AI standardized patient training method and system based on large model
Through the AI standardized patient training method based on large models, using the case library and pre-trained models for semantic matching and instant feedback, the problem of low efficiency of traditional SP training is solved, and efficient training of medical students' consultation skills is achieved.
Patent Information
- Application Number
- CN202510832317.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-10-17
AI Technical Summary
The traditional standardized patient SP training model in medical education has the problems of high labor costs and difficulty in meeting the training needs of large-scale medical students, resulting in low consultation training efficiency.
A large-scale model-based AI standardized patient training method is adopted, and standardized medical records are obtained using the case library module. Semantic matching is performed through the pre-trained medical large model, and the target answer is output in combination with the digital human image. Physical examination data is provided based on click operations to achieve instant feedback and training.
It improves the efficiency of medical students' consultation training, provides realistic training experience and instant feedback, replaces traditional SP training, and realizes comprehensive consultation skills training.
Smart Images
Figure CN120809036A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer technology, and in particular to an AI standardized patient training method and system based on a large model. BACKGROUND
[0002] In medical education, inquiry is one of the basic skills that medical students must master, which requires students to accurately collect patient medical history, symptoms and related information, laying a solid foundation for subsequent diagnosis and treatment. With the rapid development of artificial intelligence technology, especially the emergence of medical large models, it provides a new way and method for inquiry training.
[0003] Currently, in the medical education system, traditional inquiry skill training mainly relies on standardized patient (SP) training mode, which makes healthy people or patients play the role of "standardized patient" to simulate the symptoms, medical history and emotions of real patients, providing inquiry practice objects for medical students.
[0004] However, the traditional inquiry training mode relying on standardized patients (SP) requires a large amount of manpower and is limited by the number of SPs, making it difficult to meet the training needs of a large number of medical students, resulting in low inquiry training efficiency of medical students. SUMMARY
[0005] The embodiments of the present application provide an AI standardized patient training method and system based on a large model, which can improve the inquiry training efficiency of medical students.
[0006] In a first aspect, the embodiments of the present application provide an AI standardized patient training method based on a large model, which comprises:
[0007] A1: obtaining at least one patient's standardized medical record by using a case library module, wherein the standardized medical record includes doctor-patient interactive question and answer pairs, preliminary diagnosis standard answers, differential diagnosis standard answers and treatment scheme standard answers;
[0008] A2: when receiving a target question of a current user, calling a pre-trained medical large model to perform semantic matching on the target question and the doctor-patient interactive question and answer pairs;
[0009] A3: when the target question and the doctor-patient interactive question and answer pairs can be matched, outputting a target answer corresponding to the target question by an interactive module in the form of a digital person;
[0010] A4: when receiving a click operation of the current user on the digital person, calling corresponding physical examination data from the case library based on the clicked part, and returning the data to the current user by using the interactive module;
[0011] A5: Obtain the current medical history of the current user, including the current preliminary diagnosis, current differential diagnosis, and current treatment recommendations, output based on the target answer and the physical examination data, to train an AI standardized patient.
[0012] Preferably,
[0013] Before the above A1, further comprising:
[0014] Obtaining original medical record data of the at least one patient and inputting the data into the case database module, wherein the original medical record data includes: imaging reports, laboratory data and nursing records;
[0015] Using optical character recognition (OCR) technology to identify unstructured text in the original medical record data, and converting it into structured medical record text using a large medical model;
[0016] Extracting standardized fields from the structured medical record text using the natural language processing (NLP) technology of the medical big model, wherein the standardized fields include: patient demographic information, the physical examination data, and medical procedure codes;
[0017] The imaging report, the laboratory data, the nursing record and the standardized fields are mapped across modalities using multi-source data fusion technology to generate the structured standardized medical record.
[0018] Preferably,
[0019] Said A2 includes:
[0020] D1: Upon receiving the current user's target question, the pre-trained Sentence-BERT algorithm of the medical large model is called to convert the target question and each doctor's question in the doctor-patient interactive question-answer pair into a target question vector and a doctor's question vector respectively;
[0021] D2: Based on the target question vector and the doctor question vector, determine the semantic similarity between the target question and the doctor question vector using the first formula;
[0022] The first formula is:
[0023] Wherein, cosθ is the semantic similarity between the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector;
[0024] D3: Based on the semantic similarity and a preset semantic similarity threshold, determine whether there is at least one semantic similarity greater than the preset semantic similarity threshold; if so, execute step D4; otherwise, execute step D5;
[0025] D4: arranging the matched doctor-patient interactive question and answer pairs in descending order of the semantic similarity, and taking the patient answer corresponding to the first in the semantic similarity order as the target answer;
[0026] D5: generating a corresponding answer based on the principle of evidence-based medicine and returning it to the current user.
[0027] Preferably,
[0028] After the A5, further comprising:
[0029] Converting the current medical record and the standardized medical record into a current medical record vector and a standardized medical record vector, respectively;
[0030] Determining the cosine similarity of the i-th item of the current medical record vector and the standardized medical record vector by using a second formula;
[0031] The second formula is: Wherein, is the cosine similarity of the i-th item of the current medical record vector and the standardized medical record vector, M is the current medical record vector, and N is the standardized medical record vector;
[0032] Determining the total score based on the preset scoring item weight and scoring rules by using a third formula;
[0033] The third formula is:
[0034] Wherein, T is the total score, ω i is the weight of the i-th scoring item, s i is the cosine similarity of the i-th item of the current medical record and the standardized medical record, and
[0035] In a second aspect, an AI standardized patient training system based on a large model is provided, and the system comprises:
[0036] A case library module is configured to obtain a standardized medical record of at least one patient, wherein the standardized medical record comprises a doctor-patient interactive question and answer pair, a standard answer for a preliminary diagnosis, a standard answer for differential diagnosis, and a standard answer for a treatment plan;
[0037] A semantic matching module is configured to, when receiving a target question of a current user, call a pre-trained medical large model to perform semantic matching on the target question and the doctor-patient interactive question and answer pair of the case library module;
[0038] An interaction module is configured to output a target answer corresponding to the target question in a digital human form when the semantic matching module determines that the target question and the doctor-patient interactive question and answer pair can be matched, and to retrieve corresponding physical examination data from a case library based on a clicked part when receiving a click operation of the current user on the digital human and return the current user the physical examination data.
[0039] A training module is configured to obtain a current medical record including a current preliminary diagnosis, a current differential diagnosis and a current treatment suggestion output by the target answer determined by the interaction module and the physical examination data, to train an AI standardized patient.
[0040] Preferably,
[0041] Further comprising a structured processing module before the case library module.
[0042] The structured processing module is configured to perform:
[0043] Obtain the original medical record data of the at least one patient and input the case library module, wherein the original medical record data includes an image report, laboratory data and nursing records.
[0044] Utilize an optical character recognition technology OCR to recognize unstructured text in the original medical record data, and utilize a medical large model to convert the unstructured text into structured medical record text.
[0045] Utilize a natural language processing technology NLP of the medical large model to extract standardized fields in the structured medical record text, wherein the standardized fields include patient demographic information, the physical examination data and medical operation coding.
[0046] Utilize a multi-source data fusion technology to cross-modally associate and map the image report, the laboratory data, the nursing records and the standardized fields to generate a structured standardized medical record.
[0047] Preferably,
[0048] The semantic matching module is configured to perform:
[0049] D1: When receiving a target question of a current user, calling a Sentence-BERT algorithm of a pre-trained medical large model to convert the target question and each doctor question in the doctor-patient interactive question and answer pair into a target question vector and a doctor question vector, respectively.
[0050] D2: Based on the target question vector and the doctor question vector, utilizing a first formula to determine a semantic similarity of the target question and the doctor question vector.
[0051] The first formula is:
[0052] Wherein, cosθ is the semantic similarity between the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector;
[0053] D3: Based on the semantic similarity and a preset semantic similarity threshold, determine whether there is at least one semantic similarity greater than the preset semantic similarity threshold; if so, execute step D4; otherwise, execute step D5;
[0054] D4: Arrange the matched doctor-patient interactive question-answer pairs in descending order of the semantic similarity, and take the patient answer corresponding to the first patient in the semantic similarity ranking as the target answer;
[0055] D5: Generate a corresponding answer based on the principles of evidence-based medicine and return it to the current user.
[0056] Preferably,
[0057] After the training module, it further includes: a scoring module;
[0058] The scoring module is used to perform:
[0059] Converting the current medical record and the standardized medical record into a current medical record vector and a standardized medical record vector respectively;
[0060] Determine the cosine similarity of the current medical record vector and the standardized medical record vector in the i-th item using the second formula;
[0061] The second formula is: in, is the cosine similarity between the current medical record vector and the standardized medical record vector in the i-th item, M is the current medical record vector, and N is the standardized medical record vector;
[0062] Based on the preset scoring item weights and scoring criteria, the third formula is used to determine the total score;
[0063] The third formula is:
[0064] Wherein, T is the total score, ω i is the weight of the i-th scoring item, s i is the cosine similarity between the current medical record and the standardized medical record in item i, and
[0065] In a third aspect, an embodiment of the present invention provides an AI standardized patient training system based on a large model, comprising: at least one memory and at least one processor;
[0066] the at least one memory, configured to store a machine readable program;
[0067] the at least one processor, configured to invoke the machine readable program to execute the method of any one of the first aspect.
[0068] In a fourth aspect, an embodiment of the present application provides a computer readable medium, and the computer readable medium stores computer instructions, and the computer instructions, when executed by a processor, cause the processor to execute the method of any one of the first aspect.
[0069] The embodiment of the present application provides an AI standardized patient training method and system based on a large model. The method first acquires standardized medical records containing doctor-patient interactive question and answer pairs, preliminary diagnosis standard answers, differential diagnosis standard answers and treatment scheme standard answers by using a case library module, so as to provide standardized and reusable case resources for training. When a user proposes a target question, the system calls a pre-trained medical large model to perform semantic matching, matches the target question with the doctor-patient interactive question and answer pairs in the case library, and outputs the target answer in real time through the digital human image through the interactive module, so as to realize instant feedback and avoid the time loss of waiting for manual feedback in traditional training. If the user performs a clicking operation on the digital human, the system calls corresponding physical examination data from the case library based on the clicked part, so that medical students can repeatedly practice the examination and operation in the virtual environment. Finally, the current medical record including the current preliminary diagnosis, differential diagnosis and treatment suggestion is output based on the target answer and the physical examination data. Through the above process, a highly realistic examination training experience is provided, and through the standardized training process and the instant feedback mechanism, the traditional SP training is replaced, the examination skills of medical students are trained in an all-round way, and therefore the examination training efficiency of the medical students is significantly improved. BRIEF DESCRIPTION OF DRAWINGS
[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.
[0071] Figure 1 is a flowchart of an AI standardized patient training method based on a large model provided by an embodiment of the present application;
[0072] Figure 2 is a flowchart of another AI standardized patient training method based on a large model provided by an embodiment of the present application;
[0073] Figure 3is a schematic diagram of an AI standardized patient training system based on a large model provided by an embodiment of the present application;
[0074] Figure 4 is a schematic diagram of another AI standardized patient training system based on a large model provided by an embodiment of the present application;
[0075] Figure 5 is a schematic diagram of still another AI standardized patient training system based on a large model provided by an embodiment of the present application. DETAILED DESCRIPTION
[0076] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.
[0077] As shown in Figure 1 , the present application provides an AI standardized patient training method based on a large model, which can include the following steps:
[0078] Step 101: obtaining at least one patient's standardized medical record by using a case library module, wherein the standardized medical record includes doctor-patient interactive question and answer pairs, preliminary diagnosis standard answers, differential diagnosis standard answers and treatment scheme standard answers;
[0079] Step 102: when receiving a target question of a current user, calling a pre-trained medical large model to perform semantic matching on the target question and the doctor-patient interactive question and answer pairs;
[0080] Step 103: when the target question and the doctor-patient interactive question and answer pairs can be matched, outputting a target answer corresponding to the target question by a digital human image through an interaction module;
[0081] Step 104: when receiving a clicking operation of the current user on the digital human, calling corresponding physical examination data from the case library based on the clicked part and returning the data to the current user by using the interaction module;
[0082] Step 105: obtaining a current medical record including a current preliminary diagnosis, a current differential diagnosis and a current treatment suggestion output by the current user based on the target answer and the physical examination data, to train an AI standardized patient.
[0083] In the embodiment of the present application, a large model-based AI standardized patient training method is provided. First, the case library module is used to obtain standardized medical records containing doctor-patient interactive question and answer pairs, preliminary diagnosis standard answers, differential diagnosis standard answers and treatment scheme standard answers, providing standardized and reusable case resources for training. When the user raises a target question, the system calls the pre-trained medical large model for semantic matching, matches the target question with the doctor-patient interactive question and answer pairs in the case library, and outputs the target answer in real time through the digital human image through the interactive module, to realize instant feedback and avoid the time loss of waiting for manual feedback in traditional training. If the user clicks on the digital human, the system retrieves the corresponding physical examination data from the case library based on the clicked part, so that medical students can repeatedly practice the examination and examination operation in the virtual environment. Finally, the current medical record including the current preliminary diagnosis, differential diagnosis and treatment suggestion is output based on the target answer and the physical examination data. Through the above process, not only a highly realistic examination training experience is provided, but also the traditional SP training is replaced through the standardized training process and the instant feedback mechanism, realizing the all-round cultivation of the medical students' examination skills, thereby significantly improving the examination training efficiency of the medical students.
[0084] To ensure the availability and integrity of data, in an embodiment of the present application, the above embodiment further comprises, before the step 101:
[0085] Obtaining the original medical record data of the at least one patient and inputting the case library module, wherein the original medical record data includes: image report, laboratory data and nursing record;
[0086] Using optical character recognition technology OCR to recognize the unstructured text in the original medical record data, and using a medical large model to convert it into structured medical record text;
[0087] Using the natural language processing technology NLP of the medical large model to extract the standardized fields in the structured medical record text, wherein the standardized fields include: patient demographic information, physical examination data and medical operation coding;
[0088] Using multi-source data fusion technology, the image report, the laboratory data, the nursing record and the standardized field are cross-modally associated and mapped to generate structured standardized medical records.
[0089] In the embodiment of the present application, when the original medical record data of at least one patient is input into the case library module, the system first recognizes the unstructured text (such as chief complaint, history of present illness, etc. free text) in the original medical record by OCR optical character recognition technology, and performs context semantic analysis by using a medical large model, identifies the implicit logical relationship and converts it into structured medical record text. Then, by means of natural language processing technology NLP of the medical large model (such as the open source DeepSeek model), the structured medical record text is subjected to multi-level semantic analysis and medical entity recognition, and the standardized fields such as patient demographic information (such as age, gender, etc.), physical examination data (body temperature, thyroid, tonsils, etc.) and medical operation coding (ICD-10 operation classification) are accurately extracted. Through multi-level semantic analysis and medical entity recognition, the accuracy of medical record information disassembly and classification is realized, ensuring data availability, integrity and consistency of medical logic. Finally, by using multi-source data fusion technology, the image report (such as CT), laboratory data (such as blood routine value) and nursing record are cross-modally associated and mapped with the above standardized fields to form a standardized medical record containing a full-dimensional medical record portrait, and data preparation is completed for the inquiry.
[0090] In order to ensure the accuracy of the inquiry process, in an embodiment of the present application, step 102 in the above embodiment can specifically include the following steps:
[0091] D1: When receiving a target question of a current user, calling a Sentence-BERT algorithm of a pre-trained medical large model to convert the target question and each doctor question in the doctor-patient interactive question and answer pair into a target question vector and a doctor question vector, respectively;
[0092] D2: Based on the target question vector and the doctor question vector, determining a semantic similarity between the target question and the doctor question vector by using a first formula;
[0093] The first formula is:
[0094] Wherein, cosθ is the semantic similarity between the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector.
[0095] D3: Based on the semantic similarity and a preset semantic similarity threshold, determining whether there is at least one semantic similarity greater than the preset semantic similarity threshold, if yes, performing step D4, otherwise, performing step D5;
[0096] D4: Arranging the matched doctor-patient interactive question and answer pair in descending order of the semantic similarity, and taking the patient answer corresponding to the first position in the semantic similarity order as the target answer;
[0097] D5: generating a reasonable answer based on the principle of evidence-based medicine and returning it to the current user.
[0098] In an embodiment of the present application, when the system receives the target question of the current user, the Sentence-BERT algorithm built in the pre-trained medical large model is called to convert the target question and each doctor question in the doctor-patient interactive question and answer pair into a target question vector and a doctor question vector in a high-dimensional space, respectively, and then the semantic similarity between the target question vector and each doctor question vector is calculated based on the first formula. The calculated semantic similarity is compared with the preset threshold, the matching doctor-patient interactive question and answer pair is sorted in descending order of semantic similarity, and the patient answer in the question and answer pair at the top of the sorting is taken as the target answer. If all semantic similarities do not exceed the threshold, the large model generates a reasonable answer based on the principle of evidence-based medicine and returns it to the user, avoiding the occurrence of incorrect answers, while complying with the logic of realistic expression (for example, if the user deliberately asks for a difficult question "refuse treatment, please go home!", the large model should reply in a tone similar to "doctor, if I don't go for treatment, I think it will get worse. I still follow your advice and accept examination and treatment. You can rest assured that I will cooperate.") to ensure the smooth progress of the consultation.
[0099] In order to improve the medical students' consultation skills, in an embodiment of the present application, after step 105 in the above embodiment, further comprising:
[0100] converting the current medical record and the standardized medical record into a current medical record vector and a standardized medical record vector, respectively;
[0101] determining the cosine similarity of the current medical record vector and the standardized medical record vector at the i-th item by using a second formula;
[0102] The second formula is: wherein, is the cosine similarity of the current medical record vector and the standardized medical record vector at the i-th item, M is the current medical record vector, and N is the standardized medical record vector.
[0103] determining the total score by using a third formula based on the preset scoring item weight and scoring rules;
[0104] The third formula is:
[0105] wherein, T is the total score, ω i is the weight of the i-th scoring item, s i is the cosine similarity of the current medical record and the standardized medical record at the i-th item, and
[0106] In the embodiment of the present application, when the system receives the current medical record, the semantic representation capability of the medical large model is used to convert the current medical record and the standardized medical record generated in the case library module into a current medical record vector and a standardized medical record vector respectively, then the cosine similarity of the two vectors in the i-th scoring dimension is calculated using the second formula to quantify the semantic matching degree of the current medical record and the standardized medical record in this dimension; finally, based on the preset scoring item weight (such as 0.2 for opening speech and 0.1 for present history), the total score is calculated by the third formula based on the scoring rules. The above process combines the preset scoring item weight for comprehensive scoring, which can help medical students improve their interrogation skills. The scoring module serves as an "intelligent coach" for AI standardized patients, and based on the case library standardized medical record data, scoring rules and user training data, an evaluation system is constructed. Through natural language processing (NLP) and machine learning technology, dynamic evaluation, error correction and ability advancement guidance of medical decision-making are realized, covering core dimensions such as symptom collection, diagnosis reasoning, treatment suggestion and doctor-patient communication, helping medical students improve their clinical comprehensive ability.
[0107] As Figure 2 shown, in order to more clearly illustrate the technical solutions and advantages of the present application, the following will provide a detailed description of an AI standardized patient training method based on a large model, which can specifically include the following steps:
[0108] Step 201: Obtain the original medical record data of at least one patient and input it into the case library module, wherein the original medical record data includes: image report, laboratory data and nursing record;
[0109] Step 202: Use optical character recognition (OCR) to recognize the unstructured text in the original medical record data, and use a medical large model to convert it into structured medical record text;
[0110] Step 203: Use natural language processing (NLP) technology of the medical large model to extract standardized fields in the structured medical record text, wherein the standardized fields include: patient demographic information, physical examination data and medical operation coding;
[0111] Step 204: Use multi-source data fusion technology to cross-modally associate and map the image report, laboratory data, nursing record and standardized field to generate a structured standardized medical record;
[0112] Step 205: Use the case library module to obtain the standardized medical record of at least one patient, wherein the standardized medical record includes doctor-patient interactive question and answer pairs, preliminary diagnosis standard answers, differential diagnosis standard answers and treatment plan standard answers;
[0113] Step 206: Upon receiving the target question of the current user, the Sentence-BERT algorithm of the pre-trained medical large model is called to convert the target question and each doctor question in the doctor-patient interactive question and answer pair into a target question vector and a doctor question vector, respectively;
[0114] Step 207: Based on the target question vector and the doctor question vector, the semantic similarity between the target question and the doctor question vector is determined using a first formula;
[0115] Specifically, the first formula is:
[0116] Wherein, cosθ is the semantic similarity between the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector.
[0117] Step 208: Based on the semantic similarity and a preset semantic similarity threshold, it is determined whether there is at least one semantic similarity greater than the preset semantic similarity threshold. If yes, step 209 is performed, otherwise, step 215 is performed.
[0118] Step 209: The matched doctor-patient interactive question and answer pair is arranged in descending order of semantic similarity, and the patient answer corresponding to the first place in the semantic similarity order is taken as the target answer. The target question corresponding to the target answer is output by the interactive module in the form of a digital human image.
[0119] Step 210: Upon receiving a click operation of the current user on the digital human, the corresponding physical examination data is retrieved from the case library based on the clicked part, and is returned to the current user by the interactive module.
[0120] Specifically, when auxiliary examination is needed, the interactive module can send the auxiliary examination application issued by the user to the case library, and the case library queries the auxiliary examination (such as examination report) configured for the case and feeds back to the interactive module for the user to view and refer.
[0121] For example, clicking on the forehead shows the body temperature, and clicking on the knee shows the prompt word "pain".
[0122] Step 211: The current user outputs the current medical record including the current preliminary diagnosis, the current differential diagnosis and the current treatment suggestion based on the target answer and the physical examination data, to train the AI standardized patient.
[0123] Step 212: The current medical record and the standardized medical record are respectively converted into a current medical record vector and a standardized medical record vector.
[0124] Step 213: The cosine similarity of the current medical record vector and the standardized medical record vector at the i-th item is determined using a second formula.
[0125] Specifically, the second formula is: wherein, is the cosine similarity of the i-th item of the current medical record vector and the normalized medical record vector, M is the current medical record vector, and N is the normalized medical record vector;
[0126] Step 214: determining a total score based on the preset scoring item weight and scoring rules by using a third formula;
[0127] Specifically, the third formula is:
[0128] wherein, T is the total score, ω i is the weight of the i-th scoring item, s i is the cosine similarity of the i-th item of the current medical record and the normalized medical record, and
[0129] For example, the scoring categories can include "opening and self-introduction", "present illness history", "past medical history", "personal history", "family history", "questioning order and logic", and "irrelevant questions". Each category includes several scoring items, such as "opening and self-introduction" containing "language professionalism and polite attitude", and the scoring rule is "using polite language such as hello, good morning, and may I ask". This item scores 2 points. The large model module will perform semantic recognition on the communication records collected by the client one by one, and if it meets the standard, it will be counted as 2 points.
[0130] Step 215: generating a corresponding answer based on the principle of evidence-based medicine and returning it to the current user.
[0131] As shown in Figure 3 the present embodiment provides an AI standardized patient training system based on a large model, which comprises:
[0132] A case library module 301 is configured to obtain a standardized medical record of at least one patient, wherein the standardized medical record comprises doctor-patient interactive question and answer pairs, standard answers for preliminary diagnosis, standard answers for differential diagnosis, and standard answers for treatment plans.
[0133] A semantic matching module 302 is configured to, when receiving a target question of a current user, call a pre-trained medical large model to perform semantic matching on the target question and the doctor-patient interactive question and answer pairs of the case library module 301.
[0134] An interaction module 303 is configured to, when the semantic matching module determines that the target question and the doctor-patient interactive question and answer pairs can be matched, output a target answer corresponding to the target question in the form of a digital human image; and when receiving a clicking operation of the current user on the digital human, call corresponding physical examination data from the case library based on the clicked part and return it to the current user.
[0135] The training module 304 is configured to obtain the current medical record including the current preliminary diagnosis, the current differential diagnosis and the current treatment suggestion output by the physical examination data based on the target answer determined by the interaction module 303 for the current user, so as to train the AI standardized patient.
[0136] Based on Figure 3 As shown in one of the large model-based AI standardized patient training systems, as shown in Figure 4 Before the case library module 301, further comprising a structured processing module 305.
[0137] The structured processing module 305 is configured to perform:
[0138] Obtain the original medical record data of the at least one patient and input the case library module, wherein the original medical record data includes: image report, laboratory data and nursing record;
[0139] The non-structured text in the original medical record data is recognized by using an optical character recognition technology OCR, and is converted into structured medical record text by using a medical large model;
[0140] The standardized fields in the structured medical record text are extracted by using a natural language processing technology NLP of the medical large model, wherein the standardized fields include: patient demographic information, physical examination data and medical operation coding;
[0141] By using a multi-source data fusion technology, the image report, the laboratory data, the nursing record and the standardized field are cross-modally associated and mapped to generate a structured standardized medical record.
[0142] As shown in the semantic matching module 302, configured to perform: Figure 4
[0143] D1: When receiving the target question of the current user, the Sentence-BERT algorithm of the pre-trained medical large model is called to convert the target question and each doctor question in the doctor-patient interaction question and answer pair into a target question vector and a doctor question vector, respectively;
[0144] D2: Based on the target question vector and the doctor question vector, the semantic similarity of the target question and the doctor question vector is determined by using a first formula;
[0145] The first formula is:
[0146] Wherein, cosθ is the semantic similarity of the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector.
[0147] D3: determining whether there is at least one semantic similarity greater than the preset semantic similarity threshold based on the semantic similarity and the preset semantic similarity threshold, if yes, performing step D4, otherwise, performing step D5;
[0148] D4: arranging the matched doctor-patient interactive question and answer pairs in descending order of the semantic similarity, and taking the patient answer corresponding to the first in the semantic similarity order as the target answer;
[0149] D5: generating a corresponding answer based on the principle of evidence-based medicine and returning it to the current user.
[0150] Based on Figure 4 An AI standardized patient training system based on a large model is shown in FIG. 1, which comprises a training module 304. Figure 5 As shown in FIG. 1, after the training module 304, the system further comprises a scoring module 306.
[0151] The scoring module 306 is configured to perform:
[0152] convert the current medical record and the standardized medical record into a current medical record vector and a standardized medical record vector, respectively;
[0153] determine the cosine similarity of the current medical record vector and the standardized medical record vector at the ith item by using a second formula;
[0154] The second formula is:
[0155] wherein, cosφ is the cosine similarity of the current medical record vector and the standardized medical record vector at the ith item, M is the current medical record vector, and N is the standardized medical record vector.
[0156] determine the total score by using a third formula based on preset scoring item weights and scoring details;
[0157] The third formula is:
[0158] wherein, T is the total score, ω i is the weight of the ith scoring item, s i is the cosine similarity of the current medical record and the standardized medical record at the ith item, and
[0159] It can be understood that the structure of the embodiments of the present application does not constitute a specific limitation of an AI standardized patient training system based on a large model. In other embodiments of the present application, an AI standardized patient training system based on a large model can include more or fewer components than the illustration, or combine certain components, or split certain components, or different component arrangements. The illustrated components can be implemented in hardware, software, or a combination of software and hardware.
[0160] The information interaction, execution process, and the like between the units in the above device are based on the same concept as the method embodiments of the present application, and the specific content can be referred to the description in the method embodiments of the present application, which will not be described here.
[0161] The embodiments of the present application also provide an AI standardized patient training system based on a large model, comprising at least one memory and at least one processor.
[0162] The at least one memory is used to store machine readable programs.
[0163] The at least one processor is used to call the machine readable programs and execute the AI standardized patient training method based on a large model in any embodiment of the present application.
[0164] The embodiments of the present application also provide a computer readable medium, and the computer readable medium stores computer instructions, and the computer instructions make the processor execute the AI standardized patient training method based on a large model in any embodiment of the present application when executed by the processor.
[0165] Specifically, a system or device equipped with a storage medium can be provided, and the storage medium stores software program codes for realizing the functions of any one of the above embodiments, and the computer (or CPU or MPU) of the system or device reads and executes the program codes stored in the storage medium.
[0166] In this case, the program codes read from the storage medium can realize the functions of any one of the above embodiments, and therefore the program codes and the storage medium storing the program codes constitute a part of the present application.
[0167] The storage medium for providing the program codes includes a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Alternatively, the program codes can be downloaded from a server computer via a communication network.
[0168] Moreover, it should be appreciated that, not only the actual operations of part or all of the above-mentioned embodiments can be performed by executing the program codes read by the computer, but also the operating system or the like operating on the computer can be caused to perform the actual operations based on the instructions of the program codes, thereby realizing the functions of any one of the above-mentioned embodiments.
[0169] Moreover, it should be appreciated that, not only the actual operations of part or all of the above-mentioned embodiments can be performed by executing the program codes read by the computer, but also the operating system or the like operating on the computer can be caused to perform the actual operations based on the instructions of the program codes, thereby realizing the functions of any one of the above-mentioned embodiments.
[0170] The various embodiments of the present application have at least the following beneficial effects:
[0171] 1. In the embodiments of the present application, an AI standardized patient training method and system based on a large model are provided. The method first acquires standardized medical records containing doctor-patient interactive question and answer pairs, preliminary diagnosis standard answers, differential diagnosis standard answers, and treatment plan standard answers using a case library module, providing standardized and reusable case resources for training. When the user raises a target question, the system calls a pre-trained medical large model for semantic matching, matches the target question with the doctor-patient interactive question and answer pairs in the case library, and outputs the target answer in real time through the digital human image in the interactive module to realize instant feedback and avoid the time loss of waiting for manual feedback in traditional training. If the user performs a clicking operation on the digital human, the system retrieves the corresponding physical examination data from the case library based on the clicked part, allowing medical students to repeatedly practice the examination and operation in a virtual environment. Finally, the current medical record including the current preliminary diagnosis, differential diagnosis, and treatment suggestion is output based on the target answer and physical examination data. Through the above process, not only a highly realistic examination training experience is provided, but also the traditional SP training is replaced through the standardized training process and instant feedback mechanism, realizing the all-round cultivation of the medical students' examination skills, thereby significantly improving the examination training efficiency of the medical students.
[0172] 2. In the embodiment of the present application, when the original medical record data of at least one patient is input into the case library module, the system first recognizes the unstructured text in the original medical record through the OCR optical character recognition technology, and performs context semantic analysis using the medical large model to recognize the implicit logical relationship and convert it into structured medical record text. Then, with the help of the natural language processing technology NLP of the medical large model, the structured medical record text is subjected to multi-level semantic analysis and medical entity recognition, and the standardized fields such as patient demographic information and physical examination data are accurately extracted. Through multi-level semantic analysis and medical entity recognition, the accurate disassembly and classification of medical record information are realized, ensuring the data availability, integrity and consistency of medical logic. Finally, using the multi-source data fusion technology, the image report, laboratory data and nursing record are cross-modal associated and mapped with the above-mentioned standardized fields to form a standardized medical record containing a full-dimensional medical record portrait, and to prepare data for the inquiry;
[0173] 3. In the embodiment of the present application, when the system receives the target question of the current user, the pre-trained Sentence-BERT algorithm built in the medical large model is called to convert the target question and each doctor question in the doctor-patient interactive question and answer pair into a target question vector and a doctor question vector in a high-dimensional space, respectively. Then, based on the first formula, the semantic similarity of the target question vector and each doctor question vector is calculated, and the calculated semantic similarity is compared with the preset threshold. The matching doctor-patient interactive question and answer pair is sorted in descending order of semantic similarity, and the patient answer in the top question and answer pair is taken as the target answer. If all semantic similarities do not exceed the threshold, the large model generates a reasonable answer in accordance with the principles of evidence-based medicine and returns it to the user, avoiding incorrect answers, and at the same time, it should conform to the realistic expression logic to ensure the smooth progress of the inquiry.
[0174] It should be noted that not all steps and modules in the above processes and system structure diagrams are necessary, and some steps or modules can be omitted according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in each of the above embodiments can be a physical structure or a logical structure, i.e. some modules can be implemented by the same physical entity, or some modules can be implemented by multiple physical entities, or they can be implemented by some components in multiple independent devices.
[0175] In the above embodiments, a hardware unit can be implemented by mechanical means or by electrical means, or by a combination of mechanical and electrical means. For example, a hardware unit can include a permanent, dedicated electrical circuit or logic (such as a dedicated processor, FPGA, or ASIC) to perform the corresponding operation. A hardware unit can also include a programmable electrical circuit or logic (such as a general-purpose processor or other programmable processor) that can be temporarily configured by software to perform the corresponding operation. The specific implementation (mechanical, or dedicated permanent electrical circuit, or temporarily configured electrical circuit) can be determined based on cost and time considerations.
[0176] The above description is merely illustrative of the application, and is not to be taken in a limiting sense. It is being contemplated that various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the application.
Claims
1. The AI standardized patient training method based on a large model is characterized by: The method includes: A1: Obtain standardized medical records of at least one patient using a case database module, wherein the standardized medical records include doctor-patient interactive question-and-answer pairs, standard answers to preliminary diagnoses, standard answers to differential diagnoses, and standard answers to treatment plans; A2: Upon receiving the current user's target question, the pre-trained medical model is called to perform semantic matching between the target question and the doctor-patient interactive question-answer pair; A3: When the target question and the doctor-patient interactive question-answer pair are matched, the target answer corresponding to the target question is outputted by the interactive module using a digital human image; A4: When receiving a click operation on the digital human by the current user, the corresponding physical examination data is retrieved from the case database based on the clicked part, and returned to the current user using the interactive module; A5: Obtain the current medical history of the current user, including the current preliminary diagnosis, current differential diagnosis, and current treatment recommendations, output based on the target answer and the physical examination data, to train an AI standardized patient.
2. The method according to claim 1, characterized in that Before the above A1, further comprising: Obtaining original medical record data of the at least one patient and inputting the data into the case database module, wherein the original medical record data includes: imaging reports, laboratory data and nursing records; Using optical character recognition (OCR) technology to identify unstructured text in the original medical record data, and converting it into structured medical record text using a large medical model; Extracting standardized fields from the structured medical record text using the natural language processing (NLP) technology of the medical big model, wherein the standardized fields include: patient demographic information, the physical examination data, and medical procedure codes; The imaging report, the laboratory data, the nursing record and the standardized fields are mapped across modalities using multi-source data fusion technology to generate the structured standardized medical record.
3. The method according to claim 1, characterized in that Said A2 includes: D1: Upon receiving the current user's target question, the pre-trained Sentence-BERT algorithm of the medical large model is called to convert the target question and each doctor's question in the doctor-patient interactive question-answer pair into a target question vector and a doctor's question vector respectively; D2: Based on the target question vector and the doctor question vector, determine the semantic similarity between the target question and the doctor question vector using the first formula; The first formula is: Wherein, cosθ is the semantic similarity between the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector; D3: Based on the semantic similarity and a preset semantic similarity threshold, determine whether there is at least one semantic similarity greater than the preset semantic similarity threshold; if so, execute step D4; otherwise, execute step D5; D4: Arrange the matched doctor-patient interactive question-answer pairs in descending order of the semantic similarity, and take the patient answer corresponding to the first patient in the semantic similarity ranking as the target answer; D5: Generate a corresponding answer based on the principles of evidence-based medicine and return it to the current user.
4. The method according to any one of claims 1 to 3, characterized in that After A5, the method further includes: Converting the current medical record and the standardized medical record into a current medical record vector and a standardized medical record vector respectively; Determine the cosine similarity of the current medical record vector and the standardized medical record vector in the i-th item using the second formula; The second formula is: in, is the cosine similarity between the current medical record vector and the standardized medical record vector in the i-th item, M is the current medical record vector, and N is the standardized medical record vector; Based on the preset scoring item weights and scoring criteria, the third formula is used to determine the total score; The third formula is: Wherein, T is the total score, ω i is the weight of the i-th scoring item, s i is the cosine similarity between the current medical record and the standardized medical record in item i, and 5. AI standardized patient training system based on large model, characterized by: The system includes: Case database module: used to obtain standardized medical records of at least one patient, wherein the standardized medical records include doctor-patient interactive question-and-answer pairs, standard answers to preliminary diagnosis, standard answers to differential diagnosis, and standard answers to treatment plans; Semantic matching module: used for calling the pre-trained medical big model to perform semantic matching on the target question and the doctor-patient interactive question-answer pair in the case database module when receiving the target question of the current user; Interaction module: used to output the target answer corresponding to the target question using the digital human image when the semantic matching module determines that the target question and the doctor-patient interactive question-answer pair are compatible; upon receiving a click operation on the digital human by the current user, retrieve the corresponding physical examination data from the case database based on the clicked part and return it to the current user; Training module: used to obtain the target answer determined by the current user based on the interaction module and the current medical record output by the physical examination data, including the current preliminary diagnosis, current differential diagnosis and current treatment recommendations, to train AI standardized patients.
6. The system according to claim 5, characterized in that Before the case library module, it further includes: a structured processing module; The structured processing module is used to perform: Obtaining original medical record data of the at least one patient and inputting the data into the case database module, wherein the original medical record data includes: imaging reports, laboratory data and nursing records; Using optical character recognition (OCR) technology to identify unstructured text in the original medical record data, and converting it into structured medical record text using a large medical model; Extracting standardized fields from the structured medical record text using the natural language processing (NLP) technology of the medical big model, wherein the standardized fields include: patient demographic information, the physical examination data, and medical procedure codes; The imaging report, the laboratory data, the nursing record and the standardized fields are mapped across modalities using multi-source data fusion technology to generate the structured standardized medical record.
7. The system according to claim 5, characterized in that The semantic matching module is used to perform: D1: Upon receiving the current user's target question, the pre-trained Sentence-BERT algorithm of the medical large model is called to convert the target question and each doctor's question in the doctor-patient interactive question-answer pair into a target question vector and a doctor's question vector respectively; D2: Based on the target question vector and the doctor question vector, determine the semantic similarity between the target question and the doctor question vector using the first formula; The first formula is: Wherein, cosθ is the semantic similarity between the target question vector and the doctor question vector, A is the target question vector, and B is the doctor question vector; D3: Based on the semantic similarity and a preset semantic similarity threshold, determine whether there is at least one semantic similarity greater than the preset semantic similarity threshold; if so, execute step D4; otherwise, execute step D5; D4: Arrange the matched doctor-patient interactive question-answer pairs in descending order of the semantic similarity, and take the patient answer corresponding to the first patient in the semantic similarity ranking as the target answer; D5: Generate a corresponding answer based on the principles of evidence-based medicine and return it to the current user.
8. The system according to any one of claims 5 to 7, characterized in that: After the training module, it further includes: a scoring module; The scoring module is used to perform: Converting the current medical record and the standardized medical record into a current medical record vector and a standardized medical record vector respectively; Determine the cosine similarity of the current medical record vector and the standardized medical record vector in the i-th item using the second formula; The second formula is: in, is the cosine similarity between the current medical record vector and the standardized medical record vector in the i-th item, M is the current medical record vector, and N is the standardized medical record vector; Based on the preset scoring item weights and scoring criteria, the third formula is used to determine the total score; The third formula is: Wherein, T is the total score, ω i is the weight of the i-th scoring item, s i is the cosine similarity between the current medical record and the standardized medical record in item i, and 9. AI standardized patient training system based on large model, characterized by: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 4.
10. A computer-readable medium, characterized in that The computer-readable medium stores computer instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 4.