Method for training information generation model, method for generating auxiliary information and related devices

Abstract

The invention provides a method for training an information generation model, a method for generating auxiliary information and related devices, and relates to the technical field of artificial intelligence such as information generation, deep learning and intelligent medical treatment. A specific implementation mode of training the information generation model comprises the steps that a first sample medical record serves as input of the medical auxiliary information generation model, a reward function for the first sample medical record serves as supervision, and the medical auxiliary information generation model is trained in multiple rounds through a near-end strategy optimization algorithm, the reward function is determined based on a difference between a model processing result and a standard processing result corresponding to the sample medical record, and the model processing result is obtained by processing the first sample medical record by a medical auxiliary information generation model; in response to the fact that the function value of the reward function in the target training round is larger than or equal to a predetermined function value threshold value, training of the medical auxiliary information generation model is completed, and a target medical auxiliary information generation model is obtained. Therefore, the resource consumption in the training process of the medical auxiliary information generation model can be reduced, the training process can be stabilized, and the medical auxiliary information generation model with better performance can be trained.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, specifically to the field of artificial intelligence technologies such as information generation, deep learning, and smart medicine, and especially to methods for training information generation models, methods for generating auxiliary information, and devices, electronic devices, computer-readable storage media, and computer program products that match the methods. Background Art

[0002] With the continuous development of society and technology, artificial intelligence has achieved remarkable innovation and breakthroughs. Especially in the field of deep learning, with the improvement of computing power and the accumulation of big data, neural networks and processing models based on artificial intelligence technology have been widely used in various industries and are playing an increasingly important role.

[0003] In the medical field, the innovative application of artificial intelligence technology is particularly eye-catching. In the past, medical diagnosis and treatment relied on the experience and judgment of doctors. However, due to the vast amount of medical knowledge and the complexity of disease manifestations, doctors sometimes found it difficult to make accurate diagnoses or provide patients with the best treatment plans in a short period of time.

[0004] To help doctors more efficiently process large amounts of medical information, artificial intelligence (AI), particularly generative models, are becoming an increasingly important auxiliary tool. Generative models can generate multi-dimensional disease-related information from vast amounts of medical data, assisting doctors in analysis and decision-making. Therefore, the development of more efficient, stable, and high-quality models for generating such medical information is a pressing need and a critical area of ​​research. Summary of the Invention

[0005] The embodiments of the present disclosure provide a method for training an information generation model, a method and apparatus for generating auxiliary information, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] In a first aspect, an embodiment of the present disclosure proposes a method for training an information generation model, comprising: taking a first sample medical record as the input of the medical auxiliary information generation model, using a reward function for the first sample medical record as supervision, and training the medical auxiliary information generation model for multiple rounds through a proximal strategy optimization algorithm, wherein the reward function is determined based on the difference between the model processing result and the standard processing result corresponding to the sample medical record, and the model processing result is obtained by the medical auxiliary information generation model processing the first sample medical record; in response to the function value of the reward function in the target training round being greater than or equal to a predetermined function value threshold, completing the training of the medical auxiliary information generation model to obtain the target medical auxiliary information generation model.

[0007] In a second aspect, the embodiments of the present disclosure provide a device for training an information generation model, comprising: a model multi-round training unit configured to take a first sample medical record as input of a medical auxiliary information generation model, take a reward function for the first sample medical record as supervision, and train the medical auxiliary information generation model through a proximal policy optimization algorithm for multiple rounds, wherein the reward function is determined based on a difference between a model processing result and a standard processing result corresponding to the sample medical record, and the model processing result is obtained by processing the first sample medical record by the medical auxiliary information generation model; and a model training jump-out unit configured to complete training of the medical auxiliary information generation model to obtain a target medical auxiliary information generation model in response to the reward function having a function value greater than or equal to a predetermined function value threshold at a target training round.

[0008] In a third aspect, the embodiments of the present disclosure provide a method for generating auxiliary information, comprising: obtaining a to-be-processed medical record; and calling a target medical auxiliary information generation model to process the to-be-processed medical record to obtain medical auxiliary information corresponding to the to-be-processed medical record, wherein the target medical auxiliary information generation model is trained based on the method for training an information generation model of the first aspect.

[0009] In a fourth aspect, the embodiments of the present disclosure provide a device for generating auxiliary information, comprising: a medical record obtaining unit configured to obtain a to-be-processed medical record; and a generation model calling unit configured to call a target medical auxiliary information generation model to process the to-be-processed medical record to obtain medical auxiliary information corresponding to the to-be-processed medical record, wherein the target medical auxiliary information generation model is trained based on the device for training an information generation model of the second aspect.

[0010] In a fifth aspect, the embodiments of the present disclosure provide an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to implement the method for training an information generation model as described in any implementation manner of the first aspect and / or the method for generating auxiliary information as described in any implementation manner of the third aspect.

[0011] In a sixth aspect, the embodiments of the present disclosure provide a non-transitory computer-readable storage medium storing computer instructions for enabling a computer to implement the method for training an information generation model as described in any implementation manner of the first aspect and / or the method for generating auxiliary information as described in any implementation manner of the third aspect.

[0012] In the seventh aspect, an embodiment of the present disclosure provides a computer program product comprising a computer program, which, when executed by a processor, can implement the method for training information generation model as described in any implementation manner in the first aspect and / or the method for generating auxiliary information as described in any implementation manner in the third aspect.

[0013] The methods, devices, electronic devices, computer-readable storage media, and computer program products for training information generation models provided by the embodiments of the present disclosure use a first sample medical record as input to the medical auxiliary information generation model and a reward function for the first sample medical record as supervision. The medical auxiliary information generation model is trained for multiple rounds through a proximal strategy optimization algorithm, wherein the reward function is determined based on the difference between the model processing result and the standard processing result corresponding to the sample medical record, and the model processing result is obtained by the medical auxiliary information generation model processing the first sample medical record; in response to the function value of the reward function in the target training round being greater than or equal to a predetermined function value threshold, the training of the medical auxiliary information generation model is completed to obtain a target medical auxiliary information generation model.

[0014] The present disclosure can not only reduce resource consumption during the training process of the medical auxiliary information generation model, but also stabilize the training process and train a more performant medical auxiliary information generation model.

[0015] The methods, devices, electronic devices, computer-readable storage media and computer program products for generating auxiliary information provided by the embodiments of the present disclosure obtain medical records to be processed; call a target medical auxiliary information generation model to process the medical records to be processed, and obtain medical auxiliary information corresponding to the medical records to be processed, wherein the target medical auxiliary information generation model is trained based on the above-mentioned method of training information generation model.

[0016] The present disclosure can utilize the target medical auxiliary information generation model trained by the above-mentioned training information generation model method to more accurately and efficiently process medical records and generate medical auxiliary information corresponding to the medical records to be processed. In this way, it can provide more valuable medical assistance to users such as doctors.

[0017] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Other features, objects and advantages of the present disclosure will become more apparent from a reading of the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 is an exemplary system architecture in which the present disclosure may be applied; Figure 2 A flowchart of a process of training an information generation model provided by an embodiment of the present disclosure is shown. Figure 3 A flowchart of a process of maintaining annotation information of a second sample medical record provided by an embodiment of the present disclosure is shown. Figure 4 A flowchart of a process of generating auxiliary information provided by an embodiment of the present disclosure is shown. Figure 5 A flowchart of a process of training an information generation model implemented in an application scenario provided by an embodiment of the present disclosure is shown. Figure 6 A structural block diagram of an apparatus of training an information generation model provided by an embodiment of the present disclosure is shown. Figure 7 A structural block diagram of an apparatus of generating auxiliary information provided by an embodiment of the present disclosure is shown. Figure 8 A structural diagram of an electronic device suitable for executing the method of training an information generation model, the method of generating auxiliary information provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, which should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Also, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description. It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0020] In addition, in the technical solutions involved in the present disclosure, the acquisition, storage, use, processing, transportation, provision and disclosure of user personal information (for example, the "medical record" which may involve user personal information involved in the present disclosure) comply with relevant laws and regulations and do not violate public order and good customs.

[0021] Figure 1 An exemplary system architecture 100 of an embodiment of the method of training an information generation model, the method of generating auxiliary information, the apparatus, the electronic device and the computer readable storage medium to which the present disclosure can be applied is shown.

[0022] As Figure 1As shown, the system architecture 100 can include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is a medium for providing a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 can include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0023] A user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. The terminal devices 101, 102, 103 and the server 105 can be installed with various applications for realizing information communication between them, such as generative model training applications, auxiliary information generation applications, instant messaging applications, etc.

[0024] The terminal devices 101, 102, 103 and the server 105 can be hardware or software. When the terminal devices 101, 102, 103 are hardware, they can be various electronic devices with display screens, including but not limited to smartphones, tablet computers, laptop computers, desktop computers, etc. When the terminal devices 101, 102, 103 are software, they can be installed in the above-mentioned electronic devices, and can be implemented as multiple software or software modules, or as a single software or software module, which is not specifically limited here. When the server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers, or as a single server. When the server 105 is software, it can be implemented as multiple software or software modules, or as a single software or software module, which is not specifically limited here.

[0025] The server 105 can provide various services through various built-in applications. For example, in the process of training an information generation model, the server 105 can provide a generative model training application so that a user can train a target medical auxiliary information generation model through the generative model training application.

[0026] Correspondingly, the server 105 can achieve the following effects when running the generative model training application: first, obtaining the first sample medical record from the terminal devices 101, 102, and 103 through the network 104; then, the server 105 generates the model of medical auxiliary information taking the first sample medical record as the input, and the reward function for the first sample medical record as the supervision, and trains the model of medical auxiliary information through the proximal policy optimization algorithm for multiple rounds, wherein the reward function is determined based on the difference between the model processing result and the standard processing result corresponding to the sample medical record, and the model processing result is obtained by processing the first sample medical record by the model of medical auxiliary information; finally, in response to the function value of the reward function at the target training round being greater than or equal to the predetermined function value threshold, the training of the model of medical auxiliary information is completed, and the target model of medical auxiliary information is obtained.

[0027] It should be noted that in addition to being obtained from the terminal devices 101, 102, and 103 through the network 104, the first sample medical record can also be pre-stored in the server 105 locally in various ways. Therefore, when the server 105 detects that these data have been stored locally (for example, before starting the training task of the model of medical auxiliary information left over), it can choose to obtain these data directly from the local, and in this case, the exemplary system architecture 100 can also not include the terminal devices 101, 102, and 103 and the network 104.

[0028] Similarly, for the subsequent process of using the trained target model of medical auxiliary information, the server 105 can provide the application of generating auxiliary information in such a way that the user can provide the medical record to be processed to the target model of medical auxiliary information by interacting with the application of generating auxiliary information, and complete the processing of the medical record to be processed by the target model of medical auxiliary information, and obtain the corresponding processing result, such as medical auxiliary information.

[0029] Since training a medical auxiliary information generation model and deploying a medical auxiliary information generation model often need to occupy more computing resources and stronger computing power, the method for training an information generation model and the method for generating auxiliary information provided by each of the subsequent embodiments of the present disclosure are generally executed by a server 105 with stronger computing power and more computing resources, and accordingly, the method for training an information generation model and the device for generating auxiliary information are generally also arranged in the server 105. However, it should also be pointed out that when the terminal devices 101, 102, and 103 also have computing power and computing resources that meet the requirements, the terminal devices 101, 102, and 103 can also complete the above-mentioned operations performed by the server 105 through the generative model training application and the auxiliary information generation application installed thereon, and then output the same results as the server 105. Especially in the case where there are multiple terminal devices with different computing power, but the generative model training application and the auxiliary information generation application determine that the terminal device has strong computing power and has more remaining computing resources, the terminal device can be allowed to perform the above-mentioned operations, thereby appropriately reducing the computing pressure of the server 105, and accordingly, the method for training an information generation model and the device for generating auxiliary information can also be arranged in the terminal devices 101, 102, and 103 (for example, the process of training, deploying, and using the target medical auxiliary information generation model can also be implemented in the terminal devices 101, 102, and 103). In this case, the example system architecture 100 can also not include the server 105 and the network 104.

[0030] It should be understood that Figure 1 The number of terminal devices, networks, and servers in

[0031] Next, the process of training an information generation model will be discussed first.

[0032] For this purpose, please refer to Figure 2 . Figure 2 A flowchart of a process of training an information generation model provided by an embodiment of the present disclosure, which includes the process 200.

[0033] The process 200 specifically includes the following steps: Step 201: taking a first sample medical record as the input of the medical auxiliary information generation model, taking the reward function for the first sample medical record as the supervision, and training the medical auxiliary information generation model through the proximal policy optimization algorithm for multiple rounds; In an embodiment of the present disclosure, this step is intended to be executed by the execution subject of the method for training an information generation model (for example Figure 1The server 105 shown in the figure) determines the medical auxiliary information generation model, takes the first sample medical record as input, uses the reward function of the first sample medical record as supervision, and trains the medical auxiliary information generation model in multiple rounds through the proximal strategy optimization algorithm.

[0034] The first sample medical record at least includes descriptive information about symptoms, or symptom information, such as chest pain, chest tightness, shortness of breath, etc. In some embodiments, such a first sample medical record may also include information such as past medical history that needs to be considered and collected during the medical interview (e.g., medical history, allergy history), etc.

[0035] It should be noted that the first sample medical record can be obtained by the execution subject directly from a local storage device, or from a non-local storage device (such as Figure 1 The local storage device can be a data storage module provided in the execution subject, such as a server hard disk. In this case, the first sample medical record can be quickly read locally. The non-local storage device can also be any other electronic device configured to store data, such as some user terminals. In this case, the execution subject can obtain the required first sample medical record by sending a retrieval command to the electronic device.

[0036] The medical auxiliary information generation model can be a pre-trained model capable of parsing, for example, medical records. For example, the medical auxiliary information generation model can process a first sample medical record to obtain a corresponding model processing result. For example, the model processing result can be diagnostic auxiliary information and / or examination auxiliary information obtained by the medical auxiliary information generation model after parsing and analyzing the content recorded in the first sample medical record.

[0037] The above-mentioned auxiliary diagnosis information may be directly composed of the name of the disease and the relevant content of the category to which the disease belongs, so that users such as doctors can refer to the auxiliary diagnosis information to determine the possible existence of the disease.

[0038] The auxiliary examination information may be in the form of, for example, recommended examination items, such as a blood test, a computed tomography (CT) scan, etc. Thus, the auxiliary examination information allows users, such as doctors, to refer to it and decide which examination items to select.

[0039] Correspondingly, for the medical auxiliary information generation model, whether the model processing result output after processing the first sample medical record specifically includes diagnosis auxiliary information, examination auxiliary information, or both can be determined based on the pre-training and configuration of the medical auxiliary information.

[0040] In some embodiments, to enhance the capabilities of medical auxiliary information generation models, they can be constructed based on large language models (LLMs). LLMs can be trained on large amounts of text data to perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. LLMs are characterized by their large scale and typically include a large number of parameters to help them learn complex patterns in language data. These models are often based on deep learning architectures, such as transformers, which help them provide better processing performance on various NLP tasks.

[0041] For example, the medical auxiliary information generation model can actually be constructed based on the 7B model, an LLM with 7 billion parameters. The 7B model enables the medical auxiliary information generation model to record medical knowledge on a large scale and utilize this knowledge to provide more accurate and high-quality medical auxiliary information.

[0042] In some embodiments, as discussed above, the medical auxiliary information generation model may have already undergone some training. For example, the medical auxiliary information generation model may be obtained by training an initial medical auxiliary information generation model. For example, such pre-training can enable the initial medical auxiliary information generation model, after being trained and converted into a medical auxiliary information generation model, to be capable of processing, for example, a first sample medical record. For example, after the medical auxiliary information generation model receives and inputs the first sample medical record, it can process the first sample medical record accordingly and obtain the model processing results, such as the aforementioned diagnosis auxiliary information and / or examination auxiliary information.

[0043] Accordingly, in order to achieve this goal, in the above-mentioned "pre-training process", instructions (sets) can be generated to instruct and train the initial medical auxiliary information generation model, so that the initial medical auxiliary information generation model can gradually understand the actions it needs to perform and eventually "change into" a medical auxiliary information generation model.

[0044] In this way, it is possible to avoid frequently instructing the medical auxiliary information generation model on the actions that need to be performed during the subsequent training process of the medical auxiliary information generation model, that is, the process of training it into a target medical auxiliary information generation model, and to enable the target medical auxiliary information generation model to have such "instruction" execution capabilities accordingly.

[0045] For the second sample medical record used in such "pre-training," the content related to "symptoms" can be annotated using symptom annotation information, the content related to "diagnosis" can be annotated using diagnosis annotation information, and the content related to "examination" can be annotated using examination annotation information. Then, the generation instructions in the generation instruction set can instruct the initial medical auxiliary information generation model to generate diagnosis annotation information and / or examination annotation information in the annotation information based on the symptom annotation information in the annotation information.

[0046] Based on this training, the initial medical auxiliary information generation model can gradually learn, as instructed by the generated instructions, and acquire the ability to determine corresponding diagnosis annotation information and / or examination annotation information (i.e., content subsequently considered "diagnosis" or "examination") based on symptom annotation information (and, subsequently, content considered "symptoms"), thereby obtaining a medical auxiliary information generation model. Accordingly, the subsequent medical auxiliary information generation model can also identify symptom information in the input medical record and determine corresponding diagnosis auxiliary information and / or examination auxiliary information.

[0047] In other words, after being trained, the medical auxiliary information generation model can, after obtaining the input of, for example, the first sample medical record, generate corresponding diagnostic auxiliary information and / or examination auxiliary information as the above-mentioned model processing result based on the "symptoms" "considered" and "identified" by the medical auxiliary information generation model.

[0048] In some embodiments, during the process of training the initial medical auxiliary information generation model into a medical auxiliary information generation model, the generation instruction may further indicate the content format of the medical auxiliary information generation model when subsequently providing the model output results.

[0049] For example, the generation instruction may also indicate that the medical auxiliary information generation model provides the model processing results in the form of "based on XX symptoms, it is recommended to refer to YY disease, and it is recommended to perform ZZ examination."

[0050] This enables the medical auxiliary information generation model to provide model processing results in a standard text format, and subsequently enables the trained target medical auxiliary information generation model to provide medical auxiliary information to users, such as doctors, in this standard format. This enables users, such as doctors, to read and obtain medical auxiliary information (e.g., diagnostic auxiliary information and examination auxiliary information) more efficiently and accurately, improving the user experience.

[0051] In some embodiments, the second sample medical record and the above-mentioned first sample medical record may be different, so that different sample medical records can be used to complete the initial medical auxiliary information generation model and the training process of the medical auxiliary information generation model in stages to avoid overfitting.

[0052] In some embodiments, in the process of training the initial medical auxiliary information generation model based on the second sample medical record, the annotation information of the second sample medical record and the generation instruction set to obtain the medical auxiliary information generation model, it is also possible to further choose to train the initial medical auxiliary information generation model in a supervised fine-tuning manner to obtain the medical auxiliary information generation model.

[0053] Supervised Fine-Tuning (SFT) refers to retraining an already trained pre-trained model using labeled task-specific data, thereby adjusting the model's parameters to better adapt to new tasks.

[0054] Correspondingly, for example, after obtaining a pre-trained general 7B model, the executing entity can use the general 7B model as the above-mentioned initial medical auxiliary information generation model, and then use the above-mentioned second sample medical records, the annotation information of the second sample medical records and the generation instruction set to migrate the general 7B model to the "medical field" that can process medical records in an SFT manner to obtain a medical auxiliary information generation model.

[0055] Therefore, using the SFT approach allows us to use existing LLMs with strong capabilities, such as text processing, as a foundation for building a medical auxiliary information generation model. This simplifies the initial training cost and difficulty of the medical auxiliary information generation model while using the already strong initial medical auxiliary information generation model as a foundation to enhance the model's capabilities.

[0056] In an embodiment of the present disclosure, during the training of the medical auxiliary information generation model, the execution entity may choose to construct a reward function for the first sample medical record based on the difference between the model processing result and the standard processing result corresponding to the first sample medical record.

[0057] In some embodiments, the standard processing result may be provided for the first sample medical record, for example, based on manual entry, etc. For example, the standard processing result may be provided in advance for the disease and examination items involved in the first sample medical record.

[0058] In some embodiments, first sample medical records that include accurate and reliable disease diagnosis results and examination items may also be selected, and standard processing results may be provided by adding annotations to these accurate and reliable disease diagnosis results and examination items.

[0059] The reward function is then used to monitor the training process, thereby overseeing the multiple rounds of training of the medical auxiliary information generation model. For example, if the function value corresponding to the reward function is greater than or equal to a predetermined function value threshold (for example, the function value threshold can be set based on the function value corresponding to the determination that the model capabilities of the trained medical auxiliary information generation model after the current round have met the usage requirements), the execution entity can choose to skip the multiple rounds of training to complete the training process of the medical auxiliary information generation model and obtain the target medical auxiliary information generation model.

[0060] In some embodiments, such a "jump" can also be achieved by the execution entity when it determines that the function value of the reward function has reached its maximum value (for example, in some scenarios, the execution entity may consider that the function value has not changed significantly and has reached the "maximum value" when the change amplitude between a consecutive preset number of function values ​​is less than or equal to the amplitude threshold).

[0061] In an embodiment of the present disclosure, the function value corresponding to the reward function may be negatively correlated with the difference between the above-mentioned model processing result and the standard processing result, that is, the smaller the difference, the larger the function value of the reward function.

[0062] In some embodiments, the reward function may be determined based on the Jaccard distance between the model processing result and the standard processing result.

[0063] Specifically, the execution entity can determine the Jaccard distance between the model processing result and the standard processing result by the following formula (1): J : (1) in, A Processing results for the model (collection), B It is the standard processing result (collection).

[0064] Accordingly, the reward function Reward It can be determined based on the following formula (2): (2) Thus, as Jaccard distance J The smaller the value of the reward function Reward becomes, the larger the value of the reward function Reward becomes. J To determine and construct the reward function Reward , which can ignore the internal order and repeated elements of the model processing results and the standard processing results, so that the reward function can more effectively and accurately correspond to the difference between the two.

[0065] In the embodiments of the present disclosure, in the process of training the medical auxiliary information generation model in a multi-round manner, the execution subject can select to train the medical auxiliary information generation model in a multi-round manner by using a proximal policy optimization algorithm (Group Relative Policy Optimization, GRPO) to improve the training effect of the model.

[0066] For example, the GRPO-based training process can be implemented by using a clipping function provided by the following formula (3): (3) wherein, represents the probability ratio between the model processing result output in the current round and the model processing result output in the previous round, is an advantage function for measuring the quality of the model processing result under the state is a hyperparameter for limiting the variation range of the model processing result.

[0067] Correspondingly, by using , the execution subject can first interact with the environment based on the current policy in the training round to generate a batch of trajectories.

[0068] Then, the advantage function of each time step is calculated by using the time difference method or generalized advantage estimation, and then the advantage of the model output result (for example, diagnosis auxiliary information, examination auxiliary information) provided in this time is determined.

[0069] Then, the execution subject calculates the ratio between the model output result provided in this time and the previous model result, and then determines correspondingly. And by using the way of gradient descent optimization of , the training is carried out in multiple rounds.

[0070] Correspondingly, in this process, the execution subject can determine the timing of jumping out or completing the training by continuously monitoring the change of the reward function.

[0071] For this "timing", as discussed above, the execution subject can select to jump out when the reward function has a function value greater than or equal to a predetermined function value threshold in the target training round.

[0072] ​​​Step 202: in response to the reward function having a function value greater than or equal to a predetermined function value threshold at the target training round, completing the training of the medical auxiliary information generation model to obtain a target medical auxiliary information generation model.

[0073] In the embodiments of the present disclosure, if the execution subject determines that the reward function has a function value greater than or equal to a predetermined function value threshold at the target training round, the execution subject can select to perform step 202 in response to this to achieve "jumping out" and completing the training of the medical auxiliary information generation model.

[0074] The method for training an information generation model provided in the embodiments of the present disclosure takes a first sample medical record as the input of the medical auxiliary information generation model, takes the reward function for the first sample medical record as the supervision, and trains the medical auxiliary information generation model through the proximal policy optimization algorithm for multiple rounds, wherein the reward function is determined based on the difference between the model processing result and the standard processing result corresponding to the sample medical record, and the model processing result is obtained by processing the first sample medical record by the medical auxiliary information generation model; in response to the reward function having a function value greater than or equal to a predetermined function value threshold at the target training round, the training of the medical auxiliary information generation model is completed to obtain a target medical auxiliary information generation model.

[0075] In this way, not only can the resource consumption in the training process of the medical auxiliary information generation model be reduced, but also the training process can be stabilized, and a medical auxiliary information generation model with better performance can be trained.

[0076] In some embodiments, if the execution subject trains the initial medical auxiliary information generation model by using the second sample medical record (and the annotation information and the generation instruction set of the second sample medical record), in order to avoid misleading the training process due to low-quality annotation information, the execution subject can also select to detect the annotation information in the second sample medical record to optimize, update, or limit those low-quality annotation information.

[0077] Specifically, for the diagnosis annotation information, the execution subject can compare the diagnosis annotation information included in the second sample medical record with the first standard information database to determine whether the first target diagnosis annotation information not included in the first standard information database is recorded or included in the second sample medical record.

[0078] The first standard information database can record the standard name of each disease, the name of an expert, and the like. In this way, the execution subject can select to use the first standard information database to check and screen out the first target diagnosis annotation information expressed inaccurately or incorrectly in the annotation information of the second sample medical record.

[0079] Correspondingly, if the second sample medical record contains first target diagnosis annotation information that is not recorded in the first standard information database, the execution entity can respond to this by selecting standard diagnosis annotation information from the first standard information database that has the highest semantic similarity with the first target diagnosis annotation information.

[0080] Next, if the "highest semantic similarity" (i.e., the semantic similarity between the "standard diagnostic annotation information with the highest semantic similarity" and the first target diagnostic annotation information) is greater than or equal to a first similarity threshold (for example, the first similarity threshold can be set based on the standard that both can be considered to credibly express the same content), the execution entity can respond to this by choosing to replace the first target diagnostic annotation information with the standard diagnostic annotation information.

[0081] Therefore, when there is non-standard and inaccurate first target diagnostic labeling information, it can be "corrected" and "rewritten" using standard diagnostic labeling information to avoid the initial medical auxiliary information generation model learning non-standard expressions and reducing the model quality.

[0082] In some optional implementations of this embodiment, if the semantic similarity between the standard diagnostic annotation information with the highest semantic similarity and the first target diagnostic annotation information is less than the above-mentioned first similarity threshold, that is, the standard diagnostic annotation information with the highest semantic similarity cannot be accurately and reliably considered as the "first target diagnostic annotation information", then in such a case, the executing entity may choose to remove the first target diagnostic annotation information from the annotation information to avoid the initial medical auxiliary information being erroneously trained by the first target diagnostic annotation information.

[0083] In some embodiments, checking annotation information may also be processed in at least a similar manner.

[0084] Specifically, the execution entity may also respond when the target examination annotation information in the annotation information of the second sample medical record is not recorded in the second standard information database, and determine the standard examination annotation information with the highest semantic similarity to the target examination annotation information from the second standard information database.

[0085] In practice, depending on the information maintenance setup, the first standard information database and the second standard information database can be maintained in a single database or separately and independently. Accordingly, the second standard information database can record the standard names of various examination items, etc. Thus, the executing entity can choose to use the second standard information database to check and filter out inaccurate or incorrect target examination annotation information included in the annotation information of the second sample medical record.

[0086] Similarly, if target inspection annotation information exists, the execution entity determines standard inspection annotation information having the highest semantic similarity with the target inspection annotation information from the second standard information database.

[0087] Subsequently, if the semantic similarity between the standard inspection annotation information with the highest semantic similarity and the target inspection annotation information is greater than or equal to a second similarity threshold, the execution entity may choose to update the target inspection annotation information using the standard inspection annotation information.

[0088] Therefore, when there is non-standard and inaccurate target inspection annotation information, it can be "corrected" and "rewritten" using standard inspection annotation information to avoid the initial medical auxiliary information generation model learning non-standard expressions and reducing the model quality.

[0089] In some embodiments, considering that users such as doctors may often need to combine more examination items to make a diagnosis, and that examination items may often have more nicknames, in some scenarios, the value of the second similarity threshold may be different from the value of the first similarity threshold. For example, the second similarity threshold may be smaller than the first similarity threshold. In this way, the examination annotation information can have a more relaxed standard than the diagnosis annotation information, so that it can be corrected by the examination standard information as much as possible instead of being deleted.

[0090] Similarly, if the semantic similarity between the standard examination annotation information and the target examination annotation information is less than a second similarity threshold, the execution entity may also choose to remove the target examination annotation information from the annotation information to avoid the initial medical auxiliary information being incorrectly trained by the target examination annotation information.

[0091] In some embodiments, for examination items, since they are often closely related to diseases, when the target examination annotation information cannot be corrected by the examination standard information, the executing entity may also choose to further determine whether the annotation information of the second sample medical record includes second target diagnosis annotation information corresponding to the above-mentioned target examination annotation information, and if the second target diagnosis annotation information is included, use the second target diagnosis annotation information as a reference.

[0092] In other words, if such second target diagnosis annotation information exists, the execution entity may attempt to use the second target diagnosis annotation information to correct the target inspection annotation information.

[0093] For easier understanding, you can also refer to Figure 3 . Figure 3 A flowchart of a process for maintaining annotation information of a second sample medical record provided by an embodiment of the present disclosure includes process 300 .

[0094] For ease of understanding, in process 300 , only the case where “the target examination annotation information in the second sample medical record is not recorded in the second standard information database” is presented.

[0095] Process 300 includes the following steps: Step 301: In response to the target inspection annotation information in the annotation information not being recorded in the second standard information database, determining the standard inspection annotation information having the highest semantic similarity with the target inspection annotation information from the second standard information database; Specifically, in this step, the executing entity can respond as discussed above when the target examination annotation information in the annotation information of the second sample medical record is not recorded in the second standard information database, and determine the standard examination annotation information with the highest semantic similarity to the target examination annotation information from the second standard information database.

[0096] Then, if the semantic similarity between the standard inspection annotation information with the highest semantic similarity and the target inspection annotation information is greater than or equal to the second similarity threshold, the execution entity may respond by executing step 302 to “update”.

[0097] If the semantic similarity between the standard inspection annotation information with the highest semantic similarity and the target inspection annotation information is less than the second similarity threshold, the execution entity may choose to execute step 303 instead of step 302 .

[0098] Step 302: Update target inspection annotation information using standard inspection annotation information; Step 303: Check whether the annotation information includes second target diagnosis annotation information corresponding to the target inspection annotation information; Specifically, in this step, the executing entity may determine the corresponding possible diagnostic information based on the target examination annotation information. For example, the executing entity may determine the possible corresponding diagnostic information for the target examination annotation information based on its semantic understanding of the target examination annotation information, or based on reference standard examination annotation information with similar semantics to the target examination annotation information (for example, the "possible corresponding diagnostic information" may be determined based on the disease that can be detected by the reference standard examination annotation information).

[0099] It should be understood that in such a case, in order to hit the diagnostic information as much as possible and correct the inspection information, the execution entity can use a third similarity threshold that is more relaxed than the above-mentioned second similarity threshold (for example, the third similarity threshold can be smaller than the above-mentioned second similarity threshold) to determine the standard inspection annotation information that is semantically similar to the target inspection annotation information.

[0100] If the determined possible corresponding diagnostic information is also included in the annotation information, for example, the second target diagnostic annotation information in the annotation information is actually the “possibly corresponding diagnostic information”, the execution entity may continue to execute step 304 .

[0101] If such “possible corresponding diagnostic information” cannot be detected, that is, there is no “second target diagnostic annotation information” in the annotation information, the execution entity may choose to execute step 305 instead of step 304 .

[0102] Step 304: Determine updated inspection annotation information from a pre-configured diagnosis and inspection relationship database using the second target diagnosis annotation information, and replace the target inspection annotation information with the updated inspection annotation information; Specifically, in this step, if the executing entity is able to hit and query the "possible corresponding diagnostic information", that is, the "second target diagnostic labeling information", based on the above step 303, it can use the pre-configured diagnosis and inspection relationship database (for example, the diagnosis and inspection relationship database can maintain standard form diagnostic information, standard form inspection information, and the association between the two, for example, there is an association with the standard form diagnostic information E, which is used to verify and determine whether there is standard form inspection information F, G of the diagnostic information E, etc.), to find the updated inspection labeling information corresponding to the second target diagnostic labeling information and with which there is an association.

[0103] Accordingly, in this step, the execution entity may choose to replace the target inspection annotation information with the updated inspection annotation information to achieve the purpose of removing the target inspection annotation information.

[0104] Step 305: Remove the target inspection annotation information from the annotation information.

[0105] Specifically, if there is no updated inspection mark information that can replace the target inspection mark information, the execution entity may choose to directly remove the target inspection mark information to avoid interference caused by the target inspection mark information.

[0106] Thus, in this way, if the inspection annotation information may contain errors, the execution entity can first try to find the associated diagnostic annotation information (i.e., the second target diagnostic annotation information) and, if such diagnostic annotation information exists, use it to attempt to correct it. This can ensure that the amount of information in the second sample medical record is maximized while avoiding interference caused by erroneous information.

[0107] Next, based on any of the above embodiments, the execution subject can complete the training of the medical auxiliary information generation model to obtain the target medical auxiliary information generation model.

[0108] For such a target medical auxiliary information generation model, it can be used to generate auxiliary information.

[0109] For this, please refer to Figure 4 , Figure 4 A flowchart of a process for generating auxiliary information provided by an embodiment of the present disclosure includes process 400.

[0110] The process 400 specifically includes the following steps: Step 401: Obtain the medical records to be processed; In the embodiments of the present disclosure, a user can communicate with an execution entity through their terminal devices 101, 102, 103, etc. to upload or submit medical records to be processed. For example, if a user desires the execution entity to provide auxiliary medical information for reference based on the analysis of the medical records to be processed, the user can provide the medical records to the execution entity.

[0111] For example, the medical record to be processed may include symptom information filled in by the user.

[0112] Step 402: calling the target medical auxiliary information generation model to process the medical record to be processed, and obtaining the medical auxiliary information corresponding to the medical record to be processed.

[0113] In an embodiment of the present disclosure, after receiving the medical records to be processed in step 401, the execution entity may choose to call a target medical auxiliary information generation model to process the medical records to be processed. The target medical auxiliary information generation model may be the target medical auxiliary information generation model trained by, for example, processes 200 and 300.

[0114] Accordingly, the target medical auxiliary information generation model can process the pending medical records to output medical auxiliary information such as diagnosis auxiliary information and / or examination auxiliary information for the user as a reference.

[0115] Subsequently, based on different scenarios, the execution entity may present and provide the medical auxiliary information through its associated display component, such as a screen, or return and provide the medical auxiliary information to the user based on a pre-configured communication link, which will not be repeated here.

[0116] The method for generating auxiliary information provided in the embodiments of the present disclosure obtains a medical record to be processed; invokes a target medical auxiliary information generation model to process the medical record to be processed, and obtains medical auxiliary information corresponding to the medical record to be processed. Thus, the target medical auxiliary information generation model trained using the aforementioned method for training an information generation model can be used to more accurately and efficiently process medical records and generate medical auxiliary information corresponding to the medical record to be processed. This allows for more valuable medical assistance to be provided to users, such as doctors.

[0117] In order to deepen understanding, this disclosure also provides a specific implementation solution in combination with a specific application scenario. For this, please refer to Figure 5 . Figure 5 A flowchart of a process for generating a training information model in an application scenario provided by an embodiment of the present disclosure.

[0118] exist Figure 5 In the process 500 shown, for example, the execution entity of the server 105 (not directly shown in the figure) can first train the initial medical auxiliary information generation model 521 based on the sample medical record 511, the annotation information 513 and the generation instruction set 515 by executing S501 to obtain the medical auxiliary information generation model 522.

[0119] Then, after obtaining the medical auxiliary information generation model 522, the execution entity can further utilize the sample medical record 531 (the sample medical record 531 may be different from the sample medical record 511), with the reward function 532 as supervision, to execute S502, to train the medical auxiliary information generation model 522 based on the sample medical record 531 and the reward function 522 through multiple rounds of proximal strategy optimization algorithm.

[0120] Next, if in a training round, that is, the function value of the reward function 532 in the target round is greater than or equal to a predetermined function value threshold, the execution body can jump out and use the medical auxiliary information generation model 522 obtained after this round of training as the final medical auxiliary information generation model 523 to complete the training process of the generation model.

[0121] Further references Figure 6 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a device for training an information generation model. Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0122] like Figure 6As shown, the apparatus 600 for training an information generation model in this embodiment may include: a model multi-round training unit 601 and a model training exit unit 602. The model multi-round training unit 601 is configured to use a first sample medical record as input to the medical auxiliary information generation model and a reward function for the first sample medical record as supervision to train the medical auxiliary information generation model for multiple rounds through a proximal policy optimization algorithm, wherein the reward function is determined based on the difference between the model processing result and the standard processing result corresponding to the sample medical record, and the model processing result is obtained by the medical auxiliary information generation model processing the first sample medical record; and the model training exit unit 602 is configured to complete the training of the medical auxiliary information generation model in response to the function value of the reward function in the target training round being greater than or equal to a predetermined function value threshold, thereby obtaining a target medical auxiliary information generation model.

[0123] In this embodiment, in the apparatus 600 for generating a model through training information, the specific processing of the model multi-round training unit 601 and the model training jump-out unit 602 and the technical effects thereof can be referred to in the respective Figure 2 The relevant descriptions of steps 201-202 in the corresponding embodiment are not repeated here.

[0124] In some optional implementations of this embodiment, the device 600 also includes: a model initial training unit, configured to train an initial medical auxiliary information generation model based on the second sample medical record, the annotation information of the second sample medical record and a generation instruction set to obtain a medical auxiliary information generation model, wherein the generation instructions in the generation instruction set are used to instruct the initial medical auxiliary information generation model to generate diagnosis annotation information and / or examination annotation information in the annotation information based on the symptom annotation information in the annotation information.

[0125] In some optional implementations of this embodiment, the model initial training unit is further configured to train the initial medical auxiliary information generation model in a supervised fine-tuning manner based on the second sample medical record, the annotation information of the second sample medical record and the generation instruction set to obtain the medical auxiliary information generation model.

[0126] In some optional implementations of this embodiment, the device 600 also includes: a standard diagnostic labeling information query unit, configured to determine, from the first standard information database, standard diagnostic labeling information having the highest semantic similarity with the first target diagnostic labeling information in response to the first target diagnostic labeling information in the second sample medical record not being recorded in the first standard information database; and a first labeling information update unit, configured to update the first target diagnostic labeling information using the standard diagnostic labeling information in response to the semantic similarity between the standard diagnostic labeling information and the first target diagnostic labeling information being greater than or equal to a first similarity threshold.

[0127] In some optional implementations of this embodiment, the device 600 also includes: a second annotation information updating unit, configured to remove the first target diagnosis annotation information from the annotation information in response to the semantic similarity between the standard diagnosis annotation information and the first target diagnosis annotation information being less than a first similarity threshold.

[0128] In some optional implementations of this embodiment, the device 600 also includes: a standard inspection annotation information query unit, configured to determine, from the second standard information database, standard inspection annotation information having the highest semantic similarity with the target inspection annotation information in the annotation information, in response to the target inspection annotation information in the annotation information not being recorded in the second standard information database; and a third annotation information updating unit, configured to update the target inspection annotation information using the standard inspection annotation information, in response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being greater than or equal to a second similarity threshold.

[0129] In some optional implementations of this embodiment, the device 600 also includes: a fourth annotation information updating unit, configured to remove the target inspection annotation information from the annotation information in response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being less than a second similarity threshold.

[0130] In some optional implementations of this embodiment, the fourth annotation information updating unit is further configured to, in response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being less than a second similarity threshold, and the annotation information including second target diagnosis annotation information corresponding to the target inspection annotation information, determine updated inspection annotation information from a pre-configured diagnosis and inspection relationship database using the second target diagnosis annotation information, and replace the target inspection annotation information with the updated inspection annotation information.

[0131] In some optional implementations of this embodiment, the apparatus 600 further includes: a reward function determination unit configured to determine a reward function based on a Jaccard distance between a model processing result and a standard processing result.

[0132] This embodiment exists as an apparatus embodiment corresponding to the above-mentioned method embodiment. The apparatus for training the information generation model provided by this embodiment can not only reduce the resource consumption during the training process of the medical auxiliary information generation model, but also stabilize the training process and train a more performant medical auxiliary information generation model.

[0133] Further references Figure 7 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a device for generating auxiliary information. Figure 4 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0134] As Figure 7 shown, the device 700 for generating auxiliary information can include: a medical record obtaining unit 701 and a generation model calling unit 702. The medical record obtaining unit 701 is configured to obtain a to-be-processed medical record. The generation model calling unit 702 is configured to call a target medical auxiliary information generation model to process the to-be-processed medical record, to obtain medical auxiliary information corresponding to the to-be-processed medical record, where the target medical auxiliary information generation model is obtained based on the device 600.

[0135] In this embodiment, in the device 700 for generating auxiliary information: the specific processing of the medical record obtaining unit 701 and the generation model calling unit 702 and the technical effects brought by the specific processing can be respectively referred to Figure 4 The related description of steps 401-402 in the corresponding embodiment will not be repeated here.

[0136] This embodiment exists as a device embodiment corresponding to the above-mentioned method embodiment. The device for generating auxiliary information provided in this embodiment can use the target medical auxiliary information generation model obtained by the device 600 for training information generation model to more accurately and efficiently process medical records and generate medical auxiliary information corresponding to the to-be-processed medical record. Thus, more valuable medical assistance can be provided to users such as doctors.

[0137] According to embodiments of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0138] Figure 8 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit the implementations of the present disclosure described and / or claimed in this document.

[0139] As Figure 8As shown, device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. RAM 803 may also store various programs and data required for the operation of device 800. Computing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to bus 804.

[0140] Various components in device 800 are connected to I / O interface 805, including an input unit 806, such as a keyboard, mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a magnetic disk, optical disk, etc.; and a communication unit 809, such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows device 800 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0141] The computing unit 801 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the method for training information to generate a model and the method for generating auxiliary information. For example, in some embodiments, the method for training information to generate a model and the method for generating auxiliary information can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the method for training information to generate a model and the method for generating auxiliary information described above can be performed. Alternatively, in other embodiments, the computing unit 801 may be configured in any other appropriate manner (eg, by means of firmware) to execute the method of training information to generate a model and the method of generating auxiliary information.

[0142] Various embodiments of the systems and techniques described above can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0143] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0144] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0145] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0146] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0147] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. A server can be a cloud server, also known as cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of large management difficulty and weak business scalability in traditional physical host and virtual private server (VPS, Virtual Private Server) services. The server can also be a server of a distributed system, or a server combined with a blockchain.

[0148] According to the technical scheme of the embodiment of the present disclosure, the resource consumption in the training process of the medical auxiliary information generation model can be reduced, and the training process can be stabilized, and a medical auxiliary information generation model with higher performance can be trained. Moreover, the target medical auxiliary information generation model trained by the above-mentioned training information generation model can be used to more accurately and efficiently process medical records and generate medical auxiliary information corresponding to the medical records to be processed. Therefore, medical assistance can be provided to users such as doctors more valuable.

[0149] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions provided by this disclosure can be achieved. This is not limited herein.

[0150] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A method for training an information generation model, comprising: The medical auxiliary information generation model is trained multiple times using a proximal policy optimization algorithm, using a first sample medical record as input and a reward function for the first sample medical record as supervision, wherein the reward function is determined based on a difference between a model processing result obtained by processing the first sample medical record by the medical auxiliary information generation model and a standard processing result corresponding to the sample medical record; In response to the function value of the reward function in the target training round being greater than or equal to a predetermined function value threshold, the training of the medical auxiliary information generation model is completed to obtain a target medical auxiliary information generation model.

2. The method according to claim 1, further comprising: Based on the second sample medical record, the annotation information of the second sample medical record and the generation instruction set, an initial medical auxiliary information generation model is trained to obtain the medical auxiliary information generation model, wherein the generation instructions in the generation instruction set are used to instruct the initial medical auxiliary information generation model to generate diagnosis annotation information and / or examination annotation information in the annotation information based on the symptom annotation information in the annotation information.

3. The method according to claim 2, wherein: The step of training an initial medical auxiliary information generation model based on the second sample medical record, the annotation information of the second sample medical record, and the generation instruction set to obtain the medical auxiliary information generation model includes: Based on the second sample medical record, the annotation information of the second sample medical record and the generation instruction set, the initial medical auxiliary information generation model is trained in a supervised fine-tuning manner to obtain the medical auxiliary information generation model.

4. The method according to claim 2, further comprising: In response to the first target diagnosis annotation information in the second sample medical record not being recorded in the first standard information database, determining, from the first standard information database, standard diagnosis annotation information having the highest semantic similarity to the first target diagnosis annotation information; In response to the semantic similarity between the standard diagnosis labeling information and the first target diagnosis labeling information being greater than or equal to a first similarity threshold, the first target diagnosis labeling information is updated using the standard diagnosis labeling information.

5. The method according to claim 4, further comprising: In response to the semantic similarity between the standard diagnosis annotation information and the first target diagnosis annotation information being less than the first similarity threshold, the first target diagnosis annotation information is removed from the annotation information.

6. The method according to claim 2, further comprising: In response to the target inspection annotation information in the annotation information not being recorded in the second standard information database, determining, from the second standard information database, standard inspection annotation information having the highest semantic similarity to the target inspection annotation information; In response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being greater than or equal to a second similarity threshold, the target inspection annotation information is updated using the standard inspection annotation information.

7. The method according to claim 6, further comprising: In response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being less than the second similarity threshold, the target inspection annotation information is removed from the annotation information.

8. The method according to claim 7, wherein: In response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being less than the second similarity threshold, removing the target inspection annotation information from the annotation information includes: In response to the semantic similarity between the standard inspection annotation information and the target inspection annotation information being less than the second similarity threshold, and the annotation information including the second target diagnosis annotation information corresponding to the target inspection annotation information, updated inspection annotation information is determined from a pre-configured diagnosis and inspection relationship database using the second target diagnosis annotation information, and the target inspection annotation information is replaced by the updated inspection annotation information.

9. The method according to any one of claims 1 to 8, further comprising: The reward function is determined based on the Jaccard distance between the model processing result and the standard processing result.

10. A method for generating auxiliary information, comprising: Obtain pending medical records; Call the target medical auxiliary information generation model to process the medical record to be processed to obtain medical auxiliary information corresponding to the medical record to be processed, wherein the target medical auxiliary information generation model is trained based on the method of training information generation model described in any one of claims 1-9.

11. A device for training an information generation model, comprising: a model multi-round training unit, configured to use a first sample medical record as an input to a medical auxiliary information generation model, and a reward function for the first sample medical record as supervision, to train the medical auxiliary information generation model for multiple rounds through a proximal policy optimization algorithm, wherein the reward function is determined based on a difference between a model processing result and a standard processing result corresponding to the sample medical record, the model processing result being obtained by processing the first sample medical record by the medical auxiliary information generation model; The model training jump-out unit is configured to complete the training of the medical auxiliary information generation model in response to the function value of the reward function in the target training round being greater than or equal to a predetermined function value threshold, and obtain the target medical auxiliary information generation model.

12. A device for generating auxiliary information, comprising: a medical record acquisition unit, configured to acquire medical records to be processed; The generation model calling unit is configured to call the target medical auxiliary information generation model to process the medical record to be processed and obtain medical auxiliary information corresponding to the medical record to be processed, wherein the target medical auxiliary information generation model is trained based on the device of the training information generation model according to claim 11.

13. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method for generating a training information model described in any one of claims 1 to 9 and / or the method for generating auxiliary information described in claim 10.

14. A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to execute the method for generating a training information model according to any one of claims 1 to 9 and / or the method for generating auxiliary information according to claim 10.

15. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method for generating a model through training information according to any one of claims 1 to 9 and / or the method for generating auxiliary information according to claim 10.

Images