Medical interaction method and device, electronic equipment and storage medium

Abstract

The invention provides a medical interaction method and device, electronic equipment and a storage medium, and relates to the technical field of data processing, and the method comprises the steps: determining a medical interaction process tree corresponding to a target user according to user portrait information of the target user, and generating first interaction information according to a root node of the medical interaction process tree; according to the historical dialogue of the target user and the user portrait information, matching a next jump node of the root node in the medical interaction process tree to obtain a first jump node; wherein the historical dialogue comprises first interaction information and first response information of the first interaction information; inputting the user portrait information, the historical dialogue and the node information of the first jump node into a medical interaction model to obtain second interaction information continuously interacting with the target user; wherein the second interaction information comprises an open problem or a structured problem generated based on the node information of the first jump node.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a medical interaction method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the rapid development of digital medical technology, medical dialogue systems based on artificial intelligence are widely used in the fields of auxiliary diagnosis and treatment, health consultation, etc.

[0003] The existing medical dialogue system mainly adopts an open dialogue system based on a large language model, which can understand natural language and generate flexible dialogue content, and can process various unstructured user inputs. Such systems perform well in terms of dialogue naturalness and flexibility, but in the medical scenario, the dialogue content may be uncontrollable and may deviate from the medical theme. Therefore, how to realize a better medical interaction scheme has become a problem to be solved in the industry. SUMMARY

[0004] The present application provides a medical interaction method, device, electronic equipment and storage medium to solve the problem of how to realize a better medical interaction scheme in the prior art.

[0005] The present application provides a medical interaction method, comprising: According to the user portrait information of the target user, a medical interaction process tree corresponding to the target user is determined, and first interaction information is generated according to the root node of the medical interaction process tree; wherein the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree topology structure; According to the historical dialogue of the target user, the user portrait information, and the next hop node of the root node in the medical interaction process tree, a first jump node is obtained; wherein the historical dialogue includes the first interaction information and the first response information of the first interaction information; The user portrait information, the historical dialogue and the node information of the first jump node are input into a medical interaction model to obtain second interaction information for continuing interaction with the target user; Wherein, the second interaction information includes: an open question, or a structured question generated based on the node information of the first jump node.

[0006] According to the medical interaction method provided by the present application, after the step of obtaining second interaction information for continuing interaction with the target user, the method further comprises: In the case where the second interaction information includes the open question, the historical dialogue is updated according to the second interaction information and the second response information of the second interaction information; According to the updated historical dialogue, the user portrait information, in the medical interaction process tree, a next jump node of the root node is matched to obtain a second jump node; The user portrait information, the updated historical dialogue and node information of the second jump node are input into a medical interaction model to obtain third interaction information for continuing interaction with the target user; The third interaction information includes an open question or a structured question generated based on node information of the second jump node.

[0007] According to the medical interaction method provided by the application, after the step of obtaining the second interaction information for continuing interaction with the target user, the method further comprises: In the case where the second interaction information includes a structured question generated by the first jump node, the historical dialogue is updated according to the second interaction information and second response information of the second interaction information; According to the updated historical dialogue, the user portrait information, in the medical interaction process tree, a next jump node of the root node is matched to obtain a second jump node; The user portrait information, the updated historical dialogue and node information of the second jump node are input into a medical interaction model to obtain third interaction information for continuing interaction with the target user; The third interaction information includes an open question or a structured question generated based on node information of the second jump node.

[0008] According to the medical interaction method provided by the application, in the medical interaction process tree, matching a next jump node of the root node comprises: The first response information is subjected to deep semantic analysis to obtain semantic elements of the first response information; The semantic elements are fused with the user portrait information and node information of the root node to obtain a current dialogue state vector; In the medical interaction process tree, all legal downstream nodes of the root node are obtained; According to the semantic similarity and clinical relevance score between the current dialogue state vector and the semantic label of each legal downstream node, a comprehensive score of the current dialogue state vector and each legal downstream node is determined; The legal downstream node with the highest comprehensive score is selected as the next jump node of the root node to obtain a first jump node.

[0009] According to the medical interaction method provided by the application, before the step of inputting the user portrait information, the historical dialogue and the node information of the first jump node into the medical interaction model to obtain second interaction information for continuing interaction with the target user, the method further comprises the following steps: obtaining a plurality of training samples carrying reward value labels by taking an input sample of a user and interaction information corresponding to the input sample as a training sample, wherein the input sample comprises a historical dialogue sample, a user portrait information sample and a jump node sample; for any training sample, inputting the training sample into a preset model to output a reward value corresponding to the interaction information sample in the training sample; calculating a loss value based on the reward value of the training sample and the reward value label, and optimizing the preset model according to the loss value until a first preset training condition is met, stopping training and obtaining a trained reward model; The reward model is used to output the reward value of the interaction information according to the input historical dialogue, user portrait information, jump node and interaction information.

[0010] According to the medical interaction method provided by the application, after the step of obtaining the trained reward model, the method further comprises the following steps: adopting a proximal policy optimization algorithm to perform policy optimization training on the medical interaction model, wherein the optimization training uses the trained reward model as a feedback signal and takes maximizing the reward value output by the reward model as an optimization objective; continuously updating network parameters of the medical interaction model until the medical interaction model meets a second preset training condition, stopping training and obtaining the trained medical interaction model; The trained medical interaction model generates an open question or a structured question autonomously according to input user portrait information, historical dialogue and node information of a jump node under the constraint of the medical interaction flow tree.

[0011] According to the medical interaction method provided by the application, the user portrait information comprises at least one of the following: basic record information, trend evolution information and diagnosis reasoning information; The basic record information comprises at least one of the following: user identification information, historical medical record, medication record, allergy record and inspection result; The trend evolution information comprises at least one of the following: time sequence change data of physiological indicators, change record of symptom performance and tracking data of treatment effect; The diagnostic reasoning information includes at least one of the following: a risk assessment result generated based on the basic record information and the trend evolution information, abnormal prompt information, and a correlation analysis result.

[0012] The application further provides a medical interaction device, comprising the following modules: A determination module is configured to determine a medical interaction process tree corresponding to a target user according to user portrait information of the target user, and generate first interaction information according to a root node of the medical interaction process tree; wherein the medical interaction process tree comprises a plurality of nodes, and the nodes are connected by directed edges to form a tree topology structure. A matching module is configured to match a next hop node of the root node in the medical interaction process tree according to a historical dialogue of the target user, the user portrait information, to obtain a first hop node; wherein the historical dialogue comprises the first interaction information and first response information of the first interaction information. An interaction module is configured to input the user portrait information, the historical dialogue and node information of the first hop node to a medical interaction model to obtain second interaction information for continuing interaction with the target user. The second interaction information comprises an open question or a structured question generated based on the node information of the first hop node.

[0013] The application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the medical interaction method according to any one of the above when executing the computer program.

[0014] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the medical interaction method according to any one of the above.

[0015] The application further provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the medical interaction method according to any one of the above.

[0016] The medical interaction method provided by the application guarantees the subject direction of the medical interaction process and the integrity of information collection by taking the pre-set medical interaction process tree as a constraint, and overcomes the defects that pure open dialogue is easy to deviate from the theme and miss key items. Meanwhile, instead of simply and directly executing the process tree, intelligent decision is made through the medical interaction model, real-time analysis is made on the historical dialogue and portrait information of the user, and the next jump node of the process tree is combined to dynamically judge and generate a structured question based on the jump node, a targeted open question is generated for supplementary inquiry, the adaptive decision mechanism not only maintains the systematic advantage of the process tree, but also has flexibility in handling unexpected situations. Through continuous optimization of reinforcement learning, the model learns to make optimal decisions in different scenarios, avoids the mechanical rigidity of the traditional process tree system, prevents the disordered divergence of the open system, and truly realizes the intelligentization and individualization of medical dialogue, and significantly improves the quality of digital medical services and user experience. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 is a flowchart of the medical interaction method provided by the application; Figure 2 is a flowchart of the medical interaction method provided by the application; Figure 3 is a structural schematic diagram of the medical interaction device provided by the application; Figure 4 is a structural schematic diagram of the electronic device provided by the application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme in the application will be described clearly and completely in combination with the drawings in the application. Obviously, the described embodiments are some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0020] Figure 1 is a flowchart of the medical interaction method provided by the application, as shown in Figure 1 The method comprises the following steps: In step 110, a medical interaction process tree corresponding to the target user is determined according to the user portrait information of the target user, and first interaction information is generated according to a root node of the medical interaction process tree; wherein the medical interaction process tree comprises a plurality of nodes, and the nodes are connected by directed edges to form a tree topology structure. In the present application, the user portrait information refers to a set of information capable of describing and characterizing the characteristics of the target user.

[0021] In the medical scenario, the user portrait information can include but is not limited to basic record information, trend evolution information and diagnosis reasoning information. These information can be user data obtained through a compliant channel after authorization by the user.

[0022] In the present application, the medical interaction process tree is a pre-constructed knowledge structure comprising a plurality of nodes, and the nodes are connected by directed edges to form a tree topology structure.

[0023] Among them, the nodes in the medical interaction process tree represent a decision point or an information collection point in the medical interaction process, and each node can be associated with one or more medical question templates.

[0024] The directed edge represents the jump relationship between the nodes, and defines the possible path of the dialogue process. The tree topology structure ensures that different paths can be reached from the root node to different leaf nodes, covering various medical scenarios.

[0025] The determination of the medical interaction process tree can be based on various strategies. For example, a chronic disease management process tree can be selected according to the user's disease type; or a health consultation process tree can be selected according to the purpose of the visit. For example, when the user portrait information shows that the user complains of headache, the system can determine to use a headache symptom assessment process tree.

[0026] In the present application, after determining the process tree, the first interaction information is generated according to the root node of the medical interaction process tree.

[0027] The interaction information in the present application can include the reply content for the user, and can also include the questions actively proposed. The interaction information can be an interaction information for the patient after visiting the hospital in the medical scenario, or an interaction information for the patient to understand the patient information in advance before visiting the hospital.

[0028] The root node is the starting node of the process tree, and usually corresponds to the initial information collection task. The generation process of the first interaction information includes: extracting a question template from the root node, personalizing the template according to the user portrait information, and generating natural and friendly interaction text.

[0029] For example, the question template of the root node is "How long have you had your [symptoms]?", combined with the symptom information of headache in the user portrait, the first interactive information "How long have you had your headache?" is generated. Step 120, according to the history of the target user's dialogue, the user portrait information, in the medical interaction flow tree, the next jump node of the root node is matched to obtain the first jump node; wherein the history of the dialogue includes the first interactive information and the first response information of the first interactive information; In this application, the history of the dialogue records the interaction content between the target user and the system, and in this step, it at least includes the first interactive information generated in the previous step and the first response information of the user to the information. For example, the first interactive information is "What can I do for you?", and the first response information of the user is "I have been coughing recently, and I also have a runny nose."

[0030] More specifically, the next jump node is matched according to the user's preliminary response to find the most suitable next diagnosis link in the flow tree.

[0031] Specifically, the user portrait information can be combined to determine which of the multiple downstream nodes of the root node is most relevant to the current situation.

[0032] For example, the downstream nodes of the root node can include "respiratory system symptoms", "digestive system symptoms", "nervous system symptoms", etc., when the user's response is "I have been coughing recently, and I also have a runny nose." Through the matching algorithm, it can be determined that "respiratory system symptoms" is the most relevant next jump node, so it is determined as the first jump node. This matching can be realized in various ways based on keyword matching, vector similarity calculation, etc.

[0033] Step 130, inputting the user portrait information, the history of the dialogue and the node information of the first jump node into the medical interaction model to obtain the second interactive information for the target user to continue the interaction; Wherein, the second interactive information includes: an open question, or a structured question generated based on the node information of the first jump node.

[0034] In this application, the medical interaction model can be a trained artificial intelligence model that can understand multi-dimensional information input and make intelligent decisions.

[0035] The medical interaction model can be implemented based on a deep neural network architecture, such as a fine-tuned version of a pre-trained language model such as Transformer, BERT, etc.

[0036] The input of the medical interaction model includes three parts: user portrait information provides user background; historical dialogue provides context information; and node information of the first jump node provides the current process position and available question templates.

[0037] In the present application, the medical interaction model has a dual-path output mechanism, which can determine whether to generate an open question or a structured question in the current state, i.e., the second interaction information output by the medical interaction model has two possibilities: One is an open question. When the medical interaction model determines that the user's response contains information that the current node of the process tree does not anticipate but may be important, or that the existing information is not sufficient to directly enter the next structured link, it will generate an open question to ask further.

[0038] For example, an open question is generated when the model determines that more unstructured information needs to be obtained. Open questions allow users to express freely without restrictions on the answer format. For example: Can you describe the nature of the headache in detail? Is it a dull pain, a sharp pain, or other sensations? The other is a structured question generated based on the node information of the first jump node. When the medical interaction model determines that the current user's response information is within the expected range and is suitable for continuing along the process tree, it will use the node information of the first jump node to generate a specific, fixed-format question.

[0039] For example, the first jump node is "cough nature inquiry", and its question template is "Please tell me if your cough is dry or with sputum?". The model will generate a structured question: "Please tell me if your cough is dry or with sputum?".

[0040] In the present application, by using a medical interaction process tree as a constraint, the main direction of the medical interaction process and the completeness of information collection are guaranteed, overcoming the defects of pure open dialogue that easily deviates from the topic and misses key items. At the same time, instead of simply executing the process tree directly, the medical interaction model makes intelligent decisions, analyzes the user's historical dialogue and portrait information in real time, and dynamically determines to generate structured questions based on the jump node, generating targeted open questions for supplementary inquiry. This adaptive decision-making mechanism not only maintains the systematic advantage of the process tree, but also has flexibility in handling unexpected situations. Through continuous optimization of reinforcement learning, the model learns to make optimal decisions in different scenarios, avoiding the mechanical rigidity of traditional process tree systems and preventing the disordered divergence of open systems, truly realizing the intelligentization and personalization of medical dialogue, and significantly improving the quality of digital medical services and user experience.

[0041] Optionally, after the step of obtaining the second interaction information for the target user to continue the interaction, the method further comprises: If the second interaction information includes the open-ended question, the historical dialogue is updated based on the second interaction information and the second response information of the second interaction information; Based on the updated historical dialogue and user profile information, the next jump node of the root node is re-matched in the medical interaction process tree to obtain the second jump node; The user profile information, the updated historical dialogue, and the node information of the second jump node are input into the medical interaction model to obtain third interaction information for continuing to interact with the target user; The third interactive information includes: open-ended questions, or structured questions generated based on the node information of the second jump node.

[0042] In this application, firstly, the historical dialogue is updated based on the second interaction information and the second response information of the second interaction information.

[0043] The second response is the user's reply to open-ended questions, which typically contains rich unstructured information. More specifically, updating historical dialogues is not a simple patchwork, but requires maintaining the coherence and contextual integrity of the conversation.

[0044] The update process specifically involves adding the second interaction information to the dialogue history; adding the second response information to the dialogue history; updating the timestamp and round information of the dialogue; and extracting and marking key medical information.

[0045] For example, if the open-ended question is: "Can you describe the nature of your headache in detail?", and the user answers: "It's a throbbing pain, like a heartbeat, mainly in the right temple," the system will add this question-and-answer pair to the history of conversations and identify key information such as throbbing and right temple.

[0046] In this application, after a round of supplementary open-ended questions, the process did not simply return to where it had ended, but instead used the newly acquired information to conduct a comprehensive reassessment of the entire direction of the dialogue.

[0047] The supplementary information gathered from open-ended questions can be crucial, enough to alter the main direction of the entire consultation. For example, if a user initially complains of a cough, the system might initially match the "respiratory system" branch. However, if follow-up questions reveal a history of heart disease and chest tightness, the system, using the updated dialogue history, might find that the "cardiovascular system" branch has a higher priority when rematching from the root node, thus obtaining a completely new second-stepping node. If the supplementary information merely refines the original complaint, the rematching result may still be the original stepping node, ensuring the robustness of the method.

[0048] The user profile information, the updated historical dialogue, and the node information of the second jump node are input into the medical interaction model to obtain the third interaction information for continuing the interaction with the target user.

[0049] At this point, the input to the medical interaction model includes richer historical information, including initial question-and-answer data and supplementary information for open-ended questions. The model will reassess the current state and decide whether to generate open-ended or structured questions.

[0050] Therefore, in this application, the third interactive information also includes two possibilities: an open-ended question or a structured question generated based on the node information of the second jump node.

[0051] For example, after identifying a migraine assessment pathway, structured questions might be generated such as: "What of the following aura symptoms do you experience before a migraine attack? 1. Visual abnormalities 2. Sensory abnormalities 3. Language difficulties 4. No aura" In this application, a deviation-supplementation-redirection dialogue loop is constructed by introducing a mechanism for re-matching from the root node after open-ended follow-up questions. While demonstrating high flexibility, it possesses strong self-correction and path replanning capabilities, avoiding continuous free-flowing dialogue routes and promptly pulling the dialogue back to the constraints of the medical workflow tree, proceeding along the main line most consistent with clinical logic. This maintains dialogue coherence while improving the accuracy and completeness of medical information collection.

[0052] Optionally, after obtaining the second interaction information for continued interaction with the target user, the method further includes: If the second interaction information includes a structured question generated by the first jump node, the historical dialogue is updated according to the second interaction information and the second response information of the second interaction information; Based on the updated historical dialogue and user profile information, the next jump node of the first jump node is matched in the medical interaction process tree to obtain the third jump node; The user profile information, the updated historical dialogues, and the node information of the third jump node are input into the medical interaction model to obtain the fourth interaction information for continuing to interact with the target user. The fourth interactive information includes: open-ended questions, or structured questions generated based on the node information of the third jump node.

[0053] In this application, when the second interaction information includes a structured question generated by the first redirect node, the historical dialogue is updated based on the second interaction information and the second response information of the second interaction information. For example, if the structured question is "Is your cough dry or with phlegm?", and the user's second response information is "with phlegm", then the historical dialogue is updated.

[0054] Then, based on the updated historical dialogue and user profile information, the next jump node of the first jump node is matched in the medical interaction process tree to obtain the third jump node.

[0055] In this application, the starting point for matching is the first jump node, not the root node. This follows a depth-first traversal logic within a structured process. That is, when the dialogue proceeds smoothly according to the process tree nodes, the system will continue exploring downstream from the current node to delve deeper into the problem.

[0056] For example, if the current node, i.e. the first jump node, is "cough nature inquiry", and the user answers "has phlegm", the system will match downstream nodes of the "cough nature inquiry" node (such as "phlegm color inquiry", "phlegm amount inquiry", etc.). Based on the user's answer "has phlegm", the system will match "phlegm color inquiry" as the third jump node.

[0057] Then, the user profile information, the updated historical dialogue, and the node information of the third jump node are input into the medical interaction model to obtain the fourth interaction information for continuing to interact with the target user.

[0058] In this embodiment, the medical interaction model, based on the latest context and the third jump node, makes another decision to generate an open-ended question or a structured question based on the third jump node, thus obtaining the fourth interaction information.

[0059] The fourth interactive information may also be an open-ended question or a structured question generated based on the node information of the third jump node.

[0060] In this application, by steadily advancing along the structural direction of the medical interaction process tree, medical information is collected layer by layer in depth, while maintaining the ability to flexibly supplement through open-ended questions when necessary, thus ensuring the systematicness and adaptability of the medical dialogue.

[0061] Optionally, in the medical interaction flow tree, matching the next jump node of the root node includes: Perform deep semantic parsing on the first response information to obtain the semantic elements of the first response information; The semantic elements are fused with the user profile information and the node information of the root node to obtain the current dialogue state vector. In the medical interaction process tree, obtain all legal downstream nodes of the root node; Based on the semantic similarity and clinical relevance scores between the current dialogue state vector and the semantic tags of each legitimate downstream node, the comprehensive score between the current dialogue state vector and each legitimate downstream node is determined. The legitimate downstream node with the highest overall score is selected as the next hop node of the root node, thus obtaining the first hop node.

[0062] In this application, firstly, deep semantic parsing is performed on the first response information to obtain the semantic elements of the first response information.

[0063] Deep semantic parsing is a natural language understanding technology that goes beyond keyword matching. It can include named entity recognition, used to identify medical terms in responses, such as symptoms, body parts, and times. Intent recognition is used to understand the user's core needs; relation extraction is used to analyze the relationships between semantic elements, such as the location-symptom relationship between "chest" and "pain". Through parsing, unstructured user statements are transformed into a structured set of semantic elements.

[0064] Next, the semantic elements are fused with the user profile information and the node information of the root node to obtain the current dialogue state vector.

[0065] In this application, information fusion is a process of integrating multi-source information. Various fusion strategies can be employed: simple splicing, weighted fusion, attention-based fusion, or gating-based fusion.

[0066] In this application, the current dialogue state vector is a high-dimensional vector representation that encodes the complete state information of the current dialogue. The dimension of this vector can be fixed (e.g., 512-dimensional, 768-dimensional) or dynamic.

[0067] Different dimensions in a vector may correspond to different types of information: some dimensions encode symptom information, some encode time information, some encode user characteristics, etc.

[0068] Then, in the medical interaction process tree, obtain all the legitimate downstream nodes of the root node.

[0069] In this application, a legitimate downstream node refers to a child node that is directly reachable from the current node via a directed edge. The acquisition process requires traversing the process tree data structure, which can be implemented using an adjacency list, adjacency matrix, or graph database. Each node may have multiple downstream nodes, representing different branch paths.

[0070] For example, legitimate downstream nodes of the "Initial Symptom Assessment" node may include: respiratory symptoms, digestive symptoms, neurological symptoms, circulatory symptoms, etc.

[0071] More specifically, the comprehensive score of the current dialogue state vector and each legitimate downstream node is determined based on the semantic similarity and clinical relevance scores between the semantic labels of the current dialogue state vector and each legitimate downstream node.

[0072] In this application, semantic similarity can be calculated using a variety of methods: cosine similarity, Euclidean distance, dot product similarity, or neural network-based similarity measures.

[0073] Each downstream node has a preset semantic tag vector, which is used to calculate the similarity with the current dialogue state vector.

[0074] Clinical relevance scoring is a reward mechanism based on medical knowledge. This can be achieved through expert rule systems, knowledge graph reasoning, or pre-trained medical language models. Clinical relevance considers factors such as medical associations between symptoms, typical disease presentations, and diagnostic priority.

[0075] In this application, the overall score can be calculated using weighted summation, product fusion, or a more complex fusion function.

[0076] In this application, the legitimate downstream node with the highest comprehensive score is selected as the next jump node of the root node, thus obtaining the first jump node.

[0077] The selection process can be deterministic, such as directly selecting the highest score; or probabilistic, sampling based on the score distribution.

[0078] In this embodiment, the node matching mechanism enables intelligent navigation in the process tree, selecting the most suitable dialogue path based on the user's actual response and individual characteristics, thereby achieving personalized and precise medical interaction.

[0079] Optionally, before the step of inputting the user profile information, the historical dialogue, and the node information of the first jump node into the medical interaction model to obtain the second interaction information for continuing interaction with the target user, the method further includes: The user's input sample and the corresponding interaction information sample are used as a training sample to obtain multiple training samples carrying reward score labels; wherein, the input sample includes: historical dialogue samples, user profile information samples and jump node samples; For any of the training samples, the training sample is input into a preset model, and the reward score corresponding to the interaction information sample in the training sample is output. The loss value is calculated based on the reward score and reward score label of the training sample, and the preset model is optimized according to the loss value until the first preset training condition is met, then training is stopped, and a trained reward model is obtained. The reward model is used to output a reward score for the interaction information based on the input historical dialogue, user profile information, jump nodes and interaction information.

[0080] In this application, before the step of inputting the user profile information, the historical dialogue and the node information of the first jump node into the medical interaction model to obtain the second interaction information for continuing to interact with the target user, the method also includes a process of training the medical interaction model, which first requires training a reward model.

[0081] First, the user's input sample and the corresponding interaction information sample are used as a training sample to obtain multiple training samples carrying reward score labels.

[0082] The input samples consist of three components: historical dialogue samples, user profile information samples, and redirection node samples. These samples can come from real medical dialogue records, expert-annotated cases, or synthetic data generated through data augmentation techniques.

[0083] Interactive information samples are the system outputs corresponding to the input samples, and may be open-ended or structured questions. Each interactive information sample requires a corresponding quality assessment.

[0084] In this application, the reward score label is a quantitative assessment of the quality of interactive information, which can be a continuous value (such as a score of 0-1) or a discrete value (such as a rating of 1-5).

[0085] Reward score tags can be obtained in several ways: manual reward scores from medical experts, rule-based automatic reward scores, user feedback reward scores, or a combination of multiple reward score methods. Reward score dimensions may include: medical accuracy, information collection efficiency, user experience, and security.

[0086] Next, for any of the training samples, the training sample is input into a preset model, and the reward score corresponding to the interaction information sample in the training sample is output.

[0087] In this application, the pre-defined model can employ various neural network architectures: multilayer perceptron, convolutional neural network, recurrent neural network, Transformer, etc. The model's input layer needs to be able to process heterogeneous input data, and the output layer generates reward score predictions.

[0088] The forward propagation process of the model includes: input encoding, feature extraction, and reward score prediction.

[0089] Then, a loss value is calculated based on the reward score and reward score label of the training samples, and the preset model is optimized according to the loss value.

[0090] Different loss functions can be used to calculate the loss value. Mean squared error loss is used for continuous reward scores, cross-entropy loss for discrete reward scores, and ranking loss for relative reward scores, etc. The choice of loss function depends on the type of reward score label and the training objective.

[0091] The optimization process may include techniques such as learning rate scheduling, gradient pruning, and regularization to improve training stability and model generalization ability.

[0092] Finally, continue training until the first preset training condition is met, then stop training to obtain the trained reward model.

[0093] More specifically, the first preset training condition may include multiple termination criteria: reaching the maximum number of training rounds, the validation set loss no longer decreasing, etc. Reward score prediction accuracy reaches a threshold, early stop mechanism is triggered, etc. These conditions can be used individually or in combination.

[0094] The trained reward model can output the reward score of the interaction information based on the input historical dialogue, user profile information, jump nodes and interaction information.

[0095] The reward score reflects the quality and appropriateness of the generated interactive information within a specific context. The reward model will serve as a key component in subsequent reinforcement learning training, providing feedback signals for the decision-making of the medical interaction model.

[0096] Through the reward model training in this embodiment, the system acquires the ability to evaluate the quality of interactions, providing a reliable evaluation standard for subsequent strategy optimization and ensuring that the medical interaction model can generate high-quality interaction information.

[0097] Optionally, after the step of obtaining the trained reward model, the method further includes: The medical interaction model is trained using a proximal policy optimization algorithm; wherein, the optimization training uses the trained reward model as a feedback signal, with the optimization objective being to maximize the reward score output by the reward model. The network parameters of the medical interaction model are continuously updated until the medical interaction model meets the second preset training conditions. Training is then stopped, and the medical interaction model with good training assumptions is obtained. The trained medical interaction model, under the constraints of the medical interaction process tree, autonomously determines and generates open-ended or structured questions based on the input user profile information, historical dialogues, and node information of jump nodes.

[0098] In this application, the Proximal Policy Optimization (PPO) algorithm is an advanced reinforcement learning algorithm. Its core advantage is that while optimizing the policy, it ensures that the new policy does not differ too much from the old policy, thus avoiding performance crashes during the training process.

[0099] In this application, during the optimization training process, the medical interaction model plays the role of a strategy in reinforcement learning. Its task is to output an action based on the input state, namely user profile information, historical dialogue, and jump node information, namely, to decide whether to generate an open-ended question or a structured question, and to generate specific content.

[0100] The training process uses the reward model trained in the previous step as a feedback signal. Specifically, the medical interaction model interacts with a simulated user environment to generate dialogue.

[0101] For each interaction generated by the medical interaction model, it is input into the reward model along with the current context, and the reward model outputs a reward score. This reward score serves as the immediate reward signal for reinforcement learning.

[0102] The optimization objective of training is to maximize the cumulative expected value of the reward score output by the reward model. Through the PPO algorithm, the network parameters of the medical interaction model are continuously adjusted to optimize its strategy towards achieving higher cumulative rewards.

[0103] The network parameters of the medical interaction model are continuously updated until the medical interaction model meets the second preset training conditions, such as when the reward score reaches a stable plateau, at which point training is stopped, and the trained medical interaction model is obtained.

[0104] The parameter update process follows the PPO algorithm's update rules: calculating the advantage function estimate, calculating the policy ratio, applying a pruning mechanism to limit the update magnitude, and performing mini-batch updates over multiple epochs. A balance between exploration and exploitation needs to be struck during the update process; entropy regularization can be used to encourage exploration.

[0105] More specifically, the second preset training conditions may include: reaching the maximum number of training epochs, average reward convergence, policy performance metrics meeting targets, and validation set performance stability. Multiple metrics need to be monitored during training to ensure model quality.

[0106] In this application, the trained medical interaction model, under the constraints of the medical interaction process tree, can autonomously determine and generate open-ended or structured questions based on the input user profile information, historical dialogues, and node information of jump nodes.

[0107] The judgment ability of the medical interaction model is not based on hard-coded rules, but on an optimal strategy learned in the process of maximizing rewards.

[0108] For example, it will learn that generating open-ended questions will yield high rewards when there is insufficient information or high-risk signals, while generating structured questions to advance the process will yield high rewards when there is sufficient information.

[0109] In this embodiment, by employing the PPO reinforcement learning algorithm based on a reward model, the medical interaction model can evolve itself through massive trial and error in a simulated environment. Ultimately, it learns a set of intelligent decision-making strategies that dynamically balance structured consultation and flexible inquiry in complex and ever-changing medical dialogue scenarios, thereby enabling the generated dialogue to be both medically professional and humane.

[0110] Optionally, the user profile information includes at least one of the following: basic record information, trend evolution information, and diagnostic reasoning information; The basic record information includes at least one of the following: user identification information, historical medical records, medication records, allergy records, and examination and test results; The trend evolution information includes at least one of the following: time-series change data of physiological indicators, records of changes in symptom manifestations, and tracking data of treatment effects; The diagnostic reasoning information includes at least one of the following: risk assessment results, anomaly alerts, and correlation analysis results generated based on the aforementioned basic record information and trend evolution information.

[0111] In this application, all specific data information in the user profile was obtained through compliant channels with the user's authorization. User profile information includes basic record information, which is a fundamental component of a user's medical record. User identification information includes, but is not limited to: a unique identifier, basic demographic information, contact information, and emergency contact information. This information is used for user identification and basic characteristic description.

[0112] Historical medical records cover a user's past medical history: previous disease diagnoses, hospitalization records, surgical records, chronic disease management records, etc. These records can be obtained from hospital information systems, electronic medical record systems, or personal health records.

[0113] Medication records include: current medication list, historical medication records, medication dosage and frequency, medication efficacy assessment, and adverse drug reaction records. This information is crucial for evaluating treatment regimens and drug interactions.

[0114] Detailed allergy records are maintained, including: drug allergy history, food allergy history, environmental allergens, type and severity of allergic reactions, etc. Allergy information is an important safeguard for medical safety.

[0115] The examination and testing results include: laboratory test results, imaging test results, special test results, and time series of test results.

[0116] User profile information also includes trend evolution information, reflecting the dynamic changes in the user's health status.

[0117] Time-series data recording of physiological indicators document changes in key physiological parameters over time: trends in blood pressure, blood glucose fluctuations, weight changes, heart rate variability, etc. This data can be obtained through wearable devices, home medical devices, or regular checkups.

[0118] The record tracks changes in symptom presentation, including the time of first symptom onset, changes in symptom intensity, changes in symptom frequency, triggering and relieving factors, and the appearance and disappearance of accompanying symptoms.

[0119] The tracking data of treatment effects are used to evaluate the effectiveness of the treatment intervention: symptom improvement after medication, records of treatment plan adjustments, assessment of rehabilitation progress, changes in quality of life scores, etc.

[0120] User profile information also includes diagnostic reasoning information, which is the result of intelligent analysis based on basic data.

[0121] Risk assessment results are generated through risk prediction models: disease risk scores, such as cardiovascular risk scores, complication risk warnings, prognostic assessment results, and health risk stratification. These assessments can employ clinical scoring systems or machine learning models.

[0122] Anomaly alerts identify noteworthy anomalies, such as abnormal test results, warnings about abnormal symptom combinations, medication safety tips, and follow-up reminders. Anomaly detection can be based on rule engines or anomaly detection algorithms.

[0123] Association analysis revealed correlations between different medical information, such as the probability of association between symptoms and diseases, the correlation between medication and test indicators, the relationship between lifestyle and health indicators, and the association between family history and disease risk.

[0124] The user profile information system in this embodiment can provide a comprehensive understanding of a user's health status, disease history, and development trend, providing rich background information for personalized medical interactions, ensuring the relevance of interactive content and the accuracy of medical advice, and achieving truly precise medical interaction.

[0125] In one alternative embodiment, Figure 2 This is a schematic diagram of the interactive process adopted by the present invention, such as... Figure 2 As shown, the first part is the dialogue between the doctor and the patient.

[0126] In doctor-patient dialogues, doctors ask questions such as, "Have you been feeling unwell recently?"; patients then respond with information such as, "I've been feeling a bit dizzy lately."

[0127] The dialogue content is used as input and semantically encoded using natural language processing technology. This transforms the dialogue into a corresponding semantic vector representation, capturing the content and contextual information expressed in the dialogue in a quantified and structured form.

[0128] Simultaneously, the system captures patient profile information, including but not limited to: gender (male), age (24 years old), weight (85 kg), chief complaint (e.g., hypertension), blood pressure data (e.g., 175 / 110 mmHg), and other key health indicators and basic information. This information is collected through electronic medical record systems, physical monitoring devices, or patient self-reports, and is transformed into structured feature vectors after standardization.

[0129] Next, the process of determining the next jump node begins. The system fuses the aforementioned semantic encoding vector with the patient profile feature vector to form a comprehensive current dialogue state vector.

[0130] Then, the system obtains all legitimate downstream nodes of the "Initial Consultation" node, which may include, for example, "Nervous System Symptom Node", "Cardiovascular System Symptom Node" and "General Discomfort Node".

[0131] The system will calculate the comprehensive score between the current dialogue state vector and the preset semantic labels of each downstream node (this score can be combined with the results of correlation calculation and feature hierarchical processing).

[0132] In this example, because the state vector contains semantic information about "dizziness" and profile information about "hypertension," it has the highest clinical relevance and semantic similarity score with the "cardiovascular system symptom node." Therefore, the system selects the "cardiovascular system symptom node" as the next jump node.

[0133] The system inputs historical dialogues, user profile information, and identified "cardiovascular system symptom nodes" into a pre-trained medical interaction model. The model's core task is to generate the next interaction information based on this rich input.

[0134] This decision includes two possibilities: Generating structured questions: If the model determines that the current information is sufficient to proceed along the process tree, it will generate a structured question based on the selected "cardiovascular symptom node". For example, the model might generate: "Understood. Based on your blood pressure, do you experience chest tightness or palpitations when you feel dizzy?" Generate open-ended questions: If the model determines that the patient's answer "feeling a little dizzy" is too vague or may be hiding other key information, requiring further inquiry, an open-ended question will be generated. For example: "Could you describe the dizziness in detail? Is it like the world spinning or feeling lightheaded?" In this scenario, since "dizziness" is strongly correlated with "hypertension" in the profile, the model is very likely to choose the first path, generating structured questions related to the cardiovascular system in order to efficiently advance the diagnosis and treatment process.

[0135] In this embodiment, the system can not only intelligently select the most suitable path from the preset medical process tree based on the patient's real-time feedback and individual health status, ensuring the systematic nature and clinical logic of the consultation, but also flexibly handle ambiguous or unexpected information by generating open-ended questions when necessary. This achieves an organic combination of structured processes and personalized inquiry, thereby significantly improving the efficiency of doctor-patient communication and the accuracy of diagnosis and treatment.

[0136] The medical interaction device provided by the present invention is described below. The medical interaction device described below and the medical interaction method described above can be referred to in correspondence.

[0137] Figure 3 This is a schematic diagram of the medical interactive device structure provided by the present invention, such as... Figure 3 As shown, it includes: The determining module 310 is used to determine the medical interaction process tree corresponding to the target user based on the user profile information of the target user, and generate first interaction information based on the root node of the medical interaction process tree; wherein, the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree topology. The matching module 320 is used to match the next jump node of the root node in the medical interaction process tree based on the target user's historical dialogue and the user profile information to obtain the first jump node; wherein, the historical dialogue includes the first interaction information and the first response information of the first interaction information; The interaction module 330 is used to input the user profile information, the historical dialogue and the node information of the first jump node into the medical interaction model to obtain the second interaction information for continuing to interact with the target user; The second interactive information includes: open-ended questions, or structured questions generated based on the node information of the first jump node.

[0138] In this application, a pre-set medical interaction flowchart is used as a constraint to ensure the main direction of the medical interaction process and the integrity of information collection, overcoming the shortcomings of purely open-ended dialogues that are prone to deviating from the topic and omitting key items. Furthermore, instead of simply executing the flowchart directly, the medical interaction model makes intelligent decisions, analyzing the user's historical dialogues and profile information in real time, and dynamically generating structured questions based on the next jump node of the flowchart, as well as generating targeted open-ended questions for supplementary inquiry. This adaptive decision-making mechanism maintains the systematic advantages of the flowchart while possessing the flexibility to handle unexpected situations. Through continuous optimization via reinforcement learning, the model learns to make optimal decisions in different scenarios, avoiding the mechanical rigidity of traditional flowchart systems and preventing the disordered divergence of open-ended systems. This truly realizes the intelligence and personalization of medical dialogue, significantly improving the quality of digital healthcare services and user experience.

[0139] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a medical interaction method. This method includes: determining a medical interaction process tree corresponding to the target user based on the target user's user profile information, and generating first interaction information based on the root node of the medical interaction process tree; wherein the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree-like topology. Based on the target user's historical dialogues and user profile information, the next jump node of the root node is matched in the medical interaction process tree to obtain the first jump node; wherein, the historical dialogues include the first interaction information and the first response information of the first interaction information; The user profile information, the historical dialogues, and the node information of the first jump node are input into the medical interaction model to obtain the second interaction information for continuing to interact with the target user. The second interactive information includes: open-ended questions, or structured questions generated based on the node information of the first jump node.

[0140] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0141] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute the medical interaction method provided by the above methods, the method including: determining the medical interaction process tree corresponding to the target user based on the user profile information of the target user, and generating first interaction information based on the root node of the medical interaction process tree; wherein, the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree topology structure; Based on the target user's historical dialogues and user profile information, the next jump node of the root node is matched in the medical interaction process tree to obtain the first jump node; wherein, the historical dialogues include the first interaction information and the first response information of the first interaction information; The user profile information, the historical dialogues, and the node information of the first jump node are input into the medical interaction model to obtain the second interaction information for continuing to interact with the target user. The second interactive information includes: open-ended questions, or structured questions generated based on the node information of the first jump node.

[0142] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the medical interaction methods provided by the above methods. The method includes: determining a medical interaction process tree corresponding to the target user based on the user profile information of the target user, and generating first interaction information based on the root node of the medical interaction process tree; wherein the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree-like topology. Based on the target user's historical dialogues and user profile information, the next jump node of the root node is matched in the medical interaction process tree to obtain the first jump node; wherein, the historical dialogues include the first interaction information and the first response information of the first interaction information; The user profile information, the historical dialogues, and the node information of the first jump node are input into the medical interaction model to obtain the second interaction information for continuing to interact with the target user. The second interactive information includes: open-ended questions, or structured questions generated based on the node information of the first jump node.

[0143] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The 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 the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical interaction method, characterized in that, include: Based on the user profile information of the target user, the medical interaction process tree corresponding to the target user is determined, and the first interaction information is generated based on the root node of the medical interaction process tree; wherein, the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree topology. Based on the target user's historical dialogues and user profile information, the next jump node of the root node is matched in the medical interaction process tree to obtain the first jump node; wherein, the historical dialogues include the first interaction information and the first response information of the first interaction information; The user profile information, the historical dialogues, and the node information of the first jump node are input into the medical interaction model to obtain the second interaction information for continuing to interact with the target user. The second interactive information includes: open-ended questions, or structured questions generated based on the node information of the first jump node.

2. The medical interaction method according to claim 1, characterized in that, After obtaining the second interaction information for continuing interaction with the target user, the method further includes: If the second interaction information includes the open-ended question, the historical dialogue is updated based on the second interaction information and the second response information of the second interaction information; Based on the updated historical dialogue and user profile information, the next jump node of the root node is re-matched in the medical interaction process tree to obtain the second jump node; The user profile information, the updated historical dialogue, and the node information of the second jump node are input into the medical interaction model to obtain third interaction information for continuing to interact with the target user; The third interactive information includes: open-ended questions, or structured questions generated based on the node information of the second jump node.

3. The medical interaction method according to claim 1, characterized in that, After obtaining the second interaction information for continuing interaction with the target user, the method further includes: If the second interaction information includes a structured question generated by the first jump node, the historical dialogue is updated according to the second interaction information and the second response information of the second interaction information; Based on the updated historical dialogue and user profile information, the next jump node of the first jump node is matched in the medical interaction process tree to obtain the third jump node; The user profile information, the updated historical dialogues, and the node information of the third jump node are input into the medical interaction model to obtain the fourth interaction information for continuing to interact with the target user. The fourth interactive information includes: open-ended questions, or structured questions generated based on the node information of the third jump node.

4. The medical interaction method according to claim 1, characterized in that, In the medical interaction flow tree, the next jump node to match the root node includes: Perform deep semantic parsing on the first response information to obtain the semantic elements of the first response information; The semantic elements are fused with the user profile information and the node information of the root node to obtain the current dialogue state vector. In the medical interaction process tree, obtain all legal downstream nodes of the root node; Based on the semantic similarity and clinical relevance scores between the current dialogue state vector and the semantic tags of each legitimate downstream node, the comprehensive score between the current dialogue state vector and each legitimate downstream node is determined. The legitimate downstream node with the highest overall score is selected as the next hop node of the root node, thus obtaining the first hop node.

5. The medical interaction method according to claim 1, characterized in that, Before the step of inputting the user profile information, the historical dialogue, and the node information of the first redirect node into the medical interaction model to obtain the second interaction information for continuing the interaction with the target user, the method further includes: The user's input sample and the corresponding interaction information sample are used as a training sample to obtain multiple training samples carrying reward score labels; wherein, the input sample includes: historical dialogue samples, user profile information samples and jump node samples; For any of the training samples, the training sample is input into a preset model, and the reward score corresponding to the interaction information sample in the training sample is output. The loss value is calculated based on the reward score and reward score label of the training sample, and the preset model is optimized according to the loss value until the first preset training condition is met, then training is stopped, and a trained reward model is obtained. The reward model is used to output a reward score for the interaction information based on the input historical dialogue, user profile information, jump nodes and interaction information.

6. The medical interaction method according to claim 5, characterized in that, After the step of obtaining the trained reward model, the method further includes: The medical interaction model is trained using a proximal policy optimization algorithm; wherein, the optimization training uses the trained reward model as a feedback signal, with the optimization objective being to maximize the reward score output by the reward model. The network parameters of the medical interaction model are continuously updated until the medical interaction model meets the second preset training conditions. Training is then stopped, and the medical interaction model with good training assumptions is obtained. The trained medical interaction model, under the constraints of the medical interaction process tree, autonomously determines and generates open-ended or structured questions based on the input user profile information, historical dialogues, and node information of jump nodes.

7. The medical interaction method according to claim 1, characterized in that, The user profile information includes at least one of the following: basic record information, trend evolution information, and diagnostic reasoning information; The basic record information includes at least one of the following: user identification information, historical medical records, medication records, allergy records, and examination and test results; The trend evolution information includes at least one of the following: time-series change data of physiological indicators, records of changes in symptom manifestations, and tracking data of treatment effects; The diagnostic reasoning information includes at least one of the following: risk assessment results, anomaly alerts, and correlation analysis results generated based on the aforementioned basic record information and trend evolution information.

8. A medical interactive device, characterized in that, include: The determination module is used to determine the medical interaction process tree corresponding to the target user based on the user profile information of the target user, and generate first interaction information based on the root node of the medical interaction process tree; wherein, the medical interaction process tree contains multiple nodes, and the nodes are connected by directed edges to form a tree topology. The matching module is used to match the next jump node of the root node in the medical interaction process tree based on the target user's historical dialogue and the user profile information to obtain the first jump node; wherein, the historical dialogue includes the first interaction information and the first response information of the first interaction information; The interaction module is used to input the user profile information, the historical dialogue and the node information of the first jump node into the medical interaction model to obtain the second interaction information for continuing to interact with the target user; The second interactive information includes: open-ended questions, or structured questions generated based on the node information of the first jump node.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the medical interaction method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the medical interaction method as described in any one of claims 1 to 7.

Images