A medical intelligent diagnosis and treatment system

Abstract

The application provides a medical intelligent diagnosis and treatment system, obtains the multi-source medical data corresponding to a target object through a data acquisition module, and obtains a clinical data object; splits a single diagnosis and treatment task into multiple step nodes through a behavior-driven execution module, generates an execution plan according to a preset flowchart, generates a key event set according to the execution plan, calls a rule calculation module, infers the state of the target object through a model inference module, obtains an inference result, fuses the key event set and the inference result through a result fusion module, and converts the key event set and the inference result to obtain a structured clinical conclusion; finally, the structured clinical conclusion is converted into a draft document through a document generation and proofreading module, and the draft document is proofread; the problems of a large number of multi-source clinical events, inconsistent evidence quality, causal coupling and time dependence between events in the high-frequency monitoring scene of critical medicine are solved.

Description

Technical Field

[0001] This invention relates to the field of clinical decision support technology, and in particular to a medical intelligent diagnosis and treatment system. Background Technology

[0002] Against the backdrop of the global digital wave, the deep integration of artificial intelligence (AI) technology with the medical field has given rise to a new type of medical assistance tool—the AI-powered medical diagnostic system. This system utilizes AI technologies, including machine learning, deep learning, and natural language processing, to deeply mine and analyze medical data, and based on this, provides intelligent medical tools for disease prediction, assisted diagnosis, and treatment plan recommendations. The system can simulate the thought process of a human doctor, learning from massive amounts of case data and medical knowledge bases to achieve accurate identification and judgment of symptoms. Its aim is to improve the quality and efficiency of medical services, reduce the workload of medical staff, optimize the allocation of medical resources, and ultimately benefit a wide range of patients.

[0003] However, the development of intelligent medical diagnosis and treatment systems also faces many challenges: 1. In terms of data security and management, the traditional centralized data storage model has significant security risks. Multimodal physiological data contains a large amount of patient privacy information. Centralized storage is vulnerable to hacker attacks, data tampering or leakage. At the same time, the data formats of different medical service institutions are not uniform, forming "data silos" and making it difficult to achieve cross-institutional data sharing and collaborative diagnosis and treatment. 2. At the level of data processing and feature extraction, the temporal dynamics, spatial correlations and semantic complexity of cross-modal data pose a huge challenge to feature fusion. Existing methods often fail to effectively explore the intrinsic correlations between different types of data, resulting in the lack of standardization and comprehensiveness of the extracted features, which cannot provide reliable support for diagnosis and treatment decisions. 3. In terms of treatment decision-making mechanisms, traditional treatment models often employ fixed-structure decision logic, making it difficult to dynamically adjust decision paths based on changes in patients' real-time physiological data. This makes them unable to adapt to the dynamic evolution of the disease, easily leading to delays or limitations in treatment plans. Furthermore, the effectiveness verification of treatment plans lacks a unified and authoritative mechanism, and different medical entities have varying levels of acceptance of treatment plans, which may lead to non-standard treatment practices. At the same time, feedback data from patients after implementing treatment plans is not processed and applied in a timely and effective manner, failing to quickly trigger optimization and adjustments to the treatment plans, thus affecting the continuous improvement of treatment outcomes. Summary of the Invention

[0004] This invention provides a medical intelligent diagnosis and treatment system, which aims to solve the problems of large number of multi-source clinical events, inconsistent quality of evidence, and causal coupling and time dependence between events in high-frequency monitoring scenarios in critical care medicine.

[0005] To achieve the above objectives, the present invention provides a medical intelligent diagnosis and treatment system, comprising: The data acquisition module is used to acquire multi-source medical data corresponding to the target object from multiple devices to obtain clinical data objects, which include feature packages, event graphs and clinical summary objects. The behavior-driven execution module is used to break down a single diagnosis and treatment task into multiple steps and nodes, generate an execution plan based on a preset flowchart, and call the rule calculation module, model inference module, and result fusion module according to the execution plan. The rule calculation module is used to construct a candidate event pool for clinical data objects according to the execution plan, cluster the candidate events in the candidate event pool to obtain event clusters, calculate the priority of key events in each event cluster, and compare the adjusted event priorities with the key event selection threshold to generate a set of key events. The model reasoning module is used to reason about the state of the target object based on the clinical data object and the key event set input, according to the execution plan, and to obtain the reasoning result. The results fusion module is used to fuse and transform the set of key events, clinical summary objects, and inference results to obtain structured clinical conclusions. The document generation and proofreading module is used to convert structured clinical conclusions into document drafts and to proofread the document drafts.

[0006] Furthermore, the data acquisition module includes: The data acquisition unit is used to collect multi-source raw medical data of the target patient from the hospital information system and bedside equipment; The data preprocessing unit is used to preprocess multi-source raw medical data to obtain clinical data objects.

[0007] Furthermore, preprocessing of multi-source raw medical data yields clinical data objects, including: Identify and delete null values, abnormal duplicate values, and illegal characters in multi-source raw medical data to obtain the first multi-source medical data; Based on a predefined clinical data dictionary, the synonymous fields in the first multi-source medical data are standardized and mapped to obtain the first multi-source medical data after standardizing the field data. The unit conversion rules are used to unify the units in the first multi-source medical data after standardizing the field data, resulting in the second multi-source medical data; The monitoring data from the second multi-source medical data are aggregated according to a fixed time window to obtain the aggregated monitoring data. The detection data in the second multi-source medical data were processed using a near-value forward preservation and time decay weighting method to obtain the processed detection data; State intervals are constructed based on the start and end times of medical order data, operational data, and equipment operation data in the second multi-source medical data. The standardized field data, aggregated monitoring data, and state intervals are reconstructed into clinically usable events, forming an event graph as a clinical data object.

[0008] Furthermore, the behavior-driven execution module includes a process parsing unit and a scheduling unit; The process parsing unit is used to break down a single diagnosis and treatment task into multiple action nodes, and generate an execution plan based on each action node. The action nodes are connected by a directed graph. The scheduling unit is used to call the rule calculation module or model inference module to process each event in the event graph according to the type of each behavior node.

[0009] Furthermore, before constructing a candidate event pool for clinical data objects based on the execution plan, the rule calculation module also includes: Based on the execution plan, each indicator in the event graph is subjected to threshold judgment, and the threshold judgment result is obtained; Based on the threshold judgment results and the calculated short window change slope, the trend of each indicator in the time spectrum is judged to obtain the trend judgment result; The comprehensive anomaly score is determined based on the threshold judgment results and trend judgment results; Major events are identified based on the comprehensive anomaly score, resulting in a list of major events.

[0010] Furthermore, the expression for calculating the priority of key events in each event cluster is: ; in, Indicates the priority of important events. Indicates clinical severity. Indicates the level of urgency. Indicates the degree of deterioration in the trend. Indicates the degree of impact of treatment. Indicates the degree of relevance of incomplete tasks. Indicates the security level. Indicates the relevance of the role. Indicates the credibility of the event. Indicates penalty for repeated events. , , , , , , , , All represent weights.

[0011] Furthermore, the formula for calculating the threshold for selecting key events is as follows: ; in, This indicates the threshold for selecting key events. This indicates the initial threshold for selecting key events. , , , , , All represent weights. Indicates the patient's overall risk level. Indicates the degree of proximity to the shift handover time. This indicates whether life support is available. Indicates the proportion of conflicting evidence. Indicates the percentage of missing key data. This indicates items marked as of interest by clinical staff.

[0012] Furthermore, the event-coupled evidence gating fusion model includes a model registry, a multi-model adaptation unit, an event structure enhancement unit, an evidence gating unit, and a security constraint fusion unit. The multi-model adaptation unit calculates the model adaptation score of each candidate model based on the basic data of each candidate model in the model registry. The event structure enhancement unit transforms the time window features corresponding to the state of the target object into event structure enhancement vectors. Based on the model fit score, a target inference model is selected to infer the event structure enhancement vector and the inference result is obtained. The evidence gating unit and the security constraint fusion unit adjust the output weights of the target reasoning model based on the reasoning results.

[0013] Furthermore, the formula for calculating the model fit score is: ; in, This indicates the model fit score. , , , , , All represent weights. Indicates historical verification performance, Indicates the task matching degree. Indicates input integrity. Indicates the recentity of the data. Indicating model uncertainty, This indicates a security risk.

[0014] Furthermore, the process of converting structured clinical conclusions into draft documentation includes: Obtain a list of current issues, key evidence items, treatment measures, unfinished tasks, and risk warnings from the clinical summary subjects; Retrieve the corresponding document template based on the document type; Based on the slot content and preset sentence structure library in the document template, the current problem list, key evidence items, treatment measures, unfinished tasks and risk warnings are converted into a document draft.

[0015] The above-described solution of the present invention has the following beneficial effects: Compared with existing technologies, this invention acquires multi-source medical data corresponding to the target object from multiple devices through a data acquisition module to obtain a clinical data object. A behavior-driven execution module breaks down a single treatment task into multiple step nodes and generates an execution plan based on a preset flowchart. The execution plan is then used to call a rule calculation module to construct a candidate event pool for the clinical data object. Candidate events in the pool are clustered to obtain event clusters. The priority of key events in each event cluster is calculated, and after adjusting the event priorities, they are compared with a key event selection threshold to generate a set of key events. Finally, a model inference module is called, inputting the clinical data object and the set of key events. The constructed event-coupled evidence-gated fusion model infers the state of the target object, obtains the inference result, and calls the result fusion module to fuse and transform the key event set, clinical summary object, and inference result to obtain a structured clinical conclusion. Finally, the document generation and proofreading module converts the structured clinical conclusion into a document draft and proofreads the document draft. This transforms complex diagnosis and treatment tasks into a step-by-step, auditable, and intervention-friendly execution process, reduces the dependence on large-scale computing power, and solves the problems of large number of multi-source clinical events, inconsistent evidence quality, and causal coupling and time dependence between events in high-frequency monitoring scenarios in critical care medicine.

[0016] Other beneficial effects of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of an embodiment of the present invention. Detailed Implementation

[0018] To make the technical problems, solutions, and advantages of this invention clearer, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a locking connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] This invention addresses existing problems by providing a medical intelligent diagnosis and treatment system.

[0023] like Figure 1 As shown, an embodiment of the present invention provides a medical intelligent diagnosis and treatment system, comprising: The data acquisition module is used to acquire multi-source medical data corresponding to the target object from multiple devices to obtain clinical data objects, which include feature packages, event graphs and clinical summary objects. The behavior-driven execution module is used to break down a single diagnosis and treatment task into multiple steps and nodes, generate an execution plan based on a preset flowchart, and call the rule calculation module, model inference module, and result fusion module according to the execution plan. The rule calculation module is used to construct a candidate event pool for clinical data objects according to the execution plan, cluster the candidate events in the candidate event pool to obtain event clusters, calculate the priority of key events in each event cluster, and compare the adjusted event priorities with the key event selection threshold to generate a set of key events. The model reasoning module is used to reason about the state of the target object based on the clinical data object and the key event set input, according to the execution plan, and to obtain the reasoning result. The results fusion module is used to fuse and transform the set of key events, clinical summary objects, and inference results to obtain structured clinical conclusions. The document generation and proofreading module is used to convert structured clinical conclusions into document drafts and to proofread the document drafts.

[0024] Specifically, the data acquisition module includes: The data acquisition unit is used to collect multi-source raw medical data of the target patient from the hospital information system and bedside equipment; The data preprocessing unit is used to preprocess multi-source raw medical data to obtain clinical data objects.

[0025] In this embodiment of the invention, the hospital information system includes, but is not limited to, electronic medical record systems, laboratory systems, and nursing systems. Bedside equipment includes, but is not limited to, bedside monitoring equipment, ventilators, blood purification equipment, extracorporeal membrane oxygenation (ECMO) equipment, and other third-party medical equipment. In this embodiment of the invention, the data acquisition unit connects to the above systems in ways including, but not limited to, JDBC / ODBC data view reading, RESTful API calls, HL7 message interfaces, FHIR resource interfaces, medical digital imaging and communication query and feedback, Web-Socket real-time subscription, and file gateway methods. The connection method with the above devices adopts a two-layer structure of device driver adapter and unified acquisition agent. The device side can access the system through vendor software development kits, serial / network protocols, MQTT gateways, HTTP reporting, Mod-bus bus protocols, standardized industrial automation communication protocols, or middleware forwarding. Taking monitors, ventilators, and other devices as examples, the raw data frames are first obtained through the device adapter. The data acquisition unit establishes a device topic for each type of device and maps different vendor fields to a predefined parameter dictionary. For example, the PEEP, P-plat, and Vt fields in ventilators from different manufacturers will be uniformly mapped to the positive end-expiratory pressure field, plateau pressure field, and tidal volume field. If the device disconnects or data packets are lost, the agent will mark the data quality as degraded based on the heartbeat mechanism and write the disconnection event to the log.

[0026] In this embodiment of the invention, the multi-source raw medical data includes data source identifier, collection time, object identifier, bed identifier, and equipment identifier.

[0027] The specific process of preprocessing multi-source raw medical data to obtain clinical data objects in this embodiment of the invention includes: Identify and delete null values, abnormal duplicate values, and illegal characters in multi-source raw medical data to obtain the first multi-source medical data; Based on a predefined clinical data dictionary, synonymous fields in the first multi-source medical data are standardized and mapped, unifying synonymous fields to the predefined clinical data dictionary to obtain standardized field data for the first multi-source medical data; for example, "WBC", "white blood cell count", and "white blood cell count" are unified into the white blood cell count field, and "systolic blood pressure" and "SBP" are unified into the systolic blood pressure field. For fields with unit anomalies, the unit conversion rules are invoked to unify the units in the first multi-source medical data after standardizing the field data, resulting in the second multi-source medical data. The monitoring data from the second multi-source medical data are aggregated according to a fixed time window to obtain the aggregated monitoring data. The detection data in the second multi-source medical dataset are processed using a near-value forward preservation and time-decay weighting method to obtain processed detection data. For example, the effective weight of a certain test item at time t can be expressed as follows: ,in, This indicates the time interval since the most recent sample. This indicates the attenuation coefficient preset based on the project attributes; State intervals are constructed based on the start and end times of medical order data, operational data, and equipment operation data in the second multi-source medical data. The standardized field data, aggregated monitoring data, and state intervals are reconstructed into clinically usable events. For example, the medical order "continuous infusion of norepinephrine" is reconstructed into a vasoactive drug initiation event; the nursing record "prone position initiation" is reconstructed into a prone position initiation event; and the lactate test can be reconstructed into a lactate result event, forming an event graph as clinical data objects.

[0028] In this embodiment of the invention, a clinically usable event includes at least a start time, an end time, an execution source, an object of action, a status value, and an event credibility, wherein the expression for the event credibility is: ; in, Indicates the credibility of the event. Indicates the system's trust level. Indicates the timestamp integrity score. Indicates the field completeness score. , , All of these represent weighting coefficients.

[0029] To ensure accurate object identity association, this embodiment of the invention employs a triple matching mechanism—primary index + bed mapping + time window backfilling—before identifying and deleting null values, abnormal duplicate values, and illegal characters from multi-source original medical data. The specific process is as follows: First, perform primary matching based on medical record number, hospitalization number, or object identifier; If the object identifier is missing in the equipment data, a secondary mapping is performed by combining the bed number, equipment binding relationship, department information and collection time; If any doubts remain, add the item to the verification queue and prompt for manual confirmation.

[0030] Specifically, the behavior-driven execution module includes a process parsing unit and a scheduling unit; The process parsing unit is used to break down a single diagnosis and treatment task into multiple action nodes, and generate an execution plan based on each action node. The action nodes are connected by a directed graph. The scheduling unit is used to call the rule calculation module or model inference module to process each event in the event graph according to the type of each behavior node.

[0031] In this embodiment of the invention, the behavioral nodes include, but are not limited to, data observation nodes, rule judgment nodes, problem splitting nodes, evidence retrieval nodes, model reasoning nodes, and result integration nodes. Each node is connected by a directed graph, which can support sequential execution, conditional branch execution, and parallel execution of some nodes. The data structure of each node includes at least node number, node type, input format, rule set, prompt template, model identifier, threshold parameter, retry strategy, rollback action, and output format.

[0032] In this embodiment of the invention, the input to the behavior-driven execution module can be any item in the clinical data object and the diagnosis and treatment task, and its output can be a node-based execution plan, or a module call request, an exception rollback instruction, and a manual confirmation request; its specific working process is as follows: First, read the task type and task objective of the diagnosis and treatment task to determine the required set of nodes; Then, a directed graph is generated based on the input-output dependencies between nodes; The directed graph is then subjected to topological sorting, conditional branch parsing, parallel sympathetic identification, and security level verification to form an execution plan. The corresponding modules are scheduled one by one according to the execution plan. For example, when the node type is a data observation node, the clinical data object is called to determine whether the data required for the current task is complete. When the node type is a rule judgment node, the rule calculation module is called to perform threshold judgment, trend judgment, duration judgment, window count judgment, or event combination judgment. When the node type is a model inference node, the model inference module is called to select the corresponding model according to the task type and input completeness and generate inference results. When the node type is a result integration node, the result fusion module is called to fuse the results output by the rule calculation module and the inference results.

[0033] Specifically, before constructing a candidate event pool for clinical data objects based on the execution plan, the rule calculation module also includes: Based on the execution plan, each indicator in the event graph is subjected to threshold judgment, and the threshold judgment result is obtained; Based on the threshold judgment results and the calculated short window change slope, the trend of each indicator in the time spectrum is judged to obtain the trend judgment result; The comprehensive anomaly score is determined based on the threshold judgment results and trend judgment results; Major events are identified based on the comprehensive anomaly score, resulting in a list of major events.

[0034] In this embodiment of the invention, the threshold is determined as follows when performing threshold judgment on each indicator in the event graph: For any index First, a basic threshold is obtained by fusing guideline thresholds, hospital-specific rule thresholds, and historical data calibration thresholds. Its expression is: ; in, Indicates the guideline threshold. This indicates the threshold value for local rules in the hospital. This indicates the historical data calibration threshold. , , All of these indicate configurable weighting coefficients, which can be adjusted according to different departments, task types, and data quality statuses. The base threshold is then dynamically adjusted based on the individual status of the target object, data quality, and task scenario to obtain the final rule threshold. Its corrected expression is: ; in, This represents the individual patient baseline correction term, used to reflect the degree of deviation of the target patient's current indicators from their own baseline; This represents data quality correction items, used to reflect the impact of missing rate, conflict rate, time delay, and source credibility on threshold judgment; This indicates a task scenario modification item, used to reflect different requirements for rule sensitivity in shift handover, risk warning, disease progress summary, quality control review, or manual monitoring states.

[0035] Specifically, the formula for calculating the slope of the short window is: ; in, Indicates the slope of the short window change. Indicates the first in the index sequence One indicator, Indicates the first The time corresponding to each indicator.

[0036] In an embodiment of the present invention, when the first When an indicator exceeds the final rule threshold, the slope of the short window change exceeds the preset trend threshold, or the duration of the abnormality is not less than the preset duration, the indicator is marked as a trend abnormality; when the current value does not exceed the fixed threshold, but the degree of deviation from the individual baseline of the target formation reaches the preset proportion, and there is a trend of change in the same direction, it can be marked as an early change event or an event to be confirmed. If in continuous Within a window, satisfy and If it is, then it is judged as "continuously rising"; if a certain event In the time window Cumulative occurrence count If a certain indicator meets both threshold anomaly and trend anomaly criteria, its comprehensive anomaly score is calculated using the following formula: ; in, For indicator functions, This is the normalization function; For complex rules, such as "identify major events in the past 12 hours before generating a handover summary", the following conditions can be set: if a vasoactive drug initiation event occurs and lasts for more than 30 minutes, or the average MAP decreases by more than a preset proportion in the last 2 hours, or the lactate result event increases by more than a threshold in the last 6 hours, or the ventilator parameters enter the high-risk range, then these are considered major events, and a list of major events is obtained.

[0037] Specifically, the sources of candidate events in the candidate event pool include: First, physiological indicator events, such as decreased MAP, decreased SpO2, increased lactate, and decreased urine output; Second, treatment-related events, such as vasoactive drug initiation events, mechanical ventilation initiation events, continuous renal replacement therapy initiation events, prone position initiation events, and antibiotic initiation events; Thirdly, there are task status events, such as blood cultures not being sent for testing, lactate not being retested, and imaging examinations not being reported. Fourth, there are structured events in the text, such as semantic fragments in the medical record such as "shock worsened", "oxygenation deteriorated", and "plan to adjust vasopressor". Each candidate event is represented by a candidate event object, which includes at least the event number, event type, patient identifier, start time, end time, source citation, set of evidence items, original value, data quality marker, and associated task type.

[0038] In this embodiment of the invention, to avoid the same clinical event being repeatedly written into documents from multiple sources, event cluster merging is performed on candidate events in the candidate pool. If two candidate events meet the following conditions: the event type is the same or belongs to the same semantic category, and the time window overlap ratio is [not specified], then [the process is considered]. If the sources of evidence are shared or causally related, they are merged into the same event cluster. For multiple pieces of evidence within the same event cluster, the evidence with the latest time, the highest source credibility, and the highest field completeness is retained as the main evidence, and the remaining evidence is used as circumstantial evidence.

[0039] Specifically, the expression for calculating the priority of key events in each event cluster is: ; in, Indicates the priority of important events. Indicates clinical severity. Indicates the level of urgency. Indicates the degree of deterioration in the trend. Indicates the degree of impact of treatment. Indicates the degree of relevance of incomplete tasks. Indicates the security level. Indicates the relevance of the role. Indicates the credibility of the event. Indicates penalty for repeated events. , , , , , , , , All represent weights.

[0040] Before modifying the event priority, this embodiment of the invention first calculates the source system credibility, timestamp completeness, field completeness, patient identity binding credibility, cross-source consistency, data recency, and conflict penalty for each piece of evidence to obtain the credibility of that piece of evidence. The calculation expression is: ; in, Indicates the credibility of the source system. Indicates the integrity of the timestamp. Indicates the completeness of the field. This indicates the credibility of patient identity binding. Indicates cross-source consistency. Indicates the recentity of the data. Indicates conflict penalty items, , , , , , , All represent weights; Then, the credibility of the event is dynamically adjusted based on the credibility of each piece of evidence and their interrelationships. The expression is as follows: ; in, Indicates the credibility of the revised event. This indicates the number of independent pieces of evidence supporting the event cluster. This represents the cross-source consistency score. Indicates conflict of evidence as punishment. This indicates a penalty for exceeding the time limit. , , , All of these represent weighting coefficients.

[0041] Specifically, the formula for calculating the threshold for selecting key events is: ; in, This indicates the threshold for selecting key events. This indicates the initial threshold for selecting key events. , , , , , All represent weights. Indicates the patient's overall risk level. Indicates the degree of proximity to the shift handover time. This indicates whether life support is available. Indicates the proportion of conflicting evidence. Indicates the percentage of missing key data. This indicates items marked as of interest by clinical staff.

[0042] Specifically, the event-coupled evidence gating fusion model includes a model registry, a multi-model adaptation unit, an event structure enhancement unit, an evidence gating unit, and a security constraint fusion unit. The multi-model adaptation unit calculates the model adaptation score of each candidate model based on the basic data of each candidate model in the model registry. The event structure enhancement unit transforms the time window features corresponding to the state of the target object into event structure enhancement vectors. Based on the model fit score, a target inference model is selected to infer the event structure enhancement vector and the inference result is obtained. The evidence gating unit and the security constraint fusion unit adjust the output weights of the target reasoning model based on the reasoning results.

[0043] In this embodiment of the invention, the candidate models include traditional machine learning models, deep learning models, and reinforcement learning strategy models. The traditional machine learning models include logistic regression models, random forest models, extreme gradient boosting tree models, lightweight gradient boosting tree models, and support vector machine models, used to process feature packages with low missing rates and high structure. The deep learning models include temporal convolutional network models, gated recurrent unit network models, long short-term memory network models, and Transformer temporal attention models, used to process vital signs, test trends, and changes in device parameters within continuous time windows. The reinforcement learning strategy models are used at the node scheduling layer to select process actions such as continuing observation, review prompts, supplementing evidence, manual confirmation, or temporarily suspending output.

[0044] Specifically, the formula for calculating the model fit score is: ; in, This indicates the model fit score. , , , , , All represent weights. Indicates historical verification performance, Indicates the task matching degree. Indicates input integrity. Indicates the recentity of the data. Indicating model uncertainty, This indicates a security risk.

[0045] Specifically, the evidence gate unit dynamically adjusts the model's output weights based on evidence credibility and data quality, and its expression is as follows: ; in, This represents the model weights adjusted for the credibility of the evidence. Indicates cross-source consistency. Indicates recentity. Indicates the conflict rate of evidence. Indicates the missing rate of key data. , , , , All represent weights.

[0046] Specifically, this involves converting structured clinical conclusions into draft documentation, including: Obtain a list of current issues, key evidence items, treatment measures, unfinished tasks, and risk warnings from the clinical summary subjects; The corresponding document template is called according to the document type. The document template defines at least fixed paragraphs, optional paragraphs, mandatory positions, and prohibited omissions. Based on the slot content and preset sentence structure library in the document template, the current problem list, key evidence items, treatment measures, unfinished tasks and risk warnings are converted into a document draft.

[0047] Specifically, proofreading the draft document includes: The draft document is compared against real-time monitoring data, the latest medical orders, and test results item by item to determine its consistency, and the final usable result is output. This consistency comparison includes field-level and semantic-level verification. Field-level verification mainly checks whether numerical values, units, times, drug names, and treatment measures are consistent with the original evidence. Semantic-level verification uses predefined semantic rules to check for statements in the text that contradict the current reality. For example, if the document states "no vasoactive drugs were used," but a vasoactive drug activation event exists in the event table and has not yet ended, the system determines this as a semantic conflict. The verification results are preferably output as conflict item objects, each including at least the conflict type, source of evidence, conflict fragment, and suggested revision method. Only when all mandatory slots are filled and no high-level conflicts exist will the result or draft document enter the "submitable" state.

[0048] Specifically, the embodiments of the present invention also include a visualized full-process review and tracking module, which is used to transform a complete diagnosis and treatment task into a traceable, explainable, and auditable execution link, so that clinicians can clearly see how a certain conclusion, a certain document, or a certain handover result was generated, which rules and models were called in the process, which key evidence was used, whether manual revisions were made, and why the current output was ultimately formed.

[0049] Compared with existing technologies, this invention obtains multi-source medical data corresponding to a target object from multiple devices through a data acquisition module, thus obtaining a clinical data object. A behavior-driven execution module breaks down a single treatment task into multiple steps and generates an execution plan based on a preset flowchart. The execution plan is then used to call a rule calculation module to construct a candidate event pool for the clinical data object. Candidate events in the pool are clustered to obtain event clusters. The priority of key events in each event cluster is calculated, and after adjusting the event priorities, they are compared with a key event selection threshold to generate a set of key events. Finally, a model inference module is invoked to integrate the clinical data object and the set of key events. The input event-coupled evidence-gated fusion model infers the state of the target object and obtains the inference result. The result fusion module is called to fuse and transform the key event set, clinical summary object, and inference result to obtain a structured clinical conclusion. Finally, the document generation and proofreading module converts the structured clinical conclusion into a document draft and proofreads the document draft. This transforms complex diagnosis and treatment tasks into a step-by-step, auditable, and intervention-friendly execution process, reducing the dependence on large-scale computing power and solving the problems of large number of multi-source clinical events, inconsistent evidence quality, and causal coupling and time dependence between events in high-frequency monitoring scenarios in critical care medicine.

[0050] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A medical intelligent diagnosis and treatment system, characterized in that, include: The data acquisition module is used to acquire multi-source medical data corresponding to the target object from multiple devices to obtain a clinical data object, which includes a feature package, an event graph, and a clinical summary object. The behavior-driven execution module is used to break down a single diagnosis and treatment task into multiple step nodes, generate an execution plan based on a preset flowchart, and call the rule calculation module, model reasoning module, and result fusion module according to the execution plan. The rule calculation module is used to construct a candidate event pool for the clinical data object according to the execution plan, cluster the candidate events in the candidate event pool to obtain event clusters, calculate the priority of key events in each event cluster, and compare the adjusted event priorities with the key event selection threshold to generate a set of key events. The model reasoning module is used to reason about the state of the target object by inputting the clinical data object and the key event set into the event-coupled evidence gating fusion model according to the execution plan, and to obtain the reasoning result. The result fusion module is used to fuse and transform the set of key events, the clinical summary object, and the reasoning result to obtain a structured clinical conclusion. The document generation and proofreading module is used to convert the structured clinical conclusions into a document draft and to proofread the document.

2. The medical intelligent diagnosis and treatment system according to claim 1, characterized in that, The data acquisition module includes: The data acquisition unit is used to collect multi-source raw medical data of the target patient from the hospital information system and bedside equipment; The data preprocessing unit is used to preprocess the multi-source raw medical data to obtain clinical data objects.

3. The medical intelligent diagnosis and treatment system according to claim 2, characterized in that, The multi-source raw medical data is preprocessed to obtain clinical data objects, including: Identify and delete null values, abnormal duplicate values, and illegal characters in the multi-source original medical data to obtain the first multi-source medical data; Based on a predefined clinical data dictionary, the synonymous fields in the first multi-source medical data are standardized and mapped to obtain the first multi-source medical data after standardizing the field data. The unit conversion rules are used to unify the units in the first multi-source medical data after standardizing the field data, resulting in the second multi-source medical data; The monitoring data from the second multi-source medical data are aggregated according to a fixed time window to obtain the aggregated monitoring data; The detection data in the second multi-source medical data is processed by using a near-value forward preservation and time decay weighting method to obtain the processed detection data; A state interval is constructed based on the start and end times of the medical order data, operation data, and equipment operation data in the second multi-source medical data. The standardized field data, aggregated monitoring data, and state intervals are reconstructed into clinically usable events, forming an event graph as a clinical data object.

4. The medical intelligent diagnosis and treatment system according to claim 3, characterized in that, The behavior-driven execution module includes a process parsing unit and a scheduling unit; The process parsing unit is used to parse and decompose the single diagnosis and treatment task to obtain multiple behavior nodes, and generate an execution plan based on each behavior node. The behavior nodes are connected by a directed graph. The scheduling unit is used to call the rule calculation module or the model inference module to process each event in the event graph according to the type of each behavior node.

5. The medical intelligent diagnosis and treatment system according to claim 4, characterized in that, Before constructing a candidate event pool for the clinical data objects according to the execution plan, the rule calculation module further includes: Based on the execution plan, a threshold judgment is performed on each indicator in the event graph to obtain the threshold judgment result; Based on the threshold judgment result and the calculated short window change slope, a trend judgment is made for each indicator in the time spectrum to obtain the trend judgment result; A comprehensive anomaly score is determined based on the threshold judgment result and the trend judgment result; Major events are identified based on the comprehensive anomaly score, resulting in a list of major events.

6. The medical intelligent diagnosis and treatment system according to claim 5, characterized in that, The expression for calculating the priority of the most important event in each event cluster is: ； in, Indicates the priority of important events. Indicates clinical severity. Indicates the level of urgency. Indicates the degree of deterioration in the trend. Indicates the degree of impact of treatment. Indicates the degree of relevance of incomplete tasks. Indicates the security level. Indicates the relevance of the role. Indicates the credibility of the event. Indicates penalty for repeated events. , , , , , , , , All represent weights.

7. The medical intelligent diagnosis and treatment system according to claim 1, characterized in that, The calculation expression for the key event selection threshold is as follows: ； in, This indicates the threshold for selecting key events. This indicates the initial threshold for selecting key events. , , , , , All represent weights. Indicates the patient's overall risk level. Indicates the degree of proximity to the shift handover time. This indicates whether life support is available. Indicates the proportion of conflicting evidence. Indicates the percentage of missing key data. This indicates items marked as of interest by clinical staff.

8. The medical intelligent diagnosis and treatment system according to claim 1, characterized in that, The event-coupled evidence gating fusion model includes a model registry, a multi-model adaptation unit, an event structure enhancement unit, an evidence gating unit, and a security constraint fusion unit. The multi-model adaptation unit calculates the model adaptation score of each candidate model based on the basic data of each candidate model in the model registry. The event structure enhancement unit transforms the time window features corresponding to the state of the target object into an event structure enhancement vector. Based on the model adaptation score, a target inference model is selected to infer the event structure enhancement vector, and the inference result is obtained. The evidence gating unit and the security constraint fusion unit adjust the output weights of the target reasoning model based on the reasoning results.

9. The medical intelligent diagnosis and treatment system according to claim 8, characterized in that, The formula for calculating the model fit score is as follows: ； in, This indicates the model fit score. , , , , , All represent weights. Indicates historical verification performance, Indicates the task matching degree. Indicates input integrity. Indicates the recentity of the data. Indicating model uncertainty, This indicates a security risk.

10. The medical intelligent diagnosis and treatment system according to claim 1, characterized in that, The structured clinical conclusions were converted into a draft document, including: Obtain the current problem list, key evidence items, treatment measures, unfinished tasks, and risk warnings from the clinical summary object; Retrieve the corresponding document template based on the document type; Based on the slot content and preset sentence structure library in the document template, the current problem list, key evidence items, treatment measures, unfinished tasks and risk warnings are converted into a document draft.

Images