Medical diagnosis method, system and related device

Abstract

The invention discloses a medical diagnosis method and system and a related device, and the method comprises the steps: obtaining a to-be-diagnosed report, and determining a target thinking chain matched with the to-be-diagnosed report through an intelligent analysis model; wherein the intelligent analysis model is obtained by training a plurality of training data, and the training data comprises a first training report and a first training thinking chain matched with the first training report; generating a current node matched with the target thinking chain and at least one candidate answer matched with the current node; and obtaining a target reward value matched with each candidate answer, and generating a target diagnosis result matched with the to-be-diagnosed report based on the candidate answers and the target reward values matched with the candidate answers. In this way, the accuracy of medical diagnosis can be improved.

Description

Technical Field

[0001] The present application relates to the field of medical diagnosis technology, and in particular to a medical diagnosis method, system and related devices. Background Art

[0002] With the rapid development of intelligent technology, large language models are being used in an increasing number of scenarios, such as medical diagnosis. Currently, large language models for medical diagnosis rely primarily on semantic analysis of medical reports or incorporating relevant knowledge, resulting in certain limitations in diagnostic accuracy.

[0003] In view of this, how to propose a medical diagnosis method with higher accuracy has become an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem solved by this application is to provide a medical diagnosis method, system and related devices that can improve the accuracy of medical diagnosis.

[0005] In order to solve the above technical problems, a technical solution adopted in this application is: to provide a medical diagnosis method, including: obtaining a report to be diagnosed, and using an intelligent analysis model to determine a target thinking chain that matches the report to be diagnosed; wherein the intelligent analysis model is trained using multiple training data, and the training data includes a first training report and a first training thinking chain that matches the first training report; generating a current node that matches the target thinking chain, and at least one candidate answer that matches the current node; obtaining a target reward value that matches each candidate answer, and generating a target diagnosis result that matches the report to be diagnosed based on the candidate answer and the target reward value that matches it.

[0006] In order to solve the above technical problems, another technical solution adopted in this application is: to provide a medical diagnosis system, including: an acquisition module, used to obtain the report to be diagnosed, and use the intelligent analysis model to determine the target thinking chain that matches the report to be diagnosed; wherein, the intelligent analysis model is trained using multiple training data, and the training data includes a first training report and a first training thinking chain that matches the first training report; a processing module, used to generate a current node that matches the target thinking chain, and at least one candidate answer that matches the current node; a diagnosis module, used to obtain the target reward value that matches each candidate answer, and generate a target diagnosis result that matches the report to be diagnosed based on the candidate answer and the target reward value that matches it.

[0007] To solve the above technical problems, another technical solution adopted in this application is: to provide an electronic device, comprising: a memory and a processor coupled to each other, wherein the memory stores program instructions, and the processor is used to execute the program instructions to implement the method mentioned in the above technical solution.

[0008] In order to solve the above technical problems, another technical solution adopted in this application is: providing a computer-readable storage medium on which program instructions are stored, and when the program instructions are executed by a processor, the method mentioned in the above technical solution is implemented.

[0009] The beneficial effects of this application are as follows: Unlike existing technologies, the medical diagnosis method proposed in this application utilizes an intelligent analysis model to generate a target thought chain that matches the report to be diagnosed, generates a current node based on the content in the target thought chain, and expands the content of the current node to determine the target answer that matches the current node. By combining all target nodes and their matching target answers, a search path is constructed that can characterize the optimal reasoning process corresponding to the report to be diagnosed, thereby increasing the accuracy of the target diagnosis result obtained according to the search path. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without inventive efforts. Among them:

[0011] Figure 1 This is a flow chart of an embodiment of the medical diagnostic method of the present application;

[0012] Figure 2 This is a flowchart of an implementation method of the training data acquisition method of the present application;

[0013] Figure 3 yes Figure 2 The flowchart after step S203 corresponds to an embodiment;

[0014] Figure 4 yes Figure 1 Step S102 corresponds to a flow chart of another embodiment;

[0015] Figure 5 This is a flowchart of an implementation method of the reward model training method of this application;

[0016] Figure 6 This is a schematic structural diagram of an embodiment of the medical diagnostic system of the present application;

[0017] Figure 7 This is a schematic structural diagram of an embodiment of the electronic device of the present application;

[0018] Figure 8 It is a structural diagram of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them, and different embodiments can be adaptively combined. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0020] See also Figure 1 , Figure 1 1 is a flow chart of an embodiment of the medical diagnostic method of the present application, the method comprising:

[0021] S101: Obtain a report to be diagnosed, and use an intelligent analysis model to determine a target thought chain that matches the report to be diagnosed; wherein the intelligent analysis model is trained using multiple training data, and the training data includes a first training report and a first training thought chain that matches the first training report.

[0022] In one embodiment, a report to be diagnosed is obtained, input into a trained intelligent analysis model, and the intelligent analysis model is used to generate a target thought chain that matches the report to be diagnosed. The target thought chain includes intermediate reasoning steps when generating a target diagnosis result based on the report to be diagnosed.

[0023] In one implementation scenario, after obtaining the report to be diagnosed, the fine-tuned intelligent analysis model is used to perform diagnostic reasoning on the report to be diagnosed, and generate a corresponding target thinking chain. The above-mentioned intelligent analysis model is a large language model with relatively good data analysis capabilities. By inputting the report to be diagnosed into the intelligent analysis model, the intelligent analysis model plays the role of a medical professional to perform diagnostic reasoning on the report to be diagnosed, and generates a corresponding target thinking chain. Among them, the intelligent analysis model is trained using multiple training data, and the training data includes a first training report and a matching first training thinking chain. By fine-tuning the intelligent analysis model using multiple training data, the intelligent analysis model obtained after training has the ability to generate a matching target thinking chain based on the report to be diagnosed.

[0024] In a specific application scenario, the above-mentioned large language model may include but is not limited to deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and generative pre-trained Transformer models, etc., and no specific restrictions are imposed on the specific structure and specific deployment of the large language model. It should be noted that the specific structure and specific deployment of the intelligent analysis model mentioned in other embodiments of this application can refer to this embodiment.

[0025] In one implementation scenario, the medical diagnosis report is a medical-related hospitalization report, case or test report, etc.

[0026] In another embodiment, a reference knowledge base matching the report to be diagnosed is obtained. The obtained report to be diagnosed is input into a trained intelligent analysis model, which is then used to search for reference knowledge related to the report to be diagnosed from the parameter knowledge base. The intelligent analysis model is then combined with the reference knowledge to generate a target thought chain matching the report to be diagnosed.

[0027] In another embodiment, in response to the report to be diagnosed matching the related historical diagnostic report, the obtained report to be diagnosed is input into the trained intelligent analysis model, and the report to be diagnosed is analyzed using the above intelligent analysis model in combination with the reference knowledge searched in the parameter knowledge base and the above historical diagnostic report to generate a target thinking chain that matches the report to be diagnosed.

[0028] In a specific application scenario, when the report to be diagnosed is the patient's hospitalization medical record for that day, the above historical diagnosis report is the patient's historical hospitalization case within the historical time period.

[0029] S102: Generate a current node that matches the target thought chain and at least one candidate answer that matches the current node.

[0030] In one embodiment, a corresponding current node is generated for a target thought chain generated by an intelligent analysis model, and at least one matching candidate answer is generated based on the current node.

[0031] In one implementation scenario, the generated target thought chain is divided into multiple reasoning segments. For each reasoning segment, a corresponding current node is generated, and the content of the current node is expanded to obtain at least one matching candidate answer.

[0032] Specifically, in response to the target thought chain being a reasoning text, the reasoning text is divided into multiple text segments, each of which serves as a reasoning fragment. A Monte Carlo Tree Search (MCTS) algorithm is used to generate a current node corresponding to the reasoning fragment, and at least one matching candidate answer is generated based on the current node. The text segments are divided based on semantics, that is, the reasoning text is divided based on its semantics, resulting in multiple corresponding semantically complete text segments.

[0033] In another implementation scenario, in response to the target thought chain being a reasoning text, each sentence therein is taken as a text segment, thereby obtaining a corresponding plurality of reasoning fragments.

[0034] S103: Obtain a target reward value for each candidate answer match, and generate a target diagnosis result for the report to be diagnosed based on the candidate answers and their matched target reward values.

[0035] In one embodiment, for each candidate answer that matches the current node, a target reward value is predicted for each candidate answer match. The target reward value represents the confidence that the candidate answer is the expanded content corresponding to the current node. In other words, a larger target reward value indicates a higher degree of match between the path formed by the target node and the candidate answer and the accurate reasoning logic.

[0036] Furthermore, for all candidate answers corresponding to the current node, the candidate answer corresponding to the maximum target reward value is used as the target answer matched by the current node. Based on all current nodes matched by the target thought chain and their matched target answers, the target diagnosis result is obtained.

[0037] Specifically, for each current node, the current node and its corresponding target answer are concatenated into a search subpath. Following the order of the reasoning fragments corresponding to the current node in the target thought chain, all search subpaths corresponding to the current node are concatenated into a search path. Based on the search path, a target diagnosis result matching the to-be-diagnosed report is generated.

[0038] In one implementation scenario, the search path is input into an intelligent analysis model, and the search path is analyzed using the intelligent analysis model to generate a target diagnosis result that matches the report to be diagnosed.

[0039] In another embodiment, after determining the target answer corresponding to the current node, a new round of content expansion is performed using the MCTS algorithm based on the most recently generated target answer to generate the target answer under the next expansion round, until the target answers under all expansion rounds corresponding to the current node are obtained. The current node and all corresponding target answers are connected in series to obtain the corresponding search subpath. According to the search subpaths corresponding to all current nodes, a search path is constructed, and according to the search path, a target diagnosis result that matches the report to be diagnosed is generated. Among them, the process of generating the target answer under the next expansion round can refer to the corresponding embodiment described above and will not be elaborated on here.

[0040] The medical diagnosis method proposed in this application utilizes an intelligent analysis model to generate a target thought chain that matches the report to be diagnosed. Based on the content in the target thought chain, the current node is generated and the content of the current node is expanded to determine the target answer that the current node matches. Combining all target nodes and their matching target answers, a search path is constructed that characterizes the optimal reasoning process corresponding to the report to be diagnosed, resulting in a more accurate target diagnosis result based on the search path.

[0041] See also Figure 2 , Figure 2 This is a flow chart of an embodiment of the method for obtaining training data of the present application. Specifically, the steps for obtaining training data include:

[0042] S201: Acquire multiple first training reports; wherein the first training report is matched with a first diagnosis tag and a timestamp.

[0043] In one embodiment, a plurality of first training reports are acquired, where the first training reports are matched with a first diagnosis tag and a timestamp.

[0044] In one implementation scenario, the first training report is a medical-related hospitalization report, case report, or test report, and its specific report type is consistent with the report type corresponding to the to-be-tested report in the corresponding embodiment described above. The first diagnostic tag matched by the first training report includes the diagnostic result corresponding to the first diagnostic report, and the timestamp matched by the first training report indicates the generation time or detection time of the first training report.

[0045] S202: Input the first training report into the intelligent analysis model to generate a corresponding reference thinking chain.

[0046] In one embodiment, the acquired first training report is input into an intelligent analysis model, and the intelligent analysis model is used to perform diagnostic reasoning on the first training report and generate a corresponding reference thinking chain.

[0047] In one implementation scenario, a first training report is associated with at least one reference report, and the timestamp of the reference report is earlier than the timestamp of the first training report. The first training report is input into an intelligent analysis model, and the intelligent analysis model is used to obtain reference knowledge related to the first training report. Based on the reference knowledge, the reference report, and the first training report, a reference thought chain corresponding to the first training report is generated.

[0048] Specifically, at least one reference report associated with the first training report is pre-obtained, where the timestamp of the reference report is earlier than that of the first training report. Based on the report type of the first training report, a reference knowledge base matching the first training report is obtained. The first training report is input into the intelligent analysis model, and the intelligent analysis model is used to search the reference knowledge base for reference knowledge related to the first training report. Furthermore, the intelligent analysis model is used to analyze the first training report using the reference knowledge and the reference report to generate a corresponding reference thought chain.

[0049] In one specific application scenario, the first training report is a hospitalized case of a training subject. The reference report associated with the first training report is also a hospitalized medical record of the training subject, and the reference report is dated earlier than the first training report. Furthermore, in response to the first training report being a hospitalized case, the reference knowledge base is a medical-related knowledge base.

[0050] In another embodiment, in order to improve the processing efficiency of the intelligent analysis model, after the first training report is input into the intelligent analysis model, the intelligent analysis model is used to obtain reference knowledge related to the first training report, and the first training report is analyzed only in combination with the reference knowledge to generate a reference thinking chain corresponding to the first training report.

[0051] S203: Based on the reference thinking chain, obtain a reference diagnosis result corresponding to the first training report.

[0052] In one embodiment, after the reference thought chain is generated, the intelligent analysis model is used to simulate a medical expert diagnosing the first training report according to the reference thought chain to generate a predicted reference diagnosis result.

[0053] S204: In response to the reference diagnosis result being consistent with the first diagnosis label, using the reference thought chain as a first training thought chain for training the intelligent analysis model.

[0054] In one embodiment, when the reference diagnosis result is consistent with the first diagnosis label, it represents that an accurate reference diagnosis result can be obtained through reasoning through the corresponding reference thinking chain, that is, the reasoning process corresponding to the reference thinking chain is more accurate, so the reference thinking chain is used as the first training thinking chain.

[0055] Furthermore, after obtaining the first training thinking chain corresponding to the first training data, the first training data and its corresponding first training thinking chain are used as training data, and the training data is used to fine-tune the intelligent analysis model, so that the intelligent analysis model obtained after fine-tuning has the ability to generate an accurate target thinking chain based on the input diagnosis report.

[0056] Specifically, based on the training data constructed above, the intelligent analysis model is fine-tuned and trained using the supervised fine-tuning (SFT) technology until a fine-tuned intelligent analysis model is obtained.

[0057] See also Figure 3 , Figure 3 yes Figure 2 The process after step S203 corresponds to a flow chart of an embodiment. Specifically, the process after step S203 also includes:

[0058] S301: Determine whether the reference diagnosis result is consistent with the first diagnosis label.

[0059] In one embodiment, after the reference diagnosis result is obtained, the reference diagnosis result is compared with the first diagnosis label matched with the first training report to determine whether the reference diagnosis result is consistent with the first diagnosis label.

[0060] S302: In response to the reference diagnosis result being inconsistent with the first diagnosis label, obtaining supplementary information matching the first training report.

[0061] In one embodiment, when the reference diagnostic result is inconsistent with the first diagnostic label, it indicates that the intelligent analysis model cannot obtain an accurate diagnostic result based on the generated reference thinking chain. In order to correct the reference thinking chain and reference diagnostic result generated by the intelligent analysis model, supplementary information matching the first training report is obtained.

[0062] In one implementation scenario, the supplementary information includes first supplementary sub-information for supplementing the first training label and second supplementary sub-information for supplementing the generated reference thought chain. By obtaining the supplementary information to guide the reasoning process of the intelligent analysis model, the accuracy of the subsequently regenerated reference diagnosis results is improved.

[0063] In addition, it should be noted that when the reference diagnosis result is consistent with the first diagnosis label, the reference thinking chain is used as the first training thinking chain. The specific implementation process can refer to the above step S204.

[0064] S303: Input the supplementary information into the intelligent analysis model, and use the intelligent analysis model to generate an updated reference thinking chain based on the supplementary information.

[0065] In one embodiment, the acquired supplementary information is input into the intelligent analysis model, and the intelligent analysis model is prompted to update the generated reference thinking chain based on the supplementary information and the generated reference thinking chain, so that the intelligent analysis model generates an updated reference thinking chain.

[0066] S304: Based on the updated reference thinking chain, obtain the reference diagnosis result corresponding to the first training report, and return to the step of determining whether the reference diagnosis result is consistent with the first diagnosis label.

[0067] In one embodiment, based on the updated reference thinking chain, the intelligent analysis model is used to regenerate the reference diagnosis result corresponding to the first training report, and return to the above step of determining whether the reference diagnosis result is consistent with the first diagnosis label, that is, return to the above step S301.

[0068] In another embodiment, in response to the number of times the supplementary information is obtained reaching a threshold, and the reference thinking chain finally obtained by combining multiple rounds of supplementary information cannot obtain a reference diagnosis result consistent with the first diagnostic label, the update of the reference thinking chain is stopped.

[0069] See also Figure 4 , Figure 4 yes Figure 1 Step S102 corresponds to a flow chart of another embodiment. Specifically, the implementation process of step S102 includes:

[0070] S401: Acquire the current content of the target thought chain and generate the current node based on the current content.

[0071] In one embodiment, in response to inputting the diagnosis report into the intelligent analysis model, the intelligent analysis model continuously outputs a matching target thought chain, and generates a current node corresponding to the current content according to the current content of the target thought chain output by the intelligent analysis model.

[0072] Specifically, during the process of generating the target thought chain, the intelligent analysis model outputs reasoning statements one by one, combining all the output reasoning statements to form the target thought chain. For the diagnosis report, the most recently output reasoning statement of the target thought chain is used as the current content, and the corresponding current node is generated based on the current content.

[0073] In one implementation scenario, the MCTS algorithm is used to take the current content as the current node.

[0074] S402: Generate at least one candidate answer matching the current node, wherein the candidate answer includes extended content corresponding to at least one child node matching the current node.

[0075] In one embodiment, the MCTS algorithm is further used to expand the content of the obtained current node to generate at least one candidate answer that matches the current node. The candidate answer includes the expanded content that matches the current node, and the expanded content serves as a child node that matches the current node.

[0076] The above scheme uses the MCTS algorithm to generate the current node corresponding to the current content of the target thought chain, and generates the child nodes corresponding to the current node to expand the reasoning process, so as to help improve the accuracy of diagnosis.

[0077] See also Figure 5 , Figure 5 This is a flowchart of an implementation method of the reward model training method of this application. Figure 1 The target reward value for each candidate answer match in is obtained based on the trained reward model. The specific training process of the reward model includes:

[0078] S501: Obtain a second training report and generate a second training thought chain corresponding to the second training report; wherein the second training report is matched with a second diagnosis tag.

[0079] In one embodiment, a plurality of second training reports are obtained, each second training report being matched with a second diagnosis tag, and a second training thought chain corresponding to the second training report is generated.

[0080] In one implementation scenario, in response to obtaining a fine-tuned intelligent analysis model through the corresponding implementation methods described above, a second training report is input into the fine-tuned intelligent analysis model, which analyzes the second training report and generates a corresponding second training thought chain. The specific implementation process can be referred to the corresponding implementation methods described above and will not be elaborated on here.

[0081] In another embodiment, in order to save training costs, the training data used for fine-tuning the intelligent analysis model in the above corresponding embodiment is used as the second training report in this step, that is, the first training report obtained in the above corresponding embodiment and its corresponding first training thinking chain are used as the second training report in this step and its corresponding second training thinking chain.

[0082] S502: Based on the second training thinking chain, generate at least one corresponding reference search path; wherein each reference search path includes multiple training nodes.

[0083] In one embodiment, according to the second training thought chain corresponding to the obtained second training report, at least one corresponding reference search path is generated using the MCTS algorithm, and each reference search path includes a plurality of training nodes.

[0084] S503: Obtain the predicted diagnosis result corresponding to each reference search path.

[0085] In one embodiment, the second training report is subjected to inference diagnosis according to each reference search path to generate a predicted diagnosis result corresponding to each reference search path.

[0086] In one implementation scenario, the content corresponding to the last training node in each reference search path is used as the predicted diagnosis result.

[0087] In another implementation scenario, the search path is input into an intelligent analysis model, so that the intelligent analysis model is used to analyze the search path and generate a predictive diagnosis result.

[0088] S504: Determine a target search path and a matching reference reward value thereof from all reference search paths based on the predicted diagnosis result and the second diagnosis label.

[0089] In one embodiment, a reference search path corresponding to a predicted diagnostic result consistent with the second diagnostic tag is used as a target search path. A first number of predicted diagnostic results consistent with the second diagnostic tag and a second number corresponding to all predicted diagnostic results are obtained, and a reference reward value for matching the target search path is determined based on a ratio of the first number to the second number.

[0090] Specifically, for the second training report, the predicted diagnosis result corresponding to each reference search path is compared with the second diagnostic label. In response to the predicted diagnosis result corresponding to the current reference search path being consistent with the second diagnostic label, the current reference search path is set as the target search path. A first quantity corresponding to all target search paths and a second quantity corresponding to all reference search paths are obtained, and a ratio of the first quantity to the second quantity is calculated, and the ratio is used as a reference reward value for matching the target search path.

[0091] In a specific application scenario, in response to generating 10 reference search paths corresponding to the second training thinking chain, and the predicted diagnostic results corresponding to six of the reference search paths are consistent with the second diagnostic label matched by the second training thinking chain, the above six reference search paths are used as target search paths, and it is characterized that accurate diagnostic results can be obtained through the above target search paths, and the reference reward value for each target search path match is determined to be 0.6.

[0092] S505: Based on the reference reward value of the target search path matching, the reward model is trained until a trained reward model is obtained.

[0093] In one embodiment, a reward model is constructed, and the reward model is trained using the target search path and the reference reward value matched by the second training report until the reward model converges to obtain a trained reward model.

[0094] Specifically, for the target search path matched by the second training report, the reward model is used to output a predicted reward value corresponding to the target search path. The training loss of the reward model is calculated based on the predicted reward value and the reference reward value corresponding to the target search path. The training loss is used to adjust the parameters in the reward model until the reward model reaches a preset model convergence condition, at which point training is stopped and a trained reward model is obtained. Alternatively, when the number of times the reward model has been trained reaches a preset threshold, training is stopped and a trained reward model is obtained.

[0095] The above scheme trains the constructed reward model by using the target search path and reference reward value matched by the second training report, so that the obtained reward model has the ability to predict the reward value, which helps to determine the optimal path from the search path generated by the MCTS algorithm according to the predicted reward value, or select the optimal sub-node from the sub-nodes generated by the MCTS algorithm according to the predicted reward value.

[0096] In another embodiment, after obtaining the fine-tuned intelligent analysis model and the trained reward model through the above corresponding embodiments, the intelligent analysis model and the reward model are further trained in combination with the MCTS algorithm to improve the processing performance of the intelligent analysis model and the reward model.

[0097] In one implementation scenario, multiple third training reports are obtained, and the third training reports are matched with corresponding third diagnostic labels. The third training thinking chain corresponding to the third training report is generated using the fine-tuned intelligent analysis model obtained through the above corresponding implementation methods. The second reference search path corresponding to the third training thinking chain is generated using the MCTS algorithm, and the second predicted diagnosis result is generated based on the second reference search path corresponding to the third training thinking chain. In response to the second predicted diagnosis result being consistent with the third diagnostic label, the corresponding third training report is used as a positive sample; and in response to the second predicted diagnosis result being inconsistent with the third diagnostic label, the corresponding third training report is used as a negative sample. The above positive and negative samples are used to further train the fine-tuned intelligent analysis model and the trained reward model.

[0098] Specifically, based on the determined positive and negative samples, the Direct Preference Optimization (DPO) algorithm is used to obtain a supervised loss function, and the parameters in the intelligent analysis model are adjusted according to the supervised loss function until the preset stopping condition is met, thereby obtaining the final trained intelligent analysis model. Furthermore, based on the determined positive and negative samples, the reward model is used to generate corresponding reference reward values, and the mean squared error (MSE) loss function is used to calculate the training loss based on the reference reward values ​​and the third diagnostic label corresponding to the positive and negative samples, respectively. The training loss is then used to adjust the parameters of the reward model until the preset stopping condition is met, thereby obtaining the final trained reward model.

[0099] Furthermore, the intelligent analysis model and reward model finally obtained by training are used to obtain the target diagnosis result that matches the report to be diagnosed. The specific implementation process can refer to the corresponding implementation method mentioned above.

[0100] See also Figure 6 , Figure 6 1 is a schematic diagram of the structure of an embodiment of the medical diagnosis system of the present application. The medical diagnosis system includes an acquisition module 10, a processing module 20 and a diagnosis module 30 coupled to each other. Specifically:

[0101] The acquisition module 10 is used to obtain the report to be diagnosed and use the intelligent analysis model to determine the target thinking chain that matches the report to be diagnosed; wherein the intelligent analysis model is obtained by training using multiple training data, and the training data includes a first training report and a first training thinking chain that matches the first training report.

[0102] The processing module 20 is configured to generate a current node that matches a target thought chain and at least one candidate answer that matches the current node.

[0103] The diagnosis module 30 is used to obtain a target reward value for each candidate answer match, and generate a target diagnosis result for the to-be-diagnosed report match based on the candidate answers and the target reward values ​​for their matches.

[0104] In one embodiment, please continue to refer to Figure 6 The medical diagnosis system proposed in this application also includes a data acquisition module 40 coupled to the acquisition module 10. The steps of the data acquisition module 40 acquiring training data include: acquiring multiple first training reports; wherein the first training report is matched with a first diagnosis label and a timestamp; inputting the first training report into the intelligent analysis model to generate a corresponding reference thinking chain; based on the reference thinking chain, acquiring a reference diagnosis result corresponding to the first training report; in response to the reference diagnosis result being consistent with the first diagnosis label, using the reference thinking chain as the first training thinking chain for training the intelligent analysis model.

[0105] In one embodiment, the first training report is associated with at least one reference report, and the timestamp of the reference report is earlier than the timestamp of the first training report. The data acquisition module 40 inputs the first training report into the intelligent analysis model to generate a corresponding reference thinking chain, including: inputting the first training report into the intelligent analysis model, and using the intelligent analysis model to obtain reference knowledge related to the first training report; based on the reference knowledge, the reference report and the first training report, generating a reference thinking chain corresponding to the first training report.

[0106] In one embodiment, after the data acquisition module 40 obtains the reference diagnostic result corresponding to the first training report based on the reference thinking chain, it includes: determining whether the reference diagnostic result is consistent with the first diagnostic label; in response to the reference diagnostic result being inconsistent with the first diagnostic label, obtaining supplementary information that matches the first training report; inputting the supplementary information into the intelligent analysis model, and using the intelligent analysis model to generate an updated reference thinking chain based on the supplementary information; based on the updated reference thinking chain, obtaining the reference diagnostic result corresponding to the first training report, and returning to the step of determining whether the reference diagnostic result is consistent with the first diagnostic label.

[0107] In one embodiment, the processing module 20 generates a current node that matches the target thinking chain and at least one candidate answer that matches the current node, including: obtaining the current content of the target thinking chain, generating the current node based on the current content; generating at least one candidate answer that matches the current node; wherein the candidate answer includes extended content corresponding to at least one sub-node that matches the current node.

[0108] In addition, the diagnosis module 30 generates a target diagnosis result that matches the report to be diagnosed based on the candidate answers and their matching target reward values, including: taking the candidate answer corresponding to the maximum target reward value as the target answer that matches the current node; and obtaining the target diagnosis result based on all current nodes matched by the target thinking chain and their matching target answers.

[0109] In one embodiment, please continue to refer to Figure 6The medical diagnosis system proposed in this application also includes a training module 50 coupled to the processing module 20. The target reward value for matching the candidate answers is obtained based on the trained reward model. The process of training the reward model by the training module 50 includes: obtaining a second training report and generating a second training thinking chain corresponding to the second training report; wherein the second training report is matched with a second diagnosis label; based on the second training thinking chain, generating at least one corresponding reference search path; wherein each reference search path includes multiple training nodes; obtaining the predicted diagnosis result corresponding to each reference search path; based on the predicted diagnosis result and the second diagnosis label, determining the target search path and its matching reference reward value from all reference search paths; based on the reference reward value matched by the target search path, training the reward model until the trained reward model is obtained.

[0110] In one embodiment, the training module 50 determines a target search path and a reference reward value for its matching from all reference search paths based on the predicted diagnostic results and the second diagnostic label, including: taking the reference search path corresponding to the predicted diagnostic results consistent with the second diagnostic label as the target search path; obtaining a first number of predicted diagnostic results consistent with the second diagnostic label, and a second number corresponding to all predicted diagnostic results, and determining a reference reward value for the target search path matching based on the ratio of the first number to the second number.

[0111] See also Figure 7 , Figure 7 : This is a structural diagram of an embodiment of an electronic device of the present application. The electronic device includes: a memory 60 and a processor 70 coupled to each other. The memory 60 stores program instructions, and the processor 70 is used to execute the program instructions to implement the method mentioned in any of the above embodiments. Specifically, the electronic device includes but is not limited to: a desktop computer, a laptop computer, a tablet computer, a server, etc., which are not limited here. In addition, the processor 70 can also be called a CPU (Center Processing Unit). The processor 70 may be an integrated circuit chip with signal processing capabilities. The processor 70 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 70 can be implemented by an integrated circuit chip.

[0112] See also Figure 8 , Figure 8 This is a structural diagram of an embodiment of a computer-readable storage medium of the present application. The computer-readable storage medium 80 stores program instructions 90 that can be run by a processor. When the program instructions 90 are executed by the processor, the method mentioned in any of the above embodiments is implemented.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation methods described above are only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0114] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0115] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0116] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of each embodiment method of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0117] The above description is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A medical diagnosis method, characterized in that: include: Obtaining a report to be diagnosed, and determining a target thought chain matching the report to be diagnosed using an intelligent analysis model; wherein the intelligent analysis model is trained using a plurality of training data, the training data including a first training report and a first training thought chain matching the first training report; generating a current node that matches the target thought chain and at least one candidate answer that matches the current node; A target reward value for each candidate answer match is obtained, and a target diagnosis result for the report to be diagnosed is generated based on the candidate answers and the target reward values ​​for their matches.

2. The method according to claim 1, characterized in that The step of obtaining the training data includes: Acquire a plurality of the first training reports; wherein the first training reports are matched with a first diagnosis tag and a timestamp; Inputting the first training report into the intelligent analysis model to generate a corresponding reference thinking chain; Based on the reference thinking chain, obtaining a reference diagnosis result corresponding to the first training report; In response to the reference diagnosis result being consistent with the first diagnosis label, the reference thought chain is used as the first training thought chain for training the intelligent analysis model.

3. The method according to claim 2, characterized in that The first training report is associated with at least one reference report, the timestamp of the reference report is earlier than the timestamp of the first training report, and the inputting the first training report into the intelligent analysis model to generate a corresponding reference thinking chain includes: Inputting the first training report into the intelligent analysis model, and using the intelligent analysis model to obtain reference knowledge related to the first training report; Based on the reference knowledge, the reference report and the first training report, the reference thinking chain corresponding to the first training report is generated.

4. The method according to claim 2, characterized in that After obtaining the reference diagnosis result corresponding to the first training report based on the reference thinking chain, the method includes: determining whether the reference diagnostic result is consistent with the first diagnostic label; In response to the reference diagnosis result being inconsistent with the first diagnosis label, acquiring supplementary information matching the first training report; Inputting the supplementary information into the intelligent analysis model, and using the intelligent analysis model to generate an updated reference thought chain based on the supplementary information; Based on the updated reference thinking chain, a reference diagnosis result corresponding to the first training report is obtained, and the process returns to the step of determining whether the reference diagnosis result is consistent with the first diagnosis label.

5. The method according to claim 1, wherein Generating a current node that matches the target thought chain and at least one candidate answer that matches the current node includes: Acquire the current content of the target thought chain, and generate the current node based on the current content; Generate at least one candidate answer matching the current node; wherein the candidate answer includes extended content corresponding to at least one child node matching the current node; The step of generating a target diagnosis result matching the report to be diagnosed based on the candidate answers and the target reward value thereof includes: The candidate answer corresponding to the maximum target reward value is used as the target answer matched by the current node; The target diagnosis result is obtained based on all the current nodes matched by the target thinking chain and the target answers matched thereto.

6. The method according to claim 1, wherein The target reward value for matching the candidate answer is obtained based on a trained reward model, and the training process of the reward model includes: Obtaining a second training report and generating a second training thought chain corresponding to the second training report; wherein the second training report is matched with a second diagnosis tag; Based on the second training thought chain, generating at least one corresponding reference search path; wherein each reference search path includes a plurality of training nodes; Obtaining a prediction diagnosis result corresponding to each reference search path; determining a target search path and a reference reward value thereof for matching from all the reference search paths based on the predicted diagnosis result and the second diagnosis label; The reward model is trained based on the reference reward value matched by the target search path until the trained reward model is obtained.

7. The method according to claim 6, characterized in that The step of determining a target search path and a matching reference reward value thereof from all the reference search paths based on the predicted diagnosis result and the second diagnosis label includes: using the reference search path corresponding to the predicted diagnosis result consistent with the second diagnostic label as the target search path; A first number of the predicted diagnosis results consistent with the second diagnosis tag and a second number corresponding to all the predicted diagnosis results are obtained, and the reference reward value for the target search path matching is determined based on a ratio of the first number to the second number.

8. A medical diagnostic system, characterized in that: include: an acquisition module, configured to acquire a report to be diagnosed and determine a target thought chain matching the report to be diagnosed using an intelligent analysis model; wherein the intelligent analysis model is trained using a plurality of training data, the training data including a first training report and a first training thought chain matching the first training report; a processing module, configured to generate a current node matching the target thought chain and at least one candidate answer matching the current node; The diagnosis module is used to obtain a target reward value for each candidate answer match, and generate a target diagnosis result for the report to be diagnosed based on the candidate answers and the target reward values ​​for their matches.

9. An electronic device, characterized in that: include: A memory and a processor coupled to each other, wherein the memory stores program instructions, and the processor is configured to execute the program instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images