Hospital health science popularization robot service system

Abstract

The invention belongs to the technical field of intelligent medical treatment, and particularly relates to a hospital health science popularization robot service system, which receives a plurality of service requests which are initiated by appointed calling equipment and comprise patient identifiers and physical positions, calls medical record data according to the patient identifiers, and sends the medical record data to a server; the method comprises the following steps: constructing a weighted scoring rule with medical record characteristics as a main part and waiting duration and spatial distance as auxiliary parts, planning a service priority sequence of a robot, collecting oral inquiry audio of a patient and converting the oral inquiry audio into a text after the robot arrives at a specified patient position according to the planning sequence; and combining patient medical record dimensions of intention recognition, slot filling and dynamic calling to carry out context enhancement so as to reconstruct a standard clinical question, carrying out supportive evidence chain retrieval and rationality verification on preliminary answer content by introducing a medical knowledge graph, generating final health science popularization content, and outputting the final health science popularization content. And it is ensured that the finally output health science popularization content not only is personalized, but also has medical reliability and safety.

Description

Technical Field

[0001] This invention belongs to the field of smart healthcare technology, specifically a hospital health education robot service system. Background Technology

[0002] Hospital health education is an important component of the modern healthcare service system, aiming to provide patients and their families with accurate and easily understandable health information to improve their disease awareness, treatment adherence, and self-management abilities. With the development of artificial intelligence and robotics, hospital health education robots, as an emerging automated service tool, are being used to alleviate repetitive consultation work for medical staff and improve service efficiency. These robots typically integrate functions such as voice interaction, mobile navigation, and information retrieval, falling under the category of intelligent applications within healthcare informatics.

[0003] In existing technologies, hospital health education robots typically provide services in the following way. Upon receiving a service request from a patient via a calling device, the robot system plans its movement sequence based on a preset scheduling algorithm, such as first-come, first-served or shortest path priority. After arriving at the patient's location, the robot receives the patient's verbal inquiry through a speech recognition module, retrieves matching answers from a general medical knowledge base, and finally presents the query results to the patient through speech synthesis or text display.

[0004] However, the existing technical solutions have obvious technical defects: 1. The service scheduling mechanism of the existing technical solutions is too simple and cannot distinguish the degree of medical urgency of different patients' requests, which may lead to excessively long waiting times for patients with more serious conditions or urgent consultation needs.

[0005] 2. The health education content provided by existing technical solutions is usually standardized, lacks connection with individual patient medical records, cannot provide targeted and personalized health guidance, and lacks verification of the rationality of the output content. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, embodiments of the present invention provide a hospital health science popularization robot service system, which can effectively solve the problems involved in the prior art.

[0007] The objective of this invention can be achieved through the following technical solution: a hospital health science popularization robot service system, comprising: a service request receiving module, a service priority planning module, a clinical problem reconstruction module, and a science popularization answer generation module.

[0008] The service request receiving module is connected to the service priority planning module, the service priority planning module is connected to the clinical problem reconstruction module, and the clinical problem reconstruction module is connected to the popular science answer generation module.

[0009] The service request receiving module receives multiple service requests initiated by a designated calling device, which include the patient's identifier and physical location.

[0010] The service priority planning module retrieves medical records based on patient identification and constructs a weighted scoring rule based primarily on medical record characteristics and secondarily on waiting time and spatial distance, thereby planning the robot's service priority sequence.

[0011] The clinical problem reconstruction module collects audio of the patient's verbal inquiry and converts it into text after the robot arrives at the designated patient location according to the planned sequence. It then performs intent recognition and key information slot filling on the audio text and dynamically selects patient medical history dimensions related to the intent for contextual enhancement to reconstruct a standard clinical problem.

[0012] The science popularization Q&A generation module uses the robot's built-in medical question-and-answer model to provide preliminary answers to standard clinical questions. Based on the patient's intent and the selected patient medical record dimensions, it retrieves supporting evidence chains from the medical knowledge graph, verifies the reasonableness of the preliminary answers, and generates and outputs the final health science popularization content.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention constructs a dynamic risk profile and plans service priorities accordingly, enabling the robot to intelligently identify and prioritize responding to more urgent patient requests in clinical practice, breaking through the traditional first-come-first-served or simple grading service model, and significantly improving the pertinence of service response and the efficiency of medical resource allocation.

[0014] (2) This invention enhances the context by combining intent recognition, slot filling and dynamically retrieved patient medical record dimensions. It can accurately reconstruct patients’ vague and colloquial inquiries into standard clinical questions with individualized clinical backgrounds, which profoundly solves the problem of intent recognition bias caused by lack of contextual understanding in existing technologies and greatly improves the accuracy of human-computer interaction.

[0015] (3) By introducing a medical knowledge graph, this invention conducts a supporting evidence chain retrieval and rationality verification on the preliminary answer content, establishes a rigorous knowledge verification and security filtering mechanism, and ensures that the final output health science popularization content is not only personalized, but also has medical reliability and safety, effectively avoiding misleading or inappropriate suggestions that may be generated by general models. Attached Figure Description

[0016] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0017] Figure 1This is a schematic diagram of the module connection of the present invention.

[0018] Figure 2 This is a logical diagram illustrating the supporting evidence chain retrieval method of the present invention.

[0019] Figure 3 This is a logical diagram illustrating the rationality verification of the preliminary solution content in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, the present invention provides a hospital health science popularization robot service system, including: a service request receiving module, a service priority planning module, a clinical question reconstruction module, and a science popularization answer generation module.

[0022] The service request receiving module is connected to the service priority planning module, the service priority planning module is connected to the clinical problem reconstruction module, and the clinical problem reconstruction module is connected to the popular science answer generation module.

[0023] The service request receiving module receives multiple service requests initiated by a designated calling device, which include the patient's identifier and physical location.

[0024] Given that the urgency of clinical needs in a hospital environment is directly related to patient safety, and that the core value of robot services lies in prioritizing the accessibility of medical resources for critically ill patients, the service priority planning module retrieves medical records based on patient identification and constructs a weighted scoring rule based primarily on medical record characteristics and secondarily on waiting time and spatial distance, thereby planning the robot's service priority sequence.

[0025] In a preferred embodiment of the present invention, the weighted scoring rule includes: extracting disease diagnosis information and recent vital sign examination information from medical records as medical record feature items. The disease diagnosis information includes the disease name, severity level, diagnostic department, diagnosis time, and corresponding International Classification of Diseases (ICD) code. The recent vital sign examination information includes the temperature, heart rate, blood pressure, and blood oxygen saturation values ​​from the most recent examination prior to the date the request was initiated, as well as abnormal vital sign markers. The abnormal status markers refer to warning indicators automatically generated based on the examination parameter values ​​exceeding a preset safety threshold range.

[0026] A predefined set of acute and critical illnesses is defined, which includes at least acute myocardial infarction, acute stroke, and severe trauma, along with their corresponding criticality classification standards.

[0027] Determine whether the patient's disease diagnosis information belongs to a predefined set of acute and critical illnesses, and whether there are abnormal vital signs markers in the recent physical examination information. Based on the consistency between the acute and critical illnesses and the abnormal vital signs markers, the service request is divided into four priority levels: highest, high, medium, and low. The specific division logic is as follows: highest priority: the disease diagnosis belongs to a predefined set of acute and critical illnesses and there are abnormal vital signs markers.

[0028] High priority: The disease diagnosis belongs to the predefined acute and critical illness category but there are no abnormal vital signs markers.

[0029] Medium priority: The disease diagnosis does not belong to the predefined acute and critical illness category, but there are abnormal vital signs markers.

[0030] Low priority: The disease diagnosis does not belong to the predefined acute and critical illness category and there are no abnormal vital signs markers.

[0031] It should be added that the criteria for determining whether the patient's disease diagnosis information belongs to the predefined set of acute and critical illnesses are as follows: if and only if the patient's disease name and International Classification of Diseases code completely match any disease entry defined in the set, and the severity level recorded in the patient's medical records reaches or exceeds the critical level standard corresponding to that disease entry, then the patient's disease diagnosis information is determined to belong to the predefined set of acute and critical illnesses.

[0032] Assign the current multiple service requests to the corresponding priority level sequence.

[0033] For the same priority level, a basic priority score is calculated for each service request in the sequence based on the waiting time since the service request was generated and the spatial Euclidean distance between the robot's current position and the request initiation position. The specific calculation logic is as follows: the waiting time and spatial Euclidean distance are first normalized, and the two normalized parameters are linearly weighted and fused to obtain the basic priority score.

[0034] The weighting of waiting time and spatial Euclidean distance follows the principle of prioritizing clinical urgency over pathway efficiency and can be calibrated based on the nursing conditions of inpatient wards. For example, the intensive care unit uses a ratio of waiting time: 0.8 and spatial distance: 0.2. General wards use a ratio of waiting time: 0.5 and spatial distance: 0.5.

[0035] The service requests within each priority level sequence are sorted in descending order of basic priority score. The priority level sequences are then combined to generate the robot's final service priority sequence.

[0036] This invention, through the construction of dynamic risk profiles and the planning of service priorities based on them, enables robots to intelligently identify and prioritize more urgent patient requests in clinical practice. This breaks through the traditional first-come-first-served or simple tiered service model, significantly improving the targeting of service responses and the efficiency of medical resource allocation.

[0037] The clinical problem reconstruction module, after the robot arrives at the designated patient location according to the planned sequence, collects the patient's verbal inquiry audio and converts it into text. It then performs intent recognition and key information slot filling on the audio text, dynamically selects patient medical history dimensions related to the intent for contextual enhancement, and reconstructs it into a standard clinical problem.

[0038] In a preferred embodiment of the present invention, the intent recognition includes: performing word segmentation, part-of-speech tagging, and dependency parsing on the converted audio text.

[0039] Semantic role labeling identifies the predicate core and core arguments in audio text, including agent, patient, symptom, body part, and time sequence information.

[0040] The words in the audio text are matched with predefined medical domain keywords and intent associations, and the matched keywords and their associated intents are recorded as initial intent clues.

[0041] It should be noted that the data sources for the above-mentioned predefined medical domain keyword-intent association relationship are as follows: candidate medical keywords are extracted from multi-source heterogeneous medical text data using natural language processing technology and mapped to standard medical terms. The multi-source heterogeneous medical text data includes, but is not limited to, medical literature and drug instructions, electronic medical records and doctor-patient dialogue records, and professional clinical guidelines.

[0042] Medical experts label keywords with the most likely categories of consultation intent based on clinical scenarios.

[0043] Based on the frequency of keyword occurrence in different intent contexts as statistically analyzed during the system development phase, the normalized frequency of occurrence is used as the association strength between the keyword and each intent.

[0044] In summary, a structured keyword-intent-weight triplet association mapping table is formed, serving as the knowledge foundation for the keyword-intent association relationship in the medical field.

[0045] Based on the predicate core and core arguments in the audio text, perform intention logic reasoning and match the intention supported by its syntactic structure pattern as clues for reasoning intention.

[0046] The intent logic reasoning is constructed based on the semantic framework and clinical dialogue mode in the medical consultation scenario. Its execution process is as follows: when the core semantics of the predicate belongs to the treatment category and the associated patient argument type is drug, the drug consultation intent is inferred.

[0047] When the core semantics of a predicate belong to the perception or state category, and its associated arguments contain symptom descriptions, then the intention to interpret the symptoms can be inferred.

[0048] When the core semantics of a predicate belong to a relation category, and its associated arguments contain clearly defined disease or surgical entities, the intent to consult on treatment plans can be inferred.

[0049] When the core semantics of the predicate belong to the activity category, and its associated arguments or adverbs contain temporal information pointing to the postoperative stage, the intention of rehabilitation guidance can be inferred.

[0050] Prioritize inference intent clues; when there is no syntactic pattern match, use the intent with the most hits from the initial intent clues.

[0051] The intention categories include at least medication consultation, symptom explanation, treatment plan, and rehabilitation guidance.

[0052] In a preferred embodiment of the present invention, the key information slot filling includes: activating the slot template corresponding to the patient intent category, wherein the slot template defines the key information fields to be filled under its associated intent.

[0053] It should be noted that the slot template is a predefined structured data pattern for each type of query intent. It clearly defines the necessary and optional semantic roles, i.e., slots, for fully and accurately representing clinical inquiries under this type of intent, and defines the types of medical concepts that each slot is expected to fill and the possible data format constraints, such as drugs, symptoms, body parts, dosages, and times.

[0054] The design of the slot template mainly stems from two aspects: (a) clinical diagnosis and treatment guidelines and consultation paths, which abstract the core information items that need to be confirmed in standard doctor-patient communication.

[0055] (b) Analysis of historical desensitized doctor-patient dialogue data, summarizing the semantic elements that repeatedly appear in different consultation scenarios.

[0056] This design was completed during the offline preprocessing phase in the early stages of system development and directly integrated into the robot software. During runtime, the system only needs to activate and call the corresponding preset template based on the recognized intent, without needing to generate it in real time.

[0057] Extract entity content that matches the key information field from the audio text, and map the extraction results to the corresponding slots.

[0058] After the filling is completed, check whether the mandatory slots in the slot template have been filled. If there are empty mandatory slots, fill them in by combining the relevant fields in the medical records and output the filled semantic slots.

[0059] It should be added that the above-mentioned supplementary filling based on relevant fields in the medical record data is specifically executed as follows: based on the context of the current dialogue, the latest record related to the semantic type of the vacant slot is retrieved from the patient's medical record data.

[0060] If a unique record is found, its value is automatically filled into the empty slot, and a source marker for the medical record is added. For example, if a specific drug name is missing from the medication consultation intent, but the medical record records a related drug currently in use, it is filled as a candidate.

[0061] If multiple possible records are found or the information cannot be determined, the system will not automatically fill in the information. Instead, it will trigger a process of the robot asking the patient specific follow-up questions to ensure the accuracy of the information.

[0062] In a preferred embodiment of the present invention, the dynamic selection of patient medical record dimensions related to intent for context enhancement includes: dynamically loading a corresponding subset of medical record information from the medical record data according to the identified patient intent. This loading is based on core information dimensions necessary for differential diagnosis and safety assessment of different consultation types in clinical medicine, aiming to ensure the individualization and safety of subsequent answers. Specifically, when the identified intent is for drug consultation, medication records, allergy history, and disease diagnosis information are loaded to verify whether the drug is currently in use and whether there are any allergic contraindications.

[0063] When the identified intent is to interpret the symptoms, disease diagnosis information, surgical history, and medication records are loaded to determine whether the new symptoms are a possible progression of the original disease, a postoperative complication, or a side effect of the medication.

[0064] When the intent is identified as a consultation for treatment options, disease diagnosis information, surgical history, medication records, and recent physical examination information are loaded.

[0065] When the identified intent is rehabilitation guidance, the surgical history, disease diagnosis information, and recovery stage information are loaded.

[0066] The medical entities populated in the semantic slots are compared and confirmed with the loaded subset of medical record information. An entity linking operation is then performed, specifically: the medical entities populated in the semantic slots are compared with the standardized records in the loaded subset of medical record information using string or medical concept codes, such as comparing drug names with drug codes, or symptom descriptions with symptom term codes. If a complete match is found, a link is established and the entity is marked as confirmed.

[0067] If no match can be found, semantic similarity is calculated using word vectors to find corresponding or related records in a subset of medical records. For example, the patient's colloquial description of surgery is linked to the standardized laparoscopic cholecystectomy in the medical record.

[0068] Based on the matching results, a status label is attached to the slot entity. For example, a drug entity linked to a medication record is labeled as currently in use. If linked to an allergy history, it is labeled as contraindicated due to allergy.

[0069] Based on the loaded subset of medical record information and the current department environment information, semantic parsing is performed on ambiguous entities in semantic slots, and contextual disambiguation is carried out. Specifically, for slot entities with multiple interpretations, the loaded subset of medical records and the current department environment information are integrated, disambiguation rules are applied to calculate the confidence scores of each possible interpretation, the interpretation with the highest score is selected as the accurate semantic and labeled back into the slot.

[0070] The disambiguation rules include: analyzing the conformity of a candidate explanation for a slot entity from four dimensions: department matching, diagnosis association, medication association, and time sequence association. Each dimension is accompanied by a rated confidence score. If the specialty of the candidate explanation is consistent with the current robot department, it means that the department matching is successful.

[0071] If the candidate interpretation is a typical manifestation or complication of the patient's current primary diagnosis, it indicates that the diagnosis is related.

[0072] If the candidate explanation can be explained by the side effects or therapeutic effects of the drug in the patient's current medication record, it indicates that there is a correlation in medication use.

[0073] If the candidate explanation is medically reasonable in relation to the patient's recent surgical history, treatment events, or the timing of symptom onset, it indicates a temporal correlation. The determination of the reasonable attribute is mainly based on whether the time scale of the physiological or pathological process mapped by the candidate explanation is compatible with the time point or time period of the patient's actual event.

[0074] As an example, if the candidate is interpreted as a surgical complication, then check whether the onset of symptoms falls within the typical time window for postoperative complications of that type of surgery. For example, postoperative infection usually occurs within a few days to two weeks after surgery. If it falls within the window, it is considered reasonable.

[0075] If the candidate is explained as a drug side effect, check whether the onset time of the symptoms occurred after the patient started taking the drug and whether the time interval is within the known onset time range of the side effect, such as an immediate allergic reaction or a side effect that accumulates over several weeks. If so, it is considered reasonable.

[0076] If a candidate explanation is associated with the progression of a chronic disease, then examine whether the appearance of symptoms matches the expected timeline of the disease under the current treatment.

[0077] After entity linking and context disambiguation, semantic slots are labeled to form standard clinical questions that include the patient's specific clinical background.

[0078] This invention enhances context by combining intent recognition, slot filling, and dynamically retrieved patient medical records. It can accurately reconstruct vague, colloquial patient inquiries into standard clinical questions containing individualized clinical backgrounds, profoundly solving the intent recognition bias problem caused by the lack of contextual understanding in existing technologies, and significantly improving the accuracy of human-computer interaction.

[0079] The science popularization answer generation module uses the robot's built-in medical question-and-answer model to initially answer standard clinical questions. Based on the patient's intent and the selected patient medical record dimensions, it retrieves supporting evidence chains from the medical knowledge graph, verifies the rationality of the initial answer, generates the final health science popularization content, and outputs it.

[0080] In one embodiment of the present invention, the construction, hierarchy and execution process of the medical question-answering model are as follows: The medical question-answering model is constructed through offline training, and a deployable special language model is obtained through the following steps of offline training: collecting medical textbooks, clinical guidelines, drug instructions and desensitized electronic medical records, cleaning, segmenting and standardizing them, and constructing a pre-training corpus.

[0081] A structured instruction dataset was constructed, in which each data point includes a standard clinical question and its corresponding standardized answer. The standardized answers were generated by medical experts based on the aforementioned medical text, ensuring professional accuracy.

[0082] The base language model is then further pre-trained using the pre-trained corpus. Supervised fine-tuning is then performed using the instruction dataset. During the fine-tuning phase, safety rules are formatted as text instructions; if a drug is mentioned, contraindications must be checked. These instructions are then mixed with fine-tuning samples for training, enabling the model to learn to automatically apply safety constraints during the generation process.

[0083] When the model is working, it processes the input in a preset logical order: Semantic parsing and context binding layer: receives standard clinical question text, identifies the core medical entities, patient background statements and points of doubt in the question through its pre-trained and fine-tuned model parameters, and associates them with the current patient medical record dimension provided by the system.

[0084] Internal knowledge reasoning and content generation layer: Based on the bound semantic representation, the model derives answer elements from its parameterized knowledge and organizes them into coherent text. This process does not involve real-time external retrieval and relies entirely on the model's existing knowledge.

[0085] Security and compliance filtering layer: After generating the initial text, the model calls its internally integrated security rule module for rapid verification. This module detects and corrects taboo combinations and absolute assertions in the text based on keyword matching and pattern recognition.

[0086] Popular Science Format Output Layer: Finally, the model reorganizes the text according to the preset popular science writing structure and outputs it in a conversational style.

[0087] When the popular science answer generation module calls the model, it executes in the following order: inputting the standard clinical question text, along with the currently selected key information of the patient's medical record dimension, into the model.

[0088] The model processes the input through its semantic parsing and context binding layer, and then calculates and generates the original solution text through its internal knowledge reasoning and content generation layer.

[0089] The original solution text is sent to a security and compliance filtering layer, where it is scanned and modified as necessary according to preset rules.

[0090] The security-filtered text is then converted into the final format by the popular science output layer, and output as preliminary answer content.

[0091] In a preferred embodiment of the present invention, the process of constructing the medical knowledge graph includes: using diseases, drugs, symptoms, and examination indicators as entity nodes of the graph.

[0092] Define the medical relationship types between entity nodes, establish the connection edges between entity nodes in the graph, and assign confidence weights to each entity relationship in the graph. The medical relationship types include etiology relationship, treatment relationship, contraindication relationship, and examination relationship.

[0093] The confidence weight is assigned based on the source and strength of the evidence supporting the establishment of the entity relationship. The authority level of the evidence source is quantified as the benchmark weight factor, and the frequency of recurrence of the evidence in independent medical literature is used as the consistency reinforcement factor. The final confidence weight is calculated by multiplying the weight factor and the reinforcement factor.

[0094] The authority level of evidence sources is classified based on the type of medical literature and the issuing institution. For example, clinical practice guidelines, clinical pathways and drug instructions issued by medical organizations represent mandatory or core industry standards and belong to the highest level of authority. Following the principle that the higher the level, the larger the preset value, a benchmark weight factor value is preset for each level of authority.

[0095] This allows us to construct a structured medical knowledge graph that includes entity nodes, semantic relationship edges, and confidence weights.

[0096] Reference Figure 2As shown in a preferred embodiment of the present invention, the supporting evidence chain retrieval process includes: based on the identified patient intent, performing semantic similarity calculation and standardized matching between the core medical concepts extracted from the standard clinical question through semantic slot filling and entity nodes in the medical knowledge graph, thereby locating and activating one or more corresponding entity nodes in the graph to form an initial entity set.

[0097] Based on the selected patient medical record dimensions, perform context pruning on the medical knowledge graph to remove entities and associated paths that are contraindicated with the patient's safety status.

[0098] The search is centered on the query intent and the associated target entity type. A breadth-first path search is performed on the pruned subgraph to collect all reachable paths as candidate paths.

[0099] The candidate paths obtained from the search are sorted by evidence chain strength, and a preset number of candidate paths at the top of the sequence are retained as supporting evidence chains. The evidence chain strength sorting process is as follows: calculate the arithmetic mean of the confidence weights of all relation edges on the candidate path, and use it as the confidence strength of the candidate path.

[0100] Calculate the reciprocal of the number of entity nodes contained in the path, and use it as the transmission strength of the candidate path.

[0101] The sum of the confidence strength and the transmission strength is used as a comprehensive strength index, and candidate paths are sorted in descending order of this index.

[0102] Reference Figure 3 As shown in a preferred embodiment of the present invention, the rationality verification of the preliminary answer content includes: extracting key medical assertions from the preliminary answer content, wherein the medical assertions include etiological judgment, treatment suggestions, medication guidance and prognosis assessment.

[0103] The key medical assertions from the initial solutions are projected onto a graph. Specifically, each extracted key medical assertion undergoes syntactic analysis to identify its subject-relation-object triple structure. Then, the medical entities in the triples are mapped to specific nodes in the graph via entity links, and relational verbs are mapped to predefined semantic relation types in the graph. Finally, each assertion is transformed into one or more graph node-relation-graph node paths.

[0104] By tracing the paths between key medical assertions, we determine whether there are logical contradictions in the preliminary answer. If so, we directly map the preliminary answer to a low level of reasonableness.

[0105] For example, if the key medical assertion in the preliminary solution is mapped as "Drug D--[Treatment]-->Disease X", and there is an edge in the graph that reads "Drug D--[Contraindication]-->Disease X", then it is determined to be a logical contradiction.

[0106] Conversely, the extracted key medical assertions are compared and verified item by item with the supporting chain of evidence to verify the strength of evidence for each key medical assertion. Based on the overall strength of evidence for all key medical assertions, the reasonableness level of the preliminary solution is evaluated as the reasonableness verification result.

[0107] It should be added that the verification process for the strength of evidence for each of the above key medical assertions is as follows: the path segment projected by the key medical assertion into the graph is compared with the supporting chain of evidence.

[0108] If the path segment perfectly matches any sub-path in any supporting chain of evidence, then the comprehensive strength index of that chain of evidence is directly used as the evidence support strength for the key medical assertion.

[0109] If the path segment does not perfectly match any supporting chain of evidence, but can be extended through the nodes of the supporting chain of evidence in the graph, then the strength index after exponential decay of the number of extended nodes is recalculated using the length of the shortest connection path formed after extension and the confidence weight of the edges involved as its evidence support strength.

[0110] If the path segment cannot be connected through the nodes of the supporting evidence chain, a limited-range graph search centered on the assertion path segment is initiated. If a supporting path is found, its strength value is calculated as the evidence support strength. If no supporting path is found, the evidence support strength of the assertion is determined to be zero.

[0111] The process for assessing the reasonableness level of the preliminary response content is as follows: integrate the evidence support strength of all key medical assertions, take their arithmetic mean as the overall evidence support strength level, and determine the reasonableness level of the preliminary response content according to the predefined range of overall evidence support strength levels corresponding to each reasonableness level.

[0112] The rationality level is divided into three categories: low, medium, and high.

[0113] In a preferred embodiment of the present invention, the health science popularization content generation process includes: if the preliminary answer is of high rationality level, it is directly output as the final health science popularization content.

[0114] If the level of reasonableness is medium, then medical assertions in the preliminary answer that lack strong evidence will be replaced or deleted. Specifically, using the core entities in the key medical assertions lacking strong evidence as the query objects, a single-hop query will be performed along predefined substitutability relationships in the medical knowledge graph to obtain all directly related target entities with equivalent treatment or effects, forming a candidate substitutability entity set. The substitutability relationships include similar drugs, equivalent treatments, and alternative solutions.

[0115] Each entity in the candidate alternative entity set is matched with all entities in the supporting chain of evidence using either string matching or standard medical code matching.

[0116] For each successfully matched candidate alternative entity, the strength index of its supporting chain of evidence is obtained. Only candidate entities with this value higher than the strength of evidence support obtained for the original medical assertion are retained, forming a high-strength candidate set.

[0117] Each candidate entity in the high-intensity candidate set is subjected to a security conflict check with the current patient medical record dimension: if the candidate entity exists in the patient's allergy history list, it is removed.

[0118] If a candidate entity has a contraindication relationship with the patient’s current diagnosis as defined in the atlas, it is removed.

[0119] If a candidate entity has a serious interaction with a drug in the patient's current medication list, it will be removed.

[0120] If the high-strength candidate set is not empty after filtering, the candidate entity with the highest evidence chain strength value is selected from the set, and a new key medical assertion is generated based on it to replace the original assertion.

[0121] Conversely, if the high-intensity candidate set is empty, pruning is triggered.

[0122] If the level of reasonableness is low, the patient's second interaction process is initiated. The answer content is reconstructed based on the supplementary evidence retrieved. If the level of reasonableness is still low within the preset maximum number of interactions, a suggestion to transfer the case to manual processing is output.

[0123] In a preferred embodiment of the present invention, the system further includes: after the service request ends, recording the intent recognition result and patient feedback data to form an interaction log record.

[0124] Periodically filter out negative patient feedback session records from the stored interaction logs.

[0125] Analyze the audio text in the conversation records to extract unrecognized high-frequency words or ambiguous syntactic patterns that cause recognition difficulties.

[0126] Based on the analysis results, incremental optimization is performed on the keyword-intent association relationship or intent logic reasoning rules in the medical field.

[0127] This invention introduces a medical knowledge graph to retrieve supporting evidence chains and verify the rationality of preliminary answers, establishing a rigorous knowledge verification and security filtering mechanism. This ensures that the final output of health science content is not only personalized but also possesses medical reliability and safety, effectively avoiding misleading or inappropriate advice that may arise from general models.

[0128] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A hospital health education robot service system, characterized in that, include: The service request receiving module receives multiple service requests initiated by a designated calling device, which include the patient's identifier and physical location. The service priority planning module retrieves medical records based on patient identification and constructs a weighted scoring rule based primarily on medical record characteristics and secondarily on waiting time and spatial distance, thereby planning the robot's service priority sequence. The clinical problem reconstruction module collects the patient's verbal questions and converts them into text after the robot arrives at the designated patient location according to the planned sequence. It then performs intent recognition and key information slot filling on the audio text and dynamically selects patient medical history dimensions related to the intent for context enhancement to reconstruct the standard clinical problem. The science popularization Q&A generation module uses the robot's built-in medical question-and-answer model to provide preliminary answers to standard clinical questions. Based on the patient's intent and the selected patient medical record dimensions, it retrieves supporting evidence chains from the medical knowledge graph, verifies the reasonableness of the preliminary answers, and generates and outputs the final health science popularization content.

2. The hospital health education robot service system according to claim 1, characterized in that: The weighted scoring rules include: Extract disease diagnosis information and recent physical examination information from medical records as medical record feature items; Determine whether the patient's disease diagnosis information belongs to a predefined set of acute and critical illnesses, and whether there are abnormal vital signs markers in the recent physical examination information. Based on the consistency between acute and critical illnesses and abnormal vital signs markers, the service request is divided into four priority levels: highest, high, medium, and low. Assign the current multiple service requests to the corresponding priority level sequence; For the same priority level, the basic priority score of each service request in the sequence is calculated based on the waiting time since the service request was generated and the spatial Euclidean distance between the robot's current position and the request initiation position. The service requests within each priority level sequence are sorted in descending order of basic priority score. The priority level sequences are then combined to generate the robot's final service priority sequence.

3. The hospital health education robot service system according to claim 1, characterized in that: The intent recognition includes: The converted audio text is segmented, part-of-speech tagging is performed, and dependency parsing is conducted. Semantic role labeling identifies the predicate core and core arguments in audio text, including agent, patient, symptom, body part, and time sequence information. The words in the audio text are matched with predefined medical domain keywords and intent associations, and the matched keywords and their associated intents are recorded as initial intent clues. Based on the predicate core and core arguments in the audio text, perform intention logic reasoning and match the intention supported by its syntactic structure pattern as clues for reasoning intention. Prioritize inference intent clues; when there is no syntactic pattern match, use the intent with the most hits from the initial intent clues. The intent categories include at least medication consultation, symptom explanation, treatment plan, and rehabilitation guidance.

4. The hospital health education robot service system according to claim 1, characterized in that: The key information slot filling includes: Activate the slot template corresponding to the patient intent category, wherein the slot template defines the key information fields that need to be filled under its associated intent; Extract entity content that matches the key information field from the audio text, and map the extraction results to the corresponding slots; After the filling is completed, check whether the mandatory slots in the slot template have been filled. If there are empty mandatory slots, fill them in by combining the relevant fields in the medical records and output the filled semantic slots.

5. A hospital health education robot service system according to claim 4, characterized in that: The dynamic selection of patient medical record dimensions related to intent for contextual enhancement includes: Based on the identified patient intent, the corresponding subset of medical record information is dynamically loaded from the medical record data; The medical entities populated in the semantic slots are compared and confirmed with the loaded subset of medical record information, and then entity linking operations are performed. Based on the loaded subset of medical record information and the current departmental environment information, semantic parsing is performed on ambiguous entities in the semantic slots, and contextual disambiguation is carried out. After entity linking and context disambiguation, semantic slots are labeled to form standard clinical questions that include the patient's specific clinical background.

6. The hospital health education robot service system according to claim 1, characterized in that: The construction process of the medical knowledge graph includes: Diseases, drugs, symptoms, and test indicators are used as entity nodes in the graph; Define the medical relationship types between entity nodes, establish the connection edges between entity nodes in the graph, and assign confidence weights to each entity relationship in the graph. The medical relationship types include etiology relationship, treatment relationship, contraindication relationship, and examination relationship. This allows us to construct a structured medical knowledge graph that includes entity nodes, semantic relationship edges, and confidence weights.

7. A hospital health education robot service system according to claim 6, characterized in that: The supporting evidence chain retrieval process includes: Based on the identified patient intent, determine the starting set of entities in the medical knowledge graph; Based on the selected patient medical record dimensions, perform context pruning on the medical knowledge graph to remove entities and associated paths that are contraindicated with the patient's safety status; Using the predefined target entity type associated with the query intent as the exploration direction, perform path search on the pruned subgraph to find candidate paths from the starting entity set to the target entity type; The candidate paths obtained from the search are sorted by the strength of the evidence chain, and a preset number of candidate paths at the top of the sequence are retained as supporting evidence chains.

8. A hospital health education robot service system according to claim 1, characterized in that: The process of verifying the reasonableness of the preliminary answer includes: Key medical assertions are extracted from the preliminary answers, including etiological judgment, treatment recommendations, medication guidance, and prognostic assessment. The key medical assertions in the preliminary answer are projected onto the graph. By searching the paths between the key medical assertions, it is determined whether there are logical contradictions in the preliminary answer. If there are, the preliminary answer is directly mapped to a low level of reasonableness. Conversely, the extracted medical assertions are compared and verified item by item with the supporting chain of evidence to verify the strength of evidence for each medical assertion. Based on the overall strength of evidence for all medical assertions, the reasonableness level of the preliminary answer is evaluated as the reasonableness verification result. The rationality level is divided into three categories: low, medium, and high.

9. A hospital health education robot service system according to claim 8, characterized in that: The process of generating the health science content includes: If the initial answer is rated as highly reasonable, it will be directly output as the final health science content. If the level of reasonableness is medium, then medical assertions in the preliminary response that are not supported by strong evidence will be replaced or deleted. If the level of reasonableness is low, a second patient interaction process is initiated to reconstruct the answer based on the evidence retrieved from the supplementary search.

10. A hospital health education robot service system according to claim 1, characterized in that: The system also includes: After the service request is completed, the intent recognition result and patient feedback data are recorded to form an interaction log. Periodically filter out negative patient feedback conversation records from the stored interaction logs; Analyze the audio text in the conversation records to extract unrecognized high-frequency words or ambiguous syntactic patterns that cause recognition difficulties; Based on the analysis results, incremental optimization is performed on the keyword-intent association relationship or intent logic reasoning rules in the medical field.

Images