Large model-based re-diagnosis medical record generation method and system, and storage medium

Abstract

The invention provides a large-model-based re-visit medical record generation method and system and a storage medium, and the method comprises the steps: obtaining a primary visit medical record of a target patient, inputting the primary visit medical record into a large language model, and carrying out the abstract extraction to obtain a primary visit abstract; acquiring doctor-patient dialogue information of the target patient, and inputting the doctor-patient dialogue information into the large language model for abstract extraction to obtain a revise dialogue abstract; and carrying out abstract fusion on the re-diagnosis dialogue abstract and the initial diagnosis abstract to obtain a re-diagnosis medical record. According to the embodiment of the invention, the first-diagnosis medical record is input into the large language model for abstract extraction, the first-diagnosis abstract can be obtained through automatic extraction, the doctor-patient dialogue information is input into the large language model for abstract extraction, the re-diagnosis dialogue abstract can be obtained through automatic extraction, and the re-diagnosis dialogue abstract and the first-diagnosis abstract are subjected to abstract fusion. The method can automatically generate the re-visit medical record, does not need to generate the re-visit medical record in a manual input mode, and improves the efficiency of the re-visit medical record.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart medical care, and in particular to a re-visit medical record generation method and system based on a large model and a storage medium. BACKGROUND

[0002] In the medical service system, re-visit is a key link for tracking the progress of a patient's condition, adjusting the treatment plan, and evaluating the treatment effect. The re-visit medical record, as a core medical document recording the patient's symptom changes, physical sign data, diagnosis results, treatment adjustments, and medical order information during the re-visit process, is not only an important basis for medical personnel to carry out follow-up diagnosis and treatment work, but also a key support for medical quality control, medical dispute handling, medical research, and medical insurance settlement. The completeness, accuracy, and timeliness of the re-visit medical record are directly related to the quality of medical services and the safety of patient diagnosis and treatment.

[0003] In the existing re-visit medical record generation process, it is generally realized by manual input by medical personnel, resulting in low efficiency of re-visit medical record generation, which is difficult to adapt to the clinical diagnosis and treatment rhythm. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a re-visit medical record generation method and system based on a large model and a storage medium to solve the problem of low efficiency of re-visit medical record generation in the prior art.

[0005] The embodiments of the present application are implemented as follows: a re-visit medical record generation method based on a large model, the method comprising:

[0006] obtaining an initial diagnosis medical record of a target patient, and inputting the initial diagnosis medical record into a large language model for abstract extraction to obtain an initial diagnosis abstract;

[0007] obtaining medical-patient conversation information of the target patient, and inputting the medical-patient conversation information into the large language model for abstract extraction to obtain a re-visit conversation abstract;

[0008] performing abstract fusion on the re-visit conversation abstract and the initial diagnosis abstract to obtain a re-visit medical record.

[0009] Preferably, the initial diagnosis medical record is input into the large language model for abstract extraction to obtain the initial diagnosis abstract, comprising:

[0010] inputting the initial diagnosis medical record into the large language model for medical record splitting to obtain patient basic information and doctor treatment opinion content, and performing content extraction on the patient basic information according to a first preset field to obtain patient extraction information;

[0011] obtaining visit time information in the initial diagnosis medical record, and generating re-visit time information according to the visit time information;

[0012] obtain an initial diagnosis medical order document text of the target patient, and perform content extraction on the initial diagnosis medical order document text and the doctor processing opinion content according to a second preset field to obtain medical order extraction information;

[0013] combine the patient extraction information, the re-consultation time information, and the medical order extraction information to obtain the initial diagnosis summary.

[0014] Preferably, the doctor-patient conversation information is input into the large language model to obtain a re-consultation conversation summary, including:

[0015] The large language model generates a semantic summary window according to a preset character length or a preset semantic unit quantity, and an overlap rate between adjacent semantic summary windows is greater than a preset overlap rate.

[0016] According to the semantic summary window, the information of the doctor-patient conversation information is divided to obtain a window conversation segment, and the semantic of the window conversation segment is extracted to obtain a window conversation semantic.

[0017] The window conversation semantic is subjected to contradiction information elimination, and the window conversation semantic after contradiction information elimination is subjected to key information completion.

[0018] The window conversation semantic after key information completion is subjected to redundant content elimination to obtain a window revised semantic, and the window revised semantic is combined to obtain the re-consultation conversation summary.

[0019] Preferably, the window conversation semantic is subjected to contradiction information elimination, including:

[0020] Obtain object description information of a contradiction object in the window conversation semantic, the contradiction object including a patient symptom, a patient medical history, and a doctor diagnosis.

[0021] According to the object description information, it is determined whether there is a description contradiction between adjacent window conversation semantics.

[0022] If there is a description contradiction between adjacent window conversation semantics, adjacent window conversation semantics are determined as contradiction window conversations.

[0023] According to the object description information of adjacent contradiction window conversations, a contradiction type is determined, and adjacent contradiction window conversations are subjected to contradiction information elimination according to the contradiction type.

[0024] Preferably, according to the object description information of adjacent contradiction window conversations, a contradiction type is determined, and adjacent contradiction window conversations are subjected to contradiction information elimination according to the contradiction type, including:

[0025] combining object description information of adjacent dialogue of the contradictory window to obtain contradictory combination information, and matching the contradictory combination information with a contradictory query table to obtain the contradictory type;

[0026] if the contradictory type is a negative conflict type, invalidating a previous window dialogue in adjacent contradictory window dialogues;

[0027] if the contradictory type is a supplement conflict type, supplementing content of the previous window dialogue according to a latter window dialogue in adjacent contradictory window dialogues;

[0028] if the contradictory type is a trace original conflict type, determining a contradictory original dialogue in the doctor-patient dialogue information according to adjacent contradictory window dialogues, and correcting the adjacent contradictory window dialogues according to semantics of the contradictory original dialogue.

[0029] Preferably, the window dialogue semantics after the contradictory information is eliminated are supplemented with key information, including:

[0030] positioning core semantics of the window dialogue semantics after the contradictory information is eliminated according to a preset core keyword to obtain core dialogue semantics;

[0031] determining adjacent window dialogue semantics of the core dialogue semantics as candidate gap semantics, and matching gap-sensitive keywords with the candidate gap semantics;

[0032] if the gap-sensitive keywords and the candidate gap semantics do not match, positioning a gap original dialogue in the doctor-patient dialogue information according to the candidate gap semantics, and supplementing information of the candidate gap semantics according to semantics of the gap original dialogue.

[0033] Preferably, the window dialogue semantics after the key information is supplemented are subjected to redundant content elimination to obtain window corrected semantics, including:

[0034] calculating semantic similarity between adjacent window dialogue semantics after the key information is supplemented;

[0035] if the semantic similarity is greater than a similarity threshold, deleting a latter window dialogue semantic to obtain the window corrected semantics.

[0036] Another purpose of the embodiment of the application is to provide a follow-up medical record generation system based on a large model, the system comprising:

[0037] a first diagnosis extraction module configured to obtain a first diagnosis medical record of a target patient, and input the first diagnosis medical record into a large language model to perform abstract extraction to obtain a first diagnosis abstract;

[0038] The dialogue extraction module is configured to obtain medical dialogue information of the target patient, and input the medical dialogue information into the large language model to perform abstract extraction, thereby obtaining a follow-up visit dialogue abstract.

[0039] The abstract fusion module is configured to fuse the follow-up visit dialogue abstract and the initial visit abstract, thereby obtaining a follow-up visit medical record.

[0040] Preferably, the initial visit extraction module is further configured to:

[0041] input the initial visit medical record into the large language model to perform medical record splitting, thereby obtaining patient basic information and doctor treatment opinion content, and perform content extraction on the patient basic information according to a first preset field, thereby obtaining patient extraction information;

[0042] obtain visit time information in the initial visit medical record, and generate follow-up visit time information according to the visit time information;

[0043] obtain an initial visit medical order text of the target patient, and perform content extraction on the initial visit medical order text and the doctor treatment opinion content according to a second preset field, thereby obtaining medical order extraction information;

[0044] combine the patient extraction information, the follow-up visit time information and the medical order extraction information, thereby obtaining the initial visit abstract.

[0045] According to the embodiments of the present application, the initial visit abstract can be automatically extracted by inputting the initial visit medical record into the large language model to perform abstract extraction, the follow-up visit dialogue abstract can be automatically extracted by inputting the medical dialogue information into the large language model to perform abstract extraction, and the follow-up visit medical record can be automatically generated by fusing the follow-up visit dialogue abstract and the initial visit abstract, without using the manual input method to generate the follow-up visit medical record, thereby improving the efficiency of the follow-up visit medical record. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flowchart of the follow-up visit medical record generation method based on the large model provided by the first embodiment of the present application;

[0047] Figure 2 is a structural schematic diagram of the follow-up visit medical record generation system based on the large model provided by the second embodiment of the present application;

[0048] Figure 3 is a schematic diagram of the initial visit abstract extraction provided by the second embodiment of the present application;

[0049] Figure 4 is a schematic diagram of the follow-up visit dialogue abstract extraction provided by the second embodiment of the present application;

[0050] Figure 5is a structural schematic diagram of a terminal device provided by a third embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0052] In order to illustrate the technical solutions of the present application, the following will be described through specific embodiments.

[0053] Embodiment One

[0054] Please refer to Figure 1 is a flowchart of a re-visit medical record generation method based on a large model provided by a first embodiment of the present application. The re-visit medical record generation method based on a large model can be applied to any device or system. The re-visit medical record generation method based on a large model includes the following steps:

[0055] In step S10, the initial diagnosis medical record of the target patient is obtained, and the initial diagnosis medical record is input into a large language model for abstract extraction to obtain an initial diagnosis abstract.

[0056] The first preset field and the second preset field are set in the large language model. The first preset field and the second preset field can be set according to requirements. The first preset field is used for content extraction of patient basic information, and the second preset field is used for content extraction of initial diagnosis medical order document text. Based on the extraction results of the patient basic information and the initial diagnosis medical order document text, the initial diagnosis abstract is automatically spliced.

[0057] Optionally, the initial diagnosis medical record is input into the large language model for abstract extraction to obtain the initial diagnosis abstract, including:

[0058] The initial diagnosis medical record is input into the large language model for medical record splitting to obtain patient basic information and doctor's treatment opinion content, and the patient basic information is extracted according to the first preset field to obtain patient extraction information. The large language model splits the initial diagnosis medical record according to the specified field to obtain the patient basic information and the doctor's treatment opinion content. The first preset field is matched with the patient basic information through field matching, and the patient basic information is extracted based on the field matching result. The first preset field can include "current illness history-disease overview", "current illness history-drugs", "current illness history-other treatment measures", "current illness history-disease", "current illness history-surgical history", "marriage and childbirth history", "menstrual history", "personal history", etc.

[0059] Obtain the visit time information in the initial diagnosis medical record, and generate the re-visit time information according to the visit time information; wherein, the "disease summary" field information is processed separately by using a large language model, and timeline reasoning is performed to determine the re-visit time information, which includes the historical events of the last visit in the visit time information and the current re-visit time;

[0060] Obtain the initial diagnosis medical order text of the target patient, and extract the content of the initial diagnosis medical order text and the doctor's treatment opinion content according to the second preset field to obtain the medical order extraction information; wherein, the doctor's treatment opinion content and the initial diagnosis medical order text are input into the large language model, and the content is extracted according to the second preset field, which can include "treatment opinion-drug" and "treatment opinion-other suggestions";

[0061] Combine the patient extraction information, the re-visit time information and the medical order extraction information to obtain the initial diagnosis summary; wherein, the extracted field list is spliced and sorted in a given order to obtain the initial diagnosis summary.

[0062] Step S20, obtain the doctor-patient conversation information of the target patient, and input the doctor-patient conversation information into the large language model to extract the summary to obtain the re-visit conversation summary;

[0063] Wherein, a multi-round summary extraction method based on multiple windows is adopted, and a dynamic window division and multi-level summary fusion strategy are used to extract the summary of the doctor-patient conversation information, so as to solve the problems of local information fragmentation and lack of global context association in long re-visit doctor-patient conversation.

[0064] Optionally, the doctor-patient conversation information is input into the large language model to extract the summary to obtain the re-visit conversation summary, including:

[0065] The large language model generates a semantic summary window according to a preset character length or a preset semantic unit quantity; wherein, the overlap rate between adjacent semantic summary windows is greater than a preset overlap rate, and the preset character length, the preset semantic unit quantity and the preset overlap rate can be set according to requirements, for example, the preset character length can be set to 500 characters, the preset semantic unit quantity can be set to 5 rounds of conversation quantity, and the preset overlap rate can be set to 30% overlap rate;

[0066] According to the semantic summary window, the doctor-patient conversation information is divided into window conversation segments, and the window conversation semantics are extracted from the window conversation segments; wherein, the window conversation semantics corresponding to each window conversation segment are obtained by performing semantic extraction on the window conversation segments;

[0067] The window dialogue semantics are subjected to contradiction information elimination, and the window dialogue semantics subjected to contradiction information elimination are subjected to key information completion; wherein, the contradiction information elimination is performed on the window dialogue semantics to eliminate the semantic contradiction between window dialogue segments, thereby improving the accuracy of the window dialogue semantics, and the key information completion is performed on the window dialogue semantics subjected to contradiction information elimination to improve the integrity of the window dialogue semantics information;

[0068] The window dialogue semantics subjected to key information completion are subjected to redundant content elimination to obtain window revised semantics, and the window revised semantics are combined to obtain the follow-up visit dialogue summary; wherein, the redundant content elimination is performed on the window dialogue semantics subjected to key information completion to further improve the accuracy of the window dialogue semantics.

[0069] In this step, a large language model is called for each window to generate a preliminary summary and capture local key information. Based on the adjacent window summaries, multiple rounds of iteration are performed to realize contradiction information elimination (such as conflicting symptom descriptions), key information completion (such as cross-window disease evolution), redundant content elimination, and other content revision through context cross-validation. Finally, the multi-window summaries are spliced in order to generate a comprehensive and logically coherent follow-up visit dialogue summary.

[0070] Further, the contradiction information elimination on the window dialogue semantics comprises:

[0071] Object description information of a contradictory object in the window dialogue semantics is obtained; wherein, the contradictory object comprises a patient symptom, a patient medical history, and a doctor diagnosis, and the object description information is a specific description field corresponding to the contradictory object in the window dialogue semantics;

[0072] It is determined whether there is a description contradiction between adjacent window dialogue semantics according to the object description information; wherein, the object description information of the same contradictory object between adjacent window dialogue semantics is combined to obtain description combination information, the description combination information is matched with a contradiction query table, the contradiction query table stores pre-set contradiction combinations, and if the description combination information matches the contradiction query table, it is determined that there is a description contradiction between adjacent window dialogue semantics;

[0073] If there is a description contradiction between adjacent window dialogue semantics, the adjacent window dialogue semantics are determined as contradictory window dialogues.

[0074] A contradiction type is determined according to the object description information of adjacent contradictory window dialogues, and the adjacent contradictory window dialogues are subjected to contradiction information elimination according to the contradiction type; wherein, the contradiction information elimination manners adopted by different contradiction types can be the same or different.

[0075] Further, the contradiction type is determined according to the object description information of the adjacent contradiction window dialogues, and the contradiction information of the adjacent contradiction window dialogues is resolved according to the contradiction type, including:

[0076] The object description information of the adjacent contradiction window dialogues is combined to obtain contradiction combination information, and the contradiction combination information is matched with a contradiction query table to obtain the contradiction type; the contradiction query table stores a corresponding relationship between different contradiction combination information and corresponding contradiction types;

[0077] If the contradiction type is a negative conflict type, the previous window dialogue in the adjacent contradiction window dialogues is marked as invalid;

[0078] If the contradiction type is a complement contradiction type, the content of the previous window dialogue is supplemented according to the latter window dialogue in the adjacent contradiction window dialogues;

[0079] If the contradiction type is a trace original contradiction type, the contradiction original dialogue in the doctor-patient dialogue information is determined according to the adjacent contradiction window dialogues, and the adjacent contradiction window dialogues are corrected according to the semantics of the contradiction original dialogue.

[0080] Preferably, the window dialogue semantics after the contradiction information resolution is supplemented with key information, including:

[0081] The window dialogue semantics after the contradiction information resolution is positioned with core semantics according to a preset core keyword to obtain core dialogue semantics; the preset core keyword can be set according to requirements, and is used to position key window dialogue semantics to obtain core dialogue semantics;

[0082] Adjacent window dialogue semantics of the core dialogue semantics is determined as candidate gap semantics, and a gap sensitive keyword is matched with the candidate gap semantics; the gap sensitive keyword can be set according to requirements, and is used to detect whether there is sensitive information that must be carried in the candidate gap semantics;

[0083] If the gap sensitive keyword and the candidate gap semantics do not match, the gap original dialogue in the doctor-patient dialogue information is located according to the candidate gap semantics, and the candidate gap semantics is supplemented with information according to the semantics of the gap original dialogue; if the gap sensitive keyword and the candidate gap semantics do not match, it is determined that the candidate gap semantics has missing key information, and therefore, the candidate gap semantics is supplemented with information according to the semantics of the gap original dialogue to improve the integrity of the candidate gap semantics.

[0084] Optionally, the window dialogue semantics after the key information completion are subjected to redundant content elimination to obtain window revised semantics, including: calculating semantic similarity between adjacent window dialogue semantics after key information completion; if the semantic similarity is greater than a similarity threshold, deleting a latter window dialogue semantics to obtain the window revised semantics; wherein the similarity threshold can be set according to requirements.

[0085] In step S30, the follow-up visit dialogue summary and the initial visit summary are fused to obtain a follow-up visit medical record.

[0086] The initial visit summary extracted from the initial visit medical record is informationally fused with the complete follow-up visit dialogue summary generated based on multi-window summarization, and is corrected, polished and formatted by a large language model to finally generate a complete and standardized follow-up visit electronic medical record.

[0087] Specifically:

[0088] Information fusion: set field mapping rules to determine the mapping relationship between the field data of the initial visit summary and the field data of the follow-up visit dialogue summary, and the association relationship between medical terms. For example, there is a description of "rapid breathing" in the summary, and there are also test data such as "blood oxygen saturation", which can be associated in the follow-up visit electronic medical record.

[0089] Large language model assisted fusion: input the above two parts of information, call a large language model for comprehensive processing, realize multi-source information correction: through cross-field mapping rules and medical term association, realize contradiction information resolution and authoritative data optimization, medical semantic optimization: reconstruct the text logic chain to ensure that the disease description conforms to the clinical narrative specification (temporal, causal, complete), structured output control: segment, title mark and term standardization according to the requirements of medical documents.

[0090] Follow-up visit electronic medical record generation: the follow-up visit medical record data can be in pure text data format or structured data format (such as JSON, XML, etc.), which is beneficial for storage and other calls. Verification and feedback: after generation, the follow-up visit electronic medical record is automatically checked to ensure that each field content is complete, reasonable and meets expectations.

[0091] In this embodiment, by automatically extracting, intelligently generating and dynamically fusing core medical information in doctor-patient dialogue, a structured follow-up visit electronic medical record is generated with high precision and high efficiency, which fundamentally solves the problems of information one-sidedness, heavy manual workload, lack of universality and impaired doctor-patient communication in the traditional mode, and significantly improves the quality and efficiency of medical information processing.

[0092] Precise information capture: combining initial visit medical record summary generation and follow-up visit multi-round summary generation mechanism, taking into account accurate extraction of key information and fine-grained semantic capture, overcoming the adaptability defects of traditional templates to free expression.

[0093] Deep semantic understanding: Large language models have strong natural language processing capabilities, not only can adapt to spoken language and diverse expressions, but also can achieve high robustness information extraction and summary generation with low labeled data dependence.

[0094] Context full-dimension association: Fixed length window division and multi-round summary generation strategy are adopted to better capture global and local context information, avoiding information omission or repetition caused by only focusing on local.

[0095] Intelligent adaptive architecture: Reduces dependence on artificial template configuration and manual entry, supports dynamic adaptation of multiple departments and multiple diseases, and can flexibly adjust medical terminology mapping and window strategy according to different medical scenarios, realizing rapid deployment and promotion.

[0096] In this embodiment, by inputting the initial diagnosis medical record into the large language model for summary extraction, the initial diagnosis summary can be automatically extracted, by inputting the doctor-patient conversation information into the large language model for summary extraction, the reexamination conversation summary can be automatically extracted, and by fusing the reexamination conversation summary and the initial diagnosis summary, the reexamination medical record can be automatically generated without using artificial input to generate the reexamination medical record, improving the efficiency of the reexamination medical record.

[0097] Embodiment two

[0098] Please refer to Figure 2 is a structural schematic diagram of a reexamination medical record generation system 100 based on a large model provided by the second embodiment of the present application, comprising:

[0099] The initial diagnosis extraction module 10 is configured to obtain the initial diagnosis medical record of the target patient, and input the initial diagnosis medical record into the large language model for summary extraction to obtain the initial diagnosis summary.

[0100] Optionally, the initial diagnosis extraction module 10 is further configured to input the initial diagnosis medical record into the large language model for medical record splitting to obtain patient basic information and doctor's treatment opinion content, and extract the patient basic information according to a first preset field to obtain patient extraction information.

[0101] Obtain the visit time information in the initial diagnosis medical record, and generate the reexamination time information according to the visit time information;

[0102] Obtain the initial diagnosis medical record of the target patient, and input the initial diagnosis medical record into the large language model for summary extraction to obtain the initial diagnosis summary.

[0103] Combine the patient extraction information, the reexamination time information and the medical order extraction information to obtain the initial diagnosis summary.

[0104] The dialogue extraction module 11 is configured to obtain medical dialogue information of the target patient, and input the medical dialogue information into the large language model to perform abstract extraction to obtain a follow-up dialogue abstract.

[0105] Optionally, the dialogue extraction module 11 is further configured to: generate a semantic abstract window according to a preset character length or a preset semantic unit quantity, and the overlap rate between adjacent semantic abstract windows is greater than a preset overlap rate.

[0106] divide the medical dialogue information according to the semantic abstract window to obtain a window dialogue segment, and perform semantic extraction on the window dialogue segment to obtain a window dialogue semantic.

[0107] resolve contradictory information from the window dialogue semantic, and complete key information from the window dialogue semantic after resolving contradictory information.

[0108] remove redundant content from the window dialogue semantic after completing key information to obtain a window revised semantic, and combine the window revised semantic to obtain the follow-up dialogue abstract.

[0109] Further, the dialogue extraction module 11 is further configured to: obtain object description information of a contradictory object in the window dialogue semantic, the contradictory object including a patient symptom, a patient medical history, and a doctor diagnosis.

[0110] determine whether there is a description contradiction between adjacent window dialogue semantics according to the object description information.

[0111] If there is a description contradiction between adjacent window dialogue semantics, the adjacent window dialogue semantics are determined as contradictory window dialogues.

[0112] determine a contradiction type according to the object description information of adjacent contradictory window dialogues, and resolve contradictory information from adjacent contradictory window dialogues according to the contradiction type.

[0113] Further, the dialogue extraction module 11 is further configured to: combine the object description information of adjacent contradictory window dialogues to obtain contradictory combined information, and match the contradictory combined information with a contradiction query table to obtain the contradiction type.

[0114] If the contradiction type is a negative conflict type, the first window dialogue in adjacent contradictory window dialogues is invalidly marked.

[0115] If the contradiction type is a completion contradiction type, the first window dialogue is supplemented with content according to the second window dialogue in adjacent contradictory window dialogues.

[0116] If the contradiction type is a trace-back contradiction type, a contradiction original dialogue in the doctor-patient dialogue information is determined according to adjacent contradiction window dialogues, and the adjacent contradiction window dialogues are corrected according to semantics of the contradiction original dialogue.

[0117] Preferably, the dialogue extraction module 11 is further configured to: perform core semantic positioning on the window dialogue semantics after contradiction information resolution according to a preset core keyword, to obtain core dialogue semantics.

[0118] The adjacent window dialogue semantics of the core dialogue semantics are determined as candidate gap semantics, and gap-sensitive keywords are matched with the candidate gap semantics.

[0119] If the gap-sensitive keywords and the candidate gap semantics do not match, a gap original dialogue in the doctor-patient dialogue information is located according to the candidate gap semantics, and the candidate gap semantics are information-completed according to semantics of the gap original dialogue.

[0120] Optionally, the dialogue extraction module 11 is further configured to: calculate semantic similarity between adjacent window dialogue semantics after key information completion.

[0121] If the semantic similarity is greater than a similarity threshold, the last window dialogue semantics is deleted to obtain the window corrected semantics.

[0122] The abstract fusion module 12 is configured to fuse the re-visit dialogue abstract and the initial visit abstract to obtain a re-visit medical record.

[0123] Please refer to Figure 3 , the complete initial visit case information is input into a large language model for splitting processing to obtain a patient basic situation panel and a doctor treatment opinion panel, the patient basic situation panel contains patient basic situation information, and the doctor treatment opinion panel contains doctor treatment opinion content.

[0124] Please refer to Figure 4 , the doctor-patient dialogue full text is divided into window dialogue segments according to a fixed-length window, local abstracts (window dialogue semantics) are generated by performing abstract generation on the window dialogue segments, the window corrected semantics are obtained based on multi-round context correction, and structured detailed abstracts (re-visit dialogue abstracts) are obtained by performing abstract refinement on the window corrected semantics.

[0125] In this embodiment, by inputting the initial diagnosis medical record into the large language model for abstract extraction, the initial diagnosis abstract can be automatically extracted. By inputting the doctor-patient conversation information into the large language model for abstract extraction, the re-visit conversation abstract can be automatically extracted. By abstract fusion of the re-visit conversation abstract and the initial diagnosis abstract, the re-visit medical record can be automatically generated without using the artificial input method for re-visit medical record generation, thereby improving the re-visit medical record efficiency.

[0126] Embodiment Three

[0127] Figure 5 is a structural block diagram of a terminal device 2 provided by a third embodiment of the present application. As shown in Figure 5 The terminal device 2 of this embodiment includes a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program of the large model-based re-visit medical record generation method. The processor 20 executes the computer program 22 to implement the steps in each embodiment of the above-described large model-based re-visit medical record generation method.

[0128] For example, the computer program 22 can be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program 22 in the terminal device 2. The terminal device can include, but is not limited to, the processor 20 and the memory 21.

[0129] The processor 20 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0130] The memory 21 can be an internal storage unit of the terminal device 2, such as a hard drive or memory of the terminal device 2. The memory 21 can also be an external storage device of the terminal device 2, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal device 2. Furthermore, the memory 21 can include both internal and external storage units of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 can also be used to temporarily store data that has been output or will be output.

[0131] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0132] If an integrated module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. This computer-readable storage medium can be non-volatile or volatile. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of a computer-readable storage medium may be appropriately added to or subtracted from the contents as required by the legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, a computer-readable storage medium may not include electrical carrier signals and telecommunication signals.

[0133] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A large model-based follow-up medical record generation method, characterized by, The method comprises: obtaining the initial diagnosis medical record of the target patient, and inputting the initial diagnosis medical record into a large language model for abstract extraction to obtain an initial diagnosis abstract; obtaining the doctor-patient conversation information of the target patient, and inputting the doctor-patient conversation information into the large language model for abstract extraction to obtain a re-visit conversation abstract; fusing the re-visit conversation abstract and the initial diagnosis abstract to obtain a re-visit medical record.

2. The large model-based return visit medical record generation method of claim 1, wherein, inputting the initial diagnosis medical record into a large language model for abstract extraction to obtain an initial diagnosis abstract, comprising: inputting the initial diagnosis medical record into the large language model for medical record splitting to obtain patient basic information and doctor treatment opinion content, and extracting the patient basic information according to a first preset field to obtain patient extracted information; obtaining the visit time information in the initial diagnosis medical record, and generating re-visit time information according to the visit time information; obtaining the initial diagnosis medical record of the target patient, and inputting the initial diagnosis medical record into a large language model for abstract extraction to obtain an initial diagnosis abstract; combining the patient extracted information, the re-visit time information and the medical order extracted information to obtain the initial diagnosis abstract.

3. The large model-based return visit medical record generation method of claim 1, wherein, inputting the doctor-patient conversation information into the large language model for abstract extraction to obtain a re-visit conversation abstract, comprising: the large language model generates a semantic abstract window according to a preset character length or a preset semantic unit quantity, and the overlap rate between adjacent semantic abstract windows is greater than a preset overlap rate; dividing the doctor-patient conversation information according to the semantic abstract window to obtain a window conversation segment, and extracting the window conversation segment to obtain a window conversation semantic; contradiction information in the window conversation semantic is eliminated, and the window conversation semantic after contradiction information elimination is subjected to key information completion; the window conversation semantic after key information completion is subjected to redundant content elimination to obtain a window corrected semantic, and the window corrected semantic is combined to obtain the re-visit conversation abstract.

4. The large model-based return visit medical record generation method of claim 3, wherein, eliminating contradiction information in the window conversation semantic, comprising: obtaining object description information of a contradiction object in the window conversation semantic, the contradiction object including patient symptoms, patient medical history and doctor diagnosis; determining whether there is a description contradiction between adjacent window conversation semantics according to the object description information; if there is a description contradiction between adjacent window conversation semantics, adjacent window conversation semantics are determined as contradiction window conversations; determining a contradiction type according to the object description information of adjacent contradiction window conversations, and eliminating contradiction information in adjacent contradiction window conversations according to the contradiction type.

5. The large model-based return visit medical record generation method of claim 4, wherein, determining a contradiction type according to the object description information of adjacent contradiction window conversations, and eliminating contradiction information in adjacent contradiction window conversations according to the contradiction type, comprising: combining the object description information of adjacent contradiction window conversations to obtain contradiction combination information, and matching the contradiction combination information with a contradiction query table to obtain the contradiction type; If the contradiction type is a negative conflict type, a previous window dialogue adjacent to the contradictory window dialogue is invalidly marked; If the contradiction type is a complement contradiction type, a previous window dialogue adjacent to the contradictory window dialogue is content-complemented according to a subsequent window dialogue adjacent to the contradictory window dialogue; If the contradiction type is a trace-back contradiction type, a contradictory original dialogue in the doctor-patient dialogue information is determined according to the window dialogues adjacent to the contradictory window dialogue, and the window dialogues adjacent to the contradictory window dialogue are corrected according to semantics of the contradictory original dialogue.

6. The large model-based return visit medical record generation method of claim 3, wherein, The window dialogue semantics after the contradiction information is eliminated are key information-complemented, including: Core dialogue semantics are obtained by performing core semantic positioning on the window dialogue semantics after the contradiction information is eliminated according to preset core keywords; Adjacent window dialogue semantics of the core dialogue semantics are determined as candidate gap semantics, and gap-sensitive keywords are matched with the candidate gap semantics; If the gap-sensitive keywords and the candidate gap semantics are not matched, a gap original dialogue in the doctor-patient dialogue information is located according to the candidate gap semantics, and the candidate gap semantics are information-complemented according to semantics of the gap original dialogue.

7. The large model-based return visit medical record generation method of claim 3, wherein, The window dialogue semantics after the key information is complemented are redundant content-eliminated to obtain window corrected semantics, including: Semantic similarities between the window dialogue semantics after the key information is complemented are calculated; If the semantic similarity is greater than a similarity threshold, a subsequent window dialogue semantic is deleted to obtain the window corrected semantics.

8. A large model-based return visit medical record generation system, characterized by, The system includes: An initial diagnosis extraction module configured to obtain an initial diagnosis medical record of a target patient, and input the initial diagnosis medical record into a large language model to perform abstract extraction to obtain an initial diagnosis abstract; A dialogue extraction module configured to obtain doctor-patient dialogue information of the target patient, and input the doctor-patient dialogue information into the large language model to perform abstract extraction to obtain a re-visit dialogue abstract; An abstract fusion module configured to perform abstract fusion on the re-visit dialogue abstract and the initial diagnosis abstract to obtain a re-visit medical record.

9. The large model based return visit medical record generation system of claim 8, wherein, The initial diagnosis extraction module is further configured to: Input the initial diagnosis medical record into the large language model to perform medical record splitting to obtain patient basic information and doctor treatment opinion content, and perform content extraction on the patient basic information according to a first preset field to obtain patient extraction information; Obtain visit time information in the initial diagnosis medical record, and generate re-visit time information according to the visit time information; Obtain an initial diagnosis medical order text of the target patient, and perform content extraction on the initial diagnosis medical order text and the doctor treatment opinion content according to a second preset field to obtain medical order extraction information; Combine the patient extraction information, the re-visit time information, and the medical order extraction information to obtain the initial diagnosis abstract.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-9. The computer program, when executed by a processor, implements the steps of the method of any one of claims 1 to 7. The computer program, when executed by a processor, implements the steps of the method of any one of claims 1 to 7.

Images