A medical voice broadcast case summary generation and illusion suppression method and system

Abstract

The application provides a medical voice broadcast medical record abstract generation and illusion suppression method and system, and relates to the technical field of medical information and natural language processing. The system comprises: a medical record input module for locking a unique target medical record document and extracting a target text; an evidence constraint abstract generation module for generating a structured candidate abstract with evidence items; a server-side evidence anchor verification module for verifying whether the evidence items actually exist in the target text; a safe convergence module for reviewing, revising and finalizing the candidate abstract; a high-risk interception module for identifying and intercepting high-risk medical inferences in the abstract that are not supported by the target text; a voice adaptation module for converting the abstract text that has passed the safety check into a format suitable for voice broadcast; and a conservative fallback module for performing conservative reading or refusing to output when the verification fails. The application can effectively suppress illusions and improve the reliability of medical voice broadcast.

Description

Technical Field

[0001] This invention belongs to the field of medical information and natural language processing technology, and in particular relates to a method and system for generating medical record summaries and suppressing hallucinations through medical voice broadcasting. Background Technology

[0002] In clinical systems, the most common use of medical record summaries is for healthcare professionals to read on a screen. Screen reading allows doctors to slowly review the context and also allows users to return to the original text for further confirmation. However, in scenarios involving medical robots or bedside voice terminals, medical record content is directly read aloud to doctors, nurses, or other healthcare professionals. Voice reading is characterized by its short duration, inability to quickly scan, difficulty in reviewing context, and susceptibility to numerous on-site distractions.

[0003] Therefore, medical voice broadcasting has at least the following technical requirements: The broadcast should be brief; the entire original medical record should not be read aloud.

[0004] The broadcast content should be clear and should not contain a large number of HTML tags, template fields, overly dense indicator strings, or symbols that are not suitable for reading aloud.

[0005] The content of the broadcast should be conservative; medical conclusions not explicitly stated in the original text should not be presented as facts.

[0006] The content of the broadcast should be traceable, and key conclusions should be able to be traced back to the original evidence in the current target document.

[0007] There should be a mechanism to resolve failures, and a complete medical summary should not be generated when the main text is empty, too short, or lacks sufficient evidence.

[0008] The data returned by hospital electronic medical record systems typically has multiple layers of external structure, making it difficult to use directly as clean plain text. Taking a common interface as an example, the outer layer may first return JSON status codes, message fields, and encrypted or wrapped fields; the inner layer may be XML or other business wrappers; and the actual medical record text is further embedded in an HTML document structure.

[0009] The HTML document structure may also contain the following: Chapter titles, such as ward round records, medical records, and treatment recommendations.

[0010] Paragraphs or tags with data attributes are used to identify chapters.

[0011] The main text area, such as the text within the tag-article-content tag.

[0012] Template placeholder fields, blank fields, style tags, or nested inline tags.

[0013] The outer structure shared by different medical record types.

[0014] Chapters or historical fragments that are irrelevant to the current summarizing task.

[0015] If the system directly feeds the entire API response or the entire HTML document into the summarization model, the summarization input may contain shell fields, template fragments, non-target chapters, or irrelevant document content. When faced with such mixed input, the model may automatically supplement the context in order to generate complete natural language, thus creating an unacceptable risk of hallucinations in medical scenarios.

[0016] In ordinary text summarization, a model might occasionally add background information or make minor generalizations, which may simply be a matter of imprecise expression. However, in medical record summarization, the same behavior could carry significant risks. For example, the original text might only state that a certain indicator is abnormal, but the model might generalize to the presence of a certain disease; the original text might only state "continue observation," but the model might generalize to "further treatment is needed"; the original text might only state that an infection indicator is elevated, but the model might generalize to "severe infection" or "sepsis." While these statements may seem linguistically reasonable, they should not be broadcast by the robot with a fixed tone unless directly supported by the original text.

[0017] Therefore, the security of medical summaries cannot rely solely on the "do not fabricate" prompt in the model. The system also needs to establish a local verification mechanism on the server side to check the evidence items, summary conclusions, and high-risk statements output by the model, ensuring that the summary results are supported by the current target text.

[0018] The usage of text for screen reading and text for audio reading also differ. Summary suitable for screen reading may contain more parentheses, symbols, long sentences, index strings, English abbreviations or units; when this content is read aloud directly, it may result in poor auditory experience, heavy comprehension burden, and unclear focus.

[0019] Medical robots need texts that are more suitable for short listening sessions, such as retaining clear symptoms, ward round opinions, treatment suggestions, and key concerns, reducing unnecessary formatting, making common units, percentages, temperatures, frequencies, etc. more user-friendly, and controlling the overall length. Summary of the Invention

[0020] Therefore, the technical problem to be solved by the present invention is to provide a method and system for generating medical record summaries and suppressing hallucinations in medical voice broadcasting, which can avoid bringing medical record interface shells, HTML template fragments, non-target document content or medical conclusions inferred by the model into the broadcast results during the medical record summary generation process.

[0021] In a first aspect, the present invention provides a medical record summary generation and hallucination suppression system for medical voice broadcasting, comprising: The medical record input module is used to locate a unique target medical record document from the medical record list and parse the medical record content interface to extract the target text. An evidence-constrained summary generation module, connected to the medical record input module, is used to generate a structured candidate summary based on the target text. The candidate summary includes summary text, evidence items, and extractable markers. The evidence items are original text evidence in the target text that supports the conclusion of the candidate summary. The server-side evidence anchor verification module is connected to the evidence constraint summary generation module and is used to verify on the server whether the evidence item actually exists in the target text. The security convergence module is connected to the server-side evidence anchor verification module and is used to perform review, revision and final review of the candidate abstract based on the evidence items. A high-risk interception module, connected to the security convergence module, is used to identify and intercept high-risk medical inferences in the summary text that are not explicitly supported by the target text. A voice adaptation module, connected to the high-risk interception module, is used to convert the summary text processed by the high-risk interception module into a format suitable for text-to-speech broadcasting; The conservative fallback module, connected to the voice adaptation module, is used to execute a conservative reading fallback strategy or refuse output when any of the server-side evidence anchor verification module, the security convergence module, or the high-risk interception module fails.

[0022] Furthermore, the medical record input module includes: The medical record list retrieval unit is used to retrieve a list of the patient's medical records. The target identification locking unit is used to determine a unique target medical record document from the list based on the document title, type, or business rules. The medical record shell parsing unit is used to parse the outer data structure returned by the medical record content interface and extract the inner medical record text. The target chapter region positioning unit is used to locate a preset target chapter region in the internal medical record text. The main text region extraction unit is used to extract the main text region text within the target chapter region located by the target chapter region positioning unit, and use it as the target main text. The text integrity judgment unit is used to determine whether the target text is empty, too short, or contains only template fragments. If so, the conservative fallback module is triggered.

[0023] Furthermore, the main text region extraction unit is configured to: firstly extract the text within the preset main text tags in the target chapter region as the main text region text; if the preset main text tags are not matched, then backtrack to extract the text of the entire target chapter region and perform cleaning and normalization processing.

[0024] Furthermore, the evidence constraint summary generation module includes: The candidate summary generation unit is used to generate a structured candidate summary based on the target text. The structured candidate summary also includes a rejection reason. Furthermore, when it is determined that the target text is insufficient to generate a safe summary text, the extractable flag is returned as false, and the rejection reason is filled in to prohibit the fabrication of summary text. An evidence item generation unit is used to provide evidence items for key conclusions in the candidate abstract, each evidence item containing at least a phrase from the original text of the target text and its context.

[0025] Furthermore, the server-side evidence anchor verification module includes an evidence anchor verification unit, which is configured to operate independently of the evidence constraint summary generation module and perform one or more of the following verifications: verifying whether the evidence item is empty or whether its number exceeds a preset limit; verifying whether the original phrase and context in each evidence item are empty; verifying whether the normalized original phrase is contained in the normalized context; verifying whether the original phrase and context can be precisely or fuzzily matched in the target text, respectively; if any verification fails, the candidate summary is determined to be unacceptable.

[0026] Furthermore, the secure convergence module includes: The peer review decision unit is used to output the peer review results: acceptance, revision, or rejection. The revision convergence unit is used to safely converge the candidate abstract by deleting unfounded content, reducing the intensity of expression, or rewriting it into expression directly supported by the target text, based on the reviewers' comments. It prohibits the addition of new diagnoses or suggestions not supported by the original text. The final verification unit is used to re-verify the revised candidate digest back to the server-side evidence anchor verification module, and then perform final security verification.

[0027] Furthermore, the high-risk interception module includes a high-risk inference interception unit. When matching high-risk medical inference words in the abstract text with the target text, the high-risk inference interception unit uses a dynamic similarity threshold, the calculation formula of which is:

[0028] in: ith This is a dynamic similarity threshold, with a value ranging from 0 to 1; i base The preset baseline similarity threshold; C model The confidence score output by the evidence constraint summary generation module for the current candidate summary ranges from 0 to 1. E strength The strength of evidence upon which the high-risk medical inference term is based is determined by the length of the original phrase in the evidence item and its uniqueness in the original text, with a value ranging from 0 to 1; L norm The normalized value of the target text length. L norm =min(1, L doc / L ref ), L doc The target text length, L ref This is a reference length threshold. R term The preset risk level for currently high-risk medical inference terms, with a value range of 0 to 1; l 1. l 2. l 3. l 4 represents the influence coefficient of each factor, with a value ranging from 0 to 1, and satisfying the following conditions: l 1+ l 2+ l 3+ l 4 = 1; When the normalized similarity between a high-risk medical inference term and any segment in the target text is greater than or equal to 1, i th If the corresponding medical statement is explicitly supported in the target text, it is determined that the original text supports it; otherwise, it is determined to be a high-risk medical inference not supported by the original text, and the candidate abstract is rejected. Determining that the corresponding medical statement is explicitly supported in the target text includes: verifying negative words, suspected words, degree words, and consistency with contextual polarity.

[0029] Furthermore, the high-risk inference interception unit maintains a preset set of high-risk medical inference words, which are used to identify high-risk medical inferences that may exist in the summary text. Moreover, the identification and interception are based on matching after text normalization. If the summary text contains any inference word in the set of high-risk medical inference words, and the medical expression corresponding to the inference word is not explicitly supported in the normalized target text, it is determined to be a high-risk medical inference not supported by the original text and the candidate summary is rejected.

[0030] Furthermore, the conservative fallback module includes: The conservative reading catch-all unit is used to implement a conservative reading catch-all strategy, including: extracting only clearly existing objective descriptions of the illness, ward round opinions, or treatment opinions from the target text; and prohibiting any form of diagnostic upgrade, intentional completion, or medical inference of the information in the target text. The output rejection unit, connected to the conservative reading catch-all unit, is used to reject output when the conservative reading text processed by the conservative reading catch-all unit is empty, too short, or still contains high-risk expressions not supported by the target text.

[0031] Secondly, the present invention also provides a method for generating medical record summaries and suppressing hallucinations using medical voice broadcasting, comprising: Medical record input steps: Locate a unique target medical record document from the medical record list, parse the outer data structure returned by the medical record content interface to obtain the inner medical record text, locate the preset target chapter area in the inner medical record text, and extract the main text area of ​​the target chapter area as the target text. Evidence-constrained summary generation steps: Input the target text into the evidence-constrained summary generation module to generate a structured candidate summary; the candidate summary includes at least the summary text, a set of evidence items, and extractable markers; wherein, each evidence item in the set of evidence items contains at least a phrase from the original text and its context, and the evidence item is original text evidence in the target text that supports the conclusion of the candidate summary; Server-side evidence anchor verification steps: On the server side, each evidence item in the candidate summary is compared and verified with the target text to verify whether the original phrase and its context actually exist in the target text; The safe convergence process involves sequentially reviewing, revising, and confirming the candidate abstracts that pass the evidence anchor point verification. High-risk interception step: Scan the summary text that has passed the final review for high-risk medical inference words, and compare the scanned high-risk medical inference words with the target text. If the medical expression corresponding to the high-risk medical inference word is not clearly supported in the target text, it is determined that there is a high-risk medical inference that is not supported by the original text and the candidate summary is rejected. Voice adaptation step: The summary text processed by the high-risk interception step is subjected to text-to-speech adaptation processing to generate a playable summary; Conservative fallback step: If any of the steps in the medical record input step, the evidence constraint summary generation step, the server-side evidence anchor verification step, the safe convergence step, or the high-risk interception step fails, then the conservative reading fallback strategy is executed or the output is directly rejected.

[0032] Beneficial effects: (1) Single target document targeting makes the abstract object more specific. This invention first identifies a unique recordId from a list of medical records, and then extracts the main text and generates a summary based on that document. Since subsequent processes do not involve concatenating multiple documents, a one-to-one correspondence is established between the summary content and the target document.

[0033] Therefore, when the robot reads the summary, the system can indicate which medical record the summary comes from, reducing the risk of contamination across documents, time points, and chapters.

[0034] (2) Shell parsing and HTML structure extraction make the model input cleaner. This invention first parses the outer shell, then locates the target chapter area and the main text area, avoiding sending the entire content returned by the interface into the model. This reduces the amount of status fields, template fragments, style tags, and irrelevant chapters entering the summary process.

[0035] Because the input text is more focused, candidate summaries are more likely to be generated around the actual ward round text, and subsequent evidence verification is also more stable.

[0036] (3) The candidate abstract carries evidence items to make the results traceable. This invention requires the candidate summary to output `evidence_items`, and each evidence item must include the original text phrase and its context. Thus, the model must provide supporting textual evidence alongside its conclusions.

[0037] As a result, the system has shifted from generating a single summary to generating a summary with a chain of evidence, thus improving the interpretability and traceability of the summary.

[0038] (4) Server-side evidence anchor point verification can block the output of correctly formatted but false evidence. The model may return seemingly reasonable quotes and contexts, but these may not actually exist in the current text. This invention performs local anchor point validation on the server side, checking each piece of evidence for emptiness, exceeding limits, originating from the current text, and whether the quote is located within the context.

[0039] Therefore, even if the model output format is correct, the system will not accept the summary if the evidence cannot be retrieved from the current target text. This process establishes a second safety barrier after the model output.

[0040] (5) Peer review, revision and final review ensure that the abstract converges within a safe range. This invention adds peer review, revision, and final review after the candidate abstract, making abstract generation a multi-stage convergence process. Peer review can identify whether candidate abstracts need to be reduced or downgraded, revision can change unreliable expressions back to more accurate wording, and the final review confirms whether the revised results are usable.

[0041] Therefore, the system can safely converge the abstract without adding information outside the original text, thus improving usability.

[0042] (6) High-risk medical inference interception can suppress diagnostic escalation hallucinations This invention intercepts high-risk medical terms such as respiratory failure, sepsis, and severe infection, and determines whether these terms have clear evidence in the current target text. If a term appears in the abstract but not in the original text, the system rejects it.

[0043] Therefore, even if the model generates seemingly reasonable medical summaries based on abnormal indicators, these summaries should not be directly used to broadcast the results. This reduces the risk of diagnostic escalation and disease inference, which are the most sensitive aspects of medical voice broadcasting.

[0044] (7) Conservative reading and refusal to output provide a clear ending to the failure scenario. In cases of insufficient text, insufficient evidence, peer rejection, final review failure, or high-risk interception, this invention will not forcibly generate an abstract in order to return a result; instead, the system will enter a conservative reading mode or refuse to output.

[0045] Therefore, the system explicitly handles insufficient evidence as a safe failure or conservative output, avoiding the misrepresentation of a successful summary. For medical voice broadcasts, this default rejection mechanism can improve clinical acceptability.

[0046] (8) Voice broadcast adaptation makes the output more suitable for robot on-site reading. After passing security verification, this invention performs length compression, sentence structure regularization, symbol unit conversion, and colloquialization processing to make the output text suitable for short-term robot broadcasting.

[0047] Therefore, medical staff can hear the key points more quickly, reducing the comprehension burden caused by long sentences, complex formats, or stacked indicators. Attached Figure Description

[0048] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0049] Figure 1 This is a typical flowchart of existing technology; Figure 2-1 This is a schematic diagram of the input forensics and evidence constraint summary layer in Embodiment 1 of the present invention; Figure 2-2 This is a schematic diagram of the security termination and broadcast output layer of Embodiment 1 of the present invention; Figure 3-1 This is a flowchart illustrating the process of obtaining the target text in Embodiment 2 of the present invention. Figure 3-2 This is a flowchart of the evidence constraint summary of Embodiment 2 of the present invention; Figure 3-3 This is a flowchart of the security review and broadcast output process in Embodiment 2 of the present invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The principles and features of the present invention are described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other. The embodiments given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0051] Existing medical record summary solutions primarily focus on the quality of summary generation. In medical voice broadcasting scenarios, it is also necessary to simultaneously confirm whether the summary source is limited to the current target medical record, whether key conclusions can be traced back to the original evidence, and whether the broadcast content avoids including diagnoses, changes in condition, or the intent of medical orders not stated in the original text.

[0052] The first category of existing technologies is based on training a medical record summarization generation model using medical record datasets, prompt words, or evaluation models. For example, CN120354914A, "Method, Apparatus, and Computer Equipment for Automatic Medical Record Summarization Model Generation," discloses a scheme for training an automatic medical record summarization model using a medical record dataset, optimal summary prompt words, optimal evaluation prompt words, and candidate models. This scheme focuses on improving the generation and evaluation performance of the automatic medical record summarization model. CN118606717A, "Training Method, Method, and Apparatus for Medical Record Summarization Generation Model," discloses a scheme for training a model using medical record samples and prompt text, and optimizing the quality of generated medical record summaries through a prompt standard library or iterative mechanism.

[0053] The typical process of the above scheme can be summarized as follows: 1) Collect medical record samples or medical record datasets.

[0054] 2) Construct a standard library of summary prompts, evaluation prompts, or prompts.

[0055] 3) Train or optimize the medical record summary generation model.

[0056] 4) Input the medical records to be summarized into the model.

[0057] 5) Output the medical record summary text.

[0058] These approaches focus on obtaining better medical record summarization models. Their value lies in model training, suggestion optimization, and improved summary quality. However, they typically do not emphasize the boundaries of individual target documents, nor do they require the model to output the original quotes and contexts corresponding to each key conclusion. They also usually do not treat server-side evidence anchor verification, high-risk medical inference interception, and voice broadcasting as a fallback mechanism as an independent security link.

[0059] The second type of existing technology involves organizing the structured and unstructured content in electronic medical records and presenting it in the form of summaries or templates.

[0060] For example, US11581070B2, "Electronic medical record summary and presentation," addresses the summarization and presentation of information in electronic medical records, enabling users to view medical records through a summary view.

[0061] US9805160B2, titled "Automatic generation of an executive summary for a medical event in an electronic medical record," discloses a scheme for generating an executive summary from a collection of electronic medical record documents for a medical event.

[0062] The typical process of this type of solution can be summarized as follows: 1) Obtain structured information, unstructured text, or a collection of documents related to medical events from electronic medical records.

[0063] 2) Extract or organize summary information from it.

[0064] 3) Display according to templates, views, or event summaries.

[0065] 4) For medical staff to read or use in subsequent work.

[0066] These solutions have addressed the presentation of summaries of electronic medical record information, but they are primarily geared towards screen displays or summaries of medical record events, and are not primarily constrained by voice broadcasting by medical robots. They typically do not design dedicated processing mechanisms for issues such as whether the model upgrades weak statements to high-risk diagnoses, whether evidence items come from the current target text, and whether it is better not to broadcast after a strict summary fails.

[0067] The third category of existing technologies involves schemes that utilize large language models or intelligent models to assist in the generation of electronic medical records. For example, CN119763752A, "An Electronic Medical Record Generation Method, Apparatus and Electronic Equipment", discloses a scheme for generating electronic medical records based on disease summary data, sample medical records, and prompts from a large language model.

[0068] The main process of this type of solution typically includes: 1) Obtain a summary of the patient's condition, medical records, sample medical records, or other contextual information.

[0069] 2) Construct prompts or generate tasks.

[0070] 3) Call the model to generate a complete medical record or a fragment of a medical record.

[0071] 4) Display, save, or edit the generated results.

[0072] Both this type of solution and the present invention may use language models, but their technical purposes differ. The medical record generation solution emphasizes assisting in writing or generating new medical record text; the present invention emphasizes safely extracting a reportable summary from existing target medical records and prohibits the output of content not supported by the original text. In other words, the present invention constrains the model to perform short summaries and reading aloud based on existing evidence in the current document, rather than using it to generate a more complete medical record.

[0073] The fourth category of prior art relates to voice control solutions for medical devices or telemedicine devices. For example, WO2024178615A1, "Control Method, Device and System for Voice-Assisted Telemedicine Devices," discloses a solution for controlling telemedicine devices or interacting with telemedicine systems via voice assistance. This type of solution illustrates the need for voice interaction in medical scenarios, but its focus is usually on voice command recognition, remote device control, or human-computer interaction processes, rather than on structured extraction of medical records, evidence-constrained summarization, and hallucination suppression before voice broadcast.

[0074] In summary, the closest approach to existing technologies can be summarized as follows: 1) Obtain medical record text, medical record dataset, or collection of electronic medical record documents.

[0075] 2) Generate medical record summaries using rules, templates, summary models, or large language models.

[0076] 3) Present the summary results to medical staff, or use them as the result of subsequent system processing.

[0077] 4) Some solutions can improve the quality of summaries by using prompt words, evaluation models, or training samples.

[0078] This technical approach can solve the problem of condensing long medical records into shorter texts, but it typically lacks a secure consolidation mechanism for medical voice broadcasting. In particular, when the medical record interface content is complex, the boundaries of the target document are unclear, the main text fragments are insufficient, or the model outputs high-risk medical inferences, existing solutions lack a step-by-step blocking mechanism from the input boundary to the output broadcast.

[0079] Figure 1 The diagram illustrates a typical workflow that most closely resembles existing technologies. This workflow takes medical record text or datasets as input, generates summaries through a summarization model, and then directly outputs the summaries for display or broadcast. The system primarily relies on the model's summarization capabilities and lacks single-document boundary control, server-side evidence anchor verification, high-risk medical inference interception, and conservative fallback mechanisms.

[0080] Figure 1 Once the summary model generates the results, it can proceed to the display or broadcast stage. However, if the input contains non-target documents, template fragments, or incomplete segments, or if the model supplements the original text with diagnoses and suggestions to make the summary more complete, the system lacks a safety gate for further verification and output rejection.

[0081] The above technical solutions have the following problems: (1) Uncontrolled input boundaries can lead to cross-document or cross-chapter contamination. Existing medical record summarization solutions typically assume that the input is already a summarizable medical record. However, in real-world medical record interfaces, the outer layer may contain JSON, XML, and gateway wrappers, while the inner layer may contain HTML document structures. Furthermore, multiple medical records may exist for the same patient. If the system does not first identify a unique target recordId and does not distinguish between the target chapter and the main text area, it may send non-target documents, non-target chapters, template fragments, or irrelevant fields into the summarization model.

[0082] If the input boundaries are unclear, even if the summary generated by the model is fluent, it will be difficult to determine which medical record it corresponds to. As a result, the summary content may be inconsistent with the current broadcast task, or information from other time points or other chapters may be mistakenly identified as the focus of the current ward round.

[0083] (2) Abstract quality optimization cannot replace the constraint of retrospective evidence from the original text. Medical record summary model training schemes typically improve summary quality through prompts, evaluation prompts, sample training, or metric optimization. This approach helps generate more complete and natural summaries, but its evaluation focus is usually on whether the summary is fluent and covers the key points, without making the inclusion of explicit original text evidence for each key conclusion a hard constraint.

[0084] In medical voice broadcasting scenarios, if the summarized content cannot be traced back to the original text evidence, it becomes impossible to determine whether the robot is broadcasting explicit content from the original text or content supplemented by the model based on medical common sense or context. This increases the risk of unfounded conclusions being broadcast.

[0085] (3) The lack of local server-side validation will amplify the risk of model output distortion. Existing solutions often directly present the summarization model output as the displayable result. Even if the prompts instruct the model not to fabricate, it may still output content that is correctly formatted but lacks factual evidence. For example, the model might provide a seemingly reasonable phrase of evidence, but that phrase does not exist in the current text; or the quote does not match the context; or the context is content concatenated by the model.

[0086] If the system does not perform a server-side check to ensure that the quote is within the context, that the context originates from the current text, and that the number of evidence items is reasonable, then distorted model output will continue to be displayed or broadcast. Therefore, relying solely on model self-constraint is insufficient to meet the security requirements of medical voice broadcasting.

[0087] (4) When there is a lack of interception of high-risk medical inferences, weak statements are easily escalated. Medical records often contain abnormal indicators, suspected diagnoses, observations, or items requiring further investigation. To make the content more concise, summarization models may generalize these weaker statements into stronger conclusions. For example, elevated infection indicators might be summarized as severe infection, decreased oxygenation as respiratory failure, and low albumin as hypoalbuminemia.

[0088] These generalizations may alter clinical meaning from a medical semantics perspective. Without a high-risk medical inference interception mechanism, the system cannot block escalating statements not explicitly supported by the original text before they are broadcast, thus posing a risk of clinical misrepresentation.

[0089] (5) Failure scenarios lack a conservative conclusion, which may package insufficient evidence as a successful summary. In actual API calls, the medical record text may be empty, too short, contain only template fragments, or the target chapter may not be matched. If the system still hands these fragments over to the model for processing, the model may generate a seemingly complete medical record summary to meet the task requirements. From the API's perspective, the summary process returns a success; however, from a medical evidence perspective, the summary lacks sufficient supporting text.

[0090] Therefore, existing solutions, lacking a default rejection strategy, may incorrectly convert insufficient evidence into successful summary generation. In medical voice broadcasting scenarios, this is more dangerous than a typical summary failure.

[0091] (6) Abstract texts designed for reading are difficult to directly adapt to live robot reading. Existing electronic medical record summary display solutions are mostly screen-oriented and typically do not focus on controlling the length of the reading, the pronunciation of numerical units, the density of indicator strings, and the use of conversational language. If reading-oriented summaries are directly handed over to TTS (Text-to-Speech) systems, problems such as excessively long sentences, awkward unit pronunciation, scattered key points, and high listening costs for doctors may occur.

[0092] Therefore, even if an existing summary appears readable on screen, it may not be suitable for a robot to broadcast live in the ward. Medical voice broadcasting requires specialized voice-to-text adaptation after the security summary has been approved.

[0093] To address the shortcomings of existing technologies, the basic solution of this invention is as follows: First, a unique target medical record document is identified from the medical record list, and the corresponding recordId is determined. Then, the outer shell and document structure returned by the medical record content interface are parsed, prioritizing the location of target chapters such as ward round records, and the main text area is extracted. Next, only based on this main text, candidate summaries with original text evidence items are generated. These candidate summaries must simultaneously provide summary, can_extract, reject_reason, and evidence_items. The system performs local anchor point verification on the evidence items on the server side, checking the correspondence between quote, context, and the current main text. Subsequently, the summary is converged through a review, revision, and final review process, and combined with high-risk medical inference terms or rules, content such as diagnostic upgrades, disease inferences, and medical order intent completion not explicitly supported by the original text is blocked. Finally, only the results that pass the verification are adapted for voice broadcasting and their length is controlled. If the main text is insufficient, the evidence is insufficient, or the strict summary fails multiple times, the system enters a conservative reading fallback or refuses output to prevent unreliable content from entering the clinical voice broadcast.

[0094] The key features of this invention are a combination of methods including single-document boundary locking, structured extraction of target text, generation of evidence-constrained summaries, server-side evidence anchor point verification, interception of high-risk medical inferences, final review and revision, voice broadcast adaptation, and a conservative fallback approach. The specific implementation method can be adjusted based on which large model, which prompt word, or which deployment method is used.

[0095] The core idea of ​​this invention is to break down the medical record audio broadcast summary into multiple mutually constrained processing steps, avoiding the direct use of the generated result after a one-time model call. The system first determines which document to process, then determines which section of the document's text is usable, then requires candidate abstracts to provide traceable evidence, which is verified locally on the server side, followed by peer review, revision, final review, and high-risk interception, before finally generating a short text suitable for audio broadcast.

[0096] This invention includes the following processing chain: 1) The medical record list is matched and the target recordId is locked.

[0097] 2) Parsing the medical record interface shell and obtaining the HTML document body.

[0098] 3) Target chapter location and main text area extraction.

[0099] 4) Judgment of the completeness and summarability of the main text.

[0100] 5) Generation of candidate summaries with supporting evidence.

[0101] 6) Server-side evidence anchor point verification.

[0102] 7) Review, revision and final review process.

[0103] 8) Interception of high-risk medical inferences.

[0104] 9) Voice broadcast text adaptation.

[0105] 10) If a strict summary fails, a conservative reading as a fallback or a refusal to output the content will be implemented.

[0106] The above steps together form a closed loop of single-document evidence collection, evidence constraint summary, security review, voice adaptation, and failure resolution.

[0107] The following is an explanation of the relevant terms in this invention: Medical records refer to individual documents in a hospital's electronic medical record system, such as initial ward round records, progress notes, consultation records, discharge records, etc.

[0108] Target medical record document: refers to the unique medical record document that is currently allowed to enter the summary process, based on the medical record list, document type, title, time, or business rules.

[0109] recordId: A record identifier used to uniquely identify a single medical record document. In this invention, subsequent text extraction, summary generation, and broadcast adaptation are all based on documents corresponding to the same recordId.

[0110] Medical record interface shell: refers to the outer JSON, XML, gateway fields, encrypted fields, or business status fields returned by the medical record content interface. This shell is not part of the medical record body that can be directly summarized.

[0111] HTML document structure: HTML is an abbreviation for HyperText Markup Language. HTML document structure refers to the content of an electronic medical record organized using HTML or HTML-like tags, which may include paragraphs, chapters, tags, styles, template fields, and text areas.

[0112] The section of the ward round record refers to the section in the HTML document structure that corresponds to the ward round content. It can be located by section name, data attribute, tag path, or text matching rules. In this article, "section" represents a locatable section in the document.

[0113] Main text area: This refers to the core text area extracted from the target chapter of the target medical record document that can be used for summary processing. For example, in engineering implementation, the text corresponding to tag-article-content can be extracted first.

[0114] Candidate abstracts: refer to the abstracts initially generated by the system based on the target text, which have not yet passed evidence verification, peer review, final review, and high-risk interception.

[0115] Evidence item: Refers to the original text information unit used to support the conclusions of the summary. Preferably, each evidence item contains at least the original text phrase "quote" and the original text sentence or short paragraph containing that phrase, which serves as context.

[0116] Quote: The original phrases in the evidence section should come directly from the current target text and cannot be rewritten by the model or generated out of thin air.

[0117] Context: The complete sentence or short paragraph of the original text containing the quote in the evidence section should be able to be found in the current target text.

[0118] Evidence anchor point verification: This refers to the server performing local verification on the correspondence between the quote, context, and the current text, including whether the evidence is empty, whether the quantity exceeds the limit, whether the quote is located in the context, and whether both the quote and the context can be found in the current text.

[0119] Hallucination: In this invention, it refers to the model or summary module outputting medical judgments, diagnostic upgrades, changes in condition, treatment suggestions, or medical orders that are not explicitly supported by the current target text.

[0120] High-risk medical inferences: These are medical statements that, if deviated from the original evidence, could affect clinical judgment. Examples include respiratory failure, sepsis, severe infection, hypoalbuminemia, heart failure, and renal failure. The above examples are for illustrative purposes only and do not constitute a specific terminology.

[0121] Peer review: This refers to the process of determining whether a candidate abstract is directly supported by the original text, whether it needs revision, or whether it should be rejected after the candidate abstract is generated.

[0122] Revision: This refers to the process of making the candidate abstract more secure by reducing, downgrading, or rewriting it, and then making it more relevant to the original text evidence.

[0123] Final review: This refers to the stage where the revised abstract is reassessed to determine its acceptability.

[0124] Voice broadcast adaptation: This refers to further compressing, standardizing, and making the security-verified summary more conversational, so that it is suitable for robots or voice terminals to read aloud on site.

[0125] Conservative reading fallback: When strict summarization fails, the system abandons generalized diagnoses or inferential conclusions and only compiles information that is clearly present in the original text; if the compiled results are still not safe, the system refuses to output the information.

[0126] The default rejection policy (fail-closed) refers to the system's default rejection of output or entry into conservative processing when the input is insufficient, the evidence is insufficient, or the summary is unreliable, in order to avoid generating speculative content in order to return results.

[0127] Large Language Model (LLM): A language model capable of generating text output from text input. This invention can use large language models for extraction, peer review, revision, or reading aloud and transcription, but does not limit the model vendor, model name, or deployment method.

[0128] JSON: JavaScript Object Notation, is a lightweight data interchange format. This invention may use JSON to express candidate abstracts, evidence items, and peer review results, but it is not required to use this format.

[0129] TTS (Text-to-Speech) is a technology that converts text content into spoken audio. The speech adaptation in this invention can be used for text editing before TTS.

[0130] XML: Extensible Markup Language. In this invention, XML can be used as one of the wrappers or intermediate organizational forms for the content returned by the medical record interface.

[0131] DOM: DocumentObjectModel. The system can locate target chapter areas and main text areas using DOM selectors, tag attributes, or node paths.

[0132] Example 1

[0133] This embodiment is a medical record summary generation and hallucination suppression system for medical voice broadcasting. Figure 2-1 and Figure 2-2 The system structure of the present invention is illustrated. The system block diagram is divided into two parts: the input evidence collection and evidence constraint summary layer and the security conclusion and broadcast output layer. This structure starts from the access of medical record data, and proceeds through target document locking, structured text extraction, candidate summary generation, evidence verification, security review, and voice broadcast adaptation, finally outputting a broadcastable summary or a conservative fallback result.

[0134] The medical record input and boundary layer is used to ensure the uniqueness of the summary object; the text structuring layer is used to ensure that the model input is the target text and exclude interface shells or template fragments; the evidence constraint summary layer is used to establish the correspondence between the summary and the original text; the security closing layer is used to review, revise, finalize, and intercept high-risk medical inferences for candidate summaries; and the broadcast output layer is used to generate the final broadcastable text or to conservatively close the text when it is unsafe.

[0135] The functions of each module are described below: 1. Medical Record List Acquisition Unit The medical record list acquisition unit is used to retrieve a list of relevant patient medical records from the electronic medical record system or hospital interface. This unit is responsible for providing a candidate range for subsequent target document selection and does not directly generate a summary itself.

[0136] 2. Target recordId locking unit The target recordId locking unit is used to identify a unique target medical record document from the list of medical records based on the document title, document type, time, business rules, or user intent. Once the target document is locked, all subsequent text extraction, candidate summary, evidence verification, and broadcast adaptation will only be performed on the document corresponding to that recordId.

[0137] The purpose of this unit is to establish single-document boundaries and prevent the system from mixing summaries of different documents, different time points, or different chapters.

[0138] 3. Medical Record Packaging Parsing Unit The medical record shell parsing unit is used to parse the outer structure returned by the medical record content interface. For example, the outer layer may contain status codes, message fields, encrypted fields, or business data fields, while the inner layer may contain XML or HTML content. This unit extracts the actual HTML content of the medical record document from the interface shell and retains the original text for downstream structured extraction.

[0139] 4. Target Chapter Area Location Unit The target chapter region positioning unit is used to locate the chapter relevant to the current task within the HTML document structure. Preferably, the system prioritizes locating the "Ward Rounds Record" chapter; in other implementations, it can also locate target chapters such as consultation opinions, medical records, and treatment opinions.

[0140] This unit can be implemented using DOM selectors, tag attributes, chapter names, regular expression matching, or a combination of these methods. If the DOM path is not matched, alternative rules can be used for extraction; if it is still not matched, the process will proceed to rejection or conservative handling.

[0141] 5. Main Text Region Extraction Unit The main text region extraction unit is used to further extract summable main text within the target chapter region. Preferably, the system prioritizes extracting text from main text tags such as tag-article-content; if the tag does not exist, it can return to the main text of the entire chapter region and perform processing such as tag removal, whitespace removal, chapter name removal, and normalization.

[0142] The purpose of this unit is to compress the summary input from a complex HTML structure into readable, verifiable, and source-clear text.

[0143] 6. Text integrity judgment unit The text integrity judgment unit is used to determine whether the current text is suitable for entering the strict abstract process. The judgment content may include: whether the text is empty, whether the length is too short, whether it only contains template fragments, whether key paragraphs are missing, whether there is obvious truncation, and whether there is only an abnormality in the lack of context indicators, etc.

[0144] If the main text is empty, too short, or consists only of template fragments, the unit stops the subsequent rigorous summarization process and triggers a rejection output; if the main text exists but lacks clear medical instructions or key conclusions, a fixed security warning or a conservative fallback can be triggered. This reflects the security design of the default rejection policy.

[0145] 7. Candidate Abstract Generation Unit The candidate summary generation unit is used to generate structured candidate summaries based on the current target text. Unlike regular summaries, this unit requires candidate results to contain parsable fields, preferably including: The `can_extract` option indicates whether a summary can be safely extracted from the current text.

[0146] `reject_reason` is used to explain why it cannot be safely extracted.

[0147] The summary is used to carry candidate summary text.

[0148] evidence_items is used to carry the original text evidence items that support the summary.

[0149] The candidate summary generation unit can be implemented using a large language model, an extractive model, a rule template, or a hybrid approach, but its output must be included in the subsequent evidence verification and security check process and cannot be directly used as the final broadcast result.

[0150] 8. Evidence Item Generation Unit The evidence item generation unit is used to provide original textual evidence for key conclusions in the candidate abstract. Preferably, each set of evidence items includes a quote and a context, where the quote is an original text phrase and the context is a complete sentence or short paragraph from the original text containing that phrase.

[0151] The number of evidence items can be capped. For example, in one implementation, the number of evidence items is limited to no more than three groups to reduce evidence noise and force candidate summaries to focus on truly critical content.

[0152] 9. Evidence Anchor Point Verification Unit The evidence anchor verification unit executes locally on the server and independently verifies the evidence returned by the model. This unit can perform the following verifications: 1) Is evidence_items empty?

[0153] 2) Whether the number of evidence items exceeds the preset limit.

[0154] 3) Whether the quote and context in each piece of evidence are empty.

[0155] 4) Does the context contain ellipsis? If it does, the original sentence or short paragraph should be provided.

[0156] 5) Can the quote be found in the current target text?

[0157] 6) Can the context be found in the current target text?

[0158] 7) Whether the normalized quote is included in the context.

[0159] 8) Whether the autocomplete or compatibility evidence is too vague.

[0160] If the validation fails, the candidate digest is not accepted. This unit is an important security gate that distinguishes this invention from ordinary model digests.

[0161] 10. Peer review assessment unit The peer review decision unit is used to determine whether a candidate abstract is acceptable, requires revision, or should be rejected. Ideally, the review output should include three decision categories: accept, revise, and reject.

[0162] When a candidate abstract is substantially safe and supported by the original text, the review unit may return it for acceptance; when a candidate abstract has issues that can be fixed by deletion, downgrading, or rewriting, it should be returned for revision; when the original text is insufficient to support a safe abstract, it should be returned for rejection.

[0163] 11. Revise the convergence unit The revision convergence unit is used to converge candidate abstracts based on peer review comments. The preferred revision methods are deleting unsubstantiated content, reducing the intensity of expression, rewriting to a more accurate representation of the original text, or rebinding evidence items. No new diagnoses, expanded treatment recommendations, or supplementary external medical knowledge should be added during the revision process.

[0164] The revised summary still needs to be re-verified as evidence and confirmed by the final review.

[0165] 12. Final Review Confirmation Unit The final review unit is used to reassess whether the revised abstract is safe and usable. If the final review still does not accept it, the process will not proceed to voice broadcast adaptation, but will instead proceed to conservative reading or refuse output.

[0166] 13. High-risk inference interception unit The high-risk inference interception unit is used to identify high-risk medical expressions that are not explicitly stated in the original text of the abstract. This unit can maintain a high-risk vocabulary or high-risk categories, such as respiratory failure, sepsis, severe infection, hypoproteinemia, heart failure, and renal failure.

[0167] In one alternative implementation, the system performs a normalized comparison between the summary text and the current target text: if a high-risk word exists in the summary but not in the current target text, the summary is considered to contain an escalating medical inference not supported by the original text, triggering rejection or a fallback.

[0168] In one optional implementation, when the high-risk inference interception unit matches high-risk medical inference words in the summary text with the target text, it uses a dynamic similarity threshold. i th Instead of a fixed threshold, a different threshold is used. This is because different models output different confidence levels, different evidence items have different strengths, different target texts have different lengths, and different high-risk medical inference terms have different risk levels. Using a uniform threshold makes it difficult to balance the accuracy and recall of interception.

[0169] Dynamic similarity threshold i th The calculation formula is as follows:

[0170] in: i th The value range of is [0,1]. After calculation using the above formula, and then using min(1,max(0, i th )) to impose restrictions.

[0171] i base The baseline similarity threshold is a preset threshold ranging from 0.6 to 0.9. This value represents the ideal matching threshold, where the model has the highest confidence, the strongest evidence, a moderate text length, and the lowest risk of inferred words. i base It can be configured according to the actual needs of different hospitals, wards, and document types.

[0172] C model Model confidence score: The confidence score of the model used by the evidence-constrained summary generation module for the current candidate summary output, ranging from 0 to 1. This value is generated by the model's internal evaluation and represents the model's confidence in the content of its output summary. When the model confidence score is high, it indicates that the model is quite confident in the current summary. In this case, the matching threshold should be appropriately reduced, i.e., the factor (1- l 1· C model The threshold should be reduced to avoid incorrectly blocking correct summaries due to overly strict matching; when the model confidence is low, the threshold should be increased to conduct more rigorous review of the model output.

[0173] E strength Strength of Evidence: The strength of evidence upon which a high-risk medical inference relies, ranging from 0 to 1, is determined by the length of the phrase in the original text and its uniqueness within the text. Specifically, the longer the phrase and the lower its frequency of occurrence in the original text, the higher the strength of evidence. When the strength of evidence is high, it indicates that the high-risk medical inference has sufficient and unique textual support; in this case, the matching threshold should be appropriately lowered, i.e., factor (1+ l 2·(1- E strength When the strength of evidence is low, it indicates that the original text does not provide sufficient support, and the threshold should be increased to require a more precise match.

[0174] L norm Normalized value for the target text length: L norm =min(1, Ldoc / L ref ),in L doc The length of the target text characters, L ref For reference length threshold and L ref >0. When the target text is short and the information content is limited, the matching threshold should be increased (i.e., factor (1+ )). l 3·(1- L norm To avoid misjudgment due to insufficient information, when the length of the target text reaches or exceeds the reference threshold, L norm =1, this factor is 1, and has no effect on the threshold. The purpose of this design is that when the text is long enough, length will no longer be a limiting factor.

[0175] R term Risk Level: The preset risk level for currently high-risk medical inference terms, ranging from 0 to 1. A higher risk level indicates a greater impact on clinical safety if the inference term is misjudged as having supporting information from the original text. For example, respiratory failure and sepsis can be set to a higher risk level, such as 0.9, while hypoalbuminemia can be set to a lower risk level, such as 0.6. When a term's risk level is high, the matching threshold should be increased, i.e., factor (1+). l 4. R term The threshold increases, requiring stricter matching conditions; when the risk level is low, the threshold can be appropriately lowered.

[0176] l 1. l 2. l 3. l 4 represents the influence coefficients: the influence coefficients of each factor range from 0 to 1, and satisfy the following conditions: l 1+ l 2+ l 3+ l λ = 1. These coefficients control the weight of the four factors—model confidence, strength of evidence, text length, and word risk level—on the final threshold. For example, if the system designer considers word risk level to be the most important factor, λ can be increased. Set it to a larger value; if the model confidence level is considered a secondary factor, you can... l 1 is set relatively small. These coefficients can be configured through experimentation or human experience based on different application scenarios.

[0177] This formula couples the four factors together using a product form, with each factor amplifying or diminishing the other's influence, reflecting a nonlinear interaction effect. Specifically: When model confidence C model When it increases, (1- l 1· C model ) decreases, leading to i th Decrease, meaning lower the matching threshold, the more confident the model is, the easier it is to pass the match.

[0178] When the strength of evidence E strength When decreasing, (1+ l 2·(1- E strength )) increases, leading to i th Increasing the threshold means raising the matching threshold; the weaker the evidence, the more precise the match is required.

[0179] When the target text length is normalized L norm When decreasing, (1+ l 3·(1- L norm )) increases, leading to i th Increased, meaning the shorter the text, the stricter the matching requirements.

[0180] When the risk level of the word R term When it increases, (1+ l 4. R term ) increases, leading to i th Increased risk means more stringent matching requirements.

[0181] Calculate the dynamic similarity threshold i th Then, the high-risk medical inference terms are compared with any segment in the target text using normalized similarity calculation. Normalized similarity calculation includes: removing punctuation, standardizing capitalization, removing whitespace, and optionally performing stemming or synonym expansion. If the normalized similarity is greater than or equal to... i th If the corresponding medical statement is explicitly supported in the target text, it is determined that the original text supports it; otherwise, it is determined to be a high-risk medical inference not supported by the original text, and the candidate abstract is rejected. Determining that the corresponding medical statement is explicitly supported in the target text includes: verifying negative words, suspected words, degree words, and consistency with contextual polarity.

[0182] The following is a set of example parameter configurations, which can be adjusted according to specific scenarios in actual applications: benchmark threshold i base =0.75 Model confidence influence coefficient l 1 = 0.2 Influence coefficient of evidence strength l 2 = 0.2 Text length influence coefficient l 3 = 0.2 Word risk level impact coefficient l 4 = 0.4, the highest weight. Reference length threshold L ref= 500 (character count) Under the above configuration, assuming a certain scenario C model =0.9, E strength =0.8, L doc =300 (then) L norm =0.6), R term =0.8, from the calculation we can get: i th =0.75×(1-0.2×0.9)×(1+0.2×(1-0.8))×(1+0.2×(1-0.6))×(1+0.4×0.8)≈0.912 That is, when the model confidence is high (0.9) but the text is short (300 characters) and the word risk is high (0.8), the final threshold is raised to about 0.912, requiring almost a perfect match to pass.

[0183] By employing the aforementioned dynamic similarity threshold, the present invention achieves the following technical effects: Adaptive matching: The threshold is adjusted in real time based on model confidence, evidence strength, text length, and word risk level, which can adapt to different types and qualities of input and avoid the one-size-fits-all problem caused by a fixed threshold.

[0184] Strict control over high risks: For medical inference terms with high risk levels, the threshold is automatically raised, requiring more accurate matching of the original text, which reduces the possibility of high-risk diagnostic escalation hallucinations entering the voice broadcast from a mechanism perspective.

[0185] Low-risk flexible handling: For inference words with low risk level, sufficient evidence, and high model confidence, the threshold is appropriately lowered to avoid accidentally damaging reasonable summary content due to excessive strictness, thus ensuring the availability of the system.

[0186] Configurable parameters: Each influence coefficient can be flexibly configured according to different hospitals, wards, and document types, and has good portability and scenario adaptability.

[0187] 14. Voice broadcast adaptation unit The voice broadcast adaptation unit is used to organize the security-verified summary into short text suitable for TTS or robot reading. Its processing may include length compression, long sentence splitting, conversion of common medical units and symbols, standardization of percentage sign and temperature expressions, standardization of frequency expressions, preservation of key symptoms and clear instructions, and reduction of lengthy indicator strings.

[0188] This unit only processes summaries that have passed the security check and cannot bypass evidence verification.

[0189] 15. Conservative Reading Supplementary Unit The conservative readout fallback unit is used to generate a more conservative readout text when a strict summary fails. This unit only allows the compilation of clearly existing symptoms, ward round comments, changes in medical orders, treatment opinions, and key concerns in the current target text. It does not allow the addition of external knowledge, upgrading of diagnoses, or modification of weak statements into stronger judgments.

[0190] If the conservatively read text still contains high-risk inferences not supported by the original text, or if the compiled result is empty, then the output will be rejected.

[0191] 16. Reject Output Unit The output rejection unit is used to terminate abstract output in cases of insufficient input, insufficient evidence, peer rejection, failure to pass final review, high-risk interception, or failure of conservative fallback measures. Output rejection is a normal protection outcome in the medical voice broadcast security chain.

[0192] As a specific implementation method, taking "the robot broadcasting a summary of a patient's first ward round record" as an example, the system first locates the recordId corresponding to the first ward round record in the medical record list, and then calls the medical record content interface for that recordId. After parsing the interface wrapper, the system locates the "ward round record" section in the HTML document and extracts the main text area.

[0193] If the main text explicitly states the ward round opinions, treatment measures, and key concerns, the system generates a candidate summary and requests the corresponding quote and context. The server then checks whether this evidence truly exists in the current main text. If the summary expands "elevated indicators" to "severe infection" when the original text does not contain this expression, the high-risk inference interception unit rejects the summary. If the summary passes all checks, the system reorganizes it into a shorter, clearer, and more readable text for the robot to read aloud.

[0194] If the main text consists only of template fragments, the evidence items are empty, the evidence items cannot return to the current main text, or multiple rounds of strict summaries are not safe, the system will stop outputting diagnostic summaries and enter conservative reading or refuse to output.

[0195] The following parameters can be set according to the deployment environment of this invention: Maximum number of evidence items: For example, each candidate summary can provide a maximum of three sets of evidence items to ensure that the summary focuses on key evidence and reduces the risk of false evidence.

[0196] Maximum number of attempts for strict abstracts: For example, set up three rounds of strict abstract processing to cover the stages of extraction, peer review, revision, and final review; after exceeding this, conservative reading will be used.

[0197] Conservative reading attempts: For example, the fourth round can be used as a fallback for reading aloud the original text, allowing only the clear content of the original text to be read.

[0198] Paragraph segmentation length: For example, segment long paragraphs by sentence or fixed length to avoid excessively long model input or difficulty in evidence localization.

[0199] Maximum length of voice broadcast: The maximum number of characters or the maximum duration can be set according to the robot's broadcasting scenario to avoid broadcasting being too long.

[0200] High-risk medical inference term set: can be configured according to scenario, such as respiratory failure, sepsis, shock, severe infection, hypoproteinemia, heart failure, renal failure, etc.

[0201] Minimum text length or integrity threshold: Used to determine whether the text is sufficient to be included in a strict abstract.

[0202] Evidence normalization rules: used to match quotes, contexts, and body text after removing whitespace, punctuation, capitalization differences, and common formatting differences.

[0203] Rejection reason coding: can record in a structured manner situations such as insufficient text, insufficient evidence, no clear instructions, peer rejection, final rejection, high-risk overstepping, and no catch-all reason.

[0204] The above parameters all serve to ensure the secure completion of the abstract and can be configured according to different hospital document formats, robot broadcasting scenarios, and model capabilities.

[0205] Example 2

[0206] This embodiment is a method for generating medical record summaries and suppressing hallucinations through medical voice broadcasting.

[0207] Figure 3-1 , Figure 3-2 and Figure 3-3The method flow of this invention is illustrated. To avoid an excessively long single flowchart, this process is broken down into three parts: target text acquisition, evidence constraint summary, security review, and broadcast output. The process proceeds step-by-step from interface parsing, text extraction, evidence constraint summary, security review to voice output, with any failure in any critical step resulting in a conservative fallback or rejection of output.

[0208] Steps S1 to S2: Obtaining the medical record list and locking the target document The system first obtains a list of medical records for the target patient and then determines the target documents to be processed for the current task based on the title, type, time, business intent, or other matching rules. For example, in the scenario of the first ward round broadcast, the system can prioritize the records related to the first ward round and obtain a unique recordId.

[0209] Once the target document is identified, all subsequent processing revolves around that recordId. If a unique target document cannot be determined, the system should request additional conditions or refuse to proceed with the strict abstract process to avoid concatenating abstracts from across documents.

[0210] Steps S3 to S4: Medical record content interface call and shell parsing The system calls the medical record content interface based on the target recordId. After the interface returns, the system parses the outer status code, business message, encrypted fields, or XML shell, and extracts the internal medical record HTML text. If the outer status is abnormal, the business status is abnormal, or the text is empty, output is rejected.

[0211] The purpose of this step is to distinguish the content returned by the interface from the summarizable body text, and to avoid treating status fields, error messages, or shell content as the medical record body text.

[0212] Steps S5 to S6: Target chapter region location and main text region extraction The system locates the target chapter area in the HTML document. Ideally, the system locates the corresponding paragraph in the ward round record using chapter attributes or chapter name, and then extracts the main text area (tag-article-content) within that chapter area. If this main area is not found, it can return to extracting text within the chapter area; if DOM parsing fails, regular expressions or other semi-structured parsing methods can be used as backup paths.

[0213] After extraction, the system normalizes the text, such as removing chapter tags, HTML tags, repeated whitespace, and meaningless style content, to obtain the target text.

[0214] Step S7: Text Usability Assessment The system determines whether the target text is suitable for the rigorous abstracting process. Factors that can be judged include the length of the text, whether it is empty, whether it only contains template fragments, whether it lacks key content, whether there is obvious truncation, and whether it only has isolated indicators without clear disease conclusions or medical instructions.

[0215] If the main text is empty, too short, or consists only of template fragments, the system will directly refuse to output and will not call strict summary generation; if the main text exists but lacks clear medical instructions or key conclusions, it can enter a fixed safety prompt or conservative processing.

[0216] Step S8: Generate candidate summaries and evidence items The system inputs the target text into the candidate summary generation module. This module only allows output based on the current target text, and does not allow the use of external knowledge or the modification of weak statements into stronger diagnoses. Candidate results are preferably in a structured format, including at least can_extract, rejection_reason, summary, and evidence_items.

[0217] If the candidate module deems the evidence insufficient, it should return `can_extract=false` and the reason for rejection, without fabricating a summary.

[0218] Step S9: Verification of Evidence Anchor Points The system performs local validation on the `evidence_items` object on the server side. This validation bypasses the model's self-declaration and directly checks whether the evidence item can lead back to the current target text.

[0219] Preferred verification rules include: 1) The evidence item cannot be empty.

[0220] 2) The number of evidence items cannot exceed the preset limit.

[0221] 3) Neither quote nor context can be empty.

[0222] 4) Context should not be replaced by ellipses or other similar means.

[0223] 5) The quote must be found in the current text.

[0224] 6) The context must be found in the current text.

[0225] 7) The normalized quote must be included in the context-normalized text.

[0226] 8) Evidence should not be too vague, and conclusions should not be supported by general terms such as patient, condition, and treatment.

[0227] If the verification fails, the candidate summary will not be accepted and will be either read out as a fallback or the output will be rejected.

[0228] Steps S10 to S13: Review, revision and final review Once the evidence verification is successful, the system enters the peer review process. The reviewers determine whether the candidate abstract is directly supported by the original text and output "accept," "revise," or "reject."

[0229] If the reviewers deem the candidate abstract acceptable, it proceeds to the final review stage. If the reviewers believe it can be made safer through deletion, downgrading, or rewriting, it enters the revision process. The revision process can only converge based on the current target text and reviewers' comments; it cannot add new diagnoses, instructions, or inferences not present in the original text. The revised results still need to be re-verified with evidence and then proceed to the final review stage. If the reviewers or the final reviewer reject the paper, it proceeds to a conservative reading as a fallback.

[0230] In one implementation, a maximum number of attempts can be set for strict summarization, such as three rounds of strict summarization attempts; if a safe result cannot be obtained after exceeding this number, a fourth round of conservative reading is initiated as a fallback. The above number of attempts is merely an example and is not intended to limit the scope of the claims.

[0231] Step S14: Interception of High-Risk Medical Inferences Abstracts that pass the final review still need to undergo high-risk medical inference interception. The system can maintain a set of high-risk words or a set of rules, scan the abstract text, and compare it with the current target text. If a high-risk word appears in the abstract, but the word or corresponding medical expression does not appear explicitly in the current target text, it is considered an escalation inference not supported by the original text, and the abstract is rejected or moved to a conservative fallback.

[0232] This step is used to determine whether high-risk statements in the abstract are supported by the original text of the document, and is not intended to replace doctors in judging the veracity of medical facts.

[0233] Steps S15 to S16: Voice broadcast adaptation and output Once the abstract passes evidence verification, final peer review, and high-risk interception, the system will then perform voice broadcast adaptation. Adaptation processing may include controlling broadcast length, splitting excessively long sentences, standardizing unit pronunciation, converting percentage signs and temperature expressions, and reducing unsuitable formatting symbols.

[0234] The final output, a playable summary, can be read aloud by robots, bedside terminals, or TTS services.

[0235] Steps F1 to F2: Conservative reading as a safety net and refusal to output If a rigorous abstract fails at any stage—the main text, evidence stage, peer review stage, final review stage, or high-risk interception stage—the system may resort to a conservative reading as a fallback. The conservative reading only extracts content explicitly present in the original text and does not output any inferential abstracts. If the conservative reading is still empty or hits a high-risk inference not supported by the original text, the system will refuse to output an abstract and will return the reason for rejection.

[0236] Refusing to output is a key security benefit of this invention. This means that in medical voice broadcasting scenarios, refusing to broadcast without reliable evidence is safer than outputting a seemingly complete but unreliable summary.

[0237] The present invention has the following alternatives: (1) Alternatives to the target document locking method 1) Target documents can be matched by document title and document type.

[0238] 2) Target documents can be matched by document time, medical order scenario, or voice intent.

[0239] 3) Target documents can be selected through manual confirmation.

[0240] 4) The target document can be determined by combining the patient, bed, visit number, and document type.

[0241] As long as the subsequent abstracting process remains confined to a single target document, it will not deviate from the core concept of this invention.

[0242] (2) Alternatives to the structured extraction method of the main text 1) The target chapter area and the main text area can be extracted using DOM selectors.

[0243] 2) The main text can be extracted using regular expressions, tag paths, or semi-structured parsing rules.

[0244] 3) The main text can be obtained directly through the structured fields provided by the medical record system.

[0245] 4) Different target chapters can be configured for different document types, such as ward round records, consultation opinions, treatment opinions, and medical records.

[0246] (3) Alternatives to candidate abstract generation methods 1) Candidate summaries can be generated from large language models.

[0247] 2) Sentences can be extracted first by an extraction model and then compressed by a generative module.

[0248] 3) It can be generated by a combination of rule templates and models.

[0249] 4) Multiple models can be responsible for extraction, review, revision or final review respectively.

[0250] As long as the candidate abstract still needs to carry evidence items and be subject to server-side evidence verification, it falls under the implementation methods that can be covered by this invention.

[0251] (4) Alternative ways of expressing evidence 1) A binary structure of quote and context can be used.

[0252] 2) A three-part structure consisting of the original phrase, paragraph number, and the original paragraph can be used.

[0253] 3) The structure can be based on sentence index, tag path, and original text fragments.

[0254] 4) Evidence location can be represented by character offset, paragraph offset, or document node path.

[0255] (5) Alternatives to high-risk inference interception methods 1) A fixed list of high-risk terms can be used.

[0256] 2) A combination of vocabulary and rules can be used.

[0257] 3) Medical category recognition models can be used to identify diagnostic upgrade descriptions.

[0258] 4) A dual verification method using a peer review model and a local thesaurus can be adopted.

[0259] What they have in common is that they only judge whether high-risk statements in the abstract are supported by the current target text, without claiming that the system automatically judges the truth or falsehood of medical facts.

[0260] (6) Alternative solutions for voice broadcast adaptation 1) The TTS pre-text can be generated on the server side.

[0261] 2) The abstract can be adapted for secondary reading on the robot side.

[0262] 3) Different length limits can be set according to different broadcast scenarios.

[0263] 4) Different reading rules can be configured for medical units, percentage signs, temperature, frequency, English abbreviations, etc.

[0264] (7) Alternatives to the safety net approach 1) If strict summarization fails, conservative reading text can be output first.

[0265] 2) Alternatively, you can directly reject the output and return the reason for rejection.

[0266] 3) It can also be configured according to the scenario to prioritize conservative reading or directly reject reading.

[0267] 4) The reason for failure can also be sent back to the interface, prompting medical staff to view the original text.

[0268] Although the above alternative solutions differ in implementation details, as long as they still adhere to the overall mechanisms of single-document boundaries, evidence constraints, server-side verification, high-risk interception, voice adaptation, and conservative fallback, they belong to the alternative implementation directions of this invention.

[0269] This invention has the following characteristics: 1) Lock a single target medical record document with a unique recordId and prohibit the splicing of abstracts across documents.

[0270] 2) Parse the medical record interface shell and HTML document structure, and prioritize locating the target chapter area and the main text area.

[0271] 3) The usability of the main text is judged. If the main text is empty, too short, or only has template fragments, it will be rejected from the strict abstract.

[0272] 4) Candidate summaries must output both summary and evidence items; they cannot only output the conclusion.

[0273] 5) The server locally verifies the anchor relationship between the quote, context, and the current target text.

[0274] 6) Through the review, revision and final review process, the abstract is made to converge within the scope supported by the original text.

[0275] 7) Local interception of high-risk medical inferences not explicitly supported by the original text.

[0276] 8) If strict summarization fails, it will switch to conservative reading as a fallback or refuse output.

[0277] 9) The final approved summary is then adapted for voice broadcasting to form a short text suitable for TTS or robot reading.

[0278] 10) This invention protects the secure summary processing link and constraint mechanism, without limiting the specific model name, prompt text or communication protocol.

[0279] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A medical record summary generation and hallucination suppression system for medical voice broadcasting, characterized in that, include: The medical record input module is used to locate a unique target medical record document from the medical record list and parse the medical record content interface to extract the target text. An evidence-constrained summary generation module, connected to the medical record input module, is used to generate a structured candidate summary based on the target text. The candidate summary includes summary text, evidence items, and extractable markers. The evidence item is the original text evidence in the target text that supports the conclusion of the candidate abstract; The server-side evidence anchor verification module is connected to the evidence constraint summary generation module and is used to verify on the server whether the evidence item actually exists in the target text. The security convergence module is connected to the server-side evidence anchor verification module and is used to perform review, revision and final review of the candidate abstract based on the evidence items. A high-risk interception module, connected to the security convergence module, is used to identify and intercept high-risk medical inferences in the summary text that are not explicitly supported by the target text. A voice adaptation module, connected to the high-risk interception module, is used to convert the summary text processed by the high-risk interception module into a format suitable for text-to-speech broadcasting; The conservative fallback module, connected to the voice adaptation module, is used to execute a conservative reading fallback strategy or refuse output when any of the server-side evidence anchor verification module, the security convergence module, or the high-risk interception module fails.

2. The system according to claim 1, characterized in that, The medical record input module includes: The medical record list retrieval unit is used to retrieve a list of the patient's medical records. The target identification locking unit is used to determine a unique target medical record document from the list based on the document title, type, or business rules. The medical record shell parsing unit is used to parse the outer data structure returned by the medical record content interface and extract the inner medical record text. The target chapter region positioning unit is used to locate a preset target chapter region in the internal medical record text. The main text region extraction unit is used to extract the main text region text within the target chapter region located by the target chapter region positioning unit, and use it as the target main text. The text integrity judgment unit is used to determine whether the target text is empty, too short, or contains only template fragments. If so, the conservative fallback module is triggered.

3. The system according to claim 2, characterized in that, The main text region extraction unit is configured to: firstly extract the text within the preset main text tags in the target chapter region as the main text region text; if the preset main text tags are not matched, then backtrack to extract the text of the entire target chapter region and perform cleaning and normalization processing.

4. The system according to claim 1, characterized in that, The evidence constraint summary generation module includes: The candidate summary generation unit is used to generate a structured candidate summary based on the target text. The structured candidate summary also includes a rejection reason. Furthermore, when it is determined that the target text is insufficient to generate a safe summary text, the extractable flag is returned as false, and the rejection reason is filled in to prohibit the fabrication of summary text. An evidence item generation unit is used to provide evidence items for key conclusions in the candidate abstract, each evidence item containing at least a phrase from the original text of the target text and its context.

5. The system according to claim 1, characterized in that, The server-side evidence anchor verification module includes an evidence anchor verification unit, which is configured to be independent of the evidence constraint summary generation module and perform one or more of the following verifications: verifying whether the evidence item is empty or whether its number exceeds a preset upper limit; Verify whether the original phrase and context in each piece of evidence are empty; verify whether the normalized original phrase is contained within the normalized context; Verify whether the original phrase and the context can be precisely or fuzzily matched in the target text, respectively; if either verification fails, the candidate summary is deemed unacceptable.

6. The system according to claim 1, characterized in that, The secure convergence module includes: The peer review decision unit is used to output the peer review results: acceptance, revision, or rejection. The revision convergence unit is used to safely converge the candidate abstract by deleting unfounded content, reducing the intensity of expression, or rewriting it into expression directly supported by the target text, based on the reviewers' comments. It prohibits the addition of new diagnoses or suggestions not supported by the original text. The final verification unit is used to re-verify the revised candidate digest back to the server-side evidence anchor verification module, and then perform final security verification.

7. The system according to claim 1, characterized in that, The high-risk interception module includes a high-risk inference interception unit. When matching high-risk medical inference words in the abstract text with the target text, the high-risk inference interception unit uses a dynamic similarity threshold, the calculation formula of which is: in: θ th This is a dynamic similarity threshold, with a value ranging from 0 to 1; θ base The preset baseline similarity threshold; C model The confidence score output by the evidence constraint summary generation module for the current candidate summary ranges from 0 to 1. E strength The strength of evidence upon which the high-risk medical inference term is based is determined by the length of the original phrase in the evidence item and its uniqueness in the original text, with a value ranging from 0 to 1; L norm The normalized value of the target text length. L norm =min(1, L doc / L ref ), L doc The target text length, L ref This is a reference length threshold. R term The preset risk level for currently high-risk medical inference terms, with a value range of 0 to 1; λ 1. λ 2. λ 3. λ 4 represents the influence coefficient of each factor, with a value ranging from 0 to 1, and satisfying the following conditions: λ 1+ λ 2+ λ 3+ λ 4 = 1; When the normalized similarity between a high-risk medical inference term and any segment in the target text is greater than or equal to 1, θ th If the corresponding medical statement is explicitly supported in the target text, it is determined that the original text supports it; otherwise, it is determined to be a high-risk medical inference not supported by the original text, and the candidate abstract is rejected. Determining that the corresponding medical statement is explicitly supported in the target text includes: verifying negative words, suspected words, degree words, and consistency with contextual polarity.

8. The system according to claim 7, characterized in that, The high-risk inference interception unit maintains a preset set of high-risk medical inference words, which are used to identify high-risk medical inferences that may exist in the summary text. Furthermore, the identification and interception are based on matching after text normalization. If the summary text contains any inference word from the set of high-risk medical inference words, and the medical expression corresponding to the inference word is not explicitly supported in the normalized target text, it is determined to be a high-risk medical inference not supported by the original text, and the candidate summary is rejected.

9. The system according to claim 1, characterized in that, The conservative fallback module includes: The conservative reading catch-all unit is used to implement a conservative reading catch-all strategy, including: extracting only clearly existing objective descriptions of the illness, ward round opinions, or treatment opinions from the target text; and prohibiting any form of diagnostic upgrade, intentional completion, or medical inference of the information in the target text. The output rejection unit, connected to the conservative reading catch-all unit, is used to reject output when the conservative reading text processed by the conservative reading catch-all unit is empty, too short, or still contains high-risk expressions not supported by the target text.

10. A method for generating medical record summaries and suppressing hallucinations using medical voice broadcasting, characterized in that, include: Medical record input steps: Locate a unique target medical record document from the medical record list, parse the outer data structure returned by the medical record content interface to obtain the inner medical record text, locate the preset target chapter area in the inner medical record text, and extract the main text area of ​​the target chapter area as the target text. Evidence-constrained summary generation steps: Input the target text into the evidence-constrained summary generation module to generate a structured candidate summary; the candidate summary includes at least the summary text, a set of evidence items, and extractable markers; wherein, each evidence item in the set of evidence items contains at least a phrase from the original text and its context, and the evidence item is original text evidence in the target text that supports the conclusion of the candidate summary; Server-side evidence anchor verification steps: On the server side, each evidence item in the candidate summary is compared and verified with the target text to verify whether the original phrase and its context actually exist in the target text; The safe convergence process involves sequentially reviewing, revising, and confirming the candidate abstracts that pass the evidence anchor point verification. High-risk interception step: Scan the summary text that has passed the final review for high-risk medical inference words, and compare the scanned high-risk medical inference words with the target text. If the medical expression corresponding to the high-risk medical inference word is not clearly supported in the target text, it is determined that there is a high-risk medical inference that is not supported by the original text and the candidate summary is rejected. Voice adaptation step: The summary text processed by the high-risk interception step is subjected to text-to-speech adaptation processing to generate a playable summary; Conservative fallback step: If any of the steps in the medical record input step, the evidence constraint summary generation step, the server-side evidence anchor verification step, the safe convergence step, or the high-risk interception step fails, then the conservative reading fallback strategy is executed or the output is directly rejected.

Images