impose any special limitations on these. In some embodiments, the images, waveforms, and examination conclusions in the multimodal files are enhanced and interpreted to obtain structured multimodal observation data. This multimodal observation data includes, but is not limited to, examination type, anatomical location, left / right side, measurement values, abnormal signs, severity, conclusion tendency, confidence level, source page / frame number, and original fragment. This example embodiment does not impose any special limitations on these. In some embodiments, enhanced interpretation can be achieved through medical multimodal models, visual language models, specialist imaging models, electrocardiogram analysis models, and structured interfaces of hospital information systems. This example embodiment does not impose any special limitations on these. In some embodiments, test index data and multimodal observation data are jointly written into an evidence mapping table to achieve the fusion of test index data and multimodal observation data. The role of the evidence mapping table is to map the structured conclusions in the subsequently generated physical examination analysis report back to the source of evidence (including test index data and multimodal observation data). For example, the evidence mapping table includes evidenceIds (evidence ID), observationIds (multimodal observation data ID), and findingIds (test index data ID) corresponding to each conclusion in the physical examination analysis report. That is, each piece of multimodal observation data is written into the evidence mapping table as a piece of evidence, and each piece of test index data is also written into the evidence mapping table as a piece of evidence. The front end can use the evidence mapping table for traceability display, and these IDs will also be used when performing medical evidence retrieval in the future to convert clinical indicators into searchable standardized indicator items.In some embodiments, after obtaining structured multimodal observation data, operations such as confidence gating, source localization, and structured field constraints are performed to upgrade it into a structured evidence stream that runs parallel to the test index data. This stream then enters subsequent RAG retrieval enhancement generation and large model inference. Confidence gating involves scoring structured items (structured multimodal observation data). Items below a first threshold are directly excluded; those between the first and second thresholds (where the second threshold is greater than the first) enter a review queue; and only those above the second threshold enter a high-confidence pool. Only multimodal observation data in the high-confidence pool are written into the evidence mapping table and then enter subsequent RAG retrieval enhancement generation and large model inference. Source localization involves binding each structured item to original OCR evidence, such as rawText (original evidence text), evidenceIds (a list of evidence IDs), and sourceMaterial (source material, i.e., the retained original or near-original input material, such as a structured file obtained by content recognition (e.g., OCR) of a medical examination report document, such as Markdown). (Document) Structured field constraints require structured items to have fixed fields, such as indicator name, value, unit, reference range, status, confidence level, and evidence ID. By enhancing the interpretation of multimodal files such as ultrasound, imaging, and electrocardiogram data, marginal cases that were previously only indicated in materials such as CT images and currently require deliberate manual review are upgraded into traceable, referable, and searchable parallel structured multimodal observation data. By fusing multimodal observation data with test indicator data, the large model can simultaneously consider numerical anomalies, imaging / ultrasound / ECG findings, literature evidence, and conflicting evidence when generating reports, which can reduce misjudgments caused by single textual evidence.

[0092] Module 13 is used to generate at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data.

[0093] In some embodiments, the abnormal indicator items and multimodal observation data in the evidence mapping table are grouped according to organ systems. That is, abnormal indicator items and multimodal observation data under the same organ system are grouped to generate at least one abnormal evidence combination corresponding to an organ system. The abnormal evidence combination corresponding to each organ system is a combination of at least two abnormal indicator items under that organ system or a combination of at least one abnormal indicator item under that organ system and at least one multimodal observation data under that organ system, thus generating the clinical question "abnormal combination related to a certain organ system". In some embodiments, grouping according to organ systems can be achieved through organ system grouping, rule constraint search, graph algorithms (such as graph clustering, community detection, minimum connected subgraph algorithm, etc.), frequent itemset mining, knowledge graph (graph data representation) mining, graph neural networks (graph deep learning or graph machine learning), etc. In some embodiments, anomalous evidence combinations include, but are not limited to, combination identifiers, evidence identifier lists, multimodal observation data identifier lists, anomalous indicator item identifier lists, problem tags (such as "an indicator is elevated / decreased / positive"), pattern signatures (such as "indicator name.abnormal direction"), retrieval intent (fixed retrieval purpose, such as disease mapping, clinical significance, differential diagnosis, etc.), and query seeds (query seeds automatically generated from combinations of at least two anomalous indicator items, or query seeds automatically generated from combinations of at least one anomalous indicator item and at least one multimodal observation data item). This example embodiment does not specifically limit these. In some embodiments, an organ system can correspond to one or more anomalous evidence combinations. In some embodiments, the generation of anomalous evidence combinations can be achieved not only through organ system grouping but also through rule-constrained search, graph algorithms (such as graph clustering, community detection, minimum connected subgraph algorithms, etc.), frequent itemset mining, knowledge graph (graph data representation) mining, or graph neural networks (graph deep learning or graph machine learning). By combining anomalous evidence, the interpretation of a single indicator is extended to a multi-indicator combination interpretation, which can improve the recall capability when multiple mild anomalies jointly point to a certain risk direction. In some embodiments, test index data and multimodal observation data are merged in an evidence map table and enter the same inference chain through anomalous evidence combination, RAG retrieval enhancement generation, and large model context construction.

[0094] In some embodiments, the list of test indicator data, the list of multimodal observation data, and the evidence mapping table are converted into a data payload. The data payload refers to a structured JSON (structured data format) data packet transmitted between modules. Each anomaly indicator item retains the original indicator ID, the evidence identifier list, the multimodal observation data identifier list, the test indicator data identifier list, the original evidence text, the anomaly source, the anomaly priority, and the anomaly confidence level. Each multimodal observation data retains the modality type, anatomical location, discovery text, standardized discovery, measurement value, severity, confidence level, source frame, and source page. In some embodiments, based on the data payload, a set of anomalous evidence for enhanced RAG retrieval generation, a set of clinical questions (clinical questions generated from anomalous evidence), a set of RAG queries (a set of query seeds), system groupings (grouped by organ system), risk summary candidates, and syndrome retrieval packages (retrieval combinations by organ system tissues) are further generated. Then, based on this information, at least one combination of anomalous evidence corresponding to an organ system is generated, and enhanced RAG retrieval generation is performed based on the combination of anomalous evidence. This makes the subsequent retrieval independent of the original report format and does not leave CT, ultrasound, or ECG findings in an unstructured attachment state. This makes the subsequent retrieval chain independent of the original report format and image attachment format.

[0095] Module 14 is used to generate first query information based on the abnormal evidence combination, perform a first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result, generate second query information based on the first retrieval result, perform a second-stage medical evidence retrieval based on the second query information to obtain a second retrieval result, and obtain the retrieval text block matched by the first retrieval result and the second retrieval result. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions.

[0096] In some embodiments, a first query is generated based on anomalous evidence combinations. For example, based on the anomalous evidence combinations, the corresponding organ systems, and the patient's background, a first query is generated regarding clinical statements, disease hypotheses, and key points for subsequent assessment. This first query is used to search a knowledge base (including but not limited to medical textbooks, papers, guidelines, expert consensus, clinical pathways, patent literature, clinical trial registries, and institutionally built knowledge bases) to identify candidate diseases and / or risk directions. In some embodiments, a first-stage medical evidence retrieval (RAG, Retrieval Enhancement Generation) is performed based on the first query information to obtain a first search result. Then, a second query is generated based on the first search result regarding the candidate disease. This second query is used to search corpora such as guidelines, textbooks, literature, papers, or institutional knowledge bases around the candidate disease and / or risk direction to retrieve management recommendations and / or clinical evidence related to the candidate disease and / or risk direction. Finally, a second-stage medical evidence retrieval (RAG, Retrieval Enhancement Generation) is performed based on the second query information to obtain a second search result. In some embodiments, the retrieved text block refers to a searchable text block in the knowledge base, that is, the smallest searchable text block after the knowledge base document is segmented, including but not limited to the main text (text block text), parent context, source name, level, page number, PMID / URL (paper number / link), and other metadata. This example embodiment does not impose any special limitations on this. In some embodiments, the first stage of medical evidence retrieval may employ vector retrieval, and the first retrieval results may be structurally weighted according to the matching degree between indicator keywords and organ systems. The second stage may employ keyword retrieval, vector retrieval, or a hybrid retrieval, and the second retrieval results may be reordered according to the matching degree between disease keywords, indicator keywords, therapy keywords, and organ systems. In some embodiments, the same search text block may be hit by multiple search results (including the first search result and the second search result) corresponding to multiple query information (including the first query information and the second query information). When aggregating the results across query information (the combined set of search results returned by the first query information and the second query information), duplicates are removed by text block identifier (such as text block id). Search text blocks with duplicate hits are appended to the list of hit query information, the list of hit abnormal evidence combination identifiers, the list of hit evidence identifiers, and the number of hits. The score is increased according to the number of hits. Then, the sorting of search text blocks and / or the sorting of search results that hit search text blocks are adjusted according to the score, which can improve the relevance and stability of the search context.

[0097] Module 15 is used to construct context information based on the abnormal evidence combination, the first search result and the second search result, and obtain a physical examination analysis report output by the target large model based on the context information. The physical examination analysis report includes a reference number for the search text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

[0098] In some embodiments, contextual information of a large model is constructed based on anomalous evidence combinations, first search results, and second search results. Specifically, contextual information of the large model is constructed based on anomalous indicator items in the anomalous evidence combinations, multimodal observation data, candidate diseases and / or risk directions in the first search results, and management recommendations and / or clinical evidence in the second search results. In some embodiments, for the paper sources in the first or second search results, metadata such as title, year, PMID, URL, abstract, Medical Subject Headings (MeSH), keywords, therapy type, and indication are retained. For multimodal observation data, data such as examination type, image or waveform source, structured findings, confidence level, whether it supports the test indicator data, and whether manual review is required are retained. In some embodiments, the contextual information is input into the target large model to obtain a physical examination analysis report output by the target large model based on the contextual information. The target large model includes, but is not limited to, general large models, medical-specific large models, local models, and cloud models. This example embodiment does not impose special limitations on the model structure, model parameters, or model training method of the target large model. In some embodiments, the health checkup analysis report includes citation numbers (such as R1, R2, etc.) for the test indicator data, multimodal observation data, and retrieval text blocks. During the generation of the health checkup analysis report, the target large model associates the report with the test indicator data and multimodal observation data through an evidence mapping table. This citation number allows for quick retrieval of the retrieval text blocks referenced in the report, as well as quick retrieval of the associated test indicator data and multimodal observation data. In other words, the health checkup analysis report is subject to citation and association constraints, ensuring that the medical interpretations and recommendations in the report originate from medical evidence and the evidence mapping table, thus reducing unfounded interpretations or citation illusions. In some embodiments, the health checkup analysis report can be a Markdown file; the target large model first generates structured data (such as JSON) and then renders it as a Markdown file.

[0099] In some embodiments, the step of performing content recognition on a medical examination report to obtain structured test indicator data includes: performing content recognition on the medical examination report to obtain a structured file; and parsing the structured file to obtain structured test indicator data. Here, the related operations are... Figure 1The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0100] In some embodiments, obtaining a structured file by performing content recognition on the medical examination report file includes: performing content recognition on the medical examination report file, and obtaining a structured file according to a preset priority rule in response to a successful content recognition callback. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0101] In some embodiments, parsing the structured file to obtain structured inspection index data includes: parsing at least one piece of content in the structured file using a parser corresponding to the content type of the content to obtain structured inspection index data. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0102] In some embodiments, the content type includes any of the following: a table, wherein the table includes any of the following: an HTML table, a Markdown pipeline table, a table with a header, and a table without a header; a text layer; and a project detail block. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0103] In some embodiments, the content type includes a table; wherein, for at least one piece of content in the structured file, parsing the content based on a parser corresponding to the content type to obtain structured inspection index data includes: for a table in the structured file, parsing the table based on a table parser, identifying multiple columns in the table, and obtaining structured inspection index data corresponding to the multiple columns, wherein the multiple columns include at least two of the following: item column, numerical column, anomaly marker column, qualitative result column, unit column, and reference range column. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0104] In some embodiments, obtaining the structured test indicator data corresponding to the plurality of columns includes: if the plurality of columns include at least two item columns, dividing the table into multiple column groups based on the position of the item columns in the table, and obtaining the structured test indicator data corresponding to each column group. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0105] In some embodiments, for tables in the structured file, parsing the tables using a table parser to identify multiple columns in the tables and obtaining structured inspection index data corresponding to the multiple columns includes: for tables in the structured file, performing semantic recovery on the tables using a table parser to identify multiple columns in the recovered tables and obtaining structured inspection index data corresponding to the multiple columns. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0106] In some embodiments, the device is further configured to: perform anomaly determination on the inspection index data regarding the multiple anomaly sources based on priority rules corresponding to multiple anomaly sources, and determine whether the inspection index data is an anomalous indicator item, wherein the multiple anomaly sources include at least two of explicit anomaly markers, qualitative reference sources, reference range sources, and context-dependent reference range sources. Here, related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0107] In some embodiments, the device is further configured to: perform anomaly consistency recalculation on the abnormal indicator item based on the actual indicator value, the lower limit of the reference range, and the upper limit of the reference range of the abnormal indicator item. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0108] In some embodiments, the device is further configured to: if the anomaly determination source corresponding to the anomaly indicator item is a context-dependent reference range source, determine that the anomaly indicator item does not need to undergo anomaly consistency recalculation. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0109] In some embodiments, the device is further configured to: perform fact-checking on the abnormal indicator items based on the structured file corresponding to the physical examination report file. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0110] In some embodiments, the device is further configured to: determine whether the abnormal indicator item needs manual review based on the extraction confidence level and / or abnormal confidence level corresponding to the abnormal indicator item. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0111] In some embodiments, the device is further configured to: determine the extraction confidence level corresponding to the abnormal indicator item based on whether the indicator name of the abnormal indicator item matches a preset indicator vocabulary. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0112] In some embodiments, the device is further configured to: determine the anomaly confidence level corresponding to the anomaly indicator item based on the plurality of anomaly sources. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0113] In some embodiments, writing the test index data and the multimodal observation data into an evidence mapping table includes: writing the test index data into the evidence mapping table; determining the multimodal confidence of the multimodal observation data based on the model confidence, source location confidence, cross-modal consistency, and clinical relevance corresponding to the multimodal observation data; if the multimodal confidence meets a preset condition, writing the multimodal observation data into the evidence mapping table, wherein the cross-modal consistency is used to characterize whether the multimodal observation data and the test index data support each other, and the clinical relevance is used to characterize the association strength between the multimodal observation data and the abnormal indicator item. Here, related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0114] In some embodiments, generating at least one combination of anomalous evidence corresponding to an organ system based on the evidence mapping table includes: determining the organ system corresponding to the anomalous indicator item and the organ system to which the multimodal observation data belongs, based on the anomalous indicator items and multimodal observation data in the evidence mapping table; grouping the anomalous indicator items and multimodal observation data under the same organ system to generate at least one combination of anomalous evidence corresponding to an organ system. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0115] In some embodiments, the combination of anomalous evidence is a combination of at least one anomalous indicator and at least one multimodal observation data; wherein, the device is further configured to: determine the retrieval priority corresponding to the combination of anomalous evidence based on the degree of support between the at least one anomalous indicator and the at least one multimodal observation data. Here, related operations and... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0116] In some embodiments, the device is further configured to: generate target reference numbers for the test indicator data, the multimodal observation data, and the retrieval text block, and add the target reference numbers to a reference mapping table; wherein, the device is further configured to: verify whether the reference numbers in the physical examination analysis report belong to the reference mapping table, and if not, perform a first processing operation on the physical examination analysis report. Here, the related operations are similar to... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0117] In some embodiments, the device is further configured to: determine, based on the reference number, conflict information between the test indicator data referenced in the physical examination analysis report and the multimodal observation data referenced in the physical examination analysis report; if the conflict information meets preset conditions, perform a second processing operation on the physical examination analysis report. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0118] In some embodiments, the device is further configured to: generate prompt words according to a preset clinical inference path; wherein, obtaining the physical examination analysis report output by the target large model based on the context information includes: obtaining the physical examination analysis report output by the target large model based on the context information and the prompt words. Here, the related operations are... Figure 1 The embodiments shown are the same or similar, so they will not be described again, but are included here by reference.

[0119] In some embodiments, the device is further configured to: in response to an editing operation performed by a target user on the physical examination analysis report, generate third query information based on the edited physical examination analysis report, perform a third-stage medical evidence retrieval based on the third query information to obtain a third retrieval result, wherein the third query information is used to enhance evidence around the auxiliary analysis direction specified by the target user; and obtain an updated physical examination analysis report output by the target large model based on the third retrieval result.

[0120] Figure 6 Exemplary systems that can be used to implement the various embodiments described in this application are shown; such as Figure 6 As shown in some embodiments, system 300 can function as any of the devices described in each of the embodiments. In some embodiments, system 300 may include one or more computer-readable media having instructions (e.g., system memory or NVM / storage device 320) and one or more processors (e.g., one or more processors 305) coupled to the one or more computer-readable media and configured to execute the instructions to implement the module and thus perform the actions described in this application.

[0121] In one embodiment, the system control module 310 may include any suitable interface controller to provide any suitable interface to at least one of the processors 305 and / or any suitable device or component communicating with the system control module 310.

[0122] The system control module 310 may include a memory controller module 330 to provide an interface to the system memory 315. The memory controller module 330 may be a hardware module, a software module, and / or a firmware module.

[0123] System memory 315 can be used, for example, to load and store data and / or instructions for system 300. In one embodiment, system memory 315 may include any suitable volatile memory, such as suitable DRAM. In some embodiments, system memory 315 may include double data rate type quad synchronous dynamic random access memory (DDR4 SDRAM).

[0124] In one embodiment, the system control module 310 may include one or more input / output (I / O) controllers to provide interfaces to the NVM / storage device 320 and (one or more) communication interfaces 325.

[0125] For example, NVM / storage device 320 may be used to store data and / or instructions. NVM / storage device 320 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable (one or more) non-volatile storage devices (e.g., one or more hard disk drive (HDD), one or more optical disc (CD) drives, and / or one or more digital universal optical disc (DVD) drives).

[0126] NVM / storage device 320 may include storage resources that are physically part of a device on which system 300 is mounted, or that can be accessed by the device without necessarily being part of it. For example, NVM / storage device 320 may be accessed via a network through one or more communication interfaces 325.

[0127] One or more communication interfaces 325 may provide the system 300 with an interface to communicate over one or more networks and / or with any other suitable device. The system 300 may wirelessly communicate with one or more components of a wireless network in accordance with any of one or more wireless network standards and / or protocols.

[0128] In one embodiment, at least one of the processors 305 may be logically packaged with one or more controllers of the system control module 310 (e.g., memory controller module 330). In one embodiment, at least one of the processors 305 may be logically packaged with one or more controllers of the system control module 310 to form a system-in-package (SiP). In one embodiment, at least one of the processors 305 may be integrated with the logic of one or more controllers of the system control module 310 on the same die. In one embodiment, at least one of the processors 305 may be integrated with the logic of one or more controllers of the system control module 310 on the same die to form a system-on-a-chip (SoC).

[0129] In various embodiments, system 300 may be, but is not limited to, a server, workstation, desktop computing device, or mobile computing device (e.g., laptop computing device, handheld computing device, tablet computer, netbook, etc.). In various embodiments, system 300 may have more or fewer components and / or different architectures. For example, in some embodiments, system 300 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including a touchscreen display), a non-volatile memory port, multiple antennas, a graphics chip, an application-specific integrated circuit (ASIC), and a speaker.

[0130] In addition to the methods and devices described in the above embodiments, this application also provides a computer-readable storage medium storing computer code that, when executed, performs the method described in any of the preceding embodiments.

[0131] This application also provides a computer program product that, when executed by a computer device, performs the method described in any of the preceding claims.

[0132] This application also provides a computer device, the computer device comprising:

[0133] One or more processors;

[0134] Memory, used to store one or more computer programs;

[0135] When the one or more computer programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method as described in any of the preceding methods.

[0136] It should be noted that this application can be implemented in software and / or a combination of software and hardware, for example, using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device. In one embodiment, the software program of this application can be executed by a processor to implement the steps or functions described above. Similarly, the software program of this application (including related data structures) can be stored in a computer-readable recording medium, such as RAM memory, a magnetic or optical drive, a floppy disk, or similar devices. Furthermore, some steps or functions of this application can be implemented in hardware, for example, as circuitry that cooperates with a processor to perform the various steps or functions.

[0137] Furthermore, a portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0138] Communication media include media through which communication signals containing, for example, computer-readable instructions, data structures, program modules, or other data are transmitted from one system to another. Communication media can include guided transmission media (such as cables and wires (e.g., optical fibers, coaxial cables, etc.)) and wireless (unguided transmission) media capable of propagating energy waves, such as sound, electromagnetic, RF, microwave, and infrared. Computer-readable instructions, data structures, program modules, or other data can be embodied as modulated data signals in, for example, wireless media (such as carrier waves or similar mechanisms embodied as part of spread spectrum technology). The term "modulated data signal" refers to a signal whose one or more characteristics are altered or set in a manner that encodes information in the signal. Modulation can be analog, digital, or a hybrid modulation technique.

[0139] By way of example and not limitation, computer-readable storage media may include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. For example, computer-readable storage media include, but are not limited to, volatile memories such as random access memory (RAM, DRAM, SRAM); and non-volatile memories such as flash memory, various read-only memories (ROM, PROM, EPROM, EEPROM), magnetic and ferromagnetic / ferroelectric memories (MRAM, FeRAM); and magnetic and optical storage devices (hard disks, magnetic tapes, CDs, DVDs); or other media now known or hereafter developed capable of storing computer-readable information / data for use by a computer system.

[0140] Herein, one embodiment of this application includes an apparatus comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the apparatus is triggered to run a method and / or technical solution based on the foregoing embodiments of this application.

[0141] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within this application. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in the apparatus claims may also be implemented by a single unit or device in software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any particular order.

Claims

1. A method for obtaining a physical examination analysis report, wherein, The method includes: Content recognition is performed on the physical examination report document to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items; Enhanced interpretation of the multimodal file corresponding to the physical examination report file is performed to obtain structured multimodal observation data, and the test index data and the multimodal observation data are written into the evidence mapping table; Based on the evidence mapping table, at least one abnormal evidence combination corresponding to an organ system is generated, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data. First query information is generated based on the combination of abnormal evidence. First search results are obtained by performing a first-stage medical evidence retrieval based on the first query information. Second query information is generated based on the first search results. Second search results are obtained by performing a second-stage medical evidence retrieval based on the second query information. The search text blocks matched by the first search results and the second search results are obtained. The first query information is used to discover candidate diseases and / or risk directions. The second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions. Contextual information is constructed based on the abnormal evidence combination, the first search result, and the second search result to obtain a physical examination analysis report output by the target large model based on the contextual information. The physical examination analysis report includes a reference number for the searched text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

2. The method according to claim 1, wherein, The process of obtaining structured test indicator data through content recognition of the physical examination report document includes: Content recognition is performed on the medical examination report to obtain a structured file; The structured file is parsed to obtain structured test index data.

3. The method according to claim 2, wherein, The process of obtaining a structured file by performing content recognition on the physical examination report includes: The system performs content recognition on the medical examination report file. Upon successful content recognition callback, a structured file is obtained according to preset priority rules.

4. The method according to claim 2, wherein, The process of parsing the structured file to obtain structured test index data includes: For at least one piece of content in the structured file, the content is parsed based on a parser corresponding to the content type to obtain structured test index data.

5. The method according to claim 4, wherein, The content type includes any of the following: The table includes any one of the following: HTML table, Markdown pipe table, table with header, and table without header; Text layer; Project details block.

6. The method according to claim 5, wherein, The content type includes tables; Specifically, the step of parsing at least one block of content in the structured file using a parser corresponding to the content type to obtain structured inspection index data includes: For the tables in the structured file, the tables are parsed using a table parser to identify multiple columns in the tables and obtain structured inspection index data corresponding to the multiple columns. The multiple columns include at least two of the following: item column, numerical column, anomaly marker column, qualitative result column, unit column, and reference range column.

7. The method according to claim 6, wherein, Obtaining the structured test indicator data corresponding to the multiple columns includes: If the multiple columns include at least two item columns, the table is divided into multiple column groups based on the position of the item columns in the table, and structured test index data corresponding to each column group is obtained.

8. The method according to claim 6, wherein, For the tables in the structured file, a table parser is used to parse the tables, identify multiple columns in the tables, and obtain structured inspection indicator data corresponding to the multiple columns, including: For the tables in the structured file, semantic recovery is performed on the tables based on the table parser to identify multiple columns in the recovered tables and obtain the structured test index data corresponding to the multiple columns.

9. The method according to claim 1, wherein, The method further includes: Based on priority rules corresponding to multiple anomaly sources, the test index data is anomaly determined with respect to the multiple anomaly sources to determine whether the test index data is an anomaly item. The multiple anomaly sources include at least two of the following: explicit anomaly markers, qualitative reference sources, reference range sources, and context-dependent reference range sources.

10. The method according to claim 9, wherein, The method further includes: Based on the actual index value, lower limit of the reference range, and upper limit of the reference range of the abnormal index item, an abnormal consistency recalculation is performed on the abnormal index item.

11. The method according to claim 10, wherein, The method further includes: If the anomaly determination source corresponding to the anomaly indicator item is a context-dependent reference range source, it is determined that the anomaly indicator item does not need to undergo anomaly consistency recalculation.

12. The method according to claim 10, wherein, The method further includes: The abnormal indicators were fact-checked based on the structured file corresponding to the medical examination report.

13. The method according to claim 1, wherein, The method further includes: Based on the extraction confidence level and / or anomaly confidence level corresponding to the abnormal indicator item, determine whether the abnormal indicator item needs to be manually reviewed.

14. The method according to claim 13, wherein, The method further includes: The extraction confidence level of the abnormal indicator item is determined based on whether the indicator name of the abnormal indicator item matches the preset indicator term list.

15. The method according to claim 13, wherein, The method further includes: Based on the multiple sources of anomalies, determine the anomaly confidence level corresponding to the anomaly indicator item.

16. The method according to claim 1, wherein, The step of writing the test index data and the multimodal observation data into the evidence mapping table includes: Write the test index data into the evidence mapping table; The multimodal confidence level of the multimodal observation data is determined based on the model confidence level, source location confidence level, cross-modal consistency, and clinical relevance corresponding to the multimodal observation data. If the multimodal confidence level meets the preset conditions, the multimodal observation data is written into the evidence mapping table. The cross-modal consistency is used to characterize whether the multimodal observation data and the test index data support each other, and the clinical relevance is used to characterize the correlation strength between the multimodal observation data and the abnormal indicator item.

17. The method according to claim 1, wherein, The generation of at least one combination of anomalous evidence corresponding to an organ system based on the evidence mapping table includes: For the abnormal indicator items and multimodal observation data in the evidence mapping table, determine the organ system corresponding to the abnormal indicator item and the organ system to which the multimodal observation data belongs; Abnormal indicators and multimodal observation data within the same organ system are grouped to generate at least one combination of abnormal evidence corresponding to an organ system.

18. The method according to claim 1, wherein, The combination of anomalous evidence is a combination of at least one anomalous indicator and at least one multimodal observation data. The method further includes: The retrieval priority corresponding to the combination of anomalous evidence is determined based on the degree of support between the at least one anomalous indicator and the at least one multimodal observation data.

19. The method according to claim 1, wherein, The method further includes: Generate target reference numbers for the test index data, the multimodal observation data, and the search text block, and add the target reference numbers to the reference mapping table; The method further includes: Verify whether the reference number in the physical examination analysis report belongs to the reference mapping table. If not, perform the first processing operation on the physical examination analysis report.

20. The method according to claim 1, wherein, The method further includes: Based on the reference number, determine the conflict information between the test indicator data referenced in the physical examination analysis report and the multimodal observation data referenced in the physical examination analysis report. If the conflict information meets the preset conditions, perform a second processing operation on the physical examination analysis report.

21. The method according to claim 1, wherein, The method further includes: Prompt words are generated based on a pre-defined clinical inference pathway; The process of obtaining the physical examination analysis report output by the target large model based on the context information includes: Obtain the physical examination analysis report output by the target large model based on the context information and the prompt words.

22. The method according to claim 1, wherein, The method further includes: In response to the editing operation performed by the target user on the physical examination analysis report, a third query information is generated based on the edited physical examination analysis report, and a third-stage medical evidence retrieval is performed based on the third query information to obtain a third search result, wherein the third query information is used to enhance evidence around the auxiliary analysis direction specified by the target user; Obtain the updated physical examination analysis report output by the target large model based on the third search result.

23. A system for obtaining a medical examination analysis report, wherein, The system includes: The first module is used to perform content recognition on the physical examination report file to obtain structured test indicator data, wherein the test indicator data includes normal indicator items and abnormal indicator items. The second module is used to enhance the interpretation of the multimodal file corresponding to the physical examination report file, obtain structured multimodal observation data, and write the test index data and the multimodal observation data into the evidence mapping table; The third module is used to generate at least one abnormal evidence combination corresponding to an organ system based on the evidence mapping table, wherein the abnormal evidence combination is a combination of at least two abnormal indicator items or a combination of at least one abnormal indicator item and at least one multimodal observation data. The fourth module is used to generate first query information based on the abnormal evidence combination, perform a first-stage medical evidence retrieval based on the first query information to obtain a first retrieval result, generate second query information based on the first retrieval result, perform a second-stage medical evidence retrieval based on the second query information to obtain a second retrieval result, and obtain the retrieval text block matched by the first retrieval result and the second retrieval result. The first query information is used to discover candidate diseases and / or risk directions, and the second query information is used to retrieve management recommendations and / or clinical evidence related to the candidate diseases and / or risk directions. The fifth module is used to construct context information based on the abnormal evidence combination, the first search result and the second search result, and obtain a physical examination analysis report output by the target large model based on the context information. The physical examination analysis report includes a reference number for the search text block, and the physical examination analysis report is associated with the test index data and the multimodal observation data through the evidence mapping table.

24. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 22.

25. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 22.

26. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 22.

Images