Model training method and apparatus, device, and storage medium

Abstract

One or more embodiments of the specification provide a model training method, device and storage medium for training a medical report interpretation large language model, so that the trained medical report interpretation large language model can output a target medical report interpretation result corresponding to the input target medical report and target question to be answered, and can also output a reasoning process, i.e., a target thinking chain, for describing the target medical report interpretation result corresponding to the target question to be answered based on the text content in the target medical report. In the specific training process, a first training sample containing a medical report, a question to be answered and a first medical report interpretation result is first obtained; then a reference thinking chain corresponding to the first training sample is output by the target large language model under the supervision of the first medical report interpretation result; then the reference thinking chain is added to the first training sample to generate a second training sample, and the medical report interpretation large language model is trained through the second training sample.

Description

TECHNICAL FIELD

[0001] One or more embodiments of the present specification relate to the technical field of artificial intelligence, and in particular, to a model training method and device, equipment and a storage medium. BACKGROUND

[0002] In a medical scenario, medical personnel usually conducts an inquiry on a patient or uses a medical device to conduct a medical examination on the patient in order to evaluate the physical health status of the patient. The medical examination can be divided into various types, such as X-ray, computed tomography (CT), magnetic resonance imaging, ultrasonic wave, and other imaging examinations, blood tests, urine tests, fecal examinations, and other laboratory examinations, and electrocardiogram, lung function test, endoscopy, and other functional examinations. After the inquiry or medical examination on the patient, the medical personnel usually issues a corresponding medical report to record the health status of the patient, examination conclusion, diagnosis conclusion, and treatment suggestion, etc.

[0003] Since the medical report involves a large number of medical terms, professional indicators, and clinical judgment logic, etc., it is usually necessary for medical personnel with medical professional knowledge to interpret the medical report, so as to convert the medical terms, professional indicators, and clinical judgment logic, etc. in the medical report into popular and easy-to-understand descriptions, so that the patient can understand their own health status and diagnosis and treatment information. Therefore, how to automatically interpret the medical report into content understandable by the patient has become a problem to be solved. SUMMARY

[0004] One or more embodiments of the present specification provide a model training method, device, equipment and storage medium for training a medical report interpretation model capable of automatically interpreting a medical report.

[0005] In a first aspect, one or more embodiments of the present specification provide a model training method, which comprises:

[0006] obtaining a first training sample, the first training sample comprising: a medical report, a to-be-answered question, and a first medical report interpretation result obtained by answering the to-be-answered question based on the text content in the medical report;

[0007] inputting a thinking chain construction prompt word containing the first training sample into a target large language model, so that the target large language model outputs a reference thinking chain and a second medical report interpretation result under the supervision of the first medical report interpretation result, the reference thinking chain being used to describe the reasoning process of obtaining the second medical report interpretation result corresponding to the to-be-answered question based on the text content in the medical report;

[0008] if it is determined that the second medical report interpretation result is correct according to the first medical report interpretation result, the reference thought chain is added to the first training sample to generate a second training sample;

[0009] The medical report interpretation large language model is trained through the second training sample, and the medical report interpretation large language model is used for outputting a target medical report interpretation result and a target thought chain according to a target medical report and a target question to be answered input by a user, the target thought chain being used for describing a reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered based on text content in the target medical report.

[0010] In a second aspect, one or more embodiments of the present specification provide a model training apparatus, the apparatus comprising:

[0011] The sample obtaining module is configured to obtain a first training sample, the first training sample comprising a medical report, a question to be answered, and a first medical report interpretation result obtained by answering the question to be answered based on text content in the medical report;

[0012] The sample constructing module is configured to input a thought chain constructing prompt word comprising the first training sample into a target large language model, so that the target large language model outputs a reference thought chain and a second medical report interpretation result under the supervision of the first medical report interpretation result, the reference thought chain being used for describing a reasoning process of obtaining the second medical report interpretation result corresponding to the question to be answered based on the text content in the medical report; if it is determined that the second medical report interpretation result is correct according to the first medical report interpretation result, the reference thought chain is added to the first training sample to generate a second training sample;

[0013] The model training module is configured to train a medical report interpretation large language model through the second training sample, the medical report interpretation large language model being used for outputting a target medical report interpretation result and a target thought chain according to a target medical report and a target question to be answered input by a user, the target thought chain being used for describing a reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered based on text content in the target medical report.

[0014] In a third aspect, one or more embodiments of the present specification provide an electronic device, comprising a memory, a processor, and a communication interface; wherein the memory has stored thereon a computer program, when the computer program is executed by the processor, the processor can at least implement the model training method according to the first aspect.

[0015] In a fourth aspect, one or more embodiments of the present specification provide a computer-readable storage medium, having stored thereon a computer program, which, when executed by a processor of an electronic device, causes the processor to at least implement the model training method according to the first aspect.

[0016] In a fifth aspect, one or more embodiments of the present specification provide a computer program product, comprising: a computer program or instructions, which, when executed by a processor of an electronic device, causes the processor to at least implement the model training method according to the first aspect.

[0017] The model training scheme provided by one or more embodiments of the present specification is used to train a medical report interpretation large language model, and the training target is to make the medical report interpretation large language model obtained by training not only output a target medical report interpretation result corresponding to a target medical report and a target question to be answered, but also output a target thinking chain for describing a reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered based on the text content in the target medical report. Thus, on the one hand, the target medical report interpretation result can be used to answer the target question to be answered in the target medical report in a popular and easy-to-understand manner; on the other hand, the target thinking chain can be used to explain the multiple reasoning steps of the medical report interpretation large language model for obtaining the target medical report interpretation result based on the target medical report and the target question to be answered, enhance the logical relationship between the target medical report interpretation result and the target medical report and the target question to be answered perceived by the user, and improve the explainability of the target medical report interpretation result and the trust degree of the user to the target medical report interpretation result. In addition, the medical report interpretation large language model obtains the target medical report interpretation result based on the multiple reasoning steps contained in the target thinking chain, which can also improve the reasoning accuracy of the medical report interpretation large language model. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the one or more embodiments of the present specification, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present specification, and other drawings can also be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A medical report interpretation large language model usage scenario provided by one or more embodiments of the present specification;

[0020] Figure 2 A medical report interpretation interface provided by one or more embodiments of the present specification;

[0021] Figure 3 A flowchart of a model training method provided for one or more embodiments of the present specification;

[0022] Figure 4 A scene diagram of a model training process provided for one or more embodiments of the present specification Figure 1 ;

[0023] Figure 5 A scene diagram of a model training process provided for one or more embodiments of the present specification Figure 2 ;

[0024] Figure 6 A scene diagram of a model training process provided for one or more embodiments of the present specification Figure 3 ;

[0025] Figure 7 A structural diagram of a model training device provided for one or more embodiments of the present specification;

[0026] Figure 8 A structural diagram of an electronic device provided for one or more embodiments of the present specification. DETAILED DESCRIPTION

[0027] For the purposes of the present specification, technical solutions and advantages, the technical solutions of the present specification will be described in detail below with reference to the embodiments of the present specification and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present specification, not all embodiments. Based on the embodiments in the present specification, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present specification.

[0028] It should be noted that in the case of one or more embodiments of the present specification involving user information, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in one or more embodiments of the present specification are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws and regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal. In addition, the various models (including but not limited to language models or large models) involved in the present specification are in compliance with relevant laws and standards.

[0029] In addition, the step sequence in each of the following method embodiments is only an example, not a strict limitation.

[0030] The following will first explain the concepts involved in one or more embodiments of the present specification.

[0031] Large Language Mode (LLM) refers to an artificial intelligence model learned based on a deep learning framework using a large-scale data corpus, which can be used to process various tasks such as computer vision, speech recognition, machine translation, text summarization, and intelligent question answering. One or more embodiments of the present specification do not limit the number of model parameters supported by the large language model, but aim to meet actual needs.

[0032] Multi-modal large language model refers to a large language model that can process and understand multiple types (modalities) of data. Specifically, in the fields of natural language processing (NLP) and computer vision (CV), multi-modal large language models can not only process text data, but also simultaneously process image, audio, video, and other forms of data.

[0033] Chain-of-Thought (CoT) is a model reasoning method that decomposes a complex task to be solved by a large language model into a series of logically related intermediate steps, guiding the large language model to reason step by step to obtain an answer, thereby improving the reasoning ability of the model and the interpretability of the reasoning result, i.e., explicitly stating the reasoning logic of the model.

[0034] Prompt refers to the prompt information related to the task to be executed that is input to the large language model during user interaction with the large language model. The prompt can include information such as task execution context, task execution rules, input / output format, etc. The prompt is used to guide the large language model to execute the corresponding task and generate the expected result.

[0035] Supervised Fine-Tuning (SFT) refers to a process of further training a model that has been pre-trained using a certain amount of unlabeled data using a labeled data set. The purpose of supervised fine-tuning is to make the pre-trained model better adapt to the execution requirements of certain tasks and improve the execution performance on that task.

[0036] Reinforcement Learning (RL) is a machine learning method similar to supervised learning and unsupervised learning. In reinforcement learning, a large language model executes a corresponding decision-making strategy and continuously optimizes the decision-making strategy based on the execution result feedback to obtain the maximum feedback reward. The decision-making strategy corresponding to the maximum feedback reward is the target decision-making strategy learned by the large language model. The large language model can have a good task execution accuracy based on the learned target decision-making strategy.

[0037] As described above, in the medical scenario, there is a need to interpret medical reports into content understandable by patients. Based on this, the embodiments of the present specification provide a model training method for training a medical report interpretation large language model capable of automatically interpreting medical reports.

[0038] For ease of understanding, the present specification first describes the use process of the medical report interpretation large language model, and then describes the training process of the medical report interpretation large language model.

[0039] It can be understood that the medical report interpretation large language model provided by the embodiments of the present specification can be used not only by patients but also by other people (such as medical staff, etc.) except patients to realize the interpretation of various medical reports and to be applied in clinical auxiliary diagnosis or medical teaching scenarios. For ease of description, in the subsequent embodiments, the user of the medical report interpretation large language model is collectively referred to as the user.

[0040] Figure 1 A medical report interpretation large language model use scenario diagram is provided for one or more embodiments of the present specification. As shown in Figure 1 In the use scenario of the medical report interpretation large language model, a client device 110 corresponding to a user and a server device 120 deploying the medical report interpretation large language model are involved. The client device 110 and the server device 120 are in communication connection. Optionally, the client device 110 can be a notebook computer, a tablet computer, a PC, a smart phone, or a robot terminal device; and the server device 120 can be a physical server including an independent host, a virtual server, a cloud server, or a server cluster.

[0041] In an optional embodiment, the use process of the medical report interpretation large language model can be:

[0042] The client device 110 sends a medical report interpretation request to the server device 120 by calling the medical report interpretation service provided by the server device 120. The medical report interpretation request includes a target medical report to be interpreted and a target question to be answered.

[0043] The server device 120 receives the medical report interpretation request, and inputs the target medical report and the target question to be answered into the medical report interpretation large language model to generate, by using the medical report interpretation large language model: 1) a target thinking chain for describing a reasoning process of the medical report interpretation large language model to obtain a target medical report interpretation result corresponding to the target question to be answered based on text content in the target medical report; and 2) a target medical report interpretation result corresponding to the input target medical report and the target question to be answered. Then, the target medical report interpretation result and the target thinking chain are fed back to the client device 110 for display, so as to complete the interpretation of the target medical report.

[0044] Figure 1 From the perspective of device interaction between the client device 110 and the server device 120, the use process of the medical report interpretation large language model in the embodiments of the present specification is described. Next, the use process of the medical report interpretation large language model in the embodiments of the present specification is further described from the perspective of human-computer interaction between the user and the client device 110. Figure 2

[0045] Figure 2 A medical report interpretation interface schematic diagram is provided for one or more embodiments of the present specification. It is worth noting that, Figure 2 The medical report interpretation interface shown is only illustrative and is not limited thereto.

[0046] As Figure 2 shown, in the medical report interpretation interface, an interactive control for the user to input the target medical report and the target question to be answered is provided. For example, the user can call the camera of the client device 110 through the "take report" control to take the target medical report to be interpreted, so as to realize the acquisition of the image containing the target medical report, that is, the input of the target medical report; the user can also input the health problem to be asked in the input box, such as "please help me interpret this medical report?", etc., so as to realize the input of the target question to be answered.

[0047] In an optional embodiment, the user can also input the target medical report by, for example, uploading a scanned image or a photographed image of the target medical report. Of course, in addition to the data format of the image, the input target medical report can also be in a document format, such as word, pdf, etc. In the embodiments of the present specification, neither the input method of the target medical report nor the data format of the target medical report is limited.

[0048] ​The client device 110 sends a medical report interpretation request containing the target medical report and the target question to be answered to the server device 120 in response to the user input target medical report and target question to be answered, so that the server device 120 outputs the target medical report interpretation result and the target thought chain according to the medical report interpretation request through the medical report interpretation large language model.

[0049] After receiving the target medical report interpretation result and the target thought chain fed back by the server device 120 for the medical report interpretation request, the client device 110 displays the target medical report interpretation result and the target thought chain in a preset display format in the medical report interpretation interface for the user to view.

[0050] In terms of display form, optionally, the display mode of the target medical report interpretation result and the target thought chain can be customized. For example, Figure 2 The target medical report interpretation result and the target thought chain are displayed in the form of a chat message as shown in FIG. 8, wherein "AI" represents the server device 120 deployed with the medical report interpretation large language model. For another example, the target medical report interpretation result and the target thought chain can be displayed differently in visual effects.

[0051] In terms of display content, optionally, the target thought chain can contain at least one continuous reasoning step, which is used to constitute the reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered from the text content in the target medical report; the target medical report interpretation result includes but is not limited to at least one of the following interpretation results: explanation and description of the text content contained in the target medical report, health problems indicated to exist in the target medical report, and suggested solutions to the health problems. The type of interpretation result contained in the target medical report interpretation result can be customized.

[0052] For ease of understanding, for example, it is assumed that the target medical report is a "five-item hepatitis B detection report", and the target question to be answered is "Please help me interpret this medical report?". The "five-item hepatitis B detection report" contains text content recording basic information such as the patient's name and report date (Li, February 17, 2025), as well as specific values and reference ranges of five indicators HBsAg, HBsAb, HBeAg, HBeAb, and HBcAb.

[0053] Correspondingly, the target thought chain can be: this is a report of five hepatitis B tests, the patient is Li, the report date is February 17, 2025. The report lists five indicators: hepatitis B virus surface antigen HBsAg, hepatitis B virus surface antibody HBsAb, hepatitis B virus e antigen HBeAg, hepatitis B virus e antibody HBeAb, and hepatitis B virus core antibody HBcAb, each with specific values and reference ranges. First, HBsAg is negative, which means she is currently not infected with hepatitis B virus; then HBsAb is also negative, which means she has no hepatitis B surface antibody in her body, that is, she has no immunity to hepatitis B virus; further down, HBeAg, HBeAb, and HBcAb are also negative, further confirming that she has neither active hepatitis B infection nor signs of past infection. In summary, all indicators are within the normal range and no abnormalities are found. However, considering that she has no hepatitis B antibody, it means she has no protection against hepatitis B virus and may be infected if exposed to hepatitis B virus. In summary, this hepatitis B five-item test report is normal, but the lack of hepatitis B antibody is a point to note, and vaccination and good living habits are recommended.

[0054] Correspondingly, the target medical report interpretation result can be: according to the hepatitis B five-item test report you provided, all indicators are negative, which suggests the following: 1) Hepatitis B virus surface antigen (HBsAg) negative: indicates no current hepatitis B virus infection. 2) Hepatitis B virus surface antibody (HBsAb) negative: indicates that no antibodies have been produced against hepatitis B virus, possibly not having received hepatitis B vaccine or not having an effective immune response. 3) Hepatitis B virus e antigen (HBeAg), hepatitis B virus e antibody (HBeAb), and hepatitis B virus core antibody (HBcAb) are all negative: further excluding past infection history or occult infection. In summary, you are currently in a non-hepatitis B virus infection state, but lack hepatitis B antibody protection, making you a susceptible population. If you are exposed to hepatitis B virus in the future (such as blood contact, sexual contact, etc.), there is still a risk of infection, and vaccination and good living habits are recommended.

[0055] The above describes the use process of the medical report interpretation large language model provided by the embodiments of the present specification.

[0056] Based on the above description, the medical report interpretation large language model obtained by the model training method provided by the embodiments of the present specification can output the target medical report interpretation result corresponding to the input target medical report and target question to be answered, and can output the target thought chain, wherein the target thought chain is used to describe the reasoning process of the medical report interpretation large language model based on the text content in the target medical report to obtain the target medical report interpretation result corresponding to the target question to be answered.

[0057] To achieve the above training objectives, the medical report interpretation large language model is enabled to obtain a target medical report interpretation result through a target thought chain reasoning, and output the target medical report interpretation result and the target thought chain. One or more embodiments of the present specification provide a model training method, which automatically constructs a thought chain through a target large language model, thereby generating a training sample containing a medical report, a question to be answered, a medical report interpretation result and a thought chain, to train the medical report interpretation large language model.

[0058] In the embodiments of the present specification, the data format of the medical report involved is taken as an example, and the data format of the medical report is an image. In the case where the data format of the medical report is an image, the target large language model and the medical report interpretation large language model are both multi-modal large language models, which can process both the text data corresponding to the question to be answered and the image data corresponding to the medical report.

[0059] The model training method provided by one or more embodiments of the present specification will be described below.

[0060] Figure 3 The flowchart of the model training method provided by one or more embodiments of the present specification is shown in FIG. 3, which includes steps 310 to 380. Figure 3

[0061] In step 310, a first training sample is obtained, and the first training sample includes a medical report, a question to be answered, and a first medical report interpretation result obtained by answering the question to be answered based on the text content in the medical report.

[0062] As an optional method for obtaining the first training sample, a plurality of medical reports of patients, questions that the patients have asked medical staff for the medical reports, and answers to the questions provided by the medical staff based on the medical reports can be collected. The collected medical reports, questions, answers, etc. are all data that have been desensitized for patient privacy information.

[0063] The medical report contains text content, which is used to record the basic information, medical information, examination conclusion, diagnosis conclusion and treatment suggestion of the patient, and can provide a reference basis for the medical staff to answer the patient's questions.

[0064] ​For example, patient basic information such as name, date of birth, contact information, report number, etc. Visit information such as hospital / department name, test item name (e.g. "five items of hepatitis B", "blood routine", "CT scan", etc.), doctor's name (the doctor who issued the report or the diagnosing doctor), visit time / test time, etc. Test results such as test index name (e.g. "HBsAg", "white blood cell count", "blood glucose value", etc.), index measurement value (e.g. "negative", "positive", "5.6 mmol / L"), index normal value range (e.g. "0.00-5.00 pg / mL"), and whether the index is outside the normal range (e.g. "↑ elevated", "↓ decreased", "normal") etc. Diagnosis conclusion such as whether there is a disease or abnormal state (e.g. "no abnormality", "suggesting absence of hepatitis B surface antibody"), possible etiology analysis or disease assessment, etc. Treatment recommendations such as recommendations given by the doctor based on the results (e.g. "recommend hepatitis B vaccination", "review liver function"), follow-up treatment recommendations (e.g. medication, surgery, lifestyle adjustments, etc.), precautions (e.g. dietary control, avoiding high-risk behaviors, etc.).

[0065] After that, the collected data is pre-processed to verify the integrity and correctness of the collected data and to eliminate interference data. For example, in an image containing a medical report, whether the content of the medical report is clear and complete, etc. If it is not clear or complete, the data is eliminated. For example, whether the collected question is complete and related to the medical report, if the collected question is incomplete or unrelated to the medical report, the data is eliminated, such as the question "What is the weather today?" is completely unrelated to the medical report and can be eliminated.

[0066] Optionally, the above data preprocessing process can be realized by a target large language model. Specifically, the prompt word containing the data to be pre-processed and the pre-processing task description can be input into the target large language model, so that the target large language model verifies the integrity and correctness of the data to be pre-processed under the guidance of the prompt word.

[0067] By pre-processing the collected data, the accuracy of the data used for model training can be effectively guaranteed, and the problem of poor model reasoning effect caused by inaccurate data used for model training can be avoided.

[0068] Next, based on the data preprocessing result, a first training sample is constructed. The first training sample includes a medical report, a question to be answered, and a first medical report interpretation result. The question to be answered is a question that the patient has asked medical personnel about the medical report, and the first medical report interpretation result is an answer to the question given by the medical personnel based on the text content in the medical report.

[0069] The first training sample is used to construct a reference thinking chain by the target large language model, and generate a second training sample for training the medical report interpretation large language model. The second training sample includes: a medical report, a question to be answered, a first medical report interpretation result, and a reference thinking chain.

[0070] In step 320, the thinking chain construction prompt word containing the first training sample is input into the target large language model, so that the target large language model outputs the reference thinking chain and the second medical report interpretation result under the supervision of the first medical report interpretation result. The reference thinking chain is used to describe the reasoning process of obtaining the second medical report interpretation result corresponding to the question to be answered based on the text content in the medical report.

[0071] In the embodiments of the present specification, the target large language model is a large language model that reasons in the form of a thinking chain.

[0072] It can be understood that the accuracy of the reasoning result obtained by the thinking chain reasoning can reflect the accuracy of the thinking chain. Based on this, while generating the reference thinking chain corresponding to the first training sample by the target large language model, the second medical report interpretation result corresponding to the first training sample can also be generated by the target large language model.

[0073] The second medical report interpretation result is used to verify the correctness of the reference thinking chain; and the reference thinking chain is used to describe the reasoning process of obtaining the second medical report interpretation result corresponding to the question to be answered based on the text content in the medical report of the first training sample.

[0074] For ease of description, the prompt word used to guide the target large language model to generate the reference thinking chain and the second medical report interpretation result is referred to as a thinking chain construction prompt word.

[0075] In an optional embodiment, a thinking chain construction prompt word template can be constructed in advance. The thinking chain construction prompt word template includes thinking chain construction task description information, output content description information, etc., and has a slot for filling the first training sample. Therefore, after obtaining the first training sample, the first training sample can be directly filled into the preset slot in the thinking chain construction prompt word template to obtain the thinking chain construction prompt word.

[0076] For example, the thinking chain construction prompt word can be, for example:

[0077] {

[0078] “First training sample: medical report, question to be answered, first medical report interpretation result”

[0079] "Thought chain construction task description information: please use the thought chain reasoning method to answer the problem to be solved. Your answer should include multiple steps, each step consisting of three types of actions:

[0080] 1) Inner thinking: this is the step of thinking. Please note that describing comprehensive reasoning requires multiple "inner thinking" steps, and each step should start with a brief title.

[0081] 2) Final conclusion: summarize the correct reasoning in the previous "inner thinking" step and provide the final answer.

[0082] 3) Verification: verify the conclusion in the "final conclusion" step. If the conclusion is correct, end the process. If not, go back"

[0083] "Output content description information: the output format follows the following JSON structure:

[0084] "COT": {"action": "inner thinking", "title": "……", "content": "……"},

[0085] {"action": "final conclusion", "content": "……"},

[0086] {"action": "verification", "content": "……"}

[0087] }.

[0088] It is worth noting that in actual application, the thought chain construction task description information, output content description information, etc. in the thought chain construction prompt word can be customized and set, and are not limited to the above examples.

[0089] After obtaining the thought chain construction prompt word, the thought chain construction prompt word is input into the target large language model to make the target large language model generate a reference thought chain and a second medical report interpretation result under the guidance of the thought chain construction prompt word. Specifically, the target large language model generates a corresponding reference thought chain and a second medical report interpretation result based on the text content in the medical report and the problem to be solved under the supervision of the first medical report interpretation result contained in the first training sample in the thought chain construction prompt word. Wherein, the reference thought chain corresponds to the "COT" in the output content of the thought chain construction prompt word in the above example, and the second medical report interpretation result corresponds to the "final conclusion" in the output content of the thought chain construction prompt word in the above example.

[0090] Step 330, if the second medical report interpretation result is correct according to the first medical report interpretation result, the reference thought chain is added to the first training sample to generate a second training sample.

[0091] Figure 4 A scenario of a model training process provided for one or more embodiments of the present specification Figure 1 As shown in Figure 4 , after the target large language model outputs the reference thought chain and the second medical report interpretation result, further, it can be verified whether the second medical report interpretation result is correct.

[0092] Optionally, the first report interpretation result can be used as a reference to verify whether the second medical report interpretation result is correct.

[0093] If it is determined according to the first medical report interpretation result that the second medical report interpretation result is correct, it means that the reference thought chain generated by the target large language model is correct, and the reference thought chain can be added to the first training sample to generate a second training sample. Among them, the second training sample contains: medical report, question to be answered, first medical report interpretation result and reference thought chain.

[0094] If it is determined according to the first medical report interpretation result that the second medical report interpretation result is incorrect, it means that the reference thought chain generated by the target large language model is incorrect, and a new reference thought chain and a new second medical report interpretation result can be generated by the target large language model, and the new second medical report interpretation result is re-verified for correctness until the second medical report interpretation result generated by the target large language model is correct, or the verification times reach a set number threshold (such as: 4 times). Among them, optionally, for the second medical report interpretation result corresponding to a certain first training sample, if the verification times reach the set number threshold, and the verification result of the second medical report interpretation result is still incorrect, the first training sample is discarded.

[0095] In the embodiments of the present specification, for the case that the second medical report interpretation result is incorrect, a new reference thought chain and a new second medical report interpretation result can be generated by the target large language model in at least one of the following ways.

[0096] In an optional embodiment, if it is determined according to the first medical report interpretation result that the second medical report interpretation result is incorrect, the thought chain construction prompt word can be re-input into the target large language model to make the target large language model re-output a new reference thought chain and a new second medical report interpretation result under the supervision of the first medical report interpretation result.

[0097] In this scheme, by re-inputting the thought chain construction prompt word into the target large language model, the target large language model is guided to try new reasoning logic to re-generate the reference thought chain and the second medical report interpretation result with new reasoning thought, so as to explore the reference thought chain that can obtain the correct second medical report interpretation result for constructing the second training sample.

[0098] In another optional embodiment, in the process of verifying whether the second medical report interpretation result is correct with reference to the first report interpretation result, the target large language model can first output the difference information between the second medical report interpretation result and the first medical report interpretation result, and the verification result of the second medical report interpretation result indicated by the difference information. The difference information can be understood as incorrect information in the second medical report interpretation result relative to the first medical report interpretation result.

[0099] Then, if the difference information indicates that the second medical report interpretation result is incorrect, the difference information and the reference thought chain are filled into the thought chain construction prompt word; the thought chain construction prompt word containing the first training sample, the difference information and the reference thought chain is input into the target large language model, so that the target large language model adjusts the reference thought chain according to the difference information under the supervision of the first medical report interpretation result, to re-output a new reference thought chain and a new second medical report interpretation result.

[0100] In this scheme, by adding the difference information and the generated reference thought chain to the original thought chain construction prompt word, the target large language model is helped to backtrack the possible errors in the previous reasoning logic, so as to correct the possible errors in the generated reference thought chain, to regenerate a reference thought chain that can obtain a correct second medical report interpretation result, for constructing a second training sample.

[0101] Figure 4 In the illustrated scenario, the correctness verification of the second medical report interpretation result after the target large language model outputs the reference thought chain and the second medical report interpretation result is taken as an example for illustration.

[0102] In actual application, optionally, if the thought chain construction prompt word contains information for guiding the target large language model to verify the correctness of the second medical report interpretation result, such as the "verification" information contained in the thought chain construction prompt word of the above example, in addition to generating the reference thought chain and the second medical report interpretation result, the target large language model can further verify the correctness of the second medical report interpretation result based on the first medical report interpretation result, and the verification result corresponds to the output content "verification" of the thought chain construction prompt word of the above example. Therefore, the target large language model can output the verification result while outputting the reference thought chain and the second medical report interpretation result. The timing and manner of verifying the second medical report interpretation result are not limited in the embodiments of the present specification.

[0103] In the above embodiments, the correctness of the reference thought chain is verified by verifying the second medical report interpretation result with reference to the first medical report interpretation result.

[0104] To further ensure the accuracy of the reference thought chain, in an optional embodiment, after determining that the second medical report interpretation result is correct according to the first medical report interpretation result, the following can also be performed:

[0105] The target large language model outputs a semantic matching degree between the reference thought chain and the text content in the medical report of the first training sample, the question to be answered, and the first medical report interpretation result. If the semantic matching degree is greater than a set threshold, the reference thought chain is added to the first training sample to generate a second training sample; if the semantic matching degree is less than or equal to the set threshold, a new reference thought chain and a new second medical report interpretation result are output again by the target large language model.

[0106] The semantic matching degree can be a semantic matching degree between the reference thought chain and the text content in the medical report of the first training sample, the question to be answered, and the first medical report interpretation result, respectively, for example, a first semantic matching degree between the reference thought chain and the text content in the medical report of the first training sample, a second semantic matching degree between the reference thought chain and the question to be answered, and a third semantic matching degree between the reference thought chain and the first medical report interpretation result. Alternatively, different semantic matching degrees can correspond to the same or different set thresholds, and the size of the set threshold can be set by the user.

[0107] In this scheme, by determining the semantic matching degree between the reference thought chain and the text content in the medical report of the first training sample, the question to be answered, and the first medical report interpretation result, the relevance between the reference thought chain and the text content in the medical report, the question to be answered, and the first medical report interpretation result can be ensured, thereby ensuring the correctness of the second training sample obtained based on the first training sample and the reference thought chain.

[0108] In step 340, the medical report interpretation large language model is trained by the second training sample. The medical report interpretation large language model is used to output a target medical report interpretation result and a target thought chain according to a target medical report and a target question to be answered input by a user. The target thought chain is used to describe the reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered based on the text content in the target medical report.

[0109] After obtaining the second training sample, the medical report interpretation large language model to be trained is supervised and fine-tuned using the second training sample. Alternatively, to improve training efficiency, the medical report interpretation large language model can be a pre-trained model.

[0110] Understandably, the training and usage processes of the large language model for medical report interpretation are similar. Referring to the previous explanation of the model's usage, when training the model using the second training sample, the input consists of the medical report and the question to be answered contained in the second training sample, and the output is the thought process and the medical report interpretation result. For ease of description, the output of the large language model during training will be referred to as: predicted thought process and predicted medical report interpretation result.

[0111] In practical applications, the medical report interpretation big language model performs corresponding tasks under the guidance of prompt words. In the embodiments of this specification, the prompt words used to guide the medical report interpretation big language model to generate predictive thought chains and predict medical report interpretation results are referred to as: report interpretation prompt words.

[0112] In one optional embodiment, a report interpretation prompt template can be pre-built. This template includes a description of the report interpretation task, a description of the output content, and reserved slots for filling in medical reports and unanswered questions from the second training sample. Therefore, after obtaining the second training sample, the medical reports and unanswered questions from the second training sample can be filled into the preset slots in the report interpretation prompt template to obtain the report interpretation prompts.

[0113] Figure 5 A scenario illustration of a model training process provided for one or more embodiments of this specification. Figure 2 .like Figure 5 As shown, after obtaining the report interpretation prompts, the prompts, containing the medical report and the question to be answered, can be input into the medical report interpretation language model. This allows the model to output predicted medical report interpretation results and predicted thought chains. Then, a loss function is constructed with the goal of improving the first similarity between the predicted medical report interpretation result and the first medical report interpretation result, and the second similarity between the predicted thought chain and the reference thought chain. Optionally, the first similarity, second similarity, etc., can be, for example, cosine similarity, Euclidean distance, etc. Finally, the medical report interpretation language model is trained based on the loss function. For example, training of the medical report interpretation language model ends when the loss value corresponding to the loss function is less than a certain value.

[0114] In one optional embodiment, after training the medical report interpretation language model with the second training sample, the medical report interpretation language model can be further trained during its use to further improve its reasoning accuracy.

[0115] Figure 6A scenario of a model training process provided for one or more embodiments of the present specification Figure 3 As shown in Figure 6 In the use process of the medical report interpretation large language model, feedback data corresponding to the target medical report interpretation result and the target thinking chain can also be obtained from the user feedback. Then, the medical report interpretation large language model is trained again by using the feedback data.

[0116] Among them, the feedback data refers to the data corresponding to the target medical report interpretation result and the target thinking chain generated due to the failure to give the correct or expected target medical report interpretation result and the target thinking chain when performing the medical report interpretation task, that is, the so-called bad case, which refers to the bad case or the error case.

[0117] In the specific implementation process, by analyzing the feedback data, the reason why the medical report interpretation large language model fails to give the correct or expected target medical report interpretation result and the target thinking chain result when performing the medical report interpretation task can be determined to guide the further training of the medical report interpretation large language model.

[0118] Optionally, the feedback data can also include the reason why the medical report interpretation large language model fails to give the correct or expected target medical report interpretation result and the target thinking chain result, etc.

[0119] As an optional way of training the medical report interpretation large language model based on the feedback data, the target medical report interpretation result and the target thinking chain result actually output by the medical report interpretation large language model can be corrected according to the reason why the medical report interpretation large language model fails to give the correct or expected target medical report interpretation result and the target thinking chain result, such as correcting a certain reasoning and decision result in the target thinking chain. Then, based on the corrected target medical report interpretation result and the target thinking chain result, and the corresponding target medical report and target question to be answered, a third training sample for training the medical report interpretation large language model is constructed to train the medical report interpretation large language model through the third training sample.

[0120] In the present scheme, the medical report interpretation large language model is trained by using the feedback data generated by the medical report interpretation large language model when performing the medical report interpretation task, more training samples are introduced, and thus the reasoning ability of the medical report interpretation task can be better improved.

[0121] It can be understood that the medical report interpretation large language model will continuously generate feedback data in the process of being used by the user. If new feedback data is generated each time, the training of the medical report interpretation large language model is triggered, which will increase the model training cost on the one hand, and is not conducive to the improvement of the ability of the medical report interpretation large language model on the other hand.

[0122] Optionally, in the use process of the medical report interpretation large language model, the amount of data of the current generated feedback data can be counted, and in response to the amount of data of the feedback data being greater than a set data amount threshold, the medical report interpretation large language model is trained according to the feedback data, such as guiding the medical report interpretation large language model to perform reinforcement learning.

[0123] Optionally, in the use process of the medical report interpretation large language model, the time interval between the historical time point corresponding to the previous model training of the medical report interpretation large language model according to the feedback data and the current time point can also be counted, and in response to the time interval between the historical time point corresponding to the previous reinforcement learning of the medical report interpretation large language model and the current time point being greater than a set time threshold, the medical report interpretation large language model is trained according to the feedback data, such as guiding the medical report interpretation large language model to perform reinforcement learning.

[0124] In this scheme, by counting the amount of data of the feedback data, or the time interval between the time point corresponding to the previous model training of the medical report interpretation large language model according to the feedback data and the current time point, the timing of training the medical report interpretation large language model through the feedback data can be reasonably controlled, so as to not only reduce the training cost of the medical report interpretation large language model, but also be beneficial to the model capability improvement of the medical report interpretation large language model.

[0125] In summary, in the model training scheme provided by one or more embodiments of the present specification, the target large language model can automatically construct a reference thought chain based on the first training sample, thereby generating a second training sample containing a medical report, a question to be answered, a medical report interpretation result, and a reference thought chain, for training a medical report interpretation large language model. The medical report interpretation large language model trained based on the second training sample not only learns relevant medical knowledge, but also masters the ability to reason through a thought chain, so that whether it is a single target medical report or multiple target medical reports, the medical report interpretation large language model can integrate the logical relationship between the text content in the target medical report and the target question to be answered through a thought chain, ensuring the integrity of information integration and the accuracy of reasoning logic. Based on this, the accuracy of the target medical report interpretation result obtained using the reasoning steps corresponding to the target thought chain can also be improved. In addition, the output content of the medical report interpretation large language model in the embodiments of the present specification contains not only the target medical report interpretation result, but also the target thought chain, so that the user not only knows the answer to the question they want to understand, but also knows the reasoning approach of the medical report interpretation large language model, which not only enhances the interpretability of the target medical report interpretation result and the user's trust in the target medical report interpretation result, but also provides corresponding teaching functions for the user, expanding the use scenarios of the medical report interpretation large language model.

[0126] The above describes particular embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in an order other than that in the embodiments and still achieve desirable results. Additionally, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous or possible.

[0127] The model training device of one or more embodiments of the present specification will be described in detail below. Those skilled in the art can understand that these devices can be configured using commercially available hardware components by following the steps taught by the present scheme.

[0128] Figure 7 A structural schematic diagram of a model training device provided by one or more embodiments of the present specification is shown in Figure 7 The device includes a sample acquisition module 710, a sample construction module 720, and a model training module 730.

[0129] The sample acquisition module 710 is configured to acquire a first training sample, wherein the first training sample includes a medical report, a question to be answered, and a first medical report interpretation result obtained by answering the question to be answered based on the text content in the medical report.

[0130] a sample construction module 720, configured to input a thinking chain construction prompt word containing the first training sample into a target large language model, so that the target large language model outputs a reference thinking chain and a second medical report interpretation result under the supervision of the first medical report interpretation result, the reference thinking chain being used to describe a reasoning process of obtaining the second medical report interpretation result corresponding to the to-be-answered question based on text content in the medical report; and if it is determined that the second medical report interpretation result is correct according to the first medical report interpretation result, the reference thinking chain is added to the first training sample to generate a second training sample.

[0131] a model training module 730, configured to train a medical report interpretation large language model through the second training sample, the medical report interpretation large language model being used to output a target medical report interpretation result and a target thinking chain according to a target medical report and a target to-be-answered question input by a user, the target thinking chain being used to describe a reasoning process of obtaining the target medical report interpretation result corresponding to the target to-be-answered question based on text content in the target medical report.

[0132] In an optional embodiment, the sample construction module 720 is further configured to: if it is determined that the second medical report interpretation result is incorrect according to the first medical report interpretation result, re-input the thinking chain construction prompt word into the target large language model, so that the target large language model re-outputs a new reference thinking chain and a new second medical report interpretation result under the supervision of the first medical report interpretation result.

[0133] In an optional embodiment, the sample construction module 720 is further configured to: output, by the target large language model, difference information between the second medical report interpretation result and the first medical report interpretation result, and a verification result of the second medical report interpretation result indicated by the difference information; if the difference information indicates that the second medical report interpretation result is incorrect, fill the difference information and the reference thinking chain into the thinking chain construction prompt word; input the thinking chain construction prompt word containing the first training sample, the difference information and the reference thinking chain into the target large language model, so that the target large language model adjusts the reference thinking chain according to the difference information under the supervision of the first medical report interpretation result, to re-output a new reference thinking chain and a new second medical report interpretation result.

[0134] In an optional embodiment, the sample construction module 720 is further configured to: output, by the target large language model, a semantic matching degree between the reference thought chain and text content in the medical report, the question to be answered, and the first medical report interpretation result; if the semantic matching degree is greater than a set threshold, add the reference thought chain to the first training sample to generate a second training sample; and if the semantic matching degree is less than or equal to the set threshold, re-output, by the target large language model, a new reference thought chain and a new second medical report interpretation result.

[0135] In an optional embodiment, the model training module 730 is specifically configured to: input a report interpretation prompt word containing the medical report and the question to be answered into a medical report interpretation large language model, so that the medical report interpretation large language model outputs a predicted medical report interpretation result and a predicted thought chain; construct a loss function with the goal of improving a first similarity between the predicted medical report interpretation result and the first medical report interpretation result and a second similarity between the predicted thought chain and the reference thought chain; and train the medical report interpretation large language model according to the loss function.

[0136] In an optional embodiment, the model training module 730 is further configured to: during use of the medical report interpretation large language model, acquire feedback data corresponding to the target medical report interpretation result and the target thought chain and fed back by the user; in response to a data amount of the feedback data being greater than a set data amount threshold, perform reinforcement learning on the medical report interpretation large language model according to the feedback data; or in response to a time interval between a historical time point corresponding to a previous time of the reinforcement learning of the medical report interpretation large language model and a current time point being greater than a set time threshold, perform reinforcement learning on the medical report interpretation large language model according to the feedback data.

[0137] In an optional embodiment, the data format of the medical report is an image.

[0138] Figure 7 The apparatuses described above can perform the steps described in the foregoing embodiments. Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be mutually referred to. Each embodiment mainly describes differences from other embodiments. In particular, for the apparatus embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the detailed execution process and technical effects, refer to the description in the foregoing method embodiment, which will not be described here again.

[0139] In one possible design, the above Figure 7 The structure of the model training apparatus can be implemented as an electronic device, such as a server or a terminal. Figure 8As shown, the electronic device can include a memory 810, a processor 820, and a communication interface 830. The memory 810 stores a computer program, which, when executed by the processor 820, causes the processor 820 to implement at least the model training method provided in the foregoing embodiments.

[0140] The memory 810 can be implemented by any type of volatile or nonvolatile memory devices or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0141] Accordingly, one or more embodiments of the present specification also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to implement each step in the above method embodiments. The computer-readable storage medium includes volatile or non-volatile or a combination thereof, and can be removable or non-removable. Examples of the computer-readable storage medium include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital video disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium

[0142] Accordingly, the one or more embodiments of the present specification also provide a computer program product, which comprises computer programs or instructions, and when the computer programs or instructions are executed by a processor, the processor is enabled to implement each step in the above-mentioned method embodiments. It should be understood that each of the above-mentioned method processes or a combination of multiple processes can be implemented by the computer programs or instructions. In addition, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable model training devices, so that the processor of the general-purpose computer, the special-purpose computer, the embedded processor, or other programmable model training devices can be implemented as a device that implements the corresponding functions in the above-mentioned method embodiments.

[0143] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate. Some or all of the modules can be implemented to achieve the purposes of the embodiments of the present specification according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0144] From the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of a general hardware platform as necessary, and of course can also be implemented by means of a combination of hardware and software. Based on such understanding, the above technical solutions can be embodied in the form of a computer product in essence or in the form of a contribution to the prior art. The present specification can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0145] Finally, it should be noted that the terms "comprise", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of another identical element in the process, method, article or device comprising the element.

[0146] The above is only an embodiment of the present specification and is not intended to limit the present specification. The present specification can have various changes and modifications for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present specification shall be included in the scope of claims of the present specification.

Claims

1. A model training method, characterized in that, The method comprises: obtaining a first training sample, the first training sample comprising: a medical report, a question to be answered, and a first medical report interpretation result obtained by answering the question to be answered based on the text content in the medical report; inputting a thinking chain construction prompt word containing the first training sample into a target large language model, so that the target large language model outputs a reference thinking chain and a second medical report interpretation result under the supervision of the first medical report interpretation result, and the reference thinking chain is used to describe the reasoning process of obtaining the second medical report interpretation result corresponding to the question to be answered based on the text content in the medical report; if it is determined according to the first medical report interpretation result that the second medical report interpretation result is correct, then the reference thinking chain is added to the first training sample to generate a second training sample; training a medical report interpretation large language model through the second training sample, the medical report interpretation large language model being used to output a target medical report interpretation result and a target thinking chain according to a target medical report and a target question to be answered input by a user, the target thinking chain being used to describe the reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered based on the text content in the target medical report; wherein, when training the medical report interpretation large language model, a loss function used is constructed to improve a first similarity between a predicted medical report interpretation result and the first medical report interpretation result and a second similarity between a predicted thinking chain and the reference thinking chain; the predicted medical report interpretation result and the predicted thinking chain are output by the medical report interpretation large language model based on a report interpretation prompt word containing the medical report and the question to be answered.

2. The method of claim 1, wherein, The method further comprises: if it is determined according to the first medical report interpretation result that the second medical report interpretation result is incorrect, then the thinking chain construction prompt word is re-input into the target large language model, so that the target large language model re-outputs a new reference thinking chain and a new second medical report interpretation result under the supervision of the first medical report interpretation result.

3. The method of claim 1, wherein, The method further comprises: outputting, by the target large language model, difference information between the second medical report interpretation result and the first medical report interpretation result, and a verification result of the second medical report interpretation result indicated by the difference information; if the difference information indicates that the second medical report interpretation result is incorrect, then the difference information and the reference thinking chain are filled into the thinking chain construction prompt word; inputting the thinking chain construction prompt word containing the first training sample, the difference information and the reference thinking chain into the target large language model, so that the target large language model adjusts the reference thinking chain according to the difference information under the supervision of the first medical report interpretation result, to re-output a new reference thinking chain and a new second medical report interpretation result.

4. The method of claim 1, wherein, After it is determined according to the first medical report interpretation result that the second medical report interpretation result is correct, the method further comprises: output, by the target large language model, a semantic matching degree between the reference thought chain and text content in the medical report, the question to be answered, and the first medical report interpretation result; if the semantic matching degree is greater than a set threshold, adding the reference thought chain to the first training sample to generate a second training sample; if the semantic matching degree is less than or equal to a set threshold, re-outputting, by the target large language model, a new reference thought chain and a new second medical report interpretation result.

5. The method of claim 1, wherein, The method further comprises: acquiring, in a use process of the medical report interpretation large language model, feedback data corresponding to the target medical report interpretation result and the target thought chain and fed back by the user; in response to a data amount of the feedback data being greater than a set data amount threshold, performing reinforcement learning on the medical report interpretation large language model according to the feedback data; or in response to a time interval between a historical time point corresponding to a previous time of performing the reinforcement learning by the medical report interpretation large language model and a current time point being greater than a set time threshold, performing reinforcement learning on the medical report interpretation large language model according to the feedback data.

6. The method according to any one of claims 1 to 5, characterized in that, The data format of the medical report is an image.

7. A model training apparatus characterized by comprising: The device comprises: a sample acquisition module configured to acquire a first training sample, the first training sample comprising a medical report, a question to be answered, and a first medical report interpretation result obtained by answering the question to be answered based on text content in the medical report; a sample construction module configured to input a thought chain construction prompt word containing the first training sample into a target large language model, so that the target large language model outputs a reference thought chain and a second medical report interpretation result under the supervision of the first medical report interpretation result, the reference thought chain being used to describe a reasoning process of obtaining the second medical report interpretation result corresponding to the question to be answered based on the text content in the medical report; if it is determined that the second medical report interpretation result is correct according to the first medical report interpretation result, adding the reference thought chain to the first training sample to generate a second training sample; a model training module configured to train a medical report interpretation large language model by using the second training sample, the medical report interpretation large language model being configured to output a target medical report interpretation result and a target thought chain according to a target medical report and a target question to be answered input by a user, the target thought chain being used to describe a reasoning process of obtaining the target medical report interpretation result corresponding to the target question to be answered based on text content in the target medical report; wherein, in training the medical report interpretation large language model, a loss function used is constructed to improve a first similarity between a predicted medical report interpretation result and the first medical report interpretation result and a second similarity between a predicted thought chain and the reference thought chain; the predicted medical report interpretation result and the predicted thought chain are output by the medical report interpretation large language model based on report interpretation prompt words containing the medical report and the question to be answered.

8. An electronic device, comprising: comprise: A memory, a processor, and a communication interface; wherein the memory has stored thereon a computer program which, when executed by the processor, causes the processor to perform the model training method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium has stored thereon a computer program which, when executed by the processor of the electronic device, causes the processor to perform the model training method according to any one of claims 1 to 6.

Images