A safe collaborative diagnosis and treatment system for cardiovascular and cerebrovascular multi-modal data
Patent Information
- Application Number
- CN202611198695.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-08
- Publication Date
- 2026-10-27
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence medical technology and multi-agent collaborative systems, specifically to a secure collaborative diagnosis and treatment system and method for cardiovascular and cerebrovascular multimodal data, and particularly to a multi-agent collaborative architecture that dynamically intervenes in the decomposition strategy of cardiovascular and cerebrovascular specialty tasks based on triple security verification of medical data and multi-level long-term memory, and integrates clinical safety guardrails and distributed consistency protocols. Background Technology
[0002] Cardiovascular disease is the leading cause of death worldwide, and its precise diagnosis and treatment heavily rely on the comprehensive analysis of multimodal medical data. This data includes at least: electrophysiological signals (12-lead electrocardiogram, single-lead wearable electrocardiogram, photoplethysmography (PPG), Holter long-term recording), medical imaging (cardiac magnetic resonance imaging (CMR) cine / LGE / T1-T2 mapping, coronary CTA, echocardiography video stream, carotid ultrasound, fundus photography), multi-omics data (blood proteomics, metabolomics, genotype SNP arrays), structured electronic medical records (outpatient / inpatient records, medications, laboratory indicators, follow-up outcomes), and unstructured medical texts (radiology reports, electrocardiogram reports, clinical guidelines and consensus).
[0003] In recent years, significant progress has been made in basic AI models for cardiovascular and cerebrovascular diseases targeting single modalities. These include self-supervised pre-training models based on large-scale electrocardiogram data, visual-tabular multimodal models based on cardiac magnetic resonance imaging, pathological synthesis self-supervised models based on coronary CTA, and multi-omics risk scoring frameworks based on proteomics and metabolomics. However, these models are all vertical architectures oriented towards specific modalities or tasks, and a systematic approach that unifies and collaboratively reasons from electrophysiology, imaging, omics, medical records, and medical texts is lacking in clinical practice.
[0004] At the system architecture level, existing technologies mainly suffer from the following shortcomings:
[0005] 1. The security of medical data access is disconnected from agent access management. Existing IoT data access solutions (such as triple verification mechanisms based on asymmetric encryption, open-source code recognition, and API interface recognition) only address the authentication problem between devices. The verification results are not dynamically coupled with the task routing and data access permissions of the upper-layer agents. In multi-agent collaborative diagnosis and treatment scenarios, different agents lack fine-grained, dynamic access control based on verification results for accessing sensitive medical data, posing a risk of unauthorized access.
[0006] 2. Fixed task decomposition strategies cannot be automatically optimized based on historical clinical experience. Existing multi-agent collaboration frameworks (such as collaboration frameworks based on standard operating procedures (SOPs) and graph-based state management frameworks) require developers to manually orchestrate task flows for each type of cardiovascular disease. When faced with complex scenarios such as differential diagnosis of acute chest pain and multimodal assessment of heart failure, similar tasks are repeatedly decomposed, leading to delays in clinical response. Existing technologies lack the ability to automatically summarize and solidify historical task execution data, failing to establish a closed loop of "memory-decomposition-execution-solidification".
[0007] 3. The lack of collaborative mechanisms among multi-level memory systems specifically for medical scenarios. Existing general multi-agent systems operate with short-term memory, long-term contextual memory, and skill-based memory operating independently, failing to form a joint intervention capability for decision-making in diagnosis and treatment. More critically, current technologies do not incorporate medical-specific clinical safety contexts (such as contraindications, critical value thresholds, and guideline conflict rules) into the considerations for skill consolidation and memory retrieval, resulting in agents failing to fully utilize prior safety and treatment experience when performing sub-tasks.
[0008] 4. The clinical safety barrier is disconnected from the synchronization of distributed execution states. While existing distributed agent frameworks employ consensus protocols such as Raft to ensure consistency within the scheduling cluster, they primarily address the issue of state synchronization between cluster nodes. When the clinical safety barrier module rejects a treatment recommendation from an agent, this rejection result is not broadcast to all scheduler nodes in real time via the consensus protocol. This can lead to the new Leader node replaying the rejected execution path after a failure recovery.
[0009] 5. The agent's execution context is too narrow, failing to adequately consider the dynamic nature of the patient's clinical state. Current technologies often use fixed weighting ratios when integrating multi-layered memories, or only make coarse-grained adjustments based on sub-task types, without adaptively adjusting to dynamic clinical characteristics such as the patient's current critical condition, rare disease phenotype, or post-operative follow-up period. Summary of the Invention
[0010] This invention provides a secure collaborative diagnosis and treatment system and method for cardiovascular and cerebrovascular multimodal data. It aims to solve the technical problems in the prior art, such as the separation of data access security and task routing, fixed task decomposition strategies, insufficient memory system collaboration, and lack of security state synchronization, by dynamically binding medical data triple security verification with Agent permissions, establishing a multi-level memory closed-loop feedback mechanism embedded in the clinical security context, and incorporating clinical security decision results into the distributed consistency protocol log.
[0011] To achieve the above objectives, the present invention adopts the following technical solution:
[0012] A secure collaborative diagnosis and treatment system for cardiovascular and cerebrovascular multimodal data includes:
[0013] A multimodal secure access layer is configured to perform identity authentication and data traceability on accessed heterogeneous medical data sources through a triple verification mechanism. The triple verification mechanism includes: verifying the device hardware fingerprint based on an asymmetric encryption algorithm to obtain an initial identity identifier; calling the data traceability code in the traceability database based on the initial identity identifier and comparing and verifying the digital digest of the access data packet to obtain a data traceability authentication code; and calling the API interface authorization data in the model authorization database based on the data traceability authentication code and comparing and verifying the model access permissions of the requesting end to obtain a development authorization instruction code. The heterogeneous medical data sources include at least electrophysiological signals, medical images, multi-omics data, structured electronic medical records, and unstructured medical text.
[0014] A unified collaborative framework layer, coupled to the multimodal secure access layer, includes a task decomposition and routing engine. The task decomposition and routing engine maintains a capability registry, which records the unique identifier, capability tag, current task queue length, and data access permission level of each specialized agent. The development authorization command code is dynamically bound to the data access permission level, and the task decomposition and routing engine constrains the routable range of subtasks according to the data access permission level.
[0015] A multi-level long-term memory system, coupled to the unified collaborative framework layer, includes a short-term working memory area, a long-term contextual memory vector library, and a skills and process memory library.
[0016] There are N specialized agents, each containing independent domain hints, a specialized knowledge base, and a toolset. They register their capabilities with the unified collaborative framework layer through a standardized protocol, where N≥2, and the N specialized agents include at least an electrophysiological signal analysis agent, a medical image analysis agent, a multi-omics analysis agent, a structured medical record analysis agent, and a medical text reasoning agent.
[0017] The distributed Harness cluster contains multiple peer scheduler nodes. The scheduler nodes maintain the cluster state through a consistency protocol and are used to perform decentralized task orchestration on the N specialist agents.
[0018] The clinical safety barrier module includes a preset clinical mandatory rule base, which includes at least contraindication rules, critical value priority rules, and clinical guideline conflict rules.
[0019] Preferably, the task decomposition and routing engine constrains the routable range of subtasks according to the data access permission level, specifically including: when the access permission level of a certain modality data in the development authorization code is prohibited, the task decomposition and routing engine automatically marks the specialist agent node corresponding to the modality as unreachable when generating or calling the task DAG, and reroutes the subtask to other specialist agents with alternative capability tags and passed permission verification, or triggers an external consultation request.
[0020] Preferably, the skill and process memory is further configured to simultaneously record the input data pattern and output data pattern of each node in the subtask process when the subtask process is solidified into a medical skill block, and generate a parameterized DAG template; the parameterized DAG template contains configurable parameter slots, which are automatically filled according to the patient-specific parameters of the new task when called; wherein, the patient-specific parameters include at least patient identifier, genotype locus, image sequence number, test index threshold, and the set of contraindications in the clinical safety context constraints.
[0021] Preferably, the contraindication rule in the clinical mandatory rule base is configured as follows: when any specialist agent outputs a treatment suggestion that includes drugs, procedures, or examinations corresponding to the target patient's past allergy history or absolute contraindications, the suggestion is directly rejected regardless of its confidence score; the critical value priority rule is configured as follows: when any specialist agent identifies a physiological indicator that meets the clinical critical value criteria during processing, the normal execution order of the current task DAG is interrupted, the critical value processing sub-process is triggered first, and the critical value identifier is written into the consistency protocol log entry to trigger a global state change; the clinical guideline conflict rule is configured as follows: when a treatment suggestion has a directional conflict with the conclusion of a preset authoritative clinical guideline with a recommendation level of III or an evidence level of B or above, a manual review mark is triggered.
[0022] Preferably, the scheduler node in the distributed Harness cluster maintains the cluster state through the Raft protocol. The log entries of the Raft protocol record the execution state changes of each subtask in the task DAG, the corresponding medical data access audit trajectory, and the rejection records of the clinical safety guardrail module. When the scheduler node fails, the newly elected leader node restores the global execution state of the DAG and the incomplete medical data authorization links by replaying the Raft logs.
[0023] Preferably, when the specialist agent integrates the current session context, historical task experience, and medical skill blocks into a unified input context, it dynamically adjusts the weight ratio of the three components based on the medical type of the current subtask and the patient's current clinical status. The patient's current clinical status includes at least the presence of critical value alerts, whether it is a rare disease phenotype, and whether it is in the postoperative follow-up period. For diagnostic subtasks following fixed clinical pathways, the weight of the medical skill blocks is higher than that of the historical task experience. For complex reasoning subtasks requiring differential diagnosis, the weight of the historical task experience is higher than that of the medical skill blocks. For urgent subtasks with critical value alerts, the weight of the current session context in short-term working memory is higher than that of the medical skill blocks and the historical task experience.
[0024] Preferably, the scheduler node in the distributed Harness cluster adopts a logic execution and physical execution separation architecture: the scheduler node acts as a logic orchestrator, responsible for parsing clinical intent and generating a DAG containing subtask dependencies; the scheduler node acts as an execution coordinator, responsible for distributing specific execution instructions to specialist agents with corresponding capability tags and verified by the development authorization instruction code, based on the DAG, the real-time load of the specialist agent, the data access permission level, and the clinical security context constraints.
[0025] A secure collaborative diagnosis and treatment method for cardiovascular and cerebrovascular multimodal data, applied to the system described in any of the preceding claims, the method comprising:
[0026] The multimodal secure access steps involve receiving encrypted access requests from heterogeneous medical data sources, sequentially verifying the device hardware fingerprint, data traceability code, and model API interface authorization through a triple verification mechanism, generating a development authorization instruction code, and dynamically binding the development authorization instruction code to the data access permission level in the capability registry.
[0027] The user's clinical task receiving steps include receiving clinical task requests containing natural language.
[0028] The medical skill block matching step involves querying the skill and process memory to determine whether a medical skill block matches the new diagnosis and treatment task.
[0029] First execution path: If the medical skill block is successfully matched, the medical skill block is directly invoked and executed, skipping the task decomposition step;
[0030] Second execution path: If the medical skill block fails to match, query the long-term contextual memory vector library to retrieve similar historical task experience; if relevant historical task experience is retrieved, adjust the granularity of task decomposition or the routing order of subtasks according to the historical task experience before performing decomposition.
[0031] Third execution path: If the medical skill block fails to match and no relevant historical task experience is found, execute the regular task decomposition;
[0032] In the context enhancement step, before executing a subtask, the specialist agent simultaneously acquires the current session context of the short-term working memory, the historical task experience of the long-term contextual memory vector library, and the medical skill blocks of the skill and process memory library, and merges them into a unified input context.
[0033] In the distributed scheduling execution step, the distributed Harness cluster schedules the subtask to the optimal specialist agent for execution based on the current load of the specialist agent, the data access permission level, and the clinical security context constraints, and manages its lifecycle.
[0034] The clinical safety decision-making process involves collecting the execution results of each subtask, querying the clinical mandatory rule base, and making mandatory rejections or weighted decisions based on preset strategies when there are contraindications, critical values, or guideline conflicts. The rejection results are then written into the consistency protocol log entries.
[0035] In addition, the memory update step involves updating the key features of this task, execution summary, clinical outcome, and rejected treatment recommendations and their reasons for rejection to the long-term contextual memory vector library and / or the skills and processes memory library after the final diagnosis and treatment results are generated.
[0036] Preferably, in the distributed scheduling execution step, the Harness cluster uses a consistent hashing algorithm to determine a unique execution agent from a group of specialist agents with the same capability tags and who have passed permission verification, based on the capability tags required by the subtask, the real-time load information of the specialist agent, the data access permission level in the development authorization code, and the set of contraindications in the clinical security context constraints. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of the overall architecture of the secure collaborative diagnosis and treatment system for cardiovascular and cerebrovascular multimodal data according to the present invention.
[0038] Figure 2 This is a schematic diagram of the triple verification mechanism and dynamic binding process of Agent permissions in the multimodal security access layer of this invention.
[0039] Figure 3 This is a schematic diagram of the multi-level long-term memory system structure and the closed-loop intervention logic of the task decomposition engine of the present invention.
[0040] Figure 4 This is a flowchart of the automatic solidification and parameterized DAG template generation process for medical skill blocks in this invention.
[0041] Figure 5 This is a schematic diagram illustrating the decision logic and rejection results of the clinical safety guardrail module of this invention, written to the Raft log.
[0042] Figure 6 This is a schematic diagram of the Raft state synchronization and fault recovery mechanism of the distributed Harness cluster of this invention. Detailed Implementation
[0043] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0044] System Architecture Overview
[0045] like Figure 1 As shown, this system consists of the following five core subsystems:
[0046] 1. Multimodal secure access layer (100)
[0047] The multimodal security access layer is responsible for performing defense-in-depth security verification on heterogeneous medical data sources accessing the system, and dynamically coupling the verification results with the permissions and task routing of the upper-layer agent. For example... Figure 2 As shown, this layer employs a triple verification mechanism:
[0048] First layer of verification (device authentication, 110): Receives encrypted access requests from wearable ECG devices, hospital PACS systems, laboratory LIS systems, gene sequencing platforms, or doctor workstations. The system verifies the device's hardware fingerprint (such as MAC address, device certificate, TPM chip identifier) using SM2 / SM4 or RSA asymmetric encryption algorithms, and decrypts it to obtain the initial identity identifier (Initial ID). This initial identity identifier uniquely identifies the device or system from which the data originated.
[0049] The second layer of verification (data traceability authentication, 120): Based on the initial identity identifier, the system retrieves the registration data traceability code corresponding to the device from a traceability database (such as a blockchain traceability chain or a trusted timestamp database). The system calculates the digital digest (such as a SHA-256 hash) of the incoming data packet and compares it with the original digest recorded in the data traceability code. If the comparison matches, a data traceability authentication code is generated, confirming that the data has not been tampered with during transmission.
[0050] The third layer of verification (model authorization authentication, 130): Based on the data traceability authentication code, the system retrieves the corresponding API interface authorization data from the model authorization database. The system verifies whether the requesting client has permission to access a specific specialist agent or a specific analysis model. If the verification is successful, a Dev Auth Token is generated, which includes a time-limited access token, a data anonymization level identifier, and a data access permission level.
[0051] Key innovations – dynamic permission binding and routing constraints: The developed authorization command codes are not only used for data access but also dynamically bound to the data access permission level field in the capability registry within the unified collaboration framework layer. When the task decomposition and routing engine generates or invokes a task DAG, the engine queries this permission level in real time. If the access permission level for a certain modality of data (such as coronary CTA) is prohibited (e.g., the hospital's PACS system has not passed the third-level model authorization verification), the engine automatically marks the CTA image analysis Agent node as unreachable and reroutes the subtask to another specialist Agent with alternative capabilities (such as echocardiography analysis) that has passed permission verification, or triggers an external image consultation request. This achieves closed-loop control from data access verification to task routing, rather than simply allowing access after it has been granted.
[0052] 2. Unified Collaboration Framework Layer (200)
[0053] The unified collaboration framework layer includes a task decomposition and routing engine (210). The engine internally maintains a "capability registry" that records the ID, capability tags (such as "12-lead ECG arrhythmia detection", "CMR cine myocardial segmentation", "coronary CTA plaque characterization", "proteomics CVD risk score", "medical text RAG retrieval"), current load, and data access permission level of each specialty agent.
[0054] Upon receiving a clinical natural language request (such as "Please analyze the differential diagnosis of this patient's acute chest pain"), the task decomposition and routing engine first queries the skills and process memory (430):
[0055] (1) If a matching medical skill block exists (such as "Acute Chest Pain Multimodal Assessment"), the parameterized DAG template of the skill block is directly called, skipping the LLM intent recognition and subtask decomposition steps.
[0056] (2). If the skill block matching fails, the task description is converted into a query vector, and the Top-K similar historical cases in the long-term episodic memory vector library (420) are retrieved. The decomposition strategy is adjusted according to historical experience.
[0057] (3) If none of them are hit, then the lightweight medical large language model is called to perform routine task decomposition.
[0058] The communication bus (220) adopts a medical message format compliant with the HL7 FHIR standard, and achieves secure communication between agents through a publish-subscribe message queue. The message payload includes a de-identified patient identifier, task ID, capability tag, data payload, response topic, and a copy of the development authorization code.
[0059] The unified tool interface layer (230) encapsulates external tools (such as PubMed search, DrugBank drug interaction query, OMIM gene database, and hospital internal CDSS interface) into standard APIs for use by various specialty agents.
[0060] 3. Multi-level long-term memory system (400)
[0061] like Figure 3 As shown, the memory system adopts a three-tier storage architecture:
[0062] Short-term working memory (410): Based on the current doctor-patient dialogue context cache, it stores the most recent N rounds of interaction records, intermediate reasoning conclusions and matters to be confirmed in the current consultation.
[0063] Long-Term Contextual Memory Vector Library (420): Constructed based on a medical-grade vector database. Whenever a diagnostic task is completed, the system vectorizes the task's summary (including patient age and gender, chief complaint, key abnormal indicators, final diagnosis, treatment plan, follow-up outcome, rejected treatment recommendations and their reasons for rejection) and stores it in this library. Vector encoding employs a medical embedding model fine-tuned from a medical corpus. When a new task arrives, the system retrieves the Top-K most relevant historical experiences by calculating the cosine similarity between the query vector describing the task and the historical task vectors in the library.
[0064] Skills and Process Memory (430): Adopts the form of a medical process template library. When a fixed diagnosis and treatment process containing multiple sub-tasks is repeatedly executed more than a preset frequency threshold and the success rate exceeds a preset success rate threshold, the system automatically solidifies it into a callable "medical skill block".
[0065] Core Innovation – Skill Block Solidification Embedded with Clinical Safety Context: Unlike general skill blocks, the medical skill blocks in this invention, when solidified, not only record the DAG structure and parameter slots, but also simultaneously record the set of contraindication conditions that were triggered and rejected by the clinical safety guardrail module during historical execution. For example, when the "Acute Chest Pain Multimodal Assessment" skill block is solidified, it will also record the safety context that "if the patient has a history of iodine contrast agent allergy, the CTA node will be automatically blocked by the contraindication rule." When this skill block is subsequently called, the system first checks whether the current patient matches the contraindication condition set. If it does, it automatically adjusts the DAG (such as skipping the CTA node and adding an ultrasound alternative node), rather than performing a post-execution rejection after the Agent executes.
[0066] 4. N specialist agents (500)
[0067] Each specialty agent is an independent, containerized deployment process that encapsulates specific medical domain prompts, knowledge bases, and toolsets.
[0068] Electrophysiological signal analysis Agent (510): Processes 12-lead ECG, single-lead wearable ECG, PPG and Holter data, and outputs arrhythmia classification, ST segment change analysis and heart rate variability assessment.
[0069] Medical Image Analysis Agent (520): Processes CMR cine / LGE images, coronary CTA volume data, and echocardiogram video streams, and outputs ejection fraction, myocardial strain, plaque characteristics, and degree of stenosis.
[0070] Multi-omics analysis agent (530): Processes proteomic, metabolomic and genotypic data to output cardiovascular disease risk scores.
[0071] Structured Medical Record Analysis Agent (540): Processes EHR table data and analyzes outpatient / inpatient records, medication history, laboratory indicators, and follow-up outcomes.
[0072] Medical Text Reasoning Agent (550): Based on the RAG architecture, it accesses radiology reports, electrocardiogram reports, and clinical guideline consensus, and performs text summarization, guideline recommendation retrieval, and report generation.
[0073] Each agent registers its capabilities with the unified collaboration framework layer through standardized protocols (such as gRPC over mTLS).
[0074] 5. Distributed Harness Cluster (600)
[0075] The cluster consists of multiple Harness nodes, each deployed in a peer-to-peer manner. For example... Figure 6As shown, nodes elect a leader through the Raft protocol.
[0076] Core Innovation – Security Decision Results Incorporated into Raft Logs: Raft log entries not only record cluster member changes and execution status changes of each subtask in the task DAG (“Pending,” “In Execution,” “Completed,” “Failed”), but also record the corresponding medical data access audit trails and rejection records from the clinical security barrier module. When the clinical security barrier module rejects a treatment recommendation from an Agent, the rejection result and reason are written to the Raft log. When a scheduler node fails, the newly elected Leader node, by replaying the Raft logs, not only restores the global execution state of the DAG but also identifies which execution paths have been rejected by the security barrier, avoiding the replay of blocked tasks.
[0077] The Harness node is logically divided into a "logic orchestrator" and an "execution coordinator": the former parses the clinical intent to generate a DAG; the latter distributes the execution instructions to the verified Agent based on the DAG, Agent real-time load, data access permission level, and clinical security context constraints.
[0078] 6. Clinical safety railing module (700)
[0079] like Figure 5 As shown, this module has a built-in clinical mandatory rule base (710), which includes:
[0080] Contraindication rule base (711): Associates the hospital's allergy history and contraindication database with the patient's master index. If the Agent output contains a specific drug for a particular allergic patient or a specific drug recommendation for a patient with severe bradycardia, it is directly rejected, and the rejection record is written to the Raft log.
[0081] Critical Value Rule Base (712): Defines clinical critical value criteria. When the Agent identifies a critical value, it interrupts the regular DAG process, prioritizes triggering the critical value alarm subprocess, and writes the critical value identifier to the Raft log to trigger a global state change.
[0082] Guideline Conflict Rule Base (713): Built-in structured recommendations from authoritative cardiovascular disease guidelines. If the Agent's output conflicts directionally with a guideline's Level III recommendation or Level A recommendation, it is marked as a high-risk suggestion, and manual review is mandatory.
[0083] The adjudication process is as follows: Collect results and confidence levels from each agent → Query the mandatory rule base → If there is a conflict, reject the decision and write it to the log → If there is no conflict, perform a weighted vote → Output the final treatment recommendation.
[0084] Agent context fusion execution
[0085] Before executing a subtask, the scheduled specialist agent simultaneously receives input from three layers of memory:
[0086] The short-term working memory area (410) provides the context of the doctor-patient dialogue during the current consultation;
[0087] The long-term episodic memory vector library (420) provides experience of similar historical cases (including previously rejected recommendations and reasons);
[0088] The skills and process memory (430) provides the DAG structure, parameterized templates and contraindication condition set for medical skills blocks.
[0089] The agent integrates these three elements into a unified input context and dynamically adjusts the weights based on the subtask type and the patient's current clinical status.
[0090] Implementation of fixed clinical pathways 0.2 0.2 0.6 Heart failure classification assessment according to guidelines Differential Diagnosis Reasoning 0.3 0.5 0.2 Differential diagnosis of acute chest pain Critical value alarm exists 0.6 0.2 0.2 Prioritize responding to the current emergency context Rare disease phenotypes 0.2 0.6 0.2 Prioritize reference to historical experience with rare cases
[0091] Example: Multimodal differential diagnosis of acute chest pain
[0092] The user (cardiologist) entered the following request: "A 65-year-old male presented with sudden onset of chest pain for 2 hours. ECG showed ST-segment elevation in leads V1-V4 and elevated troponin I. Please provide differential diagnosis and recommendations for further treatment."
[0093] 1. Multimodal Secure Access: The hospital's PACS system, LIS system, and electrocardiograph (ECG) are accessed via triple authentication. If the hospital's CTA device fails the third-level model authorization verification (e.g., the device certificate has expired), the access permission level for the CTA mode in its development authorization code will be marked as "restricted."
[0094] 2. Dynamic Permission Binding and Routing Constraints: The task decomposition and routing engine query capabilities registry reveals that the CTA image analysis agent's data access permission level is "restricted," automatically marking the node as unreachable. The system searches for an alternative echocardiography agent with "allowed" permission level, preparing to include it in the DAG.
[0095] 3. Medical Skill Block Matching: The engine queries the skill and process memory database and finds a fixed skill block for "Acute ST-segment elevation chest pain multimodal assessment". This skill block includes: ECG precise localization analysis (electrophysiology agent), troponin dynamic trend analysis (structured medical record agent), imaging assessment node (originally CTA agent, now marked as "replaced with ultrasound agent if unreachable" due to access restrictions and security context), and contraindication screening and reperfusion strategy recommendation (text reasoning agent + security guardrail). The system directly calls this skill block, skipping LLM decomposition.
[0096] 4. Contextual Enhancement: Electrophysiological agents acquire short-term working memory (patient vital signs BP 150 / 90 mmHg, HR 98 bpm), long-term episodic memory (experience from 3 similar cases, suggesting the need to exclude aortic dissection), and skill blocks (including a set of contraindications: CTA contrast agents are contraindicated in patients with a history of iodine allergy).
[0097] 5. Distributed Scheduling: The Harness cluster schedules subtasks to various agents for parallel execution. The echocardiography agent replaces the original CTA node to perform ventricular wall motion analysis.
[0098] 6. Clinical Safety Decision: The text-based reasoning agent recommends "emergency PCI + aspirin + ticagrelor". The safety guardrail queries the contraindication rule base to confirm no conflicts; it queries the critical value rule base to confirm that although troponin is elevated, it does not reach the critical value threshold defined by the agent; there are no mandatory rule conflicts, and the final recommendation is output after weighted voting.
[0099] 7. Result Synthesis and Memory Update: The system synthesizes treatment suggestions and returns them to the doctor. In this task, the path adjustment, final diagnosis, treatment plan, and rejected suggestions (if any) for "CTA permission restricted → ultrasound substitution" are vectorized and stored in the long-term contextual memory vector library; the task execution summary and security context are updated to the statistical table in the skills and process memory library.
[0100] Explanation of reference numerals in the attached figures:
[0101] 100 Multimodal Secure Access Layer 110 Device Identity Authentication Module 120 Data Traceability Authentication Module 130 Model Authorization Authentication Module 200 Unified Collaboration Framework Layer 210 Task Decomposition and Routing Engine 220 Communication Bus 230 Unified Tool Interface Layer 400 Multi-level Long-term Memory System 410 Short-term Working Memory Area 420 Long-term Contextual Memory Vector Library 430 Skill and Process Memory Library 500 N Specialty Agents 510 Electrophysiological Signal Analysis Agent 520 Medical Image Analysis Agent 530 Multi-omics Analysis Agent 540 Structured Medical Record Analysis Agent 550 Medical Text Reasoning Agent 600 Distributed Harness Cluster 700 Clinical Safety Barrier Module 710 Clinical Mandatory Rule Base 711 Contraindication Rule Base 712 Critical Value Rule Base 713 Guideline Conflict Rule Base
Claims
1. A secure collaborative diagnosis and treatment system for cardiovascular and cerebrovascular multimodal data, characterized in that, include: A multimodal secure access layer is configured to perform identity authentication and data traceability on accessed heterogeneous medical data sources through a triple verification mechanism. The triple verification mechanism includes: verifying the device hardware fingerprint based on an asymmetric encryption algorithm to obtain an initial identity identifier; calling the data traceability code in the traceability database based on the initial identity identifier and comparing and verifying the digital digest of the access data packet to obtain a data traceability authentication code; and calling the API interface authorization data in the model authorization database based on the data traceability authentication code and comparing and verifying the model access permissions of the requesting end to obtain a development authorization instruction code. The heterogeneous medical data sources include at least electrophysiological signals, medical images, multi-omics data, structured electronic medical records, and unstructured medical text. A unified collaborative framework layer, coupled to the multimodal secure access layer, includes a task decomposition and routing engine. The task decomposition and routing engine maintains a capability registry, which records the unique identifier, capability tag, current task queue length, and data access permission level of each specialized agent. The development authorization command code is dynamically bound to the data access permission level, and the task decomposition and routing engine constrains the routable range of subtasks according to the data access permission level. A multi-level long-term memory system, coupled to the unified collaborative framework layer, includes a short-term working memory area, a long-term contextual memory vector library, and a skills and process memory library. There are N specialized agents, each containing independent domain hints, a specialized knowledge base, and a toolset. They register their capabilities with the unified collaborative framework layer through a standardized protocol, where N≥2, and the N specialized agents include at least an electrophysiological signal analysis agent, a medical image analysis agent, a multi-omics analysis agent, a structured medical record analysis agent, and a medical text reasoning agent. The distributed Harness cluster contains multiple peer scheduler nodes. The scheduler nodes maintain the cluster state through a consistency protocol and are used to perform decentralized task orchestration on the N specialist agents. The clinical safety barrier module includes a preset clinical mandatory rule base, which includes at least contraindication rules, critical value priority rules, and clinical guideline conflict rules. The skill and process memory bank is configured as follows: Monitor the execution frequency and success rate of each sub-task process in historical diagnosis and treatment tasks; When the execution frequency exceeds a preset frequency threshold and the execution success rate exceeds a preset success rate threshold, the corresponding sub-task process is solidified into a medical skill block. The medical skill block includes the directed acyclic graph structure of the sub-task process, the dependencies between sub-tasks, and the clinical safety context constraints corresponding to each node. The solidification is automatically triggered by the system without manual preset. The clinical safety context constraints include at least the set of contraindication conditions that were triggered and rejected by the clinical safety guardrail module in the historical execution of the sub-task process. When a new medical task is received, the task description of the new task is matched with the pre-defined medical skill blocks; The long-term contextual memory vector library is configured as follows: When the matching of the medical skill block fails, the task description of the new task is converted into a query vector. By calculating the similarity between the query vector and the historical task vectors stored in the database, the Top-K historical task experiences with similarity exceeding a preset threshold are retrieved. The task decomposition and routing engine is configured as follows: When the medical skill block is successfully matched, the medical skill block is directly invoked and executed, skipping the task decomposition step; When the medical skill block fails to match but the historical task experience is retrieved, the granularity of task decomposition or the routing order of subtasks is adjusted according to the historical task experience. When the medical skill block fails to match and no relevant historical task experience is found, perform regular task decomposition. The specialist agent is configured as follows: Before executing the assigned subtask, the system simultaneously receives the current session context from the short-term working memory, the historical task experience from the long-term contextual memory vector library, and the medical skill block from the skill and process memory library, and merges the three into a unified input context, and executes the subtask based on the input context. The clinical safety railing module is configured as follows: Receive treatment suggestions and their respective confidence scores from multiple specialist agents for the same subtask; If a mandatory rule conflicts with a certain treatment recommendation, the treatment recommendation is directly rejected, and the rejection result and reason are written into the consistency protocol log entry of the distributed Harness cluster. If there are no mandatory rule conflicts, the final output is determined by weighted voting based on confidence scores.
2. The system according to claim 1, characterized in that, The task decomposition and routing engine constrains the routable range of subtasks based on the data access permission level. Specifically, when the access permission level of a certain modality data in the development authorization code is prohibited, the task decomposition and routing engine automatically marks the specialist agent node corresponding to that modality as unreachable when generating or calling the task DAG, and reroutes the subtask to other specialist agents with alternative capability tags and passed permission verification, or triggers an external consultation request.
3. The system according to claim 1, characterized in that, The skill and process memory is also configured to simultaneously record the input data pattern and output data pattern of each node in the subtask process when the subtask process is solidified into a medical skill block, and generate a parameterized DAG template. The parameterized DAG template contains configurable parameter slots that are automatically filled with patient-specific parameters for a new task when invoked. The patient-specific parameters include at least the patient identifier, genotype locus, image sequence number, test index threshold, and a set of contraindications in the clinical safety context constraints.
4. The system according to claim 1, characterized in that, The contraindication rule in the clinical mandatory rule base is configured as follows: when any specialist agent outputs a treatment suggestion that includes drugs, procedures, or examinations corresponding to the target patient's past allergy history or absolute contraindications, it is directly rejected regardless of the confidence score of the treatment suggestion; the critical value priority rule is configured as follows: when any specialist agent identifies a physiological indicator that meets the clinical critical value criteria during processing, the normal execution order of the current task DAG is interrupted, the critical value processing sub-process is triggered first, and the critical value identifier is written into the consistency protocol log entry to trigger a global state change; the clinical guideline conflict rule is configured as follows: when a treatment suggestion has a directional conflict with the conclusion of a preset authoritative clinical guideline with a recommendation level of III or an evidence level of B or above, a manual review mark is triggered.
5. The system according to claim 1, characterized in that, The scheduler node in the distributed Harness cluster maintains the cluster state through the Raft protocol. The Raft protocol log entries record the execution state changes of each subtask in the task DAG, the corresponding medical data access audit trajectory, and the rejection records of the clinical safety guardrail module. When the scheduler node fails, the newly elected leader node restores the global execution state of the DAG and the incomplete medical data authorization links by replaying the Raft logs.
6. The system according to claim 1, characterized in that, When the specialist agent integrates the current session context, historical task experience, and medical skill blocks into a unified input context, it dynamically adjusts the weight ratio of the three components based on the medical type of the current subtask and the patient's current clinical status. The patient's current clinical status includes at least the presence of critical value alerts, whether it is a rare disease phenotype, and whether it is in the postoperative follow-up period. For diagnostic subtasks following fixed clinical pathways, the weight of the medical skill blocks is higher than that of historical task experience. For complex reasoning subtasks requiring differential diagnosis, the weight of historical task experience is higher than that of the medical skill blocks. For urgent subtasks with critical value alerts, the weight of the current session context in short-term working memory is higher than that of the medical skill blocks and the historical task experience.
7. The system according to claim 1, characterized in that, The scheduler node in the distributed Harness cluster adopts a logic execution and physical execution separation architecture: the scheduler node acts as a logic orchestrator, responsible for parsing clinical intent and generating a DAG containing subtask dependencies; the scheduler node acts as an execution coordinator, responsible for distributing specific execution instructions to specialist agents with corresponding capability tags and verified by the development authorization code, based on the DAG, the real-time load of the specialist agent, the data access permission level, and the clinical security context constraints.
8. A secure collaborative diagnosis and treatment method for cardiovascular and cerebrovascular multimodal data, characterized in that, Applied to the system as described in any one of claims 1-7, the method comprises: The multimodal secure access process involves receiving encrypted access requests from heterogeneous medical data sources, sequentially verifying the device hardware fingerprint, data traceability code, and model API interface authorization through a triple verification mechanism, generating a development authorization instruction code, and dynamically binding the development authorization instruction code to the data access permission level in the capability registry. The user's clinical task receiving steps include receiving clinical task requests containing natural language. The medical skill block matching step involves querying the skill and process memory to determine whether a medical skill block matches the new diagnosis and treatment task. First execution path: If the medical skill block is successfully matched, the medical skill block is directly invoked and executed, skipping the task decomposition step; Second execution path: If the medical skill block fails to match, query the long-term contextual memory vector library to retrieve similar historical task experience; if relevant historical task experience is found, adjust the granularity of task decomposition or the routing order of subtasks according to the historical task experience before performing decomposition. Third execution path: If the medical skill block fails to match and no relevant historical task experience is found, execute the regular task decomposition; In the context enhancement step, before executing a subtask, the specialist agent simultaneously acquires the current session context of the short-term working memory, the historical task experience of the long-term contextual memory vector library, and the medical skill blocks of the skill and process memory library, and merges them into a unified input context. In the distributed scheduling execution step, the distributed Harness cluster schedules the subtask to the optimal specialist agent for execution based on the current load of the specialist agent, the data access permission level, and the clinical security context constraints, and manages its lifecycle. The clinical safety decision-making process involves collecting the execution results of each subtask, querying the clinical mandatory rule base, and making mandatory rejections or weighted decisions based on preset strategies when there are contraindications, critical values, or guideline conflicts, and writing the rejection results into the consistency protocol log entries. In addition, the memory update step involves updating the key features of this task, execution summary, clinical outcome, rejected treatment recommendations and their reasons for rejection to the long-term contextual memory vector library and / or the skills and processes memory library after the final diagnosis and treatment results are generated.
9. The method according to claim 8, characterized in that, In the distributed scheduling execution step, the Harness cluster uses a consistent hashing algorithm to determine a unique execution agent from a group of specialist agents with the same capability tags and who have passed permission verification, based on the capability tags required by the subtask, the real-time load information of the specialist agent, the data access permission level in the development authorization code, and the set of contraindications in the clinical safety context constraints.
10. The system according to claim 1, characterized in that, The long-term contextual memory vector library is specifically used for: receiving the description information of the sub-task before the specialist agent executes the sub-task, converting it into a query vector, and retrieving the Top-K historical experiences with a similarity exceeding a preset threshold by calculating the cosine similarity between the query vector and the historical task vectors stored in the library, and inputting the historical experiences as part of the prompt word context into the specialist agent; wherein, the historical task experiences include multimodal data features of previous similar cases, task decomposition paths, agent scheduling sequences, final clinical outcomes, and treatment suggestions rejected by the clinical safety guardrail module and their reasons for rejection.