Medical device control system based on clinical trial protocol

Abstract

Embodiments of the present disclosure disclose a medical device control system based on a clinical trial plan. A specific embodiment of the system includes a plan generation server, a literature abstract extraction server, a data analysis server, and a medical robot, wherein: the literature abstract extraction server is configured to generate a query argument tree and a search information result set based on text information and literature request information, generate an ideal answer vector set according to the search information result set, query the ideal answer vector set to generate a search result group, and process the search result group to obtain a literature abstract; the data analysis server is configured to perform instruction conversion on query information to obtain a query instruction, and execute the query instruction to obtain an information query result; and the plan generation server is configured to generate a clinical trial plan, and send the clinical trial plan to the medical robot to perform a clinical trial operation. The embodiment can reduce system response time and waste of test samples.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to medical device control systems based on clinical trial protocols. Background Technology

[0002] A medical device control system based on a clinical trial protocol is a computer system used to assist medical researchers in controlling medical robots to execute standard clinical trial protocols. Existing medical device control systems for clinical trial protocols typically employ a combination of static protocol templates and a general large language model. User-input text information is directly filled into the protocol template, then the protocol is validated based on a basic rule base to generate the clinical trial protocol. Finally, the parameters in the protocol are directly mapped into control commands to control the medical robot to perform operations.

[0003] However, existing medical device control systems based on clinical trial protocols often suffer from the following technical problems:

[0004] The system suffers from long response times and wasted experimental samples. Static protocol templates cannot adapt to different research types or risk levels, and the system's poor scalability leads to excessively long response times when processing large-scale data. Furthermore, the lack of real-time awareness of the multimodal intervention needs of patients with complex comorbidities prevents dynamic optimization of equipment control logic based on patient characteristics, resulting in a lower success rate for medical robot operations and wasted experimental samples.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the present disclosure concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure propose a medical device control system based on a clinical trial protocol to address one or more of the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a medical device control system based on a clinical trial protocol. The system includes: a protocol generation server, a literature abstract extraction server, a data analysis server, and a medical robot, all interconnected. The literature abstract extraction server performs the following steps: in response to receiving text information and literature request information, it generates a query argument tree based on an information retrieval model, the text information, and the literature request information; it performs a pyramid-perception retrieval based on the query argument tree to obtain a retrieval result set; it generates an ideal answer vector set based on the retrieval result set; it performs a sorting query on the ideal answer vector set based on a re-ranking model and a pre-set literature database to generate a retrieval result group; and it merges and normalizes the retrieval result group to obtain a literature abstract. The data analysis server performs instruction conversion on the query information to obtain a query instruction, and executes the query instruction based on a medical information knowledge graph to obtain information query results. The protocol generation server generates a clinical trial protocol based on the text information, the literature abstract, and the information query results, and sends the clinical trial protocol to the interconnected medical robot to perform the clinical trial operation.

[0009] The various embodiments of this disclosure have the following beneficial effects: they can reduce system response time and waste of experimental samples. Specifically, the reasons for long system response time and waste of experimental samples are as follows: long system response time leads to waste of experimental samples. Static protocol templates cannot adapt to different research types or risk levels. When processing large-scale data, the system has poor scalability, resulting in excessively long system response time. Furthermore, the lack of real-time perception of the multimodal intervention needs of patients with complex comorbidities prevents dynamic optimization of device control logic based on patient characteristics, leading to a reduced success rate of medical robot operation and waste of experimental samples. Based on this, some embodiments of this disclosure include a medical device control system based on clinical trial protocols, comprising a protocol generation server, a literature abstract extraction server, a data analysis server, and a medical robot that are interconnected. First, the literature abstract extraction server performs the following steps: in response to receiving text information and literature request information, it generates a query argument tree based on an information retrieval model, the text information, and the literature request information; and performs a pyramid-perception retrieval based on the query argument tree to obtain a retrieval information result set. This generates a candidate literature set with the highest relevance to the literature request information, avoiding the processing of irrelevant literature. Based on the aforementioned retrieval results set, an ideal answer vector set is generated. This converts textual information into vector representation, improving information reliability. Using a re-ranking model and a pre-built literature database, the ideal answer vector set is sorted to generate a retrieval result set. This again converts textual information into vector representation, improving information reliability. The retrieval result set is then merged and normalized to obtain a literature abstract. This generates a concise literature abstract, reducing the amount of data transmitted to downstream servers. The data analysis server then performs instruction conversion on the query information to obtain query instructions, and executes these instructions based on a medical information knowledge graph to obtain information query results. This converts fuzzy queries into precise queries associated with the knowledge graph, avoiding the complexity of semantic parsing. Finally, the scheme generation server generates a clinical trial protocol based on the aforementioned textual information, literature abstract, and information query results, and sends the clinical trial protocol to a connected medical robot for execution. This provides a compliant clinical trial protocol and control instructions, enabling the medical robot to perform accurate operations. Ultimately, this reduces system response time and waste of experimental samples. Attached Figure Description

[0010] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0011] Figure 1 This is a schematic diagram of the structure of some embodiments of a medical device control system based on a clinical trial protocol according to the present disclosure;

[0012] Figure 2 This is a general flowchart of some embodiments of a medical device control system based on a clinical trial protocol according to the present disclosure;

[0013] Figure 3 This is a diagram of the input information interface of a medical device control system based on a clinical trial protocol according to some embodiments of the present disclosure;

[0014] Figure 4 This is a diagram showing the generation of an interface based on some embodiments of a medical device control system based on a clinical trial protocol, as disclosed herein. Detailed Implementation

[0015] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0016] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0017] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0018] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0019] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0020] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] Figure 1This is a schematic diagram of the structure of some embodiments of a medical device control system based on a clinical trial protocol according to this disclosure. The aforementioned medical device control system based on a clinical trial protocol includes a document abstract extraction server 101, a data analysis server 102, a protocol generation server 103, and a medical robot that are interconnected. See also... Figure 2 The following is a general flowchart of some embodiments of a medical device control system based on a clinical trial protocol, according to the present disclosure.

[0022] The above-mentioned document abstract extraction server 101 is used to perform the following steps:

[0023] Step one: In response to the received text information and document request information, based on the information retrieval model, the aforementioned text information, and the aforementioned document request information, a query argument tree is generated. Then, based on the aforementioned query argument tree, a pyramid-perception retrieval is performed to obtain the retrieval result set. In practice, the aforementioned document abstract extraction server can obtain the retrieval result set through the following sub-steps:

[0024] In the first sub-step, in response to receiving text information and document request information, the semantic intent is extracted by parsing the text information and document request information based on the information retrieval model. In practice, the document abstract extraction server can receive text information and document request information through an internal communication interface, and input the text information and document request information into the information retrieval model to obtain the semantic intent. The document abstract extraction server can be an intelligent retrieval server integrating the IEEE database API. The internal communication interface can be a gRPC communication interface.

[0025] The semantic intent mentioned above can be a structured set of clinical questions (PICO: Patient, Intervention, Comparison, Outcome). It can include a patient population, an intervention, a comparison, and an outcome. The patient population could be patients with spinal metastases. The intervention could be treatment of spinal metastases using a combination of bone-strengthening agents. The comparison could be the current standard of care. The outcome could be the efficacy (e.g., progression-free survival, PFS) and safety of the trial protocol. The textual information can be a background description related to the trial protocol. The document request information can be the search topic and keywords. The search results set can include document identifiers, titles, and content snippets. The information retrieval model can be a neural network model that takes textual information as input and semantic intent as output, and can include a text parsing layer, an intent encoding layer, and a retrieval output layer. The text parsing layer can convert textual information into word embedding vectors, and then use a bidirectional long short-term memory network (BiLSTM) to convert these word embedding vectors into context-aware feature sequences. The aforementioned intent encoding layer can perform deep encoding on the text feature sequence through an attention encoding mechanism to obtain a semantic intent vector. The aforementioned retrieval output layer can use a softmax function to determine the similarity score between the aforementioned semantic intent vector and the document vectors corresponding to each medical trial document in the aforementioned pre-set medical trial document database, and in response to the aforementioned similarity score exceeding a preset threshold (which can be 0.5), the aforementioned medical trial document is identified as the semantic intent.

[0026] Sub-step two involves constructing a query argument tree based on the aforementioned semantic intent. This query argument tree can be a tree structure consisting of a core clinical medical claim (root node) and sub-claims for medical support (leaf nodes). The root node can be a description of the core content of the semantic intent, such as: "The intervention has advantages over the control group in terms of endpoint indicators" for the "patient population". The leaf nodes can include mechanism-related leaf nodes, preclinical evidence leaf nodes, clinical evidence leaf nodes, and unmet needs leaf nodes. For example, the mechanism-related leaf node can be used to characterize whether the mechanism of action of the intervention (such as the combined bone-strengthening agent Zanzalintinib) is highly correlated with the pathophysiology of the target disease (such as spinal metastases). The preclinical evidence leaf node can be used to characterize preliminary data obtained before clinical trials that demonstrate the effectiveness and safety of the intervention. The clinical evidence leaf node can be data obtained during clinical trials that demonstrate the effect of the intervention on patients. The unmet needs leaf node can be used to characterize limitations of the current experimental protocol for the target patient population.

[0027] Sub-step three involves performing a pyramid-perception retrieval on the aforementioned argument tree to obtain a preliminary literature set. This pyramid-perception retrieval can be a hierarchical evidence retrieval strategy that prioritizes high-quality evidence sources and weights them. As an example, the document abstract extraction server can perform a preliminary search in a pre-built medical trial literature database (such as PubMed) based on each argument query node (leaf node) of the aforementioned argument query tree to obtain multiple high-level evidence documents (such as systematic reviews, randomized controlled trials, and the latest conference abstracts). These high-level evidence documents are then weighted and summed to obtain the preliminary literature set.

[0028] Sub-step four involves fusing the preliminary literature set with the query argument tree using a graph structure to obtain a retrieval result set. In practice, the literature abstract extraction server can extract key entities from the preliminary literature set, obtaining key literature entities and their corresponding entity relationships, and defining these key literature entities and their corresponding entity relationships as a literature entity relationship set. Then, this literature entity relationship set can be mapped to the query argument tree to construct a graph-structured retrieval result set. For example, the key literature entity describing "Zanzalintinib inhibits the MET pathway" can be associated with a mechanism-related leaf node. These key literature entities can include, but are not limited to, specific drugs, gene targets, adverse reactions, and clinical endpoints. The entity relationships can include, but are not limited to: drug A inhibits target B, target B improves disease C, and treatment significantly improves endpoint D.

[0029] Step two: Based on the above search results set, generate an ideal answer vector set. In practice, the document abstract extraction server can use hypothetical document embedding (HyDE) technology to convert the search results set into an ideal answer vector set.

[0030] Step 3: Based on the re-ranking model and the pre-set document library, the aforementioned ideal answer vector set is sorted and queried to generate a set of search results. In practice, the document abstract extraction server can generate at least one similarity score by determining the cosine similarity between each ideal answer vector in the aforementioned ideal answer vector set and each document vector in the aforementioned pre-set document library, thus obtaining a similarity score set. Then, the document abstract extraction server can use the re-ranking model to sort the aforementioned similarity score set in descending order to generate a set of search results in descending order. The re-ranking model can be a machine learning model that takes the similarity score set as input and the search results as output. For example, the re-ranking model can be the Reranker model.

[0031] Step four involves merging and normalizing the above search results to obtain the document abstract. In practice, the document abstract extraction server can obtain the document abstract through the following sub-steps:

[0032] Sub-step one: Based on the preset number of document pools and the aforementioned search result groups, generate an initial document set. In practice, the document abstract extraction server can select the first N search results from the aforementioned search result groups based on the preset number of document pools to generate the initial document set. The preset number of document pools can be a preset number N (e.g., N=20).

[0033] Sub-step two involves determining the cosine similarity between each initial document in the initial document set and the document request information to generate document similarity groups.

[0034] Sub-step three involves identifying the initial document corresponding to at least one document in the document similarity group that meets the preset relevance value condition as the final document, thus obtaining the final document group. The preset relevance value condition can be a pre-set cosine similarity threshold, such as 0.7.

[0035] Sub-step four involves fusing multi-source information from the final document group based on the aforementioned document request information to generate a document abstract. In practice, the document abstract extraction server first extracts relevant text segments from each document in the final document group according to the document request information, and then performs deduplication and terminology normalization on these relevant text segments to obtain processed text fragments, eliminating redundancy and ensuring terminology consistency. Subsequently, the document abstract extraction server inputs the processed text fragments and the aforementioned document request information into a pre-set language model to generate a logically coherent review text with standard citations, serving as the document abstract. The pre-set language model can be a Transformer model that takes the document request information as input and the document abstract as output, and may include an input layer, an attention layer, and an output layer. The input layer converts each word in the input document request information into a word vector. The attention layer determines the weights between word vectors through a self-attention mechanism and performs a non-linear transformation on the weights between the word vectors through a feedforward neural network to generate a semantic representation vector containing global contextual information. The output layer can determine the probability distribution of the above-mentioned terms using the softmax function, and generate a literature abstract based on the above probability distribution using a self-learning regression method.

[0036] The aforementioned data analysis server 102 is used to convert query information into query instructions, and to execute the query instructions based on the medical information knowledge graph to obtain information query results.

[0037] In some embodiments, the data analysis server described above is used to convert query information into instructions to obtain query instructions, and to execute the query instructions based on a medical information knowledge graph to obtain information query results. The data analysis server may be a graph database server that supports SQL conversion and privacy query protection.

[0038] In practice, the aforementioned data analysis server can obtain information query results through the following steps:

[0039] Step one: Extract the key medical concept set and logical relationships from the query information. In practice, the data analysis server can use basic word segmentation tools (such as jieba) and Named Entity Recognition (NER) technology to perform word segmentation and semantic analysis on the query information to generate the key medical concept set and logical relationships. The key medical concept set may include, but is not limited to, disease names, drug names, and examination indicators. The logical relationships can be used to represent the logical relationships between the various key medical concepts in the key medical concept set, and can be conditional relationships, temporal relationships, and logical operators (such as "AND" and "OR").

[0040] Step two: Based on the aforementioned medical information knowledge graph, each key medical concept in the aforementioned key medical concept set is associated and mapped to obtain a standard medical terminology set. In practice, the aforementioned data analysis server can use a string matching algorithm to associate and map each key medical concept in the aforementioned key medical concept set with the aforementioned medical information knowledge graph to obtain a standard medical terminology set, thereby eliminating terminology ambiguity. As an example, non-standard terms (such as "myocardial infarction") can be mapped to standard terms (such as "myocardial infarction"). The aforementioned medical information knowledge graph can be a knowledge graph stored and represented in the form of a graph structure, which can include medical concepts (such as diseases, symptoms, drugs, anatomical structures, examination indicators, etc.) and semantic relationships between medical concepts (such as "treatment," "cause," "symptom," "complication," etc.).

[0041] Step three involves conceptually reconstructing the aforementioned standard medical terminology set and logical relationships to obtain query instructions. In practice, the data analysis server can conceptually reconstruct the aforementioned standard medical terminology set based on the logical relationships to generate query expressions, and then convert these query expressions into executable query instructions (such as SQL query code). For example, the aforementioned logical relationships (such as "AND", "OR", and "NOT") can be mapped to standard Boolean logic operators (such as "AND", "OR", and "NOT").

[0042] Step four: Execute the above query command to obtain multiple target query results. In practice, firstly, the data analysis server can use a privacy protection verification module (such as a firewall) to ensure that the above query command only involves aggregation operations (such as COUNT) and does not involve individual data. Then, the data analysis server can send the above query command to a medical database (such as a hospital information system, HIS) to execute the query command and obtain multiple target query results.

[0043] Step five involves aggregating the results of the multiple target queries to obtain the query result condition values. In practice, the data analysis server described above can use aggregation functions to aggregate the results of the multiple target queries to obtain the query result condition values.

[0044] Step Six: Based on the query result condition values ​​and the medical rule knowledge base, generate information query results. In practice, the data analysis server can convert the medical rule knowledge in the medical rule knowledge base into condition-conclusion statements. Then, the data analysis server can generate information query results based on the query result condition values ​​and the medical rule knowledge. The medical rule knowledge base can be a database storing rules and knowledge related to the medical field, including but not limited to clinical guidelines, standard operating procedures, and compliance requirements. As an example, the medical rule knowledge could be IF (Study Type == "Randomized Controlled Trial" AND Primary Endpoint == "Pathological Response Rate"), THEN (Required Sample Size >= 30).

[0045] The above-mentioned solution generation server 103 is used to generate a clinical trial protocol based on the above-mentioned text information, the above-mentioned literature abstract and the above-mentioned information query results, and to send the above-mentioned clinical trial protocol to a medical robot with a communication connection to perform clinical trial operations.

[0046] In some embodiments, the protocol generation server can be used to generate a clinical trial protocol based on the aforementioned text information, the aforementioned literature abstract, and the aforementioned information query results, and to send the aforementioned clinical trial protocol to a medical robot with a communication connection to execute the clinical trial operation. The aforementioned clinical trial protocol can be a document used for clinical research, and may include, but is not limited to, research nature, literature abstract, study population, statistical methods, and research objectives. The protocol generation server can be a cloud server integrated with a natural language processing engine.

[0047] In addressing the technical problems mentioned above by adopting technical solutions, for the application scenario—low-risk observational studies (such as retrospective analysis of medical records)—using complex templates designed for high-risk interventional studies (such as Phase III trials of new drugs) often leads to risk level mismatches. This is often accompanied by the following technical issues: relying on fixed templates results in system-generated clinical trial protocols that are not adapted to the actual risk level, containing a large amount of redundant content. This causes the system to waste excessive memory and processes processing unnecessary information, leading to wasted computing resources. Furthermore, due to the mismatch between the clinical trial protocol and the actual research scenario, timing conflicts exist in the instructions generated by the medical control server when parsing the clinical trial protocol, causing the medical robot to lag when executing medical trials. Considering the following requirements for this application scenario—accurate risk level matching, high compliance requirements, and diverse research types—we have decided to adopt the following solution:

[0048] In some optional implementations of certain embodiments, the above-described protocol generation server can generate a clinical trial protocol through the following steps:

[0049] The first step, in response to the aforementioned observational study of textual information representation, involves preprocessing the textual information based on a medical large language model to obtain preprocessed textual information. In practice, the server generating the above scheme can input the textual information into the aforementioned medical large language model to generate preprocessed textual information with stop words and special characters removed. The aforementioned medical large language model can be a large language model (LLM) that cleans and standardizes medical-related textual information. As an example, the aforementioned medical large language model could be Baichuan-M1-14B.

[0050] The second step involves semantic standardization of the preprocessed text information to obtain standardized text information. In practice, the server generating the above solution can use a medical information knowledge graph to map non-standard terms (such as "myocardial infarction") in the preprocessed text information to standard terms (such as "myocardial infarction"), thereby generating standardized text information.

[0051] The third step involves extracting key parameter information from the standardized text information based on the sequence labeling model and the aforementioned information query results, thus obtaining a key parameter information set. In practice, the solution generation server can use a semantic annotation algorithm to insert the aforementioned information query results into the standardized text information to obtain enhanced text. Then, the solution generation server can use the aforementioned sequence labeling model to perform entity recognition and classification on the enhanced text to extract the key parameter set. Finally, the solution generation server can normalize the aforementioned key parameter set to obtain a key parameter information set. The key parameters in the aforementioned key parameter set may include, but are not limited to, drug dosage, study period, study type, and endpoint indicators. The aforementioned sequence labeling model can be a BERT model.

[0052] The fourth step involves generating risk level information based on the aforementioned key parameter information set and risk rule base. In practice, the above-mentioned scheme generation server can use a rule engine algorithm to match the key parameter information set and risk rule base to determine the risk score, and then generate risk level information based on preset risk rules and the aforementioned risk score. The aforementioned risk rule base can be a set of clinical trial risk grading rules defined by an institution or regulation; for example, unapproved drugs are classified as "Level 4," and observational studies without intervention are classified as "Level 1." The aforementioned preset risk rules can be rules corresponding to pre-set risk scores and risk level information; for example, "Level 1" represents low risk, and "Level 4" represents high risk.

[0053] The fifth step involves generating a protocol template based on the aforementioned risk level information and the risk template library. In practice, the protocol generation server can search the risk template library based on the risk level information to generate a protocol template. For example, for a "high-risk" study, the protocol generation server can select an interventional study template containing a detailed safety monitoring section as the protocol template. The risk template library can be a predefined set of clinical trial protocol templates.

[0054] The sixth step is to extract the core entity information from the aforementioned text information to obtain the entity information. In practice, the server generating the above solution can use named entity recognition technology to identify the core entity information in the aforementioned text information to obtain the entity information. This core entity information may include the drug name, disease name, and evaluation criteria. The evaluation criteria may be the RECIST 1.1 standard for evaluating the efficacy of treatment for solid tumors.

[0055] Step 7: Based on the aforementioned protocol template, entity information, and literature abstract, generate a clinical trial protocol and send it to the medical robot for execution of the clinical trial. In practice, the protocol generation server can merge the entity information and literature abstract with the protocol template to obtain the clinical trial protocol. Then, the protocol generation server can send the clinical trial protocol to the medical robot for execution via a preset communication interface. The clinical trial protocol can be in PDF format. The preset communication interface can be a RESTful API.

[0056] The aforementioned steps one through seven and their related content constitute an inventive point of this disclosure. Combined with the steps described below, they address the technical problem that "relying on fixed templates leads to clinical trial protocols generated by the system being unsuitable for actual risk levels, containing a large amount of redundant content, wasting excessive memory and processes processing unnecessary information, thus wasting computational resources; and due to the mismatch between the clinical trial protocol and the actual research scenario, the instructions generated by the medical control server when parsing the clinical trial protocol have timing conflicts, causing the medical robot to lag during medical trials." The reasons for this wasted computational resources and medical robot lag are: relying on fixed templates leads to clinical trial protocols generated by the system being unsuitable for actual risk levels, containing a large amount of redundant content, wasting excessive memory and processes processing unnecessary information, thus wasting computational resources; and due to the mismatch between the clinical trial protocol and the actual research scenario, the instructions generated by the medical control server when parsing the clinical trial protocol have timing conflicts, causing the medical robot to lag during medical trials. Solving these factors resolves the problems of wasted computational resources and medical robot lag. To achieve this effect, the first step involves preprocessing the aforementioned text information based on a medical large language model, in response to the observational study of text information representation. This yields preprocessed text information after removing irrelevant content. The second step involves semantic standardization of the preprocessed text information, resulting in standardized text information. This provides semantically consistent and unambiguous text information. The third step involves extracting key parameter information from the standardized text information based on a sequence labeling model and the aforementioned information query results. This provides a basis for risk assessment. The fourth step involves generating risk level information based on the key parameter information set and the risk rule base. This determines the risk level. The fifth step involves generating a protocol template based on the risk level information and the risk template library. This allows for the selection of an appropriate template based on the risk level information. The sixth step involves extracting the core entity information from the aforementioned text information. This yields the core entity information capable of generating clinical trial protocols. The seventh step involves generating a clinical trial protocol based on the aforementioned template, entity information, and literature abstract. This protocol is then sent to the medical device control server to control the associated medical robot to execute the clinical trial procedures. This process generates compliant clinical trial protocols and ultimately reduces wasted computing resources and lag in medical robot operation.

[0057] In addressing the aforementioned technical challenges by adopting technical solutions, the application scenario—patients with complex comorbidities, intolerance to standard treatments, and requiring multimodal interventions—often presents the following problems: low system resource utilization and waste of experimental samples. Using fixed templates without considering the specific circumstances of the patient population leads to clinical trial protocols that contradict the patients' true biological characteristics, resulting in excessively high error rates in clinical trial results. This forces the system to allocate significant computational resources to processing invalid information, further reducing system resource utilization. Furthermore, the mismatch between the clinical trial protocol and the patients' true biological characteristics results in structural defects in the medical device control instruction set (such as improper settings of the robotic arm's rotation angle parameters), causing sample contamination during clinical trial operations and ultimately wasting experimental samples. Considering the following requirements for this application scenario—multiple complex diseases—the following solution is adopted:

[0058] Step one: In response to the aforementioned textual information representation interventional study, and according to the aforementioned clinical trial protocol, specific features are extracted from the acquired patient medical records to obtain a patient-specific vector set. In practice, the above-mentioned protocol generation server can convert the structured and unstructured data in the patient medical records into numerical vectors to obtain the patient-specific vector set. The structured data can be examination indicators and imaging measurements. The unstructured data can be pathology report descriptions.

[0059] Step two involves performing cluster analysis on the aforementioned patient-specific vector set to obtain a set of potential molecular subtypes. In practice, the executing entity can use an unsupervised clustering algorithm to divide the patient-specific vector set into multiple clusters (potential molecular subtypes) based on the similarity between the various patient-specific vectors in the set, thus obtaining a set of potential molecular subtypes.

[0060] Step three involves performing pathway enrichment analysis on the aforementioned potential molecular subtype set to obtain a candidate biomarker set. In practice, the above-mentioned scheme generation server can use pathway enrichment analysis algorithms to identify biological pathways exhibiting significant activity differences within the potential molecular subtype set. Then, statistical methods can be used to map these biological pathways to a biological pathway database to identify significantly enriched biological pathways and determine the key molecules within these significantly enriched pathways as candidate biomarkers, thus obtaining a candidate biomarker set. The statistical method used can be the hypergeometric test.

[0061] Step four: Based on the aforementioned candidate biomarker set, generate a biomarker score set. In practice, the above scheme generation server can use statistical testing algorithms to determine the association strength between each candidate biomarker in the aforementioned candidate biomarker set and the endpoint indicator, in order to generate at least one biomarker score, thus obtaining a biomarker score set. The aforementioned statistical testing algorithm can be a Cox Proportional Hazards Model.

[0062] Step 5: Based on the aforementioned biomarker score set, rank the candidate biomarker set by importance to obtain the biomarker sequence. In practice, the above scheme generation server can sort the candidate biomarker set in descending order based on the biomarker score set to obtain the biomarker sequence.

[0063] Step six: Based on the determined optimal thresholds for the aforementioned biomarker sequences, the optimal biomarker combination is obtained. In practice, the above-mentioned scheme generation server can use an optimization algorithm to determine the optimal thresholds for the aforementioned biomarker sequences with the goal of maximizing the performance (e.g., AUC) of the predicted patient treatment response model. In response to a biomarker score exceeding the optimal threshold in the aforementioned biomarker sequence, the biomarker is identified as the optimal biomarker, thus obtaining the optimal biomarker combination. The aforementioned optimization algorithm can be a Bayesian optimization algorithm.

[0064] Step 7: Generate a patient treatment decision framework based on the optimal biomarker combination described above. In practice, the above-mentioned treatment plan generation server can formulate a treatment strategy based on the optimal biomarker combination. Then, based on the treatment knowledge base, it can determine the treatment plan with the highest similarity to each of the optimal biomarkers in the optimal biomarker combination. Finally, the above treatment strategies and treatment plans can be integrated to obtain an executable treatment logic flowchart, which serves as the patient treatment decision framework. As an example, a treatment strategy can be defined as "target-sensitive" if biomarker A is mutated and its expression level is higher than a threshold X; otherwise, it can be defined as "insensitive." The above treatment plan can be to assign a corresponding targeted drug to "target-sensitive" patients and a standard treatment method to "insensitive" patients.

[0065] Step eight: Based on the aforementioned patient treatment decision-making framework, generate a clinical trial protocol and send it to the aforementioned medical robot for execution of the clinical trial. In practice, the protocol generation server can integrate the aforementioned patient treatment decision-making framework, the aforementioned literature abstract, and the protocol template to obtain the clinical trial protocol. Then, the protocol generation server can send the clinical trial protocol to the aforementioned medical robot for execution of the clinical trial through a preset communication interface.

[0066] Steps one through eight and their related content described above constitute an inventive point of this disclosure, in conjunction with the steps below. This addresses the technical problem of "low system resource utilization and waste of experimental samples. Based on a fixed template without considering the actual situation of specific patient groups, clinical trial protocols violate the true biological characteristics of patients, resulting in an excessively high error rate in clinical trial results. This leads to the system allocating a large amount of computing resources to process invalid information, further reducing system resource utilization. Additionally, the mismatch between the clinical trial protocol and the true biological characteristics of patients results in structural defects in the medical device control instruction set (such as improper setting of the rotation angle parameters of the robotic arm), causing sample contamination when the medical robot performs clinical trial operations, thus leading to waste of experimental samples." The reasons for low system resource utilization and waste of experimental samples are: low system resource utilization and waste of experimental samples. Based on fixed templates without considering the actual circumstances of specific patient groups, clinical trial protocols often deviate from the patients' true biological characteristics, leading to excessively high error rates in clinical trial results. This forces the system to allocate significant computational resources to processing invalid information, resulting in reduced system resource utilization. Furthermore, the mismatch between the clinical trial protocol and the patients' true biological characteristics can cause structural defects in the medical device control instruction set (such as improper setting of the robotic arm's rotation angle parameters), causing sample contamination during the execution of clinical trial operations by the medical robot, thus wasting experimental samples. Addressing these factors would resolve the issues of low system resource utilization and wasted experimental samples. To achieve this, the first step involves extracting specific features from the acquired patient medical records, based on the aforementioned textual information representation interventional research and the clinical trial protocol, to obtain a patient-specific vector set. This allows multi-source data to be converted into a unified vector representation. The second step involves cluster analysis of the patient-specific vector set to obtain a set of potential molecular subtypes. This identifies molecular subtypes that are ignored by traditional classification methods but are relevant to treatment response. The third step involves pathway enrichment analysis of the potential molecular subtype set to obtain a set of candidate biomarkers. This yields candidate biomarkers with clear biological significance. The fourth step involves generating a biomarker score set based on the aforementioned candidate biomarker set. This ensures that the selected biomarkers are highly relevant to clinical trials. The fifth step involves prioritizing the candidate biomarkers based on their score set to obtain biomarker sequences. This priority order ensures the selection of the most accurate information. The sixth step involves determining the optimal biomarker combination based on the determined optimal thresholds for the biomarker sequences. This again ensures the selection of the most accurate information. The seventh step involves generating a patient treatment decision framework based on the optimal biomarker combination. This transforms biomarkers into a decision framework that includes treatment strategies.The eighth step involves generating a clinical trial protocol based on the aforementioned patient treatment decision-making framework, and then sending this protocol to the medical robot to execute the clinical trial. Ultimately, this ensures that the generated clinical trial protocol is closely related to the patient's condition, improving system resource utilization and reducing waste of trial samples.

[0067] Optionally, the above-mentioned medical device control system also includes a client 104, a working server 105, and a verification server 106.

[0068] The aforementioned client can be a front-end interface for user interaction with the system, providing guided forms and real-time feedback. The aforementioned verification server can be a compliance server integrating a multimodal verification engine. The aforementioned workflow server can be a workflow orchestration server.

[0069] The aforementioned client 104 is used to perform the following steps:

[0070] The first step involves receiving user-input text information, literature request information, and query information based on a pre-built question-and-answer text library. In practice, the client can receive these user-input text information, literature request information, and query information through a user interface. This user interface can be an interactive interface comprised of the pre-built question-and-answer text library, and may include, but is not limited to, interactive web pages and applications. The pre-built question-and-answer text library can be a text library that provides input guidance for the user, such as: "Is this study an interventional study or an observational study?". (See reference...) Figure 3 The diagram shows the input information interface of some embodiments of a medical device control system based on a clinical trial protocol according to this disclosure.

[0071] The second step is to send the aforementioned text information, document request information, and query information to the aforementioned working server. In practice, the client can send the aforementioned text information, document request information, and query information to the aforementioned working server through a preset internal communication interface. This preset internal communication interface can be a gRPC communication interface.

[0072] The aforementioned working server 105 is used to perform the following steps:

[0073] Step one: In response to receiving the aforementioned text information and document request information, the document request information is sent to the document abstract extraction server. In practice, the server can send the document request information to the abstract extraction server through a preset sending interface. This preset sending interface can be a RESTful API.

[0074] Step two: In response to receiving the above query information, send the query information to the data analysis server.

[0075] The aforementioned verification server 106 is used to generate review results and revision suggestions based on the aforementioned clinical trial protocol, and to send the review results, revision suggestions, and clinical trial protocol to the client.

[0076] In some embodiments, the verification server is used to generate review results and revision suggestions based on the clinical trial protocol, and send the review results, revision suggestions, and clinical trial protocol to the client.

[0077] In practice, the aforementioned validation server can generate review results and revision suggestions through the following steps, and then send the review results, revision suggestions, and clinical trial protocol to the client:

[0078] Step one: Based on a pre-built protocol rule base, the aforementioned clinical trial protocol is validated using risk rules to obtain risk labeling information. In practice, the validation server can use the pre-built protocol rule base to verify that the clinical trial protocol contains the necessary sections corresponding to the risk level (e.g., high-risk studies must have a data monitoring committee section), check whether the version number and date format comply with regulations, and verify whether the informed consent form contains all necessary informational elements, in order to generate risk labeling information. This risk labeling information can include the issue type (e.g., "error"), the location of the violation (e.g., section number), and the severity level (e.g., "high"). The pre-built protocol rule base can be predefined, consisting of instruction rules used to validate the compliance of clinical trial protocols.

[0079] Step two involves cross-validating the entity relationships in the aforementioned clinical trial protocol based on the drug logic knowledge graph to obtain logical labeling information. In practice, the validation server can extract a key entity set from the clinical trial protocol and cross-validate the entity relationships based on this key entity set to obtain logical labeling information. For example, the validation server can verify whether the primary endpoint defined in the "Study Objective" is consistent with the assessment time point in the "Study Flowchart," and check whether the drug dosage in the "Treatment Protocol" conflicts with the dosage adjustment criteria in the "Safety Evaluation." The logical labeling information can include inconsistent entity pairs, problem descriptions, and severity levels. The drug logic knowledge graph can be a knowledge graph stored and represented in a graph structure, and may include, but is not limited to, drugs, diseases, indications, examination indicators, related clinical procedures, treatment relationships, and assessment relationships.

[0080] Step 3: Based on a pre-defined set of principles and instructions, perform a semantic review of the aforementioned clinical trial protocol to obtain review marker information and revision suggestions. In practice, the aforementioned validation server can perform a semantic review of the aforementioned clinical trial protocol based on a pre-defined set of principles and instructions to obtain review marker information and revision suggestions. The pre-defined set of principles and instructions can be a set of instructions that includes both negative and positive instructions. For example, a negative instruction could be "Do not make validity assertions beyond what is supported by the references," and a positive instruction could be "Risk descriptions must be comprehensive and objective." The review marker information can include, but is not limited to, the issue type, issue location, and severity level. The revision suggestions can include revised text snippets and suggested explanations.

[0081] Step four involves multimodal information fusion of the aforementioned risk labeling information, logical labeling information, and review labeling information to generate a review result. In practice, the verification server can uniformly map the text representing the severity of the problem (such as "high," "medium," and "low") in the aforementioned risk labeling information, logical labeling information, and review labeling information to a predefined grading system (such as "error," "warning," and "recommendation"), obtaining a mapping information set. This mapping information set is then converted into a unified structured data format (such as JSON), resulting in a transformed mapping information set. The verification server can then sort the transformed mapping information set using predefined priority rules and deduplicate or conflicting labeling information pointing to the same location or semantic entity in the aforementioned clinical trial protocol (e.g., retaining only the label with the highest severity level), thus obtaining the review result. The predefined priority rules can be that "error" is higher than "warning," and "warning" is higher than "recommendation." The review result can be a document containing a summary list of problems, which may include the specific content, location, severity level, and possible source of labels (risk, logic, or semantics) of the problems in the aforementioned clinical trial protocol.

[0082] Step 5: Send the review results, revision suggestions, and clinical trial protocol to the client for visualization. In practice, the verification server can send the review results, revision suggestions, and clinical trial protocol to the client via the preset sending interface. This preset sending interface can be a RESTful API. See reference [link / reference]. Figure 4 The interface diagram is generated according to some embodiments of the medical device control system based on the clinical trial protocol disclosed herein.

[0083] Optionally, the above-mentioned medical device control system also includes a medical device control server 107.

[0084] In some embodiments, the medical device control server can be used to generate medical device control instructions in response to receiving the clinical trial protocol. These medical device control instructions may include, but are not limited to, instructions for collecting samples, instructions for preparing reagents, and instructions for operating the instrument. The medical device control server can be an Industrial Internet of Things (IIoT) gateway server, which can be used to control medical robots to perform related medical operations.

[0085] In addressing the technical problems mentioned above, and considering the application scenario—rare diseases with extremely small patient populations and precious experimental samples—controlling medical robots to perform complex medical experimental procedures (such as precise injection of targeted gene therapy drugs) according to clinical trial protocols often presents the following technical challenges: directly generating device control commands to control the medical robot to perform medical experimental procedures can lead to erroneous experimental results due to timing conflicts or mechanical failures during actual execution, resulting in wasted experimental samples. Given the following requirements for this application scenario—limited experimental samples and a high success rate—we have decided to adopt the following solution:

[0086] In some optional implementations of certain embodiments, the aforementioned executing entity may generate medical device control instructions through the following steps:

[0087] Step one involves performing joint semantic understanding on the aforementioned clinical trial protocol to generate a set of key medical entities and a description of the trial process. In practice, the medical device control server can use named entity recognition technology to extract key elements from the clinical trial protocol, including drug name, dosage, biological sample type, and detection indicators, to generate a set of key medical entities. Then, through relation extraction and semantic role labeling technology, it generates a description of the trial process. This description of the trial process can be a description of the trial steps, including but not limited to the steps themselves, their sequence, and the content of the experiment.

[0088] Step two involves generating a set of medical device operating parameters based on the aforementioned set of key medical entities and the medical device knowledge graph. In practice, the medical device control server can map each entity in the key medical entity set to the medical device knowledge graph, converting non-standard terms into standard medical concepts (e.g., mapping "myocardial infarction" to "heart attack"). It also converts standardized medical concepts into specific, executable device parameters (e.g., converting drug dosage and administration route into infusion pump flow rate and volume parameters), thereby generating the set of medical device operating parameters. The aforementioned medical device knowledge graph can be a knowledge base that stores medical concepts and their relationships in a graph structure, containing a knowledge graph of standard medical terminology and device specifications.

[0089] Step three: Based on the aforementioned medical device operation parameter set and the aforementioned test procedure description information, an initial medical device control instruction set is generated. In practice, the aforementioned medical device control server can convert the aforementioned medical device operation parameter set into an operation sequence based on the aforementioned test procedure description information, which serves as the initial medical device control instruction set. This initial medical device control instruction set can be a set of instructions that sequentially control the medical device to perform related operations.

[0090] Step four: Based on the aforementioned clinical trial protocol, construct a digital twin model of the target medical robot. In practice, the aforementioned medical device control server can extract key operational information from the aforementioned medical experimental protocol to construct a digital twin model of the target medical robot. This key operational information can be information related to the operation of the medical robot, and may include, but is not limited to, the type of the target robot (e.g., robotic arm, infusion pump), kinematic parameters (e.g., rotation angle), and experimental procedure information. The aforementioned digital twin model of the target medical robot can be a digital copy that completely corresponds to the real medical robot.

[0091] Step 5: Based on the initial medical device control instruction set, simulate the target medical robot digital twin model to generate a simulated performance index set and an error event set. In practice, the medical device control server can input the initial medical device control instruction set into the target medical robot digital twin model to drive it to execute complete experimental instructions, thereby obtaining the simulated performance index set and the error event set. The simulated performance instruction set can be a set of instructions composed of various instruction data from the target medical robot digital twin model during operation, including but not limited to execution time, robotic arm trajectory, and robotic arm rotation angle. The error event instruction set can be a set of instructions indicating error states encountered by the target medical robot digital twin model during operation. Error event instructions in the error event instruction set can include, but are not limited to, instruction timeouts and parameter anomalies.

[0092] Step Six: In response to the aforementioned error event set being non-empty, a command risk analysis report is generated based on the aforementioned simulated performance index set and the aforementioned error event set. In practice, the aforementioned medical device control server can use a causal discovery algorithm to analyze the aforementioned simulated performance index set and the aforementioned error event set to generate a command risk analysis report. And in response to the aforementioned error event set being empty, the aforementioned initial medical device control command set is determined as the medical device control command set, and Step Eight is executed.

[0093] Step 7: Based on the aforementioned instruction risk analysis report, the initial medical device control instruction set is strengthened to generate a new medical device control instruction set. In practice, the medical device control server can input the instruction risk analysis report into a Multi-Agent Reinforcement Learning (MARL) environment to strengthen the initial medical device control instruction set and generate a new medical device control instruction set. Then, the instruction simulation is performed again on the medical device control instruction set to obtain a set of simulation performance indicators and a set of error events. In response to the error event set being non-empty, the medical device control instruction set is determined as the initial medical device control instruction set, and steps 5-7 are executed.

[0094] Step 8: Based on the aforementioned medical device control instructions, control the associated medical robot to perform clinical trial operations. In practice, the aforementioned medical device server can send the aforementioned medical device control instructions to the associated medical robot through a preset communication protocol to control the medical robot to perform clinical trial operations. The preset communication protocol can be the OCP UA protocol based on the IEEE 11073 standard. The aforementioned clinical trial operations may include, but are not limited to, drug preparation operations and biological cell sample culture operations. The aforementioned medical robot can be an automated or semi-automated device used for clinical trials, and may include, but is not limited to, reagent modulation robots, drug preparation robots, and sample processing robots (which may be robots performing operations such as transfer, dispensing, and labeling of biological samples (such as blood and tissue sections)).

[0095] Steps one through eight and their related content, as an inventive point of this disclosure, solve the technical problem that "controlling related medical robots to perform medical trial operations through directly generated device control instructions can lead to incorrect test results due to timing conflicts or mechanical failures during actual execution, resulting in wasted test samples." The reason for this waste is that controlling related medical robots to perform medical trial operations through directly generated device control instructions can lead to incorrect test results due to timing conflicts or mechanical failures during actual execution, thus wasting test samples. Solving this problem solves the issue of wasted test samples. To achieve this, the first step involves, in response to receiving the aforementioned clinical trial protocol, performing joint semantic understanding on the protocol to generate a set of key medical entities and a trial process description. This allows the clinical trial protocol to be converted into machine-readable device control information. The second step involves generating a set of medical device operation parameters based on the aforementioned set of key medical entities and the medical device knowledge graph. This ensures the correctness of the instructions through the constraints of the medical device knowledge graph. The third step involves generating an initial set of medical device control instructions based on the aforementioned set of medical device operation parameters and the aforementioned trial process description information. Therefore, parameters can be converted into executable instructions. The fourth step involves constructing a digital twin model of the target medical robot based on the aforementioned clinical trial protocol. This provides a foundation for subsequent testing. The fifth step involves simulating the digital twin model of the target medical robot based on the initial medical device control instruction set, generating a set of simulated performance indicators and error events. This allows for the verification of the original operating instructions. The sixth step, in response to the error event set being non-empty, generates an instruction risk analysis report based on the simulated performance indicator set and the error event set. This provides a basis for adjusting the initial medical device control instruction set. The seventh step involves strengthening the initial medical device control instruction set based on the instruction risk analysis report, generating a new medical device control instruction set. This generates a robust medical device control instruction set with a low error rate. The eighth step involves controlling the associated medical robot to perform clinical trial operations based on the medical device control instruction set. This drives the medical robot to complete the trial operations, ultimately reducing the waste of experimental samples.

[0096] The various embodiments of this disclosure have the following beneficial effects: they can reduce system response time and waste of experimental samples. Specifically, the reasons for long system response time and waste of experimental samples are as follows: long system response time leads to waste of experimental samples. Static protocol templates cannot adapt to different research types or risk levels. When processing large-scale data, the system has poor scalability, resulting in excessively long system response time. Furthermore, the lack of real-time perception of the multimodal intervention needs of patients with complex comorbidities prevents dynamic optimization of device control logic based on patient characteristics, leading to a reduced success rate of medical robot operation and waste of experimental samples. Based on this, some embodiments of this disclosure include a medical device control system based on clinical trial protocols, comprising a protocol generation server, a literature abstract extraction server, a data analysis server, and a medical robot that are interconnected. First, the literature abstract extraction server performs the following steps: in response to receiving text information and literature request information, it generates a query argument tree based on an information retrieval model, the text information, and the literature request information; and performs a pyramid-perception retrieval based on the query argument tree to obtain a retrieval information result set. This generates a candidate literature set with the highest relevance to the literature request information, avoiding the processing of irrelevant literature. Based on the aforementioned retrieval results set, an ideal answer vector set is generated. This converts textual information into vector representation, improving information reliability. Using a re-ranking model and a pre-built literature database, the ideal answer vector set is sorted to generate a retrieval result set. This again converts textual information into vector representation, improving information reliability. The retrieval result set is then merged and normalized to obtain a literature abstract. This generates a concise literature abstract, reducing the amount of data transmitted to downstream servers. The data analysis server then performs instruction conversion on the query information to obtain query instructions, and executes these instructions based on a medical information knowledge graph to obtain information query results. This converts fuzzy queries into precise queries associated with the knowledge graph, avoiding the complexity of semantic parsing. Finally, the scheme generation server generates a clinical trial protocol based on the aforementioned textual information, literature abstract, and information query results, and sends the clinical trial protocol to a connected medical robot for execution. This provides a compliant clinical trial protocol and control instructions, enabling the medical robot to perform accurate operations. Ultimately, this reduces system response time and waste of experimental samples.

[0097] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A medical device control system based on a clinical trial protocol, wherein, The medical device control system includes a scheme generation server, a document abstract extraction server, a data analysis server, and a medical robot that are interconnected, wherein: The document abstract extraction server is used to perform the following steps: In response to receiving text information and document request information, a query argument tree is generated based on the information retrieval model, the text information, and the document request information, and a pyramid-perception retrieval is performed based on the query argument tree to obtain a set of retrieval information results. Based on the retrieved information result set, generate an ideal answer vector set; Based on the reordering model and a pre-built literature database, the ideal answer vector set is sorted and queried to generate a group of search results; The search results are merged and normalized to obtain a document abstract. The data analysis server is used to convert query information into instructions to obtain query instructions, and to execute the query instructions based on the medical information knowledge graph to obtain information query results; The protocol generation server is used to generate a clinical trial protocol based on the text information, the literature abstract, and the information query results, and to send the clinical trial protocol to a medical robot connected to the communication network to perform the clinical trial operation. The protocol generation server generates the clinical trial protocol through the following steps: In response to the observational study of the text information representation, the text information is preprocessed based on the medical big language model to obtain preprocessed text information; The preprocessed text information is semantically normalized to obtain normalized text information: Based on the sequence labeling model and the information query results, key parameter information of the standardized text information is extracted to obtain a key parameter information set. Based on the key parameter information set and risk rule base, risk level information is generated; Based on the risk level information and the risk template library, generate a solution template; Extract the core entity information from the text information to obtain entity information; Based on the protocol template, the entity information, and the literature abstract, a clinical trial protocol is generated, and the clinical trial protocol is sent to the medical robot to perform the clinical trial operation.

2. The system according to claim 1, wherein, The clinical trial protocol includes medical sample collection indicators and medical device operating parameters. The medical device control system also includes a medical device control server. The medical device control server is configured as follows: In response to receiving the clinical trial protocol, a set of medical device control instructions is generated, wherein the medical device control instructions include instructions for collecting samples, instructions for preparing reagents, and instructions for operating instruments; The medical robot is controlled to perform clinical trial operations according to the medical device control instruction set.

3. The system according to claim 1, wherein, The medical device control system also includes a working server, a client, and a verification server, and The client is configured as follows: Based on a pre-built question-and-answer text library, it receives text information, document request information, and query information input by users; The text information, the document request information, and the query information are sent to the working server; The working server is configured as follows: In response to receiving the text information and the document request information, the document request information is sent to the document abstract extraction server; In response to receiving the query information, the query information is sent to the data analysis server; The verification server is configured as follows: Based on the clinical trial protocol, review results and revision suggestions are generated, and the review results, revision suggestions, and the clinical trial protocol are sent to the client.

4. The system according to claim 3, wherein, The verification server is further configured as follows: Based on a pre-built protocol rule base, the clinical trial protocol is subjected to risk rule verification to obtain risk labeling information; Based on the drug logic knowledge graph, the entity relationships of the clinical trial protocol are cross-validated to obtain logical labeling information; Based on a pre-defined set of principle instructions, the clinical trial protocol is semantically reviewed to obtain review marker information and revision suggestion information. The risk marking information, the logical marking information, and the review marking information are fused using multimodal information to generate a review result; The review results, the revision recommendations, and the clinical trial protocol are sent to the client for visualization.

5. The system according to claim 1, wherein, The document abstract extraction server is configured as follows: Determine the similarity between the ideal answer vector set and each document vector in the document vector database to generate a descending similarity sequence; An initial document set is generated based on the preset number of documents in the document pool and the similarity sequence. Determine the similarity between each initial document in the initial document set and the document request information to generate document similarity groups; The initial document corresponding to at least one document similarity that meets the preset relevance value condition in the document similarity group is determined as the final document, and the final document group is obtained. Based on the document request information, the final document group is fused from multiple sources to generate a document abstract.

6. The system according to claim 1, wherein, The data analysis server is configured as follows: Extract the key medical concept set and logical relationships from the query information; Based on the medical information knowledge graph, each key medical concept in the key medical concept set is associated and mapped to obtain a standard medical terminology set; The standard medical terminology set and the logical relationships are conceptually reconstructed to obtain query instructions; Executing the query command yields multiple target query results; The results of the multiple target queries are aggregated to obtain the query result condition values; Based on the query result condition values ​​and the medical rule knowledge base, information query results are generated.

7. The system according to claim 1, wherein, The document abstract extraction server is configured as follows: Based on the information retrieval model, the text information and document request information are parsed to extract semantic intent; Construct a query argument tree based on the semantic intent; Perform pyramid-sensing retrieval on the query argument tree to obtain a preliminary document set; The preliminary literature set and the query argument tree are fused together using a graph structure to obtain a set of search results.

Images