Comprehensive disease diagnosis method and system based on specialized diagnosis agent

Abstract

The invention relates to a disease comprehensive diagnosis method and system based on a specialized diagnosis agent, and aims at multi-source data, unifies a data interface, and achieves the real-time access of the multi-source data. Performing time-space dimension aggregation based on the key clinical anchor points of the clinical event, and constructing a clinical time axis; performing lightweight multi-modal feature extraction and alignment on the multi-source data accessed in real time; matching is carried out according to the aligned features, a corresponding specialized diagnosis agent is called to carry out comprehensive diagnosis reasoning, and the weights of evidences in different periods are dynamically adjusted when the specialized diagnosis agent carries out comprehensive diagnosis reasoning; the diagnosis result is visually presented, and the presented visual diagnosis result is modified according to user input. Feature level fusion and lightweight deployment are adopted, the method can be operated on existing server resources of a hospital, cloud supercomputing is not needed, the data privacy and compliance requirements are met, the misdiagnosis rate caused by data association errors can be effectively reduced through time sequence alignment, and the diagnosis accuracy can be greatly improved.

Description

Technical Field

[0001] This invention belongs to the medical field, specifically relating to a comprehensive disease diagnosis method and system based on a specialist diagnostic intelligent agent. Background Technology

[0002] Currently, the internal information systems of top-tier hospitals are highly heterogeneous, resulting in a severe "data fragmentation in time and space" dilemma. Specifically, this manifests as follows:

[0003] Multi-source heterogeneity and distributed storage: Patient medical records (HIS), medical images (PACS), test results (LIS / EMR) and other data are scattered in systems of different vendors and standards. The data structures, storage formats and interface protocols are different, making it difficult to directly integrate and input them into a large language model for unified analysis and reasoning.

[0004] Discontinuity in Clinical Temporal Logic: A patient's visit is a continuous process over time, but existing systems lack a unified model for the timeline of clinical events. General-purpose models, unable to understand the temporal dependencies of medical events, may incorrectly associate data from different points in time (such as forcibly associating old images from two years ago with current acute symptoms), leading to diagnostic conclusions that deviate significantly from clinical reality.

[0005] Limited computing resources within the hospital: In a private hospital deployment environment, computing resources are limited, making it difficult to directly process the long sequence of raw data from massive high-resolution images and the concatenation of large segments of text, resulting in high inference latency and poor practicality.

[0006] Furthermore, existing technical solutions are mostly simple data interface calls or offline fusion after the fact, lacking the ability to perform real-time, dynamic, and clinically logical spatiotemporal alignment and deep fusion of multi-source heterogeneous data during the diagnostic reasoning process. Summary of the Invention

[0007] To address the technical deficiencies in existing technologies, the purpose of this invention is to provide a comprehensive disease diagnosis method and system based on a specialized diagnostic intelligent agent. This system can bridge heterogeneous systems such as HIS, PACS, and LIS, automatically aggregating fragmented and multimodal patient data into a structured "patient clinical lifecycle timeline" that aligns with real diagnostic thinking, achieving logical data alignment. It avoids memory overflow or computation timeouts caused by splicing raw data and enables deep fusion of high-dimensional medical image features with text and numerical features even with limited computing resources (such as local hospital servers). Furthermore, it constructs a "specialized comprehensive diagnostic intelligent agent (agent)" that automatically adjusts diagnostic focus and reasoning strategies based on the temporal evolution and content characteristics of patient data, enhancing the scenario-specific relevance of diagnosis.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] In a first aspect, the present invention discloses a comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent, the method comprising the following steps:

[0010] S1. For multi-source data, a unified data interface is provided to enable real-time access to multi-source data;

[0011] S2. Aggregate spatiotemporal dimensions based on key clinical anchor points of clinical events to construct a clinical timeline;

[0012] S3. Perform lightweight multimodal feature extraction and alignment on real-time accessed multi-source data;

[0013] S4. Match the aligned features and call the corresponding specialist diagnostic agent to perform comprehensive diagnostic reasoning. When the specialist diagnostic agent performs comprehensive diagnostic reasoning, dynamically adjust the weight of evidence at different times.

[0014] S5. Visualize the diagnostic results and modify the visualized diagnostic results based on user input.

[0015] Furthermore, in step S1, the system connects to various hospital information systems, unifies the data interface, and enables real-time data stream capture.

[0016] Furthermore, step S2 includes the following sub-steps:

[0017] S21. Construct a virtual clinical timeline using clinical events as the smallest unit;

[0018] S22. Identify key clinical anchor points in the virtual clinical timeline;

[0019] S23. Perform time-weighted attenuation correlation on key clinical anchor points of clinical events to construct a clinical timeline.

[0020] Furthermore, in step S2, clinical events include registration, ordering tests, imaging examinations, medication, and surgery. Each clinical event includes a timestamp, type, system, raw data pointer, and key entity.

[0021] Furthermore, key clinical anchors in step S2 include the date of diagnosis, the date of surgery, the start date of the chemotherapy cycle, and the date of the first appearance of abnormal indicators.

[0022] Furthermore, in step S23, a time decay function is introduced to perform time-weighted decay correlation on key clinical anchors of clinical events, so that the correlation weight between the current chief complaint and recent examination results is much higher than that of long-term historical records, thereby enabling the specialist diagnostic agent to pay attention to recent evidence and refer to past history.

[0023] Furthermore, step S3 includes the following sub-steps:

[0024] S31. Perform modal lightweight feature extraction on real-time accessed multi-source data;

[0025] S32. Align the extracted features within the same medical semantic space.

[0026] Furthermore, step S4 includes the following sub-steps:

[0027] S41. Construct and train a specialty knowledge base and a specialty diagnostic intelligence library;

[0028] S42. Match the aligned feature vector with the prompt word template in the specialty knowledge base, automatically identify the scene and dispatch the corresponding specialty diagnostic agent;

[0029] S43. The specialized diagnostic agent dynamically adjusts the weights of evidence at different times and makes diagnostic inferences accordingly.

[0030] Furthermore, in step S43, the specialist diagnostic agent not only refers to the current characteristics, but also goes back along the clinical timeline. Based on the key clinical anchors and attenuation weights of clinical events, it dynamically adjusts the degree of importance attached to evidence from different periods, generating a comprehensive diagnostic report and differential diagnosis suggestions with time dimension considerations.

[0031] Secondly, the present invention discloses a comprehensive disease diagnosis system based on a specialist diagnostic agent. The system employs a comprehensive disease diagnosis method based on a specialist diagnostic agent as described in the first aspect of the present invention and any optional embodiment thereof, to perform spatiotemporal alignment of multimodal medical data and schedule specialist diagnostic agents for diagnostic inference. The system includes:

[0032] The multi-source data access and adaptation module is used to connect with various hospital information systems. It provides a unified data interface for multi-source data and enables real-time access to multi-source data.

[0033] The clinical timeline construction module is used to aggregate spatiotemporal dimensions based on key clinical anchor points of clinical events to construct a clinical timeline.

[0034] The feature extraction and alignment module is used to perform lightweight multimodal feature extraction and alignment on real-time accessed multi-source data;

[0035] The agent scheduling and reasoning module is used to match based on the aligned features and call the corresponding specialist diagnostic agents to perform comprehensive diagnostic reasoning. When the specialist diagnostic agents perform comprehensive diagnostic reasoning, they dynamically adjust the weight of evidence at different times.

[0036] The visualization module is used to visualize the diagnostic results and modify the visualized diagnostic results based on user input.

[0037] The beneficial technical effects of the present invention are as follows: The disease comprehensive diagnosis method and system based on a specialist diagnostic intelligent agent disclosed in the present invention adopts feature-level fusion and lightweight deployment, can run on existing hospital server resources, does not need to rely on cloud supercomputing, meets data privacy and compliance requirements, and can significantly improve engineering feasibility.

[0038] By aligning time sequences, AI diagnosis becomes more consistent with clinical logic. In particular, in time-dependent scenarios such as chronic disease management and tumor follow-up, it can effectively reduce the misdiagnosis rate caused by data association errors and significantly improve diagnostic accuracy.

[0039] The solution is deeply coupled with the complex information system interfaces and business process logic within the hospital. General-purpose large-scale model vendors find it difficult to quickly replicate through standardized APIs, forming a deeply customized moat based on the hospital's private deployment scenario, which is conducive to establishing a deep competitive barrier.

[0040] The generated results are based on a structured timeline and feature alignment process, making the source of diagnostic conclusions (when and what data) more transparent, easier for doctors to understand and review, and enhancing the interpretability of the system. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating a comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent, as shown in Embodiment 1 of the present invention. Detailed Implementation

[0042] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment of the invention provides a comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent, the method comprising the following steps:

[0045] S1. Multi-source data access and adaptation: Connects to various hospital information systems (HIS, PACS, LIS, etc.), with a unified data interface, to achieve real-time / near real-time data stream capture.

[0046] S2. Aggregate spatiotemporal dimensions based on clinical event anchors to construct a clinical timeline.

[0047] Based on clinical event anchors, a spatiotemporal dimension aggregation algorithm is used to reorganize fragmented data into a structured timeline.

[0048] Step S2 includes the following sub-steps:

[0049] S21. Using clinical events as the smallest unit, model a virtual clinical timeline.

[0050] The smallest unit is defined as a "clinical event". Clinical events include things like registration, ordering tests, imaging examinations, medication, and surgery. Each clinical event contains a timestamp, type, system to which it belongs, raw data pointer, and key entities.

[0051] S22. Identify key clinical anchor points in the virtual clinical timeline;

[0052] Key nodes (such as the date of diagnosis, surgery date, chemotherapy cycle start date, and the first appearance date of abnormal indicators) in the virtual clinical timeline are automatically identified using rules and lightweight models and used as critical clinical anchors. These critical clinical anchors divide the virtual clinical timeline into stages with different clinical significance.

[0053] S23. Perform time-weighted decay correlation on key clinical anchor points of clinical events.

[0054] Introducing a time decay function ,in This is the difference between the current time and the time of the clinical event. This is the attenuation coefficient (which can be adjusted according to disease type). Current chief complaint and recent ( The correlation weight of the examination results is much higher than that of the long-term historical records, so that the diagnostic agent can simulate the doctor's logic of "emphasizing recent evidence and referring to past history".

[0055] S3. Perform lightweight multimodal feature extraction and alignment: process image, text, and numerical data separately, extract multimodal features and map them to a unified medical semantic space to align multimodal features.

[0056] Step S3 includes the following sub-steps:

[0057] S31. Perform lightweight feature extraction by modality.

[0058] Deploy a lightweight visual encoder (such as a compressed version of ViT or CNN) on the image processing end to convert DICOM images into dense feature vectors carrying information such as anatomical locations and lesion features.

[0059] Using a medical-specific NER model and encoder on the text / numerical side, entity vectors (such as drugs, diseases, and test indicators) are extracted from HIS text and LIS numerical data.

[0060] S32. Align the extracted features in a unified medical semantic space:

[0061] We design a multimodal alignment projection layer that, through training on a small-scale multimodal medical corpus, maps feature vectors from images, text, and numerical data to the same low-dimensional, dense medical semantic space. Within this medical semantic space, features from different modalities describing the same clinical concept are close in distance.

[0062] This avoids directly inputting raw pixels and long text into a large model, resulting in a significant reduction in computational load and memory usage, and achieving efficient feature-level fusion.

[0063] S4. Dynamic scheduling and reasoning of specialized comprehensive diagnostic agents: Based on the aligned features, call the corresponding specialized diagnostic agents to perform comprehensive diagnosis.

[0064] Step S4 includes the following sub-steps:

[0065] S41. Construct and train a specialty knowledge base and a specialty diagnostic intelligence library;

[0066] For multiple specific disease areas (such as cardiology, oncology, and endocrinology), "specialty diagnostic agents" and prompt word templates are constructed and trained respectively.

[0067] S42. Match the aligned feature vectors with the prompt word templates in the specialty knowledge base, automatically identify the scene, and dispatch the corresponding specialty diagnostic agent.

[0068] The aligned unified feature vectors are coarsely classified and matched with prompt word templates in the specialty knowledge base. For example, when the feature vectors contain high weights for concepts such as "abnormal cardiac ultrasound structure," "elevated BNP," and "troponin," the heart failure specialty diagnostic agent is automatically scheduled; when the features show "pulmonary nodules," "elevated tumor markers," or "pathology report keywords," the respiratory oncology specialty diagnostic agent is scheduled.

[0069] S43. Dynamically adjust the weight of evidence from different periods and make diagnostic inferences accordingly.

[0070] The scheduled specialist diagnostic agent not only refers to the current features, but also goes back along the time axis. Based on the event anchor point and decay weight, it dynamically adjusts the degree of importance attached to evidence in different periods, and generates a comprehensive diagnostic report and differential diagnosis suggestions with time dimension consideration.

[0071] S5. Results Visualization and Interaction: Visualize diagnostic results such as timeline, diagnostic conclusions, evidence, and confidence levels, and modify the presented visual diagnostic results based on user input.

[0072] Example 2

[0073] This invention provides a comprehensive disease diagnosis method system based on a specialist diagnostic agent. The system employs the comprehensive disease diagnosis method based on a specialist diagnostic agent described in Embodiment 1 and its optional embodiments to perform spatiotemporal alignment of multimodal medical data and schedule specialist diagnostic agents for diagnostic inference. The system includes:

[0074] The multi-source data access and adaptation module is used to connect with various hospital information systems (HIS, PACS, LIS, etc.), unify the data interface, and realize real-time / near real-time data stream capture.

[0075] The clinical timeline construction module is used to aggregate spatiotemporal dimensions based on clinical event anchors to construct a clinical timeline.

[0076] Based on clinical event anchors, a spatiotemporal dimension aggregation algorithm is used to reorganize fragmented data into a structured timeline.

[0077] The clinical timeline construction module includes the following sub-modules:

[0078] The Virtual Clinical Timeline Modeling Submodule is used to model a virtual clinical timeline, with clinical events as the smallest unit.

[0079] The smallest unit is defined as a "clinical event". Clinical events include things like registration, ordering tests, imaging examinations, medication, and surgery. Each clinical event contains a timestamp, type, system to which it belongs, raw data pointer, and key entities.

[0080] The key clinical anchor identification submodule is used to identify key clinical anchors in the virtual clinical timeline;

[0081] Key nodes (such as the date of diagnosis, surgery date, chemotherapy cycle start date, and the first appearance date of abnormal indicators) in the virtual clinical timeline are automatically identified using rules and lightweight models and used as critical clinical anchors. These critical clinical anchors divide the virtual clinical timeline into stages with different clinical significance.

[0082] The time-weighted decay correlation submodule is used to perform time-weighted decay correlation on key clinical anchors of clinical events.

[0083] A time decay function is introduced, where is the difference between the current time and the clinical event time, and is the decay coefficient (which can be adjusted according to the disease type). The correlation weight between the current chief complaint and recent (major) examination results is much higher than that of long-term historical records, thus enabling the diagnostic agent to simulate the doctor's logic of "emphasizing recent evidence and referring to past history".

[0084] The feature extraction and alignment module is used for lightweight multimodal feature extraction and alignment: it processes image, text, and numerical data respectively, extracts multimodal features and maps them to a unified medical semantic space to align multimodal features.

[0085] The feature extraction and alignment module includes the following sub-modules:

[0086] The feature extraction submodule is used for lightweight feature extraction based on modality.

[0087] Deploy a lightweight visual encoder (such as a compressed version of ViT or CNN) on the image processing end to convert DICOM images into dense feature vectors carrying information such as anatomical locations and lesion features.

[0088] Using a medical-specific NER model and encoder on the text / numerical side, entity vectors (such as drugs, diseases, and test indicators) are extracted from HIS text and LIS numerical data.

[0089] The feature alignment submodule is used to align the extracted features in a unified medical semantic space.

[0090] We design a multimodal alignment projection layer that, through training on a small-scale multimodal medical corpus, maps feature vectors from images, text, and numerical data to the same low-dimensional, dense medical semantic space. Within this medical semantic space, features from different modalities describing the same clinical concept are close in distance.

[0091] This avoids directly inputting raw pixels and long text into a large model, resulting in a significant reduction in computational load and memory usage, and achieving efficient feature-level fusion.

[0092] The agent scheduling and reasoning module is used for the dynamic scheduling and reasoning of specialized comprehensive diagnostic agents: based on the aligned features, it calls the corresponding specialized diagnostic agents to perform comprehensive diagnosis.

[0093] The agent scheduling and inference module includes the following sub-modules:

[0094] The Specialty Diagnostic Intelligent Body Library Construction Submodule is used to build and train the specialty knowledge base and the specialty diagnostic intelligent body library;

[0095] For multiple specific disease areas (such as cardiology, oncology, and endocrinology), "specialty diagnostic agents" and prompt word templates are constructed and trained respectively.

[0096] The Specialty Diagnostic Agent Library Scheduling Submodule is used to match the aligned feature vectors with the prompt word templates in the specialty knowledge base, automatically identify the scenario, and schedule the corresponding specialty diagnostic agents.

[0097] The aligned unified feature vectors are coarsely classified and matched with prompt word templates in the specialty knowledge base. For example, when the feature vectors contain high weights for concepts such as "abnormal cardiac ultrasound structure," "elevated BNP," and "troponin," the heart failure specialty diagnostic agent is automatically scheduled; when the features show "pulmonary nodules," "elevated tumor markers," or "pathology report keywords," the respiratory oncology specialty diagnostic agent is scheduled.

[0098] The diagnostic reasoning submodule is used to dynamically adjust the weight of evidence at different times and perform diagnostic reasoning accordingly.

[0099] The scheduled specialist diagnostic agent not only refers to the current features, but also goes back along the time axis. Based on the event anchor point and decay weight, it dynamically adjusts the degree of importance attached to evidence in different periods, and generates a comprehensive diagnostic report and differential diagnosis suggestions with time dimension consideration.

[0100] The visualization module is used to visualize and interact with the results: it visualizes diagnostic results such as timelines, diagnostic conclusions, evidence, and confidence levels, and allows modification of the visualized diagnostic results based on user input.

[0101] As can be seen from the above embodiments, the disease comprehensive diagnosis method and system based on specialized diagnostic agents disclosed in this invention dynamically aggregates multi-source heterogeneous medical data based on clinical event anchor point identification and time decay weighting, thereby constructing a virtual clinical timeline. Through lightweight feature extraction and a multimodal alignment projection layer, multimodal feature-level fusion for hospital intranets is achieved, reducing computational consumption. Clinical scenarios are automatically identified based on unified semantic features, and dedicated inference models are switched accordingly, dynamically scheduling specialized diagnostic agents. A complete integrated implementation plan is provided, from data access, timeline construction, feature alignment fusion to agent scheduling and inference.

[0102] The methods and systems described in this invention are not limited to the embodiments described in the specific implementation. Other implementation methods derived by those skilled in the art based on the technical solutions of this invention also fall within the scope of technical innovation of this invention.

Claims

1. A comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent, characterized in that: The method includes the following steps: S1. For multi-source data, a unified data interface is provided to enable real-time access to multi-source data; S2. Aggregate spatiotemporal dimensions based on key clinical anchor points of clinical events to construct a clinical timeline; S3. Perform lightweight multimodal feature extraction and alignment on real-time accessed multi-source data; S4. Match the aligned features and call the corresponding specialist diagnostic agent to perform comprehensive diagnostic reasoning. When the specialist diagnostic agent performs comprehensive diagnostic reasoning, dynamically adjust the weight of evidence at different times. S5. Visualize the diagnostic results and modify the visualized diagnostic results based on user input.

2. The comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent according to claim 1, characterized in that: In step S1, the system connects to various hospital information systems, unifies the data interface, and enables real-time data stream capture.

3. The comprehensive disease diagnosis method based on a specialized diagnostic intelligent agent according to claim 1, characterized in that, Step S2 includes the following sub-steps: S21. Construct a virtual clinical timeline using clinical events as the smallest unit; S22. Identify key clinical anchor points in the virtual clinical timeline; S23. Perform time-weighted attenuation correlation on key clinical anchor points of clinical events to construct a clinical timeline.

4. The comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent according to claim 1, characterized in that: In step S2, clinical events include registration, ordering tests, imaging examinations, medication, and surgery. Each clinical event includes a timestamp, type, system, raw data pointer, and key entity.

5. The comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent according to claim 1, characterized in that: Key clinical anchors in step S2 include the date of diagnosis, the date of surgery, the start date of the chemotherapy cycle, and the date of the first appearance of abnormal indicators.

6. A comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent according to claim 3, characterized in that: In step S23, a time decay function is introduced to perform time-weighted decay correlation on key clinical anchors of clinical events, so that the correlation weight between the current chief complaint and recent examination results is much higher than that of long-term historical records, thereby making the specialist diagnostic agent pay attention to recent evidence and refer to past history.

7. A comprehensive disease diagnosis method based on a specialized diagnostic intelligent agent according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Perform modal lightweight feature extraction on real-time accessed multi-source data; S32. Align the extracted features within the same medical semantic space.

8. A comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent according to claim 1, characterized in that, Step S4 includes the following sub-steps: S41. Construct and train a specialty knowledge base and a specialty diagnostic intelligence library; S42. Match the aligned feature vector with the prompt word template in the specialty knowledge base, automatically identify the scene and dispatch the corresponding specialty diagnostic agent; S43. The specialized diagnostic agent dynamically adjusts the weights of evidence at different times and makes diagnostic inferences accordingly.

9. A comprehensive disease diagnosis method based on a specialist diagnostic intelligent agent according to claim 8, characterized in that: In step S43, the specialist diagnostic agent not only refers to the current characteristics, but also goes back along the clinical timeline. Based on the key clinical anchors and attenuation weights of clinical events, it dynamically adjusts the degree of importance attached to evidence in different periods, and generates a comprehensive diagnostic report and differential diagnosis suggestions with time dimension considerations.

10. A comprehensive disease diagnosis system based on a specialist diagnostic intelligent agent, characterized in that, The system employs a comprehensive disease diagnosis method based on a specialist diagnostic agent, as described in any one of claims 1-9, to perform spatiotemporal alignment of multimodal medical data and schedule the specialist diagnostic agent for diagnostic inference. The system includes: The multi-source data access and adaptation module is used to connect with various hospital information systems. It provides a unified data interface for multi-source data and enables real-time access to multi-source data. The clinical timeline construction module is used to aggregate spatiotemporal dimensions based on key clinical anchor points of clinical events to construct a clinical timeline. The feature extraction and alignment module is used for lightweight multimodal feature extraction and alignment of real-time accessed multi-source data; The agent scheduling and reasoning module is used to match based on the aligned features and call the corresponding specialist diagnostic agents to perform comprehensive diagnostic reasoning. When the specialist diagnostic agents perform comprehensive diagnostic reasoning, they dynamically adjust the weight of evidence at different times. The visualization module is used to visualize the diagnostic results and modify the visualized diagnostic results based on user input.

Images