Evidence boundary constraint-based critical illness diagnosis and treatment intelligent agent assisted decision-making method and system

CN123033704APending Publication Date: 2026-11-03XIANGYA HOSPITAL CENT SOUTH UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611531358.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-30
Publication Date
2026-11-03

AI Technical Summary

Technical Problem

Existing medical AI systems struggle to maintain consistency and security in the diagnosis and treatment of acute and critical illnesses. They are unable to identify critical missing evidence and cannot reliably reuse treatment pathways after receiving feedback from doctors and corrections from experts, resulting in skipping steps and potential safety risks.

Method used

The critical care diagnosis and treatment process is modeled as a structured diagnosis and treatment path diagram. Through multi-medical intelligent agent collaboration, standardized diagnosis and treatment status alignment, temporal interactive memory, and pre-execution evidence boundary arbitration, candidate diagnosis and treatment actions are generated and evidence boundary verification is performed. Structured evaluation items and clinical auxiliary decision-making are output.

Benefits of technology

It improves the safety, consistency, and interpretability of decision support for the diagnosis and treatment of acute and critical illnesses, enhances its adaptability in complex and dynamic clinical scenarios, and avoids directly outputting diagnosis and treatment recommendations when there is insufficient evidence or no risk assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123033704A_ABST
    Figure CN123033704A_ABST
Patent Text Reader

Abstract

The application relates to the field of medical information technology and provides an evidence boundary constraint-based critical illness diagnosis and treatment intelligent agent auxiliary decision-making method and system. The critical illness diagnosis and treatment process is converted from a traditional natural language question and answer into a dynamic decision-making process constrained by a structured diagnosis and treatment path graph. Natural language input by a doctor and structured medical record information are converted into standardized diagnosis and treatment states. Path recognition, disease assessment, evidence review, safety risk judgment and final coordination are respectively completed through multiple functional modules. Historical doctor feedback and expert error correction are converted into controllable constraints through time sequence interaction memory. Each candidate diagnosis and treatment action is checked before execution through an evidence boundary arbitration mechanism, and action results such as supplementary questions, waiting, blocking or execution are output, so that the safety and explainability of the method are improved while the consistency of the diagnosis and treatment process is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of medical information technology, specifically to a method and system for intelligent agent-assisted decision-making in the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints. Background Technology

[0002] The diagnosis and treatment of acute and critical illnesses are typically characterized by rapid changes in condition, short treatment windows, incomplete information, strong reliance on evidence, and high safety risks. Taking post-infectious shock, acute heart failure, and post-cardiopulmonary resuscitation syndrome as examples, physicians need to comprehensively assess multiple sources of information, including mean arterial pressure, lactate, urine output, skin perfusion, cardiac function, volume responsiveness, risk of pulmonary edema, response to vasoactive drugs, and bedside ultrasound results, while the patient's vital signs are constantly changing, to determine the patient's current stage of diagnosis and treatment and decide whether to conduct further examinations, wait for reassessment, implement treatment, or prevent high-risk procedures.

[0003] Existing medical AI systems mostly focus on medical question answering, single-round diagnostic suggestions, medical record summaries, or knowledge retrieval. These systems typically input natural language text from doctors directly into a large language model, which then generates medical interpretations or treatment suggestions. While this approach provides some language generation and medical knowledge integration capabilities, it still has significant limitations in critical care workflow control scenarios.

[0004] Because the diagnosis and treatment of acute and critical illnesses is not a single-round question-and-answer session, but a multi-round process that progresses continuously as the patient's condition changes, each treatment action often needs to meet specific evidence conditions, branching conditions, and safety conditions. For example, in a shock scenario, whether to administer fluids depends not only on hypotension, but also on volume responsiveness, cardiac function status, risk of pulmonary edema, and reassessment results. If the system cannot represent the diagnosis and treatment process as a structured path diagram, it is prone to problems such as skipping steps, inconsistent path progression, and untraceable outputs.

[0005] Furthermore, since the information entered by doctors is often incomplete, the system needs to determine which fields are known, which fields are missing, and which missing fields would affect the safety of treatment actions. Traditional question-and-answer models usually do not explicitly distinguish between "permissible actions" and "actions that require supplementary evidence," and therefore may directly give fluid resuscitation, vasopressor, cardiotonic, or other treatment suggestions when information on lactate, cardiac function, volume responsiveness, or contraindications is lacking.

[0006] Furthermore, clinical use generates continuous feedback from doctors, experts, and evaluations. For example, a doctor might point out that certain types of shock patients must receive lactate supplementation before fluid resuscitation, that rapid fluid resuscitation should not be initiated directly in certain risk states of pulmonary edema, or that certain treatment actions require additional cardiac protection measures. If this feedback is stored solely in natural language log format, the system will struggle to reliably reuse it in subsequent similar cases. Conversely, directly overriding the original rules with feedback could introduce erroneous or low-confidence feedback into the treatment pathway, creating new safety risks.

[0007] Therefore, there is an urgent need for an auxiliary decision-making scheme for the dynamic diagnosis and treatment process of acute and critical illnesses. This scheme should not only be able to understand the doctor's input, but also maintain the current path status, identify key missing evidence, coordinate the opinions of multiple medical agents, use historical feedback to form controllable constraints, and complete the verification of evidence boundaries and safety conditions before the execution of each candidate treatment action. This would improve the safety, interpretability, process consistency and continuous optimization capabilities of the auxiliary decision-making process. Summary of the Invention

[0008] To address the aforementioned technical issues, this application provides a method and system for intelligent agent-assisted decision-making in the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints.

[0009] To solve the above problems, this application is implemented as follows:

[0010] Firstly, this application provides a decision-making aid method for acute and critical care diagnosis and treatment based on evidence boundary constraints, including: Acquire the clinical input information for the current round, which includes one or more of the following: natural language case descriptions entered by doctors, structured medical record fields, vital signs, laboratory test results, imaging test results, bedside examination results, current treatment records, treatment response records, and output results from at least one external medical model; The clinical input information is analyzed to extract the structured update content of the current round and its corresponding semantic intent type. Based on the semantic intent type and the structured update content, the diagnosis and treatment task type and the acute and critical care diagnosis and treatment path type are determined. According to the type of acute and critical care treatment pathway, load the corresponding structured treatment pathway diagram and map the clinical input information into a standardized treatment status; the standardized treatment status includes at least the current treatment pathway, current pathway node, set of known clinical fields, set of missing clinical fields, set of fields to be reviewed, set of candidate pathway nodes, set of candidate treatment actions, set of risk markers, and historical pathway progress records; The standardized diagnosis and treatment status is input into the path advancement module, disease assessment module, evidence review module, security verification module, and result coordination module, respectively. Each module analyzes the data from different dimensions and outputs the analysis results. The result coordination module maps and restructures the analysis results from different dimensions output by each module according to the data structure of the assessment items to generate structured assessment items. Based on the current treatment path, current path node, and known clinical field set in the standardized treatment state, retrieve historical treatment information including historical doctor feedback records, historical expert error correction information, historical evaluation feedback records, and historical similar case information; convert the historical treatment information that meets the preset activation conditions into structured constraint units; project the structured constraint units into runtime treatment path constraints that act on the subsequent candidate action generation stage; Based on the current path node, the candidate path node set, the structured evaluation items, and the runtime treatment path constraints, a candidate treatment action set is generated; the candidate treatment action set includes one or more of the following: supplementary information action, waiting for re-evaluation action, blocking operation action, execution suggestion action, and path switching prompt action; For each candidate diagnosis and treatment action in the candidate diagnosis and treatment action set, obtain the evidence boundary configuration corresponding to the candidate diagnosis and treatment action; determine the final arbitration action based on the runtime diagnosis and treatment path constraints and the evidence boundary configuration; Clinical auxiliary output information is generated based on the final arbitration action and pushed to the clinical terminal; and upon receiving a doctor's confirmation instruction, a doctor's correction instruction, or new clinical input information, the standardized diagnosis and treatment status and the historical doctor feedback record are updated based on the doctor's confirmation instruction, doctor's correction instruction, or new clinical input information, forming a closed-loop iteration.

[0011] Secondly, this application provides an intelligent agent-assisted decision-making system for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints, comprising: The acquisition module is used to acquire the clinical input information for the current round. The clinical input information includes one or more of the following: natural language case descriptions entered by doctors, structured medical record fields, vital signs, laboratory test results, imaging test results, bedside examination results, current treatment records, treatment response records, and output results from at least one external medical model. The parsing and determination module is used to parse the clinical input information, extract the structured update content of the current round and its corresponding semantic intent type, and determine the diagnosis and treatment task type and the acute and critical care diagnosis and treatment path type based on the semantic intent type and the structured update content. The loading and mapping module is used to load the corresponding structured treatment path diagram according to the type of acute and critical care treatment path, and map the clinical input information into a standardized treatment status; the standardized treatment status includes at least the current treatment path, the current path node, the set of known clinical fields, the set of missing clinical fields, the set of fields to be reviewed, the set of candidate path nodes, the set of candidate treatment actions, the set of risk markers, and the historical path progress record; The evaluation module is used to input the standardized diagnosis and treatment status into the path advancement module, disease evaluation module, evidence review module, security verification module, and result coordination module, respectively. Each module analyzes the data from different dimensions and outputs the analysis results. The result coordination module maps and restructures the analysis results from different dimensions output by each module according to the evaluation item data structure to generate structured evaluation items. The interaction module is used to retrieve historical treatment information, including historical doctor feedback records, historical expert error correction information, historical evaluation feedback records, and historical similar case information, based on the current treatment path, current path node, and known clinical field set in the standardized treatment state; convert the historical treatment information that meets the preset activation conditions into structured constraint units; and project the structured constraint units into runtime treatment path constraints that act on the subsequent candidate action generation stage. The generation module is used to generate a set of candidate diagnostic and treatment actions based on the current path node, the set of candidate path nodes, the structured evaluation items, and the runtime diagnostic and treatment path constraints; the set of candidate diagnostic and treatment actions includes one or more of the following: supplementary information action, waiting for re-evaluation action, blocking operation action, execution suggestion action, and path switching prompt action; The arbitration module is used to obtain the evidence boundary configuration corresponding to each candidate diagnosis and treatment action in the candidate diagnosis and treatment action set; and to determine the final arbitration action based on the runtime diagnosis and treatment path constraints and the evidence boundary configuration. The feedback generation module is used to generate clinical auxiliary output information based on the final arbitration action and push it to the clinical terminal; and after receiving the doctor's confirmation instruction, doctor's correction instruction or new clinical input information, it updates the standardized diagnosis and treatment status and the historical doctor feedback record based on the doctor's confirmation instruction, doctor's correction instruction or new clinical input information to form a closed-loop iteration.

[0012] Thirdly, this application provides a terminal device including a processor and a memory, wherein the memory stores a program or instructions executable on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0013] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0014] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the steps of the method described in the first aspect.

[0015] In a sixth aspect, this application provides a computer program product stored in a storage medium, the program product being executed by at least one processor to implement the steps of the method described in the first aspect.

[0016] Compared with the prior art, this application has the following beneficial effects: By abstracting the clinical diagnosis and treatment process into a structured diagnosis and treatment path diagram, and uniformly mapping the natural language case descriptions, structured medical record fields, vital signs, laboratory test results, bedside examination data, and external medical model outputs input by doctors into standardized diagnosis and treatment states, multiple functional modules are invoked for collaborative evaluation based on the standardized diagnosis and treatment states to obtain disease identification results, path progression results, evidence completeness results, and safety risk results. Simultaneously, doctor feedback, expert error correction information, and constraints from historical similar cases are activated based on historical interaction information and projected into runtime diagnosis and treatment path constraints. For generated candidate diagnosis and treatment actions, evidence boundaries are further calculated, and action arbitration is performed between supplementary information, awaiting review, blocking operations, and execution suggestions based on safe evidence, complete evidence, soft evidence, contraindications, and path consistency requirements. This avoids directly outputting diagnosis and treatment suggestions when evidence is insufficient, paths are inconsistent, or risks are not assessed, improving the safety, process consistency, and interpretability of decision support for acute and critical care diagnosis and treatment, and enhancing the adaptability of the proposed method in complex and dynamic clinical scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the intelligent agent-assisted decision-making method for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints provided in this application. Figure 2 This is a schematic diagram of the structure of an intelligent agent-assisted decision-making system for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints, provided in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. Additionally, the use of "and / or" in this application indicates at least one of the connected objects, such as A and / or B and / or C, representing seven possibilities: including A alone, B alone, C alone, and the presence of both A and B, both B and C, both A and C, and the presence of A, B, and C.

[0021] In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0022] The following describes the intelligent agent-assisted decision-making method for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints provided in this application.

[0023] Understandably, this application involves modeling the diagnosis and treatment processes of acute and critical illnesses such as shock, acute heart failure, post-cardiopulmonary resuscitation syndrome, and sepsis-related hypoperfusion as structured treatment pathways. It combines multi-medical agent collaboration, standardized treatment status alignment, temporal interactive memory, and pre-execution evidence boundary arbitration. After the doctor inputs patient information, the system jointly assesses the current patient status, pathway nodes, missing evidence, candidate treatment actions, and safety risks, thereby outputting clinical auxiliary decision-making results such as supplementary questioning, awaiting reassessment, blocking high-risk procedures, or executing treatment recommendations. This application is applicable to intensive care units, emergency departments, cardiovascular acute and critical care, infection-related shock, in-hospital cardiopulmonary resuscitation management, intelligent medical record systems, clinical pathway management systems, and medical multi-agent auxiliary decision-making platforms. It can also be extended to other medical scenarios requiring dynamic decision control according to structured processes.

[0024] This application provides an intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints. This method transforms the traditional natural language question-and-answer process into a dynamic decision-making process constrained by a structured treatment path diagram. It converts the natural language input by doctors and structured medical record information into standardized treatment states. Multiple functional modules are used to complete path identification, disease assessment, evidence review, safety risk judgment, and final coordination. Historical doctor feedback and expert corrections are transformed into controllable constraints through temporal interactive memory. An evidence boundary arbitration mechanism performs pre-execution verification on each candidate treatment action and outputs results such as supplementary questioning, waiting, blocking, or execution. This ensures consistency in the treatment process while improving the security and interpretability of the method.

[0025] Furthermore, by integrating treatment pathways, clinical status, multifunctional module evaluation results, and historical feedback into a structured representation system, this application enables stable and controllable decision output in complex, variable, and incomplete information-rich acute and critical care scenarios, avoiding the uncertainty and potential risks of traditional large language models under unconstrained conditions.

[0026] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the intelligent agent-assisted decision-making method for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints provided in this application. Figure 1 The evidence boundary constraint-based intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment shown can be executed by terminal devices such as computers. This method can also be deployed in hospital information systems, critical care clinical decision support platforms, emergency auxiliary diagnosis and treatment systems, or medical big language model service platforms.

[0027] The system corresponding to the method of this application can be implemented through a local server, a hospital private cloud, an edge computing device or other computing environment. That is, the intelligent agent-assisted decision-making system for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints provided in this application can be implemented through a local server, a hospital private cloud, an edge computing device or other computing environment.

[0028] like Figure 1 As shown, the intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints provided in this application may include the following steps: Step 101: Obtain the clinical input information for the current round.

[0029] In this embodiment, the clinical input information includes one or more of the following: natural language case descriptions entered by the doctor, structured medical record fields, vital signs, laboratory test results, imaging test results, bedside examination results, current treatment records, treatment response records, and output results from at least one external medical model.

[0030] In some embodiments, the natural language case description may be expressed as: The patient's MAP is 58 mmHg, urine output is 0.2 ml / kg / h, skin is mottled, bedside lung ultrasound shows diffuse B-lines, and no lactate, LVEF, PLR-VTI or pressor response is currently recorded. Please determine the next course of action.

[0031] In some embodiments, structured medical record fields include: patient age, gender, body temperature, heart rate, respiratory rate, blood pressure, mean arterial pressure, blood oxygen saturation, urine output, lactate, blood gas analysis results, creatinine, white blood cell count, procalcitonin, central venous oxygen saturation, carbon dioxide gap, left ventricular ejection fraction, lung B-line, aortic velocity time integral, passive leg raise test results, fluid resuscitation response, and vasoactive drug dosage.

[0032] In some embodiments, the output of the external medical model may include sepsis risk scores, acute heart failure identification results, EEG abnormality identification results, organ function deterioration risk, low perfusion risk, shock classification results, or other medical auxiliary model results. The method in this application does not directly use the external model output as the final treatment recommendation, but rather as part of the standardized treatment status, subject to subsequent path diagram constraints and evidence boundary arbitration mechanisms.

[0033] In this embodiment, the system receives natural language case descriptions, structured medical record fields, vital signs, laboratory test results, bedside examination results, current treatment records, treatment response records, and output results from external medical models from doctors. The system then performs unified parsing and preprocessing on the above multi-source heterogeneous data, including medical entity recognition, field standardization, unit normalization, and synonym mapping, thereby forming a unified input representation that can be used for subsequent processing.

[0034] Step 102: Analyze the clinical input information, extract the structured update content of the current round and its corresponding semantic intent type, and determine the diagnosis and treatment task type and the acute and critical care diagnosis and treatment path type based on the semantic intent type and the structured update content.

[0035] The types of diagnostic and treatment tasks include one of the following: general medical Q&A, diagnostic and treatment pathway initiation, diagnostic and treatment pathway advancement, diagnostic and treatment pathway switching, supplementary information entry, treatment suggestion request, or doctor feedback update; the types of acute and critical care diagnostic and treatment pathways include one or more of the following: shock pathway, acute heart failure pathway, post-cardiopulmonary resuscitation syndrome pathway, and sepsis-related hypoperfusion pathway.

[0036] In some embodiments, a semantic routing agent can be used to identify the task type of the current clinical input.

[0037] In some embodiments, determining the diagnostic task type and the acute and critical care diagnostic pathway type based on the semantic intent type and the structured updated content includes: Based on the semantic intent type and the currently accumulated standardized diagnosis and treatment status, determine the diagnosis and treatment task type corresponding to the current round; Based on the structured update content, the semantic intent type, and the currently accumulated standardized diagnosis and treatment status, determine the acute and critical care diagnosis and treatment path type corresponding to the current round.

[0038] In some embodiments, when the input contains a new description of a patient's condition and meets the entry conditions of a certain acute and critical care pathway, it can be identified as a pathway initiation request; when the input contains review indicators, treatment responses, or new examination results, it can be identified as a pathway advancement request; when the input contains a state that is inconsistent with the current pathway but highly related to another disease pathway, it can be identified as a pathway switching request; when the input contains a doctor's affirmative, negative, or corrective opinions on the output of the method of this application, it can be identified as a feedback update request.

[0039] In some embodiments, while identifying the task type, the possible treatment pathway for the current patient can be determined. For example, when a patient presents with decreased mean arterial pressure, decreased urine output, increased lactate, mottled skin, or other signs of hypoperfusion, the shock pathway can be identified; when a patient presents with dyspnea, B-lines in the lungs, decreased left ventricular ejection fraction, volume overload, or elevated natriuretic peptides, the acute heart failure pathway can be identified; and when a patient presents with post-cardiopulmonary resuscitation spontaneous circulation recovery, hypoxia, hypotension, altered mental status, or a need for target temperature management, the post-cardiopulmonary resuscitation syndrome pathway can be identified.

[0040] In some embodiments, if multiple paths match simultaneously, multiple candidate paths can be retained, and path priority can be determined in subsequent evaluation and evidence boundary arbitration stages.

[0041] In this embodiment, the input intent of the current round is identified, and it is determined whether the current task belongs to general question and answer, path initiation, path advancement, path switching, supplementary information entry, treatment suggestion request, or feedback update request. At the same time, it can also identify the possible treatment path corresponding to the current patient, such as the shock path, acute heart failure path, or post-cardiopulmonary resuscitation syndrome path; when multiple candidate paths exist, multiple candidate paths can be retained and proceed to the subsequent joint evaluation stage.

[0042] Step 103: Load the corresponding structured treatment path diagram according to the type of acute and critical care treatment path, and map the clinical input information to a standardized treatment status.

[0043] The structured treatment pathway diagram includes pathway nodes, node types, node entry conditions, node exit conditions, required evidence fields for nodes, branching conditions, candidate treatment actions, conditions for waiting for re-evaluation, contraindication conditions, and safety protection conditions. The standardized treatment status includes at least the current treatment pathway, current pathway nodes, set of known clinical fields, set of missing clinical fields, set of fields to be re-evaluated, set of candidate pathway nodes, set of candidate treatment actions, set of risk markers, and historical pathway progress records.

[0044] In this embodiment, by modeling the acute and critical care diagnosis and treatment process as a structured diagnosis and treatment path diagram, the method provided in this application can advance the process within an executable and traceable path framework, reducing the risks of cross-path jumps and premature execution of steps caused by the free generation of large language models, and improving the consistency of the diagnosis and treatment process.

[0045] In some embodiments, a corresponding structured diagnostic path graph can be loaded based on the identified path type. The structured diagnostic path graph can be represented using YAML, JSON, database tables, graph databases, or other structured files, and each structured diagnostic path graph includes multiple nodes and directed edges between nodes.

[0046] In some embodiments, the node type includes: The entry node is used to represent the entry conditions of a path; The ask node is used to represent the evidence fields that need to be supplemented. The decision node is used to indicate path branch decisions; The action node is used to represent the diagnostic and treatment actions that can be performed or suggested. The wait node is used to indicate waiting for test results, waiting for treatment response, or waiting for re-evaluation. The exit node is used to indicate the end of a stage of the path or to output a summary.

[0047] Each node includes at least a node identifier, node name, node type, entry conditions, exit conditions, required evidence fields, candidate actions, taboo conditions, and security protection conditions. Edges between nodes represent path progression relationships. Each edge can contain branching conditions, such as an indicator exceeding a threshold, the existence of a risk, a positive test result, or a poor treatment response.

[0048] Taking shock pathways as an example, a structured treatment pathway diagram can include nodes for identifying hypoperfusion, assessing hemodynamic status, evaluating volume responsiveness, assessing fluid resuscitation strategies, evaluating vasopressors, assessing flow rates, and reviewing the results. The required fields for each node can include mean arterial pressure, lactate, urine output, capillary refill time, central venous oxygen saturation, carbon dioxide gap, left ventricular ejection fraction, lung B-line, aortic velocity-time integral, and passive leg raise test results.

[0049] In this embodiment, by maintaining the current path, current node, known fields, missing fields, and candidate actions through standardized diagnosis and treatment status, the method provided in this application can continuously update the patient status in multiple rounds of interaction, and realize cross-round information accumulation and dynamic decision-making.

[0050] In some embodiments, the known clinical field set is used to record currently obtained patient indicators and status; the missing clinical field set is used to record indicators required for the current pathway but not yet obtained; the field set to be reviewed is used to record fields that need to wait for treatment response or examination return; the candidate path node set is used to record the next node that can be reached from the current node; the candidate treatment action set is used to record the treatment action that may be output at present; and the risk marker set is used to record risk states such as pulmonary edema risk, abnormal cardiac function, volume overload, hypoperfusion deterioration, and high-dose vasoactive drug requirement.

[0051] In some embodiments, natural language input can be mapped to standardized fields through methods such as medical entity recognition, field mapping, unit normalization, synonym normalization, and path field matching. For example, "low blood pressure," "MAP 58," and "mean arterial pressure 58 mmHg" can all be mapped to the MAP field; "oliguria" and "urine output 0.2 ml / kg / h" can all be mapped to the urine output field; and "diffuse B-line lung ultrasound" and "excessive pulmonary edema" can all be mapped to pulmonary edema risk-related fields.

[0052] In this embodiment, by constructing a standardized diagnosis and treatment status, the consistency between the current round and the historical rounds can be continuously maintained, avoiding the need to start the judgment from scratch in each round of dialogue.

[0053] Step 104: Input the standardized diagnosis and treatment status into the path advancement module, disease assessment module, evidence review module, security verification module, and result coordination module respectively. Each module analyzes from different dimensions and outputs the analysis results. The result coordination module maps and restructures the analysis results from different dimensions output by each module according to the data structure of the assessment items to generate structured assessment items.

[0054] The structured evaluation items include at least one or more of the following: path matching results, candidate path nodes, missing evidence fields, risk warnings, taboo conditions, suggested actions, and action confidence.

[0055] In this embodiment, the multi-functional module collaboration mechanism enables different functional modules to complete path advancement, disease identification, evidence review, security risk identification, and final fusion, thereby improving the sufficiency of review and the reliability of decision-making in complex acute and critical illness scenarios.

[0056] The aforementioned path advancement module is used to determine the currently reachable candidate path nodes based on the current treatment path map, the current path nodes, and the set of known fields.

[0057] The aforementioned disease assessment modules are used to determine whether a patient meets the criteria for shock, acute heart failure, post-cardiopulmonary resuscitation syndrome, or other disease states. Different disease assessment modules can provide disease matching results and severity assessments, respectively.

[0058] The aforementioned evidence review module is used to check whether the evidence required for candidate treatment actions is complete. For example, when the candidate action is fluid resuscitation, the evidence review module checks whether there is evidence of volume responsiveness, cardiac function, pulmonary edema risk, and perfusion parameters; when the candidate action is vasopressor support, it checks whether there is evidence of hypotension, hypoperfusion, and volume status assessment.

[0059] The aforementioned safety verification module is used to identify contraindications and high-risk conditions. For example, when there is a significant risk of pulmonary edema, the safety verification module provides a risk warning for rapid fluid resuscitation; when there is severe hypotension and hypoperfusion, it provides a risk warning for delayed treatment.

[0060] The above-mentioned results coordination module integrates the outputs of the above modules, transforming unstructured opinions into candidate actions, missing fields, risk warnings, and path advancement suggestions in a unified format.

[0061] In some embodiments, the path advancement module, disease assessment module, evidence review module, security verification module, and result coordination module can be understood as path advancement intelligent agents, disease assessment intelligent agents, evidence review intelligent agents, security verification intelligent agents, and result coordination intelligent agents, respectively, and can realize their respective functions.

[0062] Step 105: Based on the current treatment path, current path node, and known clinical field set in the standardized treatment state, retrieve historical treatment information including historical doctor feedback records, historical expert error correction information, historical evaluation feedback records, and historical similar case information; convert the historical treatment information that meets the preset activation conditions into structured constraint units; project the structured constraint units into runtime treatment path constraints that act on the subsequent candidate action generation stage.

[0063] In this embodiment, a temporal interactive memory mechanism can be introduced to store doctor feedback, expert error correction, and experience from similar historical cases. This mechanism does not directly cover the treatment pathway map, but rather transforms feedback into structured memory atoms, which then participate in the current round of decision-making after retrieval, activation scoring, conflict detection, and constraint projection.

[0064] In some embodiments, when a doctor confirms, denies, supplements, or corrects the output of the method of this application, the feedback type is first parsed. For example, if a doctor points out that "this risk of hypotension combined with pulmonary edema cannot be directly and rapidly resuscitated with fluids", the feedback can be parsed as a shielding or protective condition against rapid resuscitation; if a doctor points out that "lactic acid must be administered before resuscitation", the feedback can be parsed as adding the required field; if a doctor points out that "this answer can be used as a standard expression", it can be parsed as a reusable answer pattern.

[0065] In some embodiments, the memory atom includes: AddReq is used to add necessary evidence fields to a path node or candidate action; AddMask is used to mask a high-risk action in a specific context; AddGuard is used to add protective conditions to a specific action; ReuseAnswer is used to save reusable positive response patterns; Correction is used to record error correction information for historical system outputs.

[0066] In some embodiments, similar memory atoms are retrieved based on standardized treatment states in the current round, and an activation score is calculated. The activation score can be determined based on state similarity, memory confidence, number of supports, number of conflicts, and temporal relevance. A higher number of supports, higher confidence, more similar states, and more recent times result in a higher activation score; a higher number of conflicts results in a lower activation score.

[0067] In this embodiment, the doctor's feedback and expert error correction are transformed into structured constraints through a temporal interactive memory mechanism. After activation scoring, confidence assessment and conflict detection, the constraints are projected into the runtime path, thereby enabling controllable reuse and continuous optimization of feedback experience.

[0068] In some embodiments, when multiple memory atoms conflict with each other, conflict information can be recorded instead of directly overwriting the original constraint with the last feedback. For example, if there have been multiple expert feedbacks indicating "lactic acid must be added before performing fluid replacement", while a new single feedback indicates "lactic acid can be performed directly without lactic acid", although the conflict will be recorded, the original safety constraint will not be revoked if the new feedback is insufficient.

[0069] When a memory atom meets the activation condition, it can be projected as a runtime diagnostic path constraint. For example, AddReq is projected as a candidate action to add a required field; AddMask is projected as a candidate action masking rule; and AddGuard is projected as an action protection condition. In this way, experience can be reused while preventing error feedback from bypassing security controls.

[0070] In this embodiment, the conflict feedback handling mechanism avoids simply using the last feedback or directly overwriting the original path. It can perform conservative handling when multiple feedbacks contradict each other, thereby reducing the risk of erroneous feedback.

[0071] In some embodiments, the structured constraint unit includes at least one of the following: The first type of constraint unit used to add the required fields; The second type of constraint unit is used to shield high-risk actions; A third type of constraint unit used to increase the conditions for motion protection; The fourth type of constraint unit is used for reusing positive response patterns; The fifth type of constraint unit is used to record error correction information.

[0072] Step 106: Based on the current path node, the candidate path node set, the structured evaluation item, and the runtime diagnosis and treatment path constraints, generate a candidate diagnosis and treatment action set; the candidate diagnosis and treatment action set includes one or more of the following: supplementary information action, waiting for re-evaluation action, blocking operation action, execution suggestion action, and path switching prompt action.

[0073] Step 107: For each candidate diagnosis and treatment action in the candidate diagnosis and treatment action set, obtain the evidence boundary configuration corresponding to the candidate diagnosis and treatment action; based on the runtime diagnosis and treatment path constraints and the evidence boundary configuration, determine the final arbitration action.

[0074] In this embodiment, an evidence boundary arbitration mechanism is used to determine whether a candidate treatment action meets the requirements of safe evidence, complete evidence, contraindications, and path consistency before execution. This avoids directly outputting treatment actions when there is insufficient evidence or risks, thereby improving the safety of the method in this application.

[0075] In some embodiments, evidence boundaries can be calculated for candidate diagnostic and treatment actions. These evidence boundaries are used to determine whether a candidate diagnostic or treatment action is executable in the current state, requires further information, needs to be awaited for review, or should be blocked.

[0076] For each candidate diagnostic or treatment action, the corresponding set of safety evidence, complete evidence, and soft evidence can be determined.

[0077] The set of safety evidence conditions represents the minimum safety conditions that must be met before a candidate action can be performed. For example, for fluid resuscitation, safety evidence may include the presence of a significant risk of pulmonary edema, the presence of clues to volume responsiveness, and the presence of a risk of severe cardiac dysfunction. If the safety evidence is not met, the action cannot be performed directly.

[0078] A complete set of evidence conditions represents the complete set of conditions that a candidate action should meet when proceeding according to the standard pathway. For example, for fluid strategy decisions in a shock pathway, complete evidence may include MAP, lactate, urine output, volume responsiveness, cardiac function assessment, and rehydration response records. If the complete evidence is not met but the safety risks are manageable, supplementary information will be prioritized or a review will be requested.

[0079] The set of soft evidence conditions represents evidence that helps improve the quality of judgment but does not necessarily constitute an absolute block. For example, certain review indicators, trend indicators, or auxiliary examination results can serve as soft evidence to determine whether it is necessary to wait or expedite supplementation.

[0080] Understandably, this application also checks taboo conditions and path consistency. Taboo conditions are used to determine whether a candidate action poses a high risk in the current state; path consistency is used to determine whether a candidate action is a permitted next action of the current path node. If an action is linguistically reasonable but is not a legal action permitted by the current path node or the current branch, the action can be filtered, downgraded to a supplementary question, or trigger a path switching judgment.

[0081] In some embodiments, the evidence boundary configuration includes at least a safe evidence condition set, a complete evidence condition set, and a soft evidence condition set; wherein, the safe evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the minimum safe execution conditions, the complete evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the standard path advancement conditions, and the soft evidence condition set is used to determine whether further supplementary information is needed, whether to wait for examination results, or whether to conduct dynamic re-evaluation; The determination of the final arbitration action based on the runtime treatment path constraints and the evidence boundary configuration includes: When the candidate diagnostic or treatment action does not meet the requirements of the set of safety evidence conditions or hits the set of contraindications, the blocking operation action will be determined as the final arbitration action. When the candidate diagnosis and treatment action meets the requirements of the set of safety evidence conditions but does not meet the requirements of the set of complete evidence conditions, and there are unobtained soft evidence fields or fields to be reviewed, the supplementary information action or the waiting for review action will be determined as the final arbitration action. When the candidate treatment action meets the requirements of the safety evidence condition set, the complete evidence condition set, and the path consistency condition set, the execution suggestion action will be determined as the final arbitration action. When the candidate treatment action does not meet the requirements of the current path consistency condition set but matches other acute and critical care treatment paths, the path switching prompt action will be determined as the final arbitration action.

[0082] In this embodiment, by combining runtime constraints and evidence boundaries in arbitration, it can be ensured that reasonable actions matching safety requirements can be output under different levels of evidence completeness, taking into account both decision-making safety and efficiency in critical care scenarios.

[0083] Step 108: Generate clinical auxiliary output information based on the final arbitration action and push it to the clinical terminal; and after receiving the doctor's confirmation instruction, doctor's correction instruction or new clinical input information, update the standardized diagnosis and treatment status and the historical doctor feedback record based on the doctor's confirmation instruction, doctor's correction instruction or new clinical input information to form a closed-loop iteration.

[0084] In some embodiments, the clinical auxiliary output information includes one or more of the following: current treatment pathway identifier, current pathway node identifier, explanation of pathway advancement basis, list of key indicators that need to be supplemented, high-risk actions that are not recommended to be performed at this time and the reasons therefor, executable treatment recommendations, waiting or re-evaluation time recommendations, and safety reminder information.

[0085] In this embodiment, a consultation field optimization mechanism is used to prioritize asking for key fields that are most likely to unlock subsequent paths, have the greatest impact on treatment safety, and have the lowest acquisition cost when there are many missing fields. This improves interaction efficiency and reduces invalid inquiries.

[0086] This application supports dynamic interruption, rollback, and switching of the diagnosis and treatment process. When the patient's condition deteriorates or new risk signals appear, the method of this application can relocate the path node and adjust the next action, thereby improving the adaptability to complex clinical changes.

[0087] In addition, by introducing a structured output and decision-making basis recording mechanism, this application ensures that each auxiliary suggestion has a traceable source of evidence and path basis, thereby enhancing the interpretability and clinical acceptability of the method in this application.

[0088] Furthermore, this application is also compatible with the output of external medical models and can re-validate them through path constraints and evidence boundary mechanisms, thereby avoiding the direct driving of high-risk decisions by external model results and improving the overall robustness and safety.

[0089] In some embodiments, when the arbitration result is supplementary information, the key fields that need to be supplemented can be output, and the relationship between these fields and pathway advancement or treatment safety can be explained; when the arbitration result is pending review, the examination results, treatment response, or review time that needs to be waited for can be output; when the arbitration result is blocking operation, the blocking action, the contraindications triggered, and the reasons for the risk can be explained; when the arbitration result is to execute treatment, treatment recommendations that conform to the current pathway and evidence boundaries can be output, and review conditions and safety monitoring indicators can be prompted.

[0090] In some embodiments, doctors can confirm, modify, or correct the output. Doctor feedback can be written into a temporal interactive memory and used in subsequent similar cases after retrieval, activation, conflict detection, and evidence boundary filtering.

[0091] This application provides an intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints. It abstracts the clinical diagnosis and treatment process into a structured diagnosis and treatment path diagram and uniformly maps the doctor's input of natural language case descriptions, structured medical record fields, vital signs, laboratory test results, bedside examination data, and external medical model outputs into standardized diagnosis and treatment states. Based on these standardized states, multiple functional modules are invoked for collaborative evaluation to obtain disease identification results, path progression results, evidence completeness results, and safety risk results. Simultaneously, doctor feedback, expert correction information, and constraints from historical similar cases are activated based on historical interaction information and projected into runtime diagnosis and treatment path constraints. For generated candidate diagnosis and treatment actions, evidence boundaries are further calculated, and action arbitration is performed between supplementary information, awaiting review, blocking operations, and execution suggestions based on safe evidence, complete evidence, soft evidence, contraindications, and path consistency requirements. This avoids directly outputting diagnosis and treatment suggestions when evidence is insufficient, paths are inconsistent, or risks are not assessed, improving the safety, process consistency, and interpretability of auxiliary decision-making for acute and critical care diagnosis and treatment, and enhancing the adaptability of this method in complex and dynamic clinical scenarios.

[0092] The following is a detailed explanation of the application process of this application in patients at risk of post-infectious shock complicated by pulmonary edema.

[0093] The patient is a 68-year-old male who presented with shock after infection. The doctor entered the following information: "Patient's MAP 58 mmHg, urine output 0.2 ml / kg / h, mottled skin; bedside lung ultrasound shows diffuse B-lines; no lactate, LVEF, PLR-VTI, or pressor response recorded; please determine the next steps."

[0094] First, the clinical input acquisition steps were performed to identify a MAP of 58 mmHg, a urine output of 0.2 ml / kg / h, the presence of skin mottledness, and diffuse B-lines on bedside lung ultrasound. At the same time, missing records were identified for lactate, LVEF, PLR-VTI, and pressor response.

[0095] Then, the semantic routing agent determines that the input is a treatment suggestion request and identifies that the patient has hypotension, low urine output, and abnormal skin perfusion, meeting the entry conditions for the shock pathway. The shock diagnosis and treatment pathway map is loaded, and the current state is positioned at the relevant nodes for hypoperfusion identification and initial hemodynamic assessment.

[0096] During the multi-agent parallel assessment phase, the shock assessment agent determined the patient's condition to be consistent with shock based on a MAP ≤ 65 mmHg, decreased urine output, and mottled skin. The acute heart failure assessment agent and the safety guardian agent identified diffuse B-lines, suggesting a risk of pulmonary edema or volume overload. The evidence review agent found a lack of records for lactate, LVEF, PLR-VTI, volume responsiveness, and pressor response, concluding that there was insufficient evidence to directly formulate a rapid fluid resuscitation strategy.

[0097] During the temporal interactive memory activation phase, historical similar cases and expert feedback can be retrieved, including: lactate should be administered before key treatments in shock cases; rapid fluid resuscitation is not advisable in cases of hypotension with pulmonary edema risk; fluid resuscitation actions require evidence of volume responsiveness or cardiac function. This feedback can be converted into AddReq, AddMask, and AddGuard memory atoms. AddReq requires the addition of the lactate field; AddMask disables rapid fluid resuscitation actions in the current state; AddGuard adds volume responsiveness and pulmonary edema risk protection conditions to fluid resuscitation actions.

[0098] During the boundary assessment stage, candidate actions can be evaluated. These include rapid fluid resuscitation, lactate supplementation and cardiac function assessment, volume responsiveness assessment, consideration of vasopressor support, and waiting or reassessment. Given the patient's risk of pulmonary edema and the lack of crucial evidence such as lactate, LVEF, and PLR-VTI, it can be determined that "direct rapid fluid resuscitation" does not meet the requirements for safety and complete evidence; therefore, it is downgraded to providing supplementary information and risk warnings. Simultaneously, it is suggested that vasopressor support be assessed while maintaining perfusion targets.

[0099] Finally, the following auxiliary suggestions can be output: "The patient is currently in shock, exhibiting hypotension, low urine output, and abnormal skin perfusion. However, bedside lung ultrasound reveals diffuse B-lines, suggesting a risk of pulmonary edema or volume overload. In the absence of lactate, LVEF, PLR-VTI, or other evidence of volume responsiveness, rapid fluid resuscitation is not recommended. Prioritize lactate resuscitation or confirmation, cardiac function assessment, volume responsiveness indicators, and vasopressor response records; simultaneously maintain MAP and organ perfusion targets, and assess the need for vasopressor support such as norepinephrine. If subsequent confirmation of a positive volume response and manageable pulmonary edema risk, a small-volume trial of fluid resuscitation can be considered with dynamic reassessment; if the volume response is negative or pulmonary water overload is significant, continued fluid resuscitation should be avoided, and a vasopressor, cardiotonic, or de-load strategy should be evaluated." This embodiment demonstrates that the present application can simultaneously identify hypoperfusion and pulmonary edema risk in shock pathways, and downgrade or block candidate actions based on missing evidence, pathway constraints, and historical feedback, avoiding direct recommendation of rapid fluid resuscitation based solely on shock labels.

[0100] The following section provides a detailed explanation of the application process of this application in scenarios where acute heart failure and shock pathways intersect.

[0101] The doctor entered the following: "The patient suddenly developed dyspnea, is orthopnea, has a blood pressure of 86 / 50 mmHg, a heart rate of 122 bpm, and moist rales in the lungs. Bedside ultrasound shows decreased LVEF and multiple B-lines in the lungs. Urine output is decreased, and lactate levels have not yet been reported. Please determine the next steps." Upon receiving input, the semantic routing agent identifies whether the input belongs to a path initiation or treatment suggestion request. It simultaneously identifies that both acute heart failure and shock pathways may match: on the one hand, the patient presents with dyspnea, pulmonary rales, decreased LVEF, and multiple B lines, consistent with acute heart failure; on the other hand, the patient presents with hypotension, tachycardia, and decreased urine output, suggesting low perfusion risk.

[0102] It can load acute heart failure and shock pathways and construct standardized diagnostic and treatment statuses. Known fields include blood pressure, heart rate, dyspnea, pulmonary rales, decreased LVEF, lung B-lines, and decreased urine output; missing fields include lactate, volume responsiveness, blood gas, blood oxygenation status, vasoactive drug response, and dynamic reassessment indicators.

[0103] During the multi-agent parallel assessment phase, the acute heart failure assessment agent identified the patient as exhibiting acute heart failure and pulmonary edema; the shock assessment agent identified the patient as having a low perfusion risk and needed to be assessed for entry into the shock-related treatment node; the safety guardian agent suggested that when hypotension and pulmonary edema coexist, the conventional diuretic or rapid fluid resuscitation strategies from a single pathway should be avoided; and the evidence review agent indicated a lack of evidence for reassessment of lactate, blood oxygen, volume status, and hemodynamics.

[0104] The temporal interactive memory module retrieved historical feedback: In cases of acute heart failure with hypotension, diuresis should not be administered directly following the standard acute heart failure pathway; in cases of hypotension complicated by pulmonary edema, perfusion parameters and evidence of cardiac function / volume status should be supplemented; if hypoperfusion is present, vasopressor or cardiac support should be evaluated. This can be converted into AddGuard and AddReq constraints, requiring supplementation of lactate, oxygenation status, volume responsiveness, and perfusion parameters before giving recommendations for diuresis, fluid resuscitation, or vasopressor administration.

[0105] The arbitration module evaluates candidate actions. For "rapid fluid resuscitation," safety evidence is insufficient due to pulmonary edema and decreased LVEF; for "direct high-dose diuresis," caution is required due to the risks of hypotension and low perfusion; for "replenishing lactate, blood oxygen, and volume status with bedside dynamic assessment," the current supplementary and reassessment criteria are met. This ultimately outputs supplementary key evidence and cautious management recommendations, rather than directly executing a single treatment action.

[0106] This embodiment illustrates that the present application can handle the intersections and conflicts between different acute and critical care pathways, and avoid inappropriate treatment recommendations through pathway switching, evidence boundaries and safety guarding mechanisms when acute heart failure and shock risks coexist.

[0107] The following section provides a detailed explanation of how this application transforms doctor feedback into temporal memory and activates it in subsequent cases.

[0108] This embodiment is used to illustrate the physician feedback memory mechanism of this application.

[0109] In one round of decision support for shock cases, the system output "Consider fluid resuscitation." The doctor's feedback was as follows: "The patient has obvious B-line in the lungs and no PLR-VTI or cardiac function assessment. Fluid resuscitation cannot be recommended directly. At least fluid responsiveness and evidence of cardiac function should be tested first." After receiving the feedback, the system determines that it belongs to the category of doctor error correction feedback. The feedback parsing unit breaks down the feedback into three structured constraints: First, when there are obvious B-lines in the lungs, direct fluid resuscitation should be restricted; Second, evidence of volume responsiveness is required for fluid resuscitation. Third, fluid resuscitation requires evidence of cardiac function assessment.

[0110] The memory atomization unit writes these to the AddMask and AddReq memory atoms, respectively. AddMask is used to block direct fluid resuscitation in the state of "obvious B-lines in the lungs and no assessment of volume responsiveness"; AddReq is used to add fields such as PLR-VTI and cardiac function assessment to the fluid resuscitation procedure.

[0111] The subsequent input for another patient was: "The patient's MAP was 60 mmHg, urine output was decreased, and lung ultrasound showed increased B lines, but no LVEF or PLR was recorded." The system retrieved a historical memory atom similar to the current state. Because this memory atom is highly similar to the current state and originates from doctor error correction feedback, its confidence level meets the threshold, therefore it is activated and projected as a runtime constraint. Based on this, the arbitration module downgrades "direct fluid resuscitation" to "resuscitation based on evidence of volume responsiveness and cardiac function," and indicates a risk of pulmonary edema.

[0112] This embodiment illustrates that the present application can transform doctor feedback into reusable, activatable, and conflict-handling structured constraints, enabling the system to automatically avoid repeating errors in subsequent similar cases, while not directly covering the original diagnosis and treatment pathway.

[0113] The aforementioned arbitration module can be understood as the evidence boundary arbitration module.

[0114] See Figure 2 , Figure 2 This is a schematic diagram of the structure of an intelligent agent-assisted decision-making system for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints, provided in an embodiment of this application. Figure 2 As shown, system 200 includes: The acquisition module 201 is used to acquire the clinical input information of the current round. The clinical input information includes one or more of the following: natural language case descriptions entered by doctors, structured medical record fields, vital signs, laboratory test results, imaging test results, bedside examination results, current treatment records, treatment response records, and output results of at least one external medical model. The parsing and determination module 202 is used to parse the clinical input information, extract the structured update content of the current round and its corresponding semantic intent type, and determine the diagnosis and treatment task type and the acute and critical care diagnosis and treatment path type based on the semantic intent type and the structured update content. The loading mapping module 203 is used to load the corresponding structured treatment path diagram according to the type of acute and critical care treatment path, and map the clinical input information into a standardized treatment status; the standardized treatment status includes at least the current treatment path, current path node, set of known clinical fields, set of missing clinical fields, set of fields to be reviewed, set of candidate path nodes, set of candidate treatment actions, set of risk markers, and historical path progress records; The evaluation module 204 is used to input the standardized diagnosis and treatment status into the path advancement module, disease evaluation module, evidence review module, security verification module and result coordination module respectively, and each module analyzes from different dimensions and outputs analysis results; the result coordination module maps and restructures the analysis results of different dimensions output by each module according to the evaluation item data structure to generate structured evaluation items. The interaction module 205 is used to retrieve historical treatment information, including historical doctor feedback records, historical expert error correction information, historical evaluation feedback records, and historical similar case information, based on the current treatment path, current path node, and known clinical field set in the standardized treatment state; convert the historical treatment information that meets the preset activation conditions into structured constraint units; and project the structured constraint units into runtime treatment path constraints that act on the subsequent candidate action generation stage. The generation module 206 is used to generate a set of candidate diagnostic and treatment actions based on the current path node, the set of candidate path nodes, the structured evaluation items, and the runtime diagnostic and treatment path constraints; the set of candidate diagnostic and treatment actions includes one or more of the following: supplementary information action, waiting for re-evaluation action, blocking operation action, execution suggestion action, and path switching prompt action; Arbitration module 207 is used to obtain the evidence boundary configuration corresponding to each candidate diagnosis and treatment action in the candidate diagnosis and treatment action set; and determine the final arbitration action based on the runtime diagnosis and treatment path constraints and the evidence boundary configuration. The feedback generation module 208 is used to generate clinical auxiliary output information based on the final arbitration action and push it to the clinical terminal; and after receiving the doctor's confirmation instruction, doctor's correction instruction or new clinical input information, it updates the standardized diagnosis and treatment status and the historical doctor feedback record based on the doctor's confirmation instruction, doctor's correction instruction or new clinical input information to form a closed loop iteration.

[0115] Optionally, the evidence boundary configuration includes at least a safe evidence condition set, a complete evidence condition set, and a soft evidence condition set; wherein, the safe evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the minimum safe execution conditions, the complete evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the standard path advancement conditions, and the soft evidence condition set is used to determine whether further supplementary information is needed, whether to wait for examination results, or whether to conduct dynamic re-evaluation; The arbitration module 207 is specifically used for: When the candidate diagnostic or treatment action does not meet the requirements of the set of safety evidence conditions or hits the set of contraindications, the blocking operation action will be determined as the final arbitration action. When the candidate diagnosis and treatment action meets the requirements of the set of safety evidence conditions but does not meet the requirements of the set of complete evidence conditions, and there are unobtained soft evidence fields or fields to be reviewed, the supplementary information action or the waiting for review action will be determined as the final arbitration action. When the candidate treatment action meets the requirements of the safety evidence condition set, the complete evidence condition set, and the path consistency condition set, the execution suggestion action will be determined as the final arbitration action. When the candidate treatment action does not meet the requirements of the current path consistency condition set but matches other acute and critical care treatment paths, the path switching prompt action will be determined as the final arbitration action.

[0116] Optionally, the parsing and determining module 202 is specifically used for: Based on the semantic intent type and the currently accumulated standardized diagnosis and treatment status, determine the diagnosis and treatment task type corresponding to the current round; Based on the structured update content, the semantic intent type, and the currently accumulated standardized diagnosis and treatment status, determine the acute and critical care diagnosis and treatment path type corresponding to the current round.

[0117] Optionally, the structured constraint unit includes at least one of the following: The first type of constraint unit used to add the required fields; The second type of constraint unit is used to shield high-risk actions; A third type of constraint unit used to increase the conditions for motion protection; The fourth type of constraint unit is used for reusing positive response patterns; The fifth type of constraint unit is used to record error correction information.

[0118] Optionally, the clinical auxiliary output information includes one or more of the following: current treatment pathway identifier, current pathway node identifier, explanation of pathway advancement basis, list of key indicators that need to be supplemented, high-risk actions that are not recommended to be performed at this time and the reasons therefor, executable treatment recommendations, waiting or re-evaluation time recommendations, and safety reminder information.

[0119] Optionally, the diagnosis and treatment task type includes one of the following: general medical Q&A, diagnosis and treatment pathway initiation, diagnosis and treatment pathway advancement, diagnosis and treatment pathway switching, supplementary information entry, treatment suggestion request, or doctor feedback update; the acute and critical care diagnosis and treatment pathway type includes one or more of the following: shock pathway, acute heart failure pathway, post-cardiopulmonary resuscitation syndrome pathway, and sepsis-related hypoperfusion pathway.

[0120] The intelligent agent-assisted decision-making system for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints provided in this application can achieve the goals outlined in this application. Figure 1 The various processes in the method embodiments, and the ways to achieve the same beneficial effects, will not be repeated here to avoid repetition.

[0121] like Figure 3As shown, this application also provides a terminal device, including a processor 301 and a memory 302. The memory 302 stores a program or instructions that can run on the processor 301. When the program or instructions are executed by the processor 301, they implement the various steps of the above-described embodiment of the intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0122] It should be noted that the terminal device in this application can be a terminal or other devices besides a terminal. For example, the terminal device can be a tablet computer, a laptop computer, etc., and this application does not make any specific limitation.

[0123] This application also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the various processes of the above-described embodiment of the intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints, and achieve the same technical effect. To avoid repetition, these will not be described again here.

[0124] The processor is the processor in the terminal device described in the above embodiments. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (Read-Only Memory). Only memory (ROM), random access memory (RAM), magnetic disks or optical disks, etc.

[0125] This application also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0126] It should be understood that the chip mentioned in this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0127] This application provides a computer program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of the above-described embodiment of the intelligent agent-assisted decision-making method for acute and critical care diagnosis and treatment based on evidence boundary constraints, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0128] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as a read-only memory). The device includes a number of instructions in a ROM (random access memory), RAM (magnetic disk), or optical disk to cause a terminal (which may be a computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0130] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A decision-making aid method for acute and critical care diagnosis and treatment based on evidence boundary constraints, characterized in that, include: Acquire the clinical input information for the current round, which includes one or more of the following: natural language case descriptions entered by doctors, structured medical record fields, vital signs, laboratory test results, imaging test results, bedside examination results, current treatment records, treatment response records, and output results from at least one external medical model; The clinical input information is analyzed to extract the structured update content of the current round and its corresponding semantic intent type. Based on the semantic intent type and the structured update content, the diagnosis and treatment task type and the acute and critical care diagnosis and treatment path type are determined. According to the type of acute and critical care treatment pathway, load the corresponding structured treatment pathway diagram and map the clinical input information into a standardized treatment status; the standardized treatment status includes at least the current treatment pathway, current pathway node, set of known clinical fields, set of missing clinical fields, set of fields to be reviewed, set of candidate pathway nodes, set of candidate treatment actions, set of risk markers, and historical pathway progress records; The standardized diagnosis and treatment status is input into the path advancement module, disease assessment module, evidence review module, security verification module, and result coordination module, respectively. Each module analyzes the data from different dimensions and outputs the analysis results. The result coordination module maps and restructures the analysis results from different dimensions output by each module according to the data structure of the assessment items to generate structured assessment items. Based on the current treatment path, current path node, and known clinical field set in the standardized treatment state, retrieve historical treatment information including historical doctor feedback records, historical expert error correction information, historical evaluation feedback records, and historical similar case information; convert the historical treatment information that meets the preset activation conditions into structured constraint units; project the structured constraint units into runtime treatment path constraints that act on the subsequent candidate action generation stage; Based on the current path node, the candidate path node set, the structured evaluation items, and the runtime treatment path constraints, a candidate treatment action set is generated; the candidate treatment action set includes one or more of the following: supplementary information action, waiting for re-evaluation action, blocking operation action, execution suggestion action, and path switching prompt action; For each candidate diagnostic action in the candidate diagnostic action set, obtain the evidence boundary configuration corresponding to the candidate diagnostic action; Based on the runtime treatment path constraints and the evidence boundary configuration, the final arbitration action is determined; Clinical auxiliary output information is generated based on the final arbitration action and pushed to the clinical terminal; Upon receiving a doctor's confirmation instruction, a doctor's correction instruction, or new clinical input information, the standardized diagnosis and treatment status and the historical doctor feedback record are updated according to the doctor's confirmation instruction, doctor's correction instruction, or new clinical input information, forming a closed-loop iteration.

2. The method according to claim 1, characterized in that, The evidence boundary configuration includes at least a safe evidence condition set, a complete evidence condition set, and a soft evidence condition set; wherein, the safe evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the minimum safe execution conditions, the complete evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the standard path advancement conditions, and the soft evidence condition set is used to determine whether further supplementary information is needed, whether to wait for the examination results, or whether to conduct a dynamic review. The determination of the final arbitration action based on the runtime treatment path constraints and the evidence boundary configuration includes: When the candidate diagnostic or treatment action does not meet the requirements of the set of safety evidence conditions or hits the set of contraindications, the blocking operation action will be determined as the final arbitration action. When the candidate diagnosis and treatment action meets the requirements of the set of safety evidence conditions but does not meet the requirements of the set of complete evidence conditions, and there are unobtained soft evidence fields or fields to be reviewed, the supplementary information action or the waiting for review action will be determined as the final arbitration action. When the candidate treatment action meets the requirements of the safety evidence condition set, the complete evidence condition set, and the path consistency condition set, the execution suggestion action will be determined as the final arbitration action. When the candidate treatment action does not meet the requirements of the current path consistency condition set but matches other acute and critical care treatment paths, the path switching prompt action will be determined as the final arbitration action.

3. The method according to claim 1, characterized in that, The process of determining the diagnostic task type and the acute and critical care diagnostic pathway type based on the semantic intent type and the structured updated content includes: Based on the semantic intent type and the currently accumulated standardized diagnosis and treatment status, determine the diagnosis and treatment task type corresponding to the current round; Based on the structured update content, the semantic intent type, and the currently accumulated standardized diagnosis and treatment status, determine the acute and critical care diagnosis and treatment path type corresponding to the current round.

4. The method according to any one of claims 1 to 3, characterized in that, The structured constraint unit includes at least one of the following: The first type of constraint unit used to add the required fields; The second type of constraint unit is used to shield high-risk actions; A third type of constraint unit used to increase the conditions for motion protection; The fourth type of constraint unit is used for reusing positive response patterns; The fifth type of constraint unit is used to record error correction information.

5. The method according to any one of claims 1 to 3, characterized in that, The clinical auxiliary output information includes one or more of the following: current treatment pathway identifier, current pathway node identifier, explanation of the basis for pathway advancement, list of key indicators that need to be supplemented, high-risk actions that are not recommended to be performed at this time and the reasons therefor, executable treatment recommendations, waiting or re-evaluation time recommendations, and safety reminder information.

6. The method according to any one of claims 1 to 3, characterized in that, The types of diagnostic and treatment tasks include one of the following: general medical Q&A, diagnostic and treatment pathway initiation, diagnostic and treatment pathway advancement, diagnostic and treatment pathway switching, supplementary information entry, treatment suggestion request, or doctor feedback update; the types of acute and critical care diagnostic and treatment pathways include one or more of the following: shock pathway, acute heart failure pathway, post-cardiopulmonary resuscitation syndrome pathway, and sepsis-related hypoperfusion pathway.

7. A decision-making aid system for the diagnosis and treatment of acute and critical illnesses based on evidence boundary constraints, characterized in that, include: The acquisition module is used to acquire the clinical input information for the current round. The clinical input information includes one or more of the following: natural language case descriptions entered by doctors, structured medical record fields, vital signs, laboratory test results, imaging test results, bedside examination results, current treatment records, treatment response records, and output results from at least one external medical model. The parsing and determination module is used to parse the clinical input information, extract the structured update content of the current round and its corresponding semantic intent type, and determine the diagnosis and treatment task type and the acute and critical care diagnosis and treatment path type based on the semantic intent type and the structured update content. The loading and mapping module is used to load the corresponding structured treatment path diagram according to the type of acute and critical care treatment path, and map the clinical input information into a standardized treatment status; the standardized treatment status includes at least the current treatment path, the current path node, the set of known clinical fields, the set of missing clinical fields, the set of fields to be reviewed, the set of candidate path nodes, the set of candidate treatment actions, the set of risk markers, and the historical path progress record; The evaluation module is used to input the standardized diagnosis and treatment status into the path advancement module, disease evaluation module, evidence review module, security verification module, and result coordination module, respectively. Each module analyzes the data from different dimensions and outputs the analysis results. The result coordination module maps and restructures the analysis results from different dimensions output by each module according to the evaluation item data structure to generate structured evaluation items. The interaction module is used to retrieve historical treatment information, including historical doctor feedback records, historical expert error correction information, historical evaluation feedback records, and historical similar case information, based on the current treatment path, current path node, and known clinical field set in the standardized treatment state; convert the historical treatment information that meets the preset activation conditions into structured constraint units; and project the structured constraint units into runtime treatment path constraints that act on the subsequent candidate action generation stage. The generation module is used to generate a set of candidate diagnostic and treatment actions based on the current path node, the set of candidate path nodes, the structured evaluation items, and the runtime diagnostic and treatment path constraints; the set of candidate diagnostic and treatment actions includes one or more of the following: supplementary information action, waiting for re-evaluation action, blocking operation action, execution suggestion action, and path switching prompt action; The arbitration module is used to obtain the evidence boundary configuration corresponding to each candidate diagnosis and treatment action in the candidate diagnosis and treatment action set. Based on the runtime treatment path constraints and the evidence boundary configuration, the final arbitration action is determined; A feedback generation module is used to generate clinical auxiliary output information based on the final arbitration action and push it to the clinical terminal; Upon receiving a doctor's confirmation instruction, a doctor's correction instruction, or new clinical input information, the standardized diagnosis and treatment status and the historical doctor feedback record are updated according to the doctor's confirmation instruction, doctor's correction instruction, or new clinical input information, forming a closed-loop iteration.

8. The system according to claim 7, characterized in that, The evidence boundary configuration includes at least a safe evidence condition set, a complete evidence condition set, and a soft evidence condition set; wherein, the safe evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the minimum safe execution conditions, the complete evidence condition set is used to verify whether the candidate diagnosis and treatment action meets the standard path advancement conditions, and the soft evidence condition set is used to determine whether further supplementary information is needed, whether to wait for the examination results, or whether to conduct a dynamic review. The arbitration module is specifically used for: When the candidate diagnostic or treatment action does not meet the requirements of the set of safety evidence conditions or hits the set of contraindications, the blocking operation action will be determined as the final arbitration action. When the candidate diagnosis and treatment action meets the requirements of the set of safety evidence conditions but does not meet the requirements of the set of complete evidence conditions, and there are unobtained soft evidence fields or fields to be reviewed, the supplementary information action or the waiting for review action will be determined as the final arbitration action. When the candidate treatment action meets the requirements of the safety evidence condition set, the complete evidence condition set, and the path consistency condition set, the execution suggestion action will be determined as the final arbitration action. When the candidate treatment action does not meet the requirements of the current path consistency condition set but matches other acute and critical care treatment paths, the path switching prompt action will be determined as the final arbitration action.

9. The system according to claim 7, characterized in that, The parsing and determination module is specifically used for: Based on the semantic intent type and the currently accumulated standardized diagnosis and treatment status, determine the diagnosis and treatment task type corresponding to the current round; Based on the structured update content, the semantic intent type, and the currently accumulated standardized diagnosis and treatment status, determine the acute and critical care diagnosis and treatment path type corresponding to the current round.

10. A terminal device, characterized in that, It includes a processor and a memory, wherein the memory stores a program or instructions executable on the processor, the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 6.