A traditional chinese medicine multi-task large model intelligent diagnosis and treatment system and method

Abstract

The application discloses a traditional Chinese medicine multi-task large model intelligent diagnosis and treatment system and method, and relates to the technical field of artificial intelligence and large models. The system comprises an end-side data acquisition module, a cross-modal feature fusion module, a cloud-side syndrome differentiation intention understanding module, a cloud-edge collaborative scheduling module, an edge-side lightweight interaction module, a hierarchical diagnosis and treatment decision module, a cross-scene adaptation module and a closed-loop self-learning iteration module. The application realizes efficient allocation of computing power through a cloud-edge collaborative architecture, improves the response speed of the edge side while ensuring deep diagnosis and treatment reasoning. The introduced knowledge transfer mechanism enables the model to quickly adapt to different clinical disease syndrome scenes, improving the versatility of the system. Multi-modal fusion and sequential task decomposition significantly enhance the intelligence of human-computer interaction, enabling the system to more accurately understand the syndrome differentiation intention and optimizing the accuracy and robustness of the whole process of traditional Chinese medicine diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large model technology, specifically to a multi-task large model intelligent diagnosis and treatment system and method for traditional Chinese medicine. Background Technology

[0002] With the development of artificial intelligence and human-computer interaction, the intelligentization of traditional Chinese medicine (TCM) has become an industry trend. However, the TCM diagnosis and treatment environment is complex, involving the collection and comprehensive analysis of multimodal information from observation, auscultation, inquiry, and palpation. Currently, TCM intelligent systems face challenges such as the diversity of disease types and uneven distribution of computing resources, resulting in low levels of intelligence in edge devices and difficulty in handling complex diagnostic and treatment tasks. Furthermore, existing models have poor transferability, making it difficult to achieve a balance between deep reasoning in the cloud and rapid interaction at the edge. Therefore, how to build a cloud-edge collaborative system that integrates perception, cognition, reasoning, and decision-making, and possesses efficient knowledge transfer capabilities, has become a key bottleneck in the digital transformation of TCM.

[0003] Existing TCM diagnosis and treatment systems mainly adopt offline architectures based on simple rule bases or single deep learning models. Their drawbacks are as follows: First, the lack of cloud-edge collaboration mechanisms makes it difficult to deploy complex, large-scale models in resource-constrained treatment settings (edge ​​devices), resulting in slow human-computer interaction responses. Second, the single task processing mode prevents the sequential parallel processing of consultation, syndrome differentiation, prescription, and efficacy evaluation. Third, the models lack deep predictive capabilities regarding TCM treatment intentions; the fusion of multimodal information such as tongue and pulse images and case texts is limited to feature stitching, leading to significant susceptibility of syndrome differentiation accuracy to individual differences and insufficient robustness. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-task large-scale intelligent diagnosis and treatment system and method for traditional Chinese medicine, which solves the problems of lack of cloud-edge collaboration mechanism, single task processing mode, and lack of deep predictive ability of model to predict the diagnosis and treatment intention of traditional Chinese medicine.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multi-task large-scale intelligent diagnosis and treatment system for traditional Chinese medicine includes an edge data acquisition module, a cross-modal feature fusion module, a cloud-based syndrome differentiation intent understanding module, a cloud-edge collaborative scheduling module, an edge-end lightweight interaction module, a hierarchical diagnosis and treatment decision module, a cross-scenario adaptation module, and a closed-loop self-learning iteration module.

[0007] The edge data acquisition module is used to collect multi-source TCM multimodal data from patients in real time. The multimodal data includes tongue image data, facial color data, consultation voice data, and case text data. The edge data acquisition module has a built-in multi-algorithm feature parsing unit. It uses the visual Transformer feature extraction algorithm to extract visual features from the tongue image data and facial color data, uses the Wav2Vec 2.0 voice representation algorithm to process voice features from the consultation voice data, and uses a bidirectional coding representation model to perform structured parsing of the case text data, thus completing the preliminary feature extraction and standardization preprocessing of the multimodal raw data.

[0008] The cross-modal feature fusion module is connected to the edge data acquisition module and has a built-in adaptive cross-modal attention mechanism. It is used to perform heterogeneous feature fusion on preprocessed visual features, speech features, and text structured features, and to convert multi-source heterogeneous multimodal data into a unified structured semantic embedding space and output standardized semantic information.

[0009] The cloud-based syndrome differentiation intention understanding module is equipped with a large-scale TCM vertical model. The TCM vertical model adopts a hybrid expert architecture to receive the structured semantic embedding information, dynamically predict the patient's pathogenesis and syndrome intention, and complete the TCM syndrome differentiation analysis.

[0010] The cloud-edge collaborative scheduling module has a built-in task allocation strategy based on neural architecture search, which is used to perform hierarchical scheduling of diagnosis and treatment tasks, allocate computationally intensive deep dialectical reasoning tasks to the cloud for execution, and migrate real-time human-computer consultation and interaction logic tasks in a lightweight manner.

[0011] The edge lightweight interaction module uses a quantized LLaMA architecture to build a lightweight model and migrates cloud interaction logic capabilities through a knowledge distillation algorithm. It is used to run real-time consultation interaction tasks in a low-computing-power edge environment and achieve millisecond-level diagnosis and treatment command response.

[0012] The hierarchical diagnosis and treatment decision module has a built-in sequential task hierarchical decomposition mechanism and a thinking chain reasoning algorithm, which is used to automatically break down the overall TCM diagnosis and treatment process into four orderly and connected execution links: consultation guidance, syndrome differentiation analysis, prescription recommendation, and efficacy evaluation, so as to complete the output of the entire process of diagnosis and treatment decision;

[0013] The cross-scenario adaptation module integrates a low-rank adaptive fine-tuning algorithm, which is used to transfer specialized knowledge based on general TCM diagnosis and treatment knowledge, quickly adapt to the disease and syndrome patterns of different clinical specialties, and realize cross-scenario diagnosis and treatment applications.

[0014] The closed-loop self-learning iterative module adopts a multi-task learning framework to achieve real-time data synchronization and dynamic model evolution between the cloud and the edge. It uses reinforcement learning algorithms to align preferences with real-time diagnosis and treatment feedback from the edge, continuously optimizing the diagnosis and treatment decision logic of the cloud-based TCM vertical model, thus forming a self-learning closed-loop diagnosis and treatment system.

[0015] Furthermore, the Wav2Vec 2.0 speech representation algorithm is adapted to TCM consultation speech scenarios, and optimizes the modeling of patient consultation speech features such as uneven speech speed, dialect accent, and colloquial expression, and accurately extracts core semantic features of consultation speech such as patient complaints and symptom descriptions.

[0016] Furthermore, the adaptive cross-modal attention mechanism can dynamically adjust the weight ratio of different modal features, automatically increase the weight of effective modal features and weaken the weight of ineffective interference features according to the patient diagnosis and treatment scenario, thereby optimizing the accuracy of heterogeneous data fusion.

[0017] Furthermore, the hybrid expert architecture of the TCM vertical category model includes a general TCM knowledge expert submodule, a specialized disease and syndrome expert submodule, and a syndrome differentiation and reasoning expert submodule. These submodules work collaboratively, dynamically activating the corresponding expert submodule based on input semantic information to achieve accurate prediction of pathogenesis and syndromes.

[0018] The edge lightweight model optimizes the native LLaMA architecture by using model quantization and knowledge distillation, simplifying parameters and accelerating inference. While retaining the core consultation and interaction capabilities, it is adapted to low-computing edge devices such as mobile phones and portable medical terminals.

[0019] Furthermore, the task allocation strategy of the neural architecture search makes intelligent decisions based on the computing power resource thresholds of the cloud and edge, task latency requirements, and computational complexity parameters to achieve optimal cloud-edge allocation of diagnostic and treatment tasks.

[0020] This invention also provides the following technical solutions:

[0021] A multi-task large-scale intelligent diagnosis and treatment method for traditional Chinese medicine, applied to the system described in claim 1, includes the following steps:

[0022] S1. Real-time acquisition and preprocessing of multimodal data from the end device: Real-time acquisition of multi-source TCM data such as patient tongue appearance, facial color, consultation voice, and case text through the end device. Visual features of tongue appearance and facial color are extracted using the visual Transformer feature extraction algorithm. The consultation voice is processed using the Wav2Vec 2.0 voice representation algorithm to obtain voice representation features. The case text is structured using a bidirectional coding representation model to output multi-dimensional preprocessed feature data.

[0023] S2. Adaptive cross-modal feature fusion: The preprocessed visual, speech, and text heterogeneous features are fused through an adaptive cross-modal attention mechanism to eliminate feature differences between different modal data and map heterogeneous data to a structured semantic embedding space to generate standardized diagnostic semantic information.

[0024] S3, Cloud-based Large Model for Understanding Syndrome Differentiation Intent: Based on a large model of TCM vertical categories deployed in the cloud and using a hybrid expert architecture, it receives standardized semantic information, dynamically infers the patient's pathogenesis, accurately identifies the patient's syndrome intent, and outputs syndrome differentiation results.

[0025] S4. Neural Architecture Search-Driven Cloud-Edge Collaborative Task Scheduling: By using a task allocation strategy based on neural architecture search, diagnostic tasks are classified according to computing power. High-computing-power-consuming deep dialectical reasoning tasks are kept in the cloud for execution, while low-latency human-computer consultation interaction logic is migrated to the edge lightweight model through a knowledge distillation algorithm.

[0026] S5, Millisecond-level response with low computing power at the edge: The edge-based lightweight model is built on the quantized LLaMA architecture and runs the migrated consultation interaction logic to achieve millisecond-level response to diagnosis and treatment commands in a low computing power terminal environment, thus completing real-time human-computer interaction consultation.

[0027] S6. Hierarchical Sequential Diagnosis and Treatment Decision Execution: By combining a hierarchical decomposition mechanism of sequential tasks with a thinking chain reasoning algorithm, the diagnosis and treatment process is broken down into four sequential execution links: consultation guidance, syndrome differentiation analysis, prescription and drug recommendation, and efficacy evaluation. Intelligent diagnosis and treatment decisions and outputs are completed step by step.

[0028] S7. Cross-specialty scenario adaptive adaptation: The model knowledge transfer is completed through the low-rank adaptive fine-tuning algorithm. Based on the general TCM diagnosis and treatment model, it is adapted to the disease and syndrome characteristics and diagnosis and treatment rules of different clinical specialties to realize the implementation of diagnosis and treatment in multiple scenarios.

[0029] S8, Cloud-Edge Closed-Loop Self-Learning Iteration: Through a multi-task learning framework, data synchronization between the cloud and the edge is achieved, real-time diagnosis and treatment feedback data from the edge is collected, reinforcement learning algorithms are used to align diagnosis and treatment preferences, and the diagnosis and treatment decision logic of the large cloud model is continuously iterated and optimized to achieve autonomous learning and capability upgrade of the system.

[0030] Furthermore, the visual Transformer feature extraction algorithm is optimized for TCM tongue and facial color diagnosis scenarios, focusing on core TCM diagnostic features such as tongue color, tongue coating, facial complexion, and facial texture, eliminating irrelevant image interference features, and achieving accurate extraction of TCM visual diagnostic features.

[0031] Furthermore, the low-rank adaptive fine-tuning algorithm only fine-tunes a small number of core parameters of the model, without the need for full parameter reconstruction, and quickly completes the transfer and adaptation of knowledge of different specialties and diseases, reducing the cost of cross-scenario deployment.

[0032] Furthermore, the reinforcement learning algorithm uses clinical diagnosis accuracy, prescription suitability, and patient efficacy feedback as reward functions to continuously align the model's diagnosis preferences and iteratively optimize the accuracy of syndrome differentiation and the rationality of diagnosis and treatment decisions.

[0033] Furthermore, the multi-task learning framework simultaneously supports multiple parallel tasks such as data synchronization, model updates, task scheduling, and feedback optimization, ensuring that the cloud-based model iterative upgrades do not affect the normal operation of real-time diagnosis and treatment services on the edge.

[0034] The beneficial effects of this invention are as follows: This application achieves efficient allocation of computing power through a cloud-edge collaborative architecture, improving edge response speed while ensuring deep diagnostic reasoning. The introduced knowledge transfer mechanism enables the model to quickly adapt to different clinical disease scenarios, improving the system's versatility. Multimodal fusion and sequential task decomposition significantly enhance the intelligence of human-computer interaction, enabling the system to more accurately understand diagnostic intentions and optimizing the accuracy and robustness of the entire TCM diagnosis and treatment process.

[0035] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the overall process of an embodiment of the present invention. Detailed Implementation

[0037] The technical solution of the present invention 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 the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0039] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0040] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0041] Please see Figure 1 A preferred embodiment of this application shows a multi-task large-scale intelligent diagnosis and treatment system for traditional Chinese medicine, including an edge data acquisition module, a cross-modal feature fusion module, a cloud-based syndrome differentiation intent understanding module, a cloud-edge collaborative scheduling module, an edge-end lightweight interaction module, a hierarchical diagnosis and treatment decision module, a cross-scenario adaptation module, and a closed-loop self-learning iteration module.

[0042] The edge data acquisition module is used to collect multi-source TCM multimodal data from patients in real time. The multimodal data includes tongue image data, facial color data, consultation voice data, and case text data. The edge data acquisition module has a built-in multi-algorithm feature parsing unit. It uses the visual Transformer feature extraction algorithm to extract visual features from tongue image data and facial color data, uses the Wav2Vec 2.0 voice representation algorithm to process voice features from consultation voice data, and uses a bidirectional coding representation model to perform structured parsing of case text data, thus completing the preliminary feature extraction and standardization preprocessing of the multimodal raw data.

[0043] The cross-modal feature fusion module connects to the edge data acquisition module and has a built-in adaptive cross-modal attention mechanism. It is used to perform heterogeneous feature fusion on preprocessed visual features, speech features, and text structured features, and convert multi-source heterogeneous multimodal data into a unified structured semantic embedding space to output standardized semantic information.

[0044] The cloud-based diagnostic intent understanding module is equipped with a large-scale TCM vertical model. This TCM vertical model adopts a hybrid expert architecture to receive structured semantic embedded information, dynamically predict the patient's pathogenesis and syndrome intent, and complete TCM diagnostic analysis.

[0045] The cloud-edge collaborative scheduling module has a built-in task allocation strategy based on neural architecture search, which is used to schedule diagnosis and treatment tasks in a hierarchical manner, allocate computationally intensive deep dialectical reasoning tasks to the cloud for execution, and migrate real-time human-computer consultation and interaction logic tasks in a lightweight manner.

[0046] The edge lightweight interaction module uses a quantized LLaMA architecture to build a lightweight model and migrates cloud interaction logic capabilities through a knowledge distillation algorithm. It is used to run real-time consultation interaction tasks in low-computing edge environments and achieve millisecond-level diagnosis and treatment command response.

[0047] The hierarchical diagnosis and treatment decision module has a built-in sequential task hierarchical decomposition mechanism and thinking chain reasoning algorithm, which is used to automatically break down the overall TCM diagnosis and treatment process into four orderly and connected execution links: consultation guidance, syndrome differentiation analysis, prescription recommendation, and efficacy evaluation, and complete the output of diagnosis and treatment decisions throughout the entire process.

[0048] The cross-scenario adaptation module integrates a low-rank adaptive fine-tuning algorithm, which is used to transfer specialized knowledge based on general TCM diagnosis and treatment knowledge, quickly adapt to the disease and syndrome patterns of different clinical specialties, and realize cross-scenario diagnosis and treatment applications.

[0049] The closed-loop self-learning iterative module adopts a multi-task learning framework to achieve real-time data synchronization and dynamic model evolution between the cloud and the edge. It uses reinforcement learning algorithms to align preferences with real-time diagnosis and treatment feedback from the edge, continuously optimizing the diagnosis and treatment decision logic of the cloud-based TCM vertical model, thus forming a self-learning closed-loop diagnosis and treatment system.

[0050] The Wav2Vec 2.0 speech representation algorithm is adapted to TCM consultation speech scenarios. It optimizes the modeling of patient speech characteristics such as uneven speech speed, regional accents, and colloquial expressions, and accurately extracts core semantic features of patient speech such as chief complaints and symptom descriptions.

[0051] The adaptive cross-modal attention mechanism can dynamically adjust the weight ratio of different modal features, automatically increase the weight of effective modal features and weaken the weight of ineffective interference features according to the patient diagnosis and treatment scenario, and optimize the accuracy of heterogeneous data fusion.

[0052] The hybrid expert architecture of the TCM vertical category big data model includes a general TCM knowledge expert submodule, a specialized disease and syndrome expert submodule, and a syndrome differentiation and reasoning expert submodule. These submodules work collaboratively, dynamically activating the corresponding expert submodule based on the input semantic information to achieve accurate prediction of pathogenesis and syndrome.

[0053] The edge-lightweight model optimizes the native LLaMA architecture by simplifying parameters and accelerating inference through model quantization and knowledge distillation. While retaining the core consultation and interaction capabilities, it is adapted to low-computing edge devices such as mobile phones and portable medical terminals.

[0054] The task allocation strategy of neural architecture search makes intelligent decisions based on cloud and edge computing resource thresholds, task latency requirements, and computational complexity parameters to achieve optimal cloud-edge allocation of diagnostic and treatment tasks.

[0055] This invention also provides the following technical solutions:

[0056] A multi-task large-scale intelligent diagnosis and treatment method for traditional Chinese medicine, applied to the system of claim 1, includes the following steps:

[0057] S1. Real-time acquisition and preprocessing of multimodal data from the end device: Real-time acquisition of multi-source TCM data such as patient tongue appearance, facial color, consultation voice, and case text through the end device. Visual features of tongue appearance and facial color are extracted using the visual Transformer feature extraction algorithm. The consultation voice is processed using the Wav2Vec 2.0 voice representation algorithm to obtain voice representation features. The case text is structured using a bidirectional coding representation model to output multi-dimensional preprocessed feature data.

[0058] S2. Adaptive cross-modal feature fusion: The preprocessed visual, speech, and text heterogeneous features are fused through an adaptive cross-modal attention mechanism to eliminate feature differences between different modal data and map heterogeneous data to a structured semantic embedding space to generate standardized diagnostic semantic information.

[0059] S3, Cloud-based Large Model for Understanding Syndrome Differentiation Intent: Based on a large model of TCM vertical categories deployed in the cloud and using a hybrid expert architecture, it receives standardized semantic information, dynamically infers the patient's pathogenesis, accurately identifies the patient's syndrome intent, and outputs syndrome differentiation results.

[0060] S4. Neural Architecture Search-Driven Cloud-Edge Collaborative Task Scheduling: By using a task allocation strategy based on neural architecture search, diagnostic tasks are classified according to computing power. High-computing-power-consuming deep dialectical reasoning tasks are kept in the cloud for execution, while low-latency human-computer consultation interaction logic is migrated to the edge lightweight model through a knowledge distillation algorithm.

[0061] S5, Millisecond-level response with low computing power at the edge: The edge-based lightweight model is built on the quantized LLaMA architecture and runs the migrated consultation interaction logic to achieve millisecond-level response to diagnosis and treatment commands in a low computing power terminal environment, thus completing real-time human-computer interaction consultation.

[0062] S6. Hierarchical Sequential Diagnosis and Treatment Decision Execution: By combining a hierarchical decomposition mechanism of sequential tasks with a thinking chain reasoning algorithm, the diagnosis and treatment process is broken down into four sequential execution links: consultation guidance, syndrome differentiation analysis, prescription and drug recommendation, and efficacy evaluation. Intelligent diagnosis and treatment decisions and outputs are completed step by step.

[0063] S7. Cross-specialty scenario adaptive adaptation: The model knowledge transfer is completed through the low-rank adaptive fine-tuning algorithm. Based on the general TCM diagnosis and treatment model, it is adapted to the disease and syndrome characteristics and diagnosis and treatment rules of different clinical specialties to realize the implementation of diagnosis and treatment in multiple scenarios.

[0064] S8, Cloud-Edge Closed-Loop Self-Learning Iteration: Through a multi-task learning framework, data synchronization between the cloud and the edge is achieved, real-time diagnosis and treatment feedback data from the edge is collected, reinforcement learning algorithms are used to align diagnosis and treatment preferences, and the diagnosis and treatment decision logic of the large cloud model is continuously iterated and optimized to achieve autonomous learning and capability upgrade of the system.

[0065] The visual Transformer feature extraction algorithm is optimized for TCM tongue and facial color diagnosis scenarios. It focuses on the core diagnostic features of TCM, such as tongue color, tongue coating, facial complexion, and facial texture, and removes irrelevant image interference features to achieve accurate extraction of TCM visual diagnostic features.

[0066] The low-rank adaptive fine-tuning algorithm only fine-tunes a small number of core parameters of the model, without the need for full parameter reconstruction, and quickly completes the transfer and adaptation of knowledge of different specialties and diseases, reducing the cost of cross-scenario deployment.

[0067] The reinforcement learning algorithm uses clinical diagnosis and treatment accuracy, prescription and drug suitability, and patient efficacy feedback as reward functions to continuously align the model's diagnosis and treatment preferences and iteratively optimize the accuracy of syndrome differentiation and the rationality of diagnosis and treatment decisions.

[0068] The multi-task learning framework synchronously supports multiple parallel tasks such as data synchronization, model updates, task scheduling, and feedback optimization, ensuring that the cloud-based model is iterated and upgraded without affecting the normal operation of real-time diagnosis and treatment services on the edge.

[0069] This patent protects a cloud-edge collaborative reasoning architecture based on a large-scale model of traditional Chinese medicine. Its core lies in utilizing a multi-task parallel strategy, dynamically scheduling cloud-based expert models and lightweight edge models through neural networks and knowledge transfer mechanisms. The system employs a sequential task hierarchical decomposition mechanism to transform the complex "questioning, identification, treatment, and evaluation" process into executable semantic instructions. Simultaneously, the patent protects a multimodal fusion model that can accurately capture diagnostic intent and achieve efficient reasoning and rapid cross-scenario deployment on heterogeneous large-scale model platforms.

[0070] The cloud-edge collaborative architecture and knowledge transfer mechanism proposed in this method are not only applicable to the syndrome differentiation and treatment of traditional Chinese medicine, but can also be extended to multimodal diagnosis and treatment tasks such as chronic disease management, TCM preventive medicine, rehabilitation guidance, and telemedicine interaction, showing broad application prospects. In the future, it can be widely promoted and applied in fields such as smart hospitals, primary healthcare institutions, home health monitoring, and smart wearable devices.

[0071] In summary, this invention provides a multi-task large-scale intelligent diagnosis and treatment system and method for Traditional Chinese Medicine (TCM). This method achieves efficient allocation of computing power through a cloud-edge collaborative architecture, improving edge response speed while ensuring deep diagnostic reasoning. The introduced knowledge transfer mechanism enables the model to quickly adapt to different clinical disease scenarios, enhancing the system's versatility. Multimodal fusion and sequential task decomposition significantly enhance the intelligence of human-computer interaction, enabling the system to more accurately understand diagnostic intentions and optimizing the accuracy and robustness of the entire TCM diagnosis and treatment process.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A multi-task large-scale intelligent diagnosis and treatment system for traditional Chinese medicine, characterized in that, It includes a terminal data acquisition module, a cross-modal feature fusion module, a cloud-based dialectical intent understanding module, a cloud-edge collaborative scheduling module, a lightweight edge-terminal interaction module, a hierarchical diagnosis and treatment decision-making module, a cross-scenario adaptation module, and a closed-loop self-learning iteration module; The edge data acquisition module is used to collect multi-source TCM multimodal data from patients in real time. The multimodal data includes tongue image data, facial color data, consultation voice data, and case text data. The edge data acquisition module has a built-in multi-algorithm feature parsing unit. It uses the visual Transformer feature extraction algorithm to extract visual features from the tongue image data and facial color data, uses the Wav2Vec 2.0 voice representation algorithm to process voice features from the consultation voice data, and uses a bidirectional coding representation model to perform structured parsing of the case text data, thus completing the preliminary feature extraction and standardization preprocessing of the multimodal raw data. The cross-modal feature fusion module is connected to the edge data acquisition module and has a built-in adaptive cross-modal attention mechanism. It is used to perform heterogeneous feature fusion on preprocessed visual features, speech features, and text structured features, and to convert multi-source heterogeneous multimodal data into a unified structured semantic embedding space and output standardized semantic information. The cloud-based syndrome differentiation intention understanding module is equipped with a large-scale TCM vertical model. The TCM vertical model adopts a hybrid expert architecture to receive the structured semantic embedding information, dynamically predict the patient's pathogenesis and syndrome intention, and complete the TCM syndrome differentiation analysis. The cloud-edge collaborative scheduling module has a built-in task allocation strategy based on neural architecture search, which is used to perform hierarchical scheduling of diagnosis and treatment tasks, allocate computationally intensive deep dialectical reasoning tasks to the cloud for execution, and migrate real-time human-computer consultation and interaction logic tasks in a lightweight manner. The edge lightweight interaction module uses a quantized LLaMA architecture to build a lightweight model and migrates cloud interaction logic capabilities through a knowledge distillation algorithm. It is used to run real-time consultation interaction tasks in a low-computing-power edge environment and achieve millisecond-level diagnosis and treatment command response. The hierarchical diagnosis and treatment decision module has a built-in sequential task hierarchical decomposition mechanism and a thinking chain reasoning algorithm, which is used to automatically break down the overall TCM diagnosis and treatment process into four orderly and connected execution links: consultation guidance, syndrome differentiation analysis, prescription recommendation, and efficacy evaluation, so as to complete the output of the entire process of diagnosis and treatment decision; The cross-scenario adaptation module integrates a low-rank adaptive fine-tuning algorithm, which is used to transfer specialized knowledge based on general TCM diagnosis and treatment knowledge, quickly adapt to the disease and syndrome patterns of different clinical specialties, and realize cross-scenario diagnosis and treatment applications. The closed-loop self-learning iterative module adopts a multi-task learning framework to achieve real-time data synchronization and dynamic model evolution between the cloud and the edge. It uses reinforcement learning algorithms to align preferences with real-time diagnosis and treatment feedback from the edge, continuously optimizing the diagnosis and treatment decision logic of the cloud-based TCM vertical model, thus forming a self-learning closed-loop diagnosis and treatment system.

2. The intelligent diagnosis and treatment system for traditional Chinese medicine with a large multi-task model as described in claim 1, characterized in that, The Wav2Vec 2.0 speech representation algorithm is adapted to TCM consultation speech scenarios. It optimizes and models the speech characteristics of patients with uneven speaking speed, regional accents, and colloquial expressions, and accurately extracts the core semantic features of patient consultation speech such as chief complaints and symptom descriptions.

3. The intelligent diagnosis and treatment system for traditional Chinese medicine with a large multi-task model as described in claim 1, characterized in that, The adaptive cross-modal attention mechanism can dynamically adjust the weight ratio of different modal features, automatically increase the weight of effective modal features and weaken the weight of ineffective interference features according to the patient diagnosis and treatment scenario, and optimize the fusion accuracy of heterogeneous data.

4. The intelligent diagnosis and treatment system for traditional Chinese medicine with a large multi-task model as described in claim 1, characterized in that, The hybrid expert architecture of the TCM vertical category model includes a general TCM knowledge expert submodule, a specialized disease and syndrome expert submodule, and a syndrome differentiation and reasoning expert submodule. These submodules work collaboratively, dynamically activating the corresponding expert submodule based on input semantic information to achieve accurate prediction of pathogenesis and syndromes. The edge lightweight model optimizes the native LLaMA architecture by using model quantization and knowledge distillation, simplifying parameters and accelerating inference. While retaining the core consultation and interaction capabilities, it is adapted to low-computing edge devices such as mobile phones and portable medical terminals.

5. The intelligent diagnosis and treatment system for traditional Chinese medicine with a large multi-task model as described in claim 1, characterized in that, The task allocation strategy for neural architecture search makes intelligent decisions based on cloud and edge computing resource thresholds, task latency requirements, and computational complexity parameters to achieve optimal cloud-edge allocation of diagnostic and treatment tasks.

6. The intelligent diagnosis and treatment method of traditional Chinese medicine multi-task large model as described in claim 1, characterized in that, The system described in claim 1 is applied to include the following steps: S1. Real-time acquisition and preprocessing of multimodal data from the end device: Real-time acquisition of multi-source TCM data such as patient tongue appearance, facial color, consultation voice, and case text through the end device. Visual features of tongue appearance and facial color are extracted using the visual Transformer feature extraction algorithm. The consultation voice is processed using the Wav2Vec 2.0 voice representation algorithm to obtain voice representation features. The case text is structured using a bidirectional coding representation model to output multi-dimensional preprocessed feature data. S2. Adaptive cross-modal feature fusion: The preprocessed visual, speech, and text heterogeneous features are fused through an adaptive cross-modal attention mechanism to eliminate feature differences between different modal data and map heterogeneous data to a structured semantic embedding space to generate standardized diagnostic semantic information. S3, Cloud-based Large Model for Understanding Syndrome Differentiation Intent: Based on a large model of TCM vertical categories deployed in the cloud and using a hybrid expert architecture, it receives standardized semantic information, dynamically infers the patient's pathogenesis, accurately identifies the patient's syndrome intent, and outputs syndrome differentiation results. S4. Neural Architecture Search-Driven Cloud-Edge Collaborative Task Scheduling: By using a task allocation strategy based on neural architecture search, diagnostic tasks are classified according to computing power. High-computing-power-consuming deep dialectical reasoning tasks are kept in the cloud for execution, while low-latency human-computer consultation interaction logic is migrated to the edge lightweight model through a knowledge distillation algorithm. S5, Millisecond-level response with low computing power at the edge: The edge-based lightweight model is built on the quantized LLaMA architecture and runs the migrated consultation interaction logic to achieve millisecond-level response to diagnosis and treatment commands in a low computing power terminal environment, thus completing real-time human-computer interaction consultation. S6. Hierarchical Sequential Diagnosis and Treatment Decision Execution: By combining a hierarchical decomposition mechanism of sequential tasks with a thinking chain reasoning algorithm, the diagnosis and treatment process is broken down into four sequential execution links: consultation guidance, syndrome differentiation analysis, prescription and drug recommendation, and efficacy evaluation. Intelligent diagnosis and treatment decisions and outputs are completed step by step. S7. Cross-specialty scenario adaptive adaptation: The model knowledge transfer is completed through the low-rank adaptive fine-tuning algorithm. Based on the general TCM diagnosis and treatment model, it is adapted to the disease and syndrome characteristics and diagnosis and treatment rules of different clinical specialties to realize the implementation of diagnosis and treatment in multiple scenarios. S8, Cloud-Edge Closed-Loop Self-Learning Iteration: Through a multi-task learning framework, data synchronization between the cloud and the edge is achieved, real-time diagnosis and treatment feedback data from the edge is collected, reinforcement learning algorithms are used to align diagnosis and treatment preferences, and the diagnosis and treatment decision logic of the large cloud model is continuously iterated and optimized to achieve autonomous learning and capability upgrade of the system.

7. The intelligent diagnosis and treatment method of traditional Chinese medicine with a multi-task large model as described in claim 6, characterized in that, The visual Transformer feature extraction algorithm is optimized for TCM tongue and facial color diagnosis scenarios. It focuses on the core diagnostic features of TCM, such as tongue color, tongue coating, facial complexion, and facial texture, and removes irrelevant image interference features to achieve accurate extraction of TCM visual diagnostic features.

8. The intelligent diagnosis and treatment method of traditional Chinese medicine multi-task large model as described in claim 6, characterized in that, The low-rank adaptive fine-tuning algorithm only fine-tunes a small number of core parameters of the model, without the need for full parameter reconstruction, and quickly completes the transfer and adaptation of knowledge of different specialties and diseases, reducing the cost of cross-scenario deployment.

9. The intelligent diagnosis and treatment method of traditional Chinese medicine multi-task large model as described in claim 6, characterized in that, The reinforcement learning algorithm uses clinical diagnosis accuracy, prescription suitability, and patient efficacy feedback as reward functions to continuously align the model's diagnosis preferences and iteratively optimize the accuracy of syndrome differentiation and the rationality of diagnosis and treatment decisions.

10. The intelligent diagnosis and treatment method of traditional Chinese medicine multi-task large model as described in claim 6, characterized in that, The multi-task learning framework simultaneously supports multiple parallel tasks such as data synchronization, model updates, task scheduling, and feedback optimization, ensuring that the cloud-based model is iterated and upgraded without affecting the normal operation of real-time diagnosis and treatment services on the edge.

Images