Intelligent auxiliary diagnosis and treatment system based on bionic memory evolution mechanism

Abstract

The application discloses an intelligent auxiliary diagnosis and treatment system based on a bionic memory evolution mechanism, comprising a data processing module, a memory evolution engine, a bionic memory center and a dynamic retrieval module; the data processing module is configured to perform structured processing on multi-modal medical data to obtain structured data; a short-term memory layer is configured to process an initial database to obtain a complete database; a middle-term memory layer is configured to process the complete database to obtain a memory mode; the memory evolution engine is configured to update the weight of the memory mode, transfer the updated memory mode to a long-term memory layer, and form a final database stored in the long-term memory layer; and the dynamic retrieval module performs retrieval in the final database according to query information, sends the retrieval result to an inference model, and generates a final suggestion. The application performs plasticity update through a multi-level memory structure simulating the human memory consolidation process.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and smart healthcare, and in particular to an intelligent assisted diagnosis and treatment system based on a biomimetic memory evolution mechanism. Background Technology

[0002] Currently, medical AI-assisted diagnostic systems have moved from technological exploration to large-scale clinical application. In particular, medical AI-assisted diagnostic systems based on general large language models or static knowledge graphs mainly enable the diagnostic system to understand and generate natural language by structuring medical knowledge into entity relationship networks or by using massive amounts of text for pre-training, thereby enabling diagnostic reasoning. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent assisted diagnosis and treatment system based on a biomimetic memory evolution mechanism, in order to solve the technical problems of rigid knowledge structure and difficulty in knowledge updating in existing medical assisted diagnostic systems. The various technical effects of the preferred technical solutions provided in this application are detailed below.

[0004] To achieve the above objectives, this application provides the following technical solutions:

[0005] This application provides an intelligent assisted diagnosis and treatment system based on a biomimetic memory evolution mechanism, comprising a data processing module, a memory evolution engine, a biomimetic memory center, and a dynamic retrieval module. The biomimetic memory center includes a short-term memory layer, a medium-term memory layer, and a long-term memory layer. The data processing module is configured to perform structured processing on multimodal medical data to obtain structured data including text fragments, medical entities, and entity relationships. The structured data is used to construct an initial database, which includes a vector library storing the text fragments and a knowledge graph constructed based on the medical entities and their relationships. The short-term memory layer is configured to perform pattern processing on the initial database. The system performs completion and rapid associative retrieval processing to obtain a complete database. The intermediate memory layer is configured to abstract treatment paths from the entity relationships in the complete database and cluster the medical entities to generate concept nodes, thus obtaining memory patterns. The memory evolution engine is configured to update the weights of the memory patterns and transfer the updated memory patterns to the long-term memory layer to form a final database stored in the long-term memory layer. The dynamic retrieval module is configured to receive query information, perform a retrieval in the final database of the long-term memory layer using a dynamic weighted scoring algorithm based on the query information, and send the retrieval results to the inference model to generate final suggestions.

[0006] In some embodiments, the short-term memory layer is configured to simulate pattern completion by activating nodes in the neighborhood of the knowledge graph and to adjust synaptic strength in real time based on the activation frequency of the input signal.

[0007] In some embodiments, the short-term memory layer is further configured to automatically combine activated nodes and relationships to form a temporary active subgraph, and to mark the temporary active subgraph as a state to be consolidated after repeated activation or feedback.

[0008] In some embodiments, the intermediate memory layer is configured to extract treatment paths reflecting clinical logic from nodes of the knowledge graph based on the entity relationships, aggregate the medical entities into concept nodes based on the embedding features of the medical entities or the neighborhood similarity of the graph structure, and integrate medical knowledge fragments, treatment paths, and concept nodes to construct the memory pattern as an associated subgraph.

[0009] In some embodiments, the memory evolution engine is configured to transfer memory nodes in the intermediate memory layer with a weight greater than a first preset threshold to the long-term memory layer, and the memory evolution engine reduces the weight of memory nodes in the long-term memory layer with a weight less than a second preset threshold.

[0010] In some embodiments, the memory evolution engine is configured to collect feedback values, calculate weight coefficients based on preset weights or confidence levels, and update the weights of memory nodes in each memory layer based on the feedback values ​​and the weight coefficients.

[0011] In some embodiments, the intelligent assisted diagnosis and treatment system based on the biomimetic memory evolution mechanism may further include a case memory sub-database, which is linked to the final database of the long-term memory layer by calculating semantic similarity.

[0012] In some embodiments, the dynamic weighted scoring algorithm retrieves data from the final database based on semantic similarity, structural importance, and memory strength.

[0013] In some embodiments, the data processing module is configured to perform time-series calibration on the structured data to generate a sequence of knowledge units with timestamps.

[0014] In some embodiments, the short-term memory layer calculates a score for the knowledge unit sequence based on text similarity, graph structure similarity, temporal relevance, and credibility.

[0015] Implementing one of the technical solutions described above in this application has the following advantages or beneficial effects: In this application, structured data for constructing an initial database is first obtained through a data processing module. Then, the initial database is preprocessed using a short-term memory layer to obtain a complete database. Next, the complete database is processed using a medium-term memory layer to form a memory pattern. Finally, the memory evolution engine transfers the weighted and updated memory pattern to a long-term memory layer, forming the final database for retrieval. In this scenario, by constructing a multi-level memory structure that simulates the memory consolidation process from the hippocampus to the neocortex and then to the cerebral cortex—that is, a bionic memory center—and by enabling plastic updates to the database through activation and feedback, the medical information in the database becomes more accurate and stable when the structured data is transferred and stored to higher-level memory layers. This improves the accuracy, interpretability, and personalization of the final recommendations. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0017] Figure 1 This is a structural block diagram of an intelligent assisted diagnosis and treatment system based on a biomimetic memory evolution mechanism, according to an embodiment of this application.

[0018] In the figure: 1. Intelligent auxiliary diagnosis and treatment system based on bionic memory evolution mechanism; 10. Data processing module; 20. Memory evolution engine; 30. Bionic memory center; 31. Short-term memory layer; 32. Medium-term memory layer; 33. Long-term memory layer; 40. Dynamic retrieval module; 50. Case memory sub-library. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments and depict various exemplary embodiments that may be adopted to implement this application. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of this application disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of this application.

[0020] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "multiple" means two or more. The terms "connected" and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, integral connections, mechanical connections, electrical connections, communication connections, direct connections, indirect connections through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0021] To illustrate the technical solutions described in this application, specific embodiments are provided below, showing only the parts related to the embodiments of this application.

[0022] like Figure 1 As shown, this application provides an intelligent auxiliary diagnosis and treatment system 1 based on a bionic memory evolution mechanism, which may include a data processing module 10, a memory evolution engine 20, a bionic memory center 30 and a dynamic retrieval module 40. The bionic memory center 30 includes a short-term memory layer 31, a medium-term memory layer 32 and a long-term memory layer 33.

[0023] The data processing module 10 can be configured to perform structured processing on multimodal medical data to obtain structured data including text fragments, medical entities, and entity relationships. This structured data can be used to construct an initial database, which may include a vector library storing text fragments and a knowledge graph constructed based on medical entities and entity relationships.

[0024] The short-term memory layer 31 can be configured to perform pattern completion and fast associative retrieval processing on the initial database to obtain a complete database. The medium-term memory layer 32 is configured to abstract the diagnosis and treatment path of the entity relationship in the complete database and perform clustering processing on medical entities to generate concept nodes to obtain a memory pattern.

[0025] The memory evolution engine 20 can be configured to update the weights of memory patterns and transfer the updated memory patterns to the long-term memory layer 33 to form a final database stored in the long-term memory layer 33.

[0026] The dynamic retrieval module 40 can be configured to receive query information, perform a retrieval in the final database of the long-term memory layer 33 based on the query information using a dynamic weighted scoring algorithm, send the retrieval results to the inference model, and generate the final suggestion.

[0027] In some embodiments, the data processing module 10 can connect to an external HIS (Hospital Information System) / LIS (Laboratory Information System) to collect multimodal medical data, which may include electronic medical records, imaging reports, surgical records, laboratory indicators, and follow-up information. The hospital information system is a comprehensive information system covering all hospital business and operational management, while the laboratory information system is a specialized system for the management of laboratory operations. These systems can provide a large amount of structured knowledge such as disease diagnostic criteria and drug information, as well as electronic medical record texts and laboratory reports, which can be used to build databases.

[0028] In some embodiments, the data processing module 10 can utilize natural language processing technology to perform medical entity recognition and relation extraction on multimodal medical data, obtaining structured data. This structured data may include text chunks, identified medical entities, standard codes (ICD / LOINC / SNOMED), extracted entity relations (i.e., entity-relationship-entity), source information, timestamps, and trust scores. The knowledge graph contains node attributes (type, name, version) and edge attributes (relation, weight). The text chunks can be used to construct a vector library, while the extracted medical entities and entity relations can be used to construct the knowledge graph, assemble evidence chains, perform causal consistency checks, and serve as the basis for "edge weight updates / pruning" during the evolutionary stage. The clearer the relational structure, the more accurate the subsequent causal inference and dynamic updates. Information extraction technology can be used to identify medical entities and entity relations; this technology can employ rule-based methods or Open Information Extraction (OpenIE) and other similar implementations.

[0029] In some embodiments, the data processing module 10 can be configured to perform time-series calibration on structured data to generate a sequence of knowledge units with timestamps. The knowledge unit sequence can include multiple structured text fragments ordered by time or relationship. For example, it could be symptom → test results → imaging findings → diagnosis, where each element contains a text fragment, a medical entity, entity relationships, and source information. By carrying timestamps, a timeline and development context can be added to the knowledge unit sequence, facilitating subsequent retrieval and update operations. After scoring and filtering in subsequent processing, the knowledge unit sequence can serve as a chain of evidence required for reasoning.

[0030] In some embodiments, the bionic memory center 30 can be based on a combination mechanism of vector library and knowledge graph, wherein in the knowledge graph, nodes can represent knowledge units (such as medical entities), edges (synapses) can represent entity relationships, and the strength (trust) of a node and the weight (weight) of an edge correspond to the synaptic strength, and plasticity updates are performed by activating nodes and feedback. The bionic memory center 30 is used for hierarchical management and evolution of medical knowledge.

[0031] The short-term memory layer 31 in this embodiment simulates the "rapid association, pattern completion, and memory consolidation" functions of the CA3 region of the hippocampus. The core functions of the CA3 region are pattern completion, auto-associative recall, and fast plasticity. In biological mechanisms, the CA3 region of the hippocampus supports rapid storage of input signals, associative retrieval, and complete information recovery with partial cues through its dense recurrent connections. Specifically, pattern completion refers to recovering complete associative memories with partial input; fast associative retrieval refers to associative recall based on recurrent networks; and fast plasticity refers to rapidly forming short-term memory traces of new information.

[0032] In some embodiments, the short-term memory layer 31 is configured to complete the knowledge graph by simulating the neighborhood diffusion of activated nodes, and to adjust the synaptic strength in real time according to the activation frequency of the input signal. The short-term memory layer 31 can also be configured to automatically combine activated nodes and relationships to form a temporary active subgraph, and to mark the temporary active subgraph as a state to be consolidated after repeated activation or feedback.

[0033] Specifically, in order to achieve the above functions, the working mechanism of the short-term memory layer 31 can include the following four aspects: the neighborhood propagation mechanism of the graph structure, the fast update mechanism of node / edge weights, the temporary subgraph construction mechanism, and the temporary subgraph solidification mechanism.

[0034] The neighborhood propagation mechanism of graph structures corresponds to the diffusion of the neighborhood of activated nodes in the knowledge graph. For example, the Personalized PageRank algorithm is used to simulate the pattern completion ability of the CA3 region of the hippocampus to recover the complete clinical context from partial clues.

[0035] The node / edge weight fast update mechanism corresponds to the real-time adjustment of synaptic strength based on the activation frequency of the input signal, simulating the rapid plasticity of the CA3 region of the hippocampus under short-term experience.

[0036] The temporary subgraph construction mechanism automatically combines recently activated nodes and relationships to form a temporary active subgraph, corresponding to the short-term cache of the CA3 region of the hippocampus, i.e., the formation of short-term memory;

[0037] The temporary subgraph consolidation mechanism corresponds to marking the temporary active subgraph as a state to be consolidated after repeated activation or feedback, and promoting it to long-term memory layer 33, simulating the consolidation transfer from hippocampus to cortex.

[0038] Through the above working mechanism, the short-term memory layer 31 can improve and complete the knowledge graph in the initial database, thereby obtaining a more accurate and complete database.

[0039] In some embodiments, the short-term memory layer 31 can receive a sequence of knowledge units from the data processing module 10 as a temporary workspace to cache the clinical context information most relevant to the current diagnostic task. Specifically, the short-term memory layer 31 can infer the clinical context information of the knowledge unit sequence based on various structured signals.

[0040] In some embodiments, the structured signal may include entity hierarchy, time sequence, source credibility, and neighborhood structure in the knowledge graph. Specifically, for example, the entity hierarchy may be the order of disease—symptom—treatment; the time sequence may be the examination time and disease course; the source credibility may refer to guideline > case summary > OCR (Optical Character Recognition) extracted text; and the neighborhood structure in the knowledge graph may refer to whether a coherent diagnosis-treatment chain is formed. Through the above structured signal, the short-term memory layer 31 can reconstruct the clinical semantics represented by the knowledge unit sequence, such as "bile duct stones → causing obstruction → leading to elevated bilirubin → recommending ERCP (endoscopic retrograde cholangiopancreatography)", etc.

[0041] In some embodiments, the short-term memory layer 31 can calculate a score for the knowledge unit sequence based on text similarity, graph structure similarity, temporal relevance, and credibility. The knowledge unit sequence with the highest score can serve as evidence for generation and causal verification, and the most relevant clinical context information, i.e., the chain of evidence, can be selected.

[0042] In some embodiments, the formula for calculating the score of a knowledge unit sequence is as follows:

[0043] score=α×text_sim+β×graph_sim+γ×time_weight+δ×trust_score,

[0044] Where text_sim represents semantic matching score, graph_sim represents graph connectivity score, time_weight represents temporal relevance, trust_score represents knowledge reliability, and α, β, γ, and δ are all preset fixed weights. Furthermore, the fixed weighted ranking of short-term memory layer 31 can be replaced with a reinforcement learning (RL) strategy, thereby directly learning "which retrieval methods are more likely to be adopted by doctors".

[0045] The intermediate memory layer 32 in this embodiment can simulate the memory consolidation process from the hippocampus to the neocortex. In biological mechanisms, the hippocampus is responsible for rapidly storing recent information, while the neocortex is responsible for long-term, stable, and structured knowledge storage. The information transfer between the two is the memory consolidation process. The intermediate memory layer 32 can periodically process the content in the short-term memory layer 31 and form candidate memory patterns through methods such as text summarization, treatment pathway abstraction, and symptom clustering.

[0046] In some embodiments, the intermediate memory layer 32 can summarize from the complete database to obtain fragments of medical knowledge. In this case, it is possible to summarize from multiple recently activated knowledge units, extract key information, and obtain more compact fragments of medical knowledge.

[0047] In some embodiments, the intermediate memory layer 32 can extract treatment paths reflecting clinical logic from nodes of the knowledge graph, namely the aforementioned medical entity nodes, based on entity relationships. Then, based on the embedding features of medical entities (i.e., vectorized representations of entities) or the neighborhood similarity of the structure of the knowledge graph, medical entities are aggregated into concept nodes. Finally, text fragments (i.e., the source of the aforementioned medical knowledge fragments), treatment paths, and concept nodes are integrated to construct a memory pattern as an associated subgraph.

[0048] Specifically, the entity relationship can be symptom → examination → diagnosis → treatment. The treatment path can be used to reflect clinical logic, and highly relevant symptoms or medical entities can be aggregated into more robust concept nodes. The processed results are organized into a small-scale association subgraph, which is the memory pattern. The memory pattern is used to represent the structured understanding of a disease, symptom combination, or treatment chain in this application embodiment, and will be stored and retrieved as stable knowledge in the long-term memory layer 33.

[0049] A memory node can refer to the smallest unit of knowledge in a knowledge graph, including medical entities such as symptoms, examinations, and disease names, as well as document fragment nodes, such as summarized knowledge fragments. A memory state can refer to a set of attributes of a memory node, including memory strength, frequency of occurrence, usage timestamp, and connection strength with other nodes, used to reflect the importance of that knowledge in the system. A memory pattern can refer to an associative subgraph formed by processing in the intermediate memory layer 32, consisting of multiple memory nodes and key relationships, and is a structured expression of medical problems, such as etiological chains and diagnostic rules.

[0050] The long-term memory layer 33 of this application can simulate the cerebral cortex. In some embodiments, the long-term memory layer 33 can store highly stable medical knowledge that has been verified and reinforced by the memory evolution engine 20, which may include typical treatment pathways, knowledge of complications, surgical procedure experience, and complete typical case atlases.

[0051] The embodiments of this application utilize a multi-level memory structure of a biomimetic memory center 30, which significantly outperforms traditional models in modeling complex multimodal time-series data, such as continuous imaging changes and fluctuations in biochemical indicators. This improves the accuracy of disease progression prediction by approximately 15% and enhances the precision of time-series information processing.

[0052] In summary, this embodiment of the application temporarily stores recently generated or activated knowledge fragments in the short-term memory layer 31 in the medium-term memory layer 32, and periodically filters out high-value content based on factors such as weight, usage frequency, and consistency, and then processes and writes it into the long-term memory layer 33. This mechanism enables the intelligent assisted diagnosis and treatment system 1 based on the biomimetic memory evolution mechanism to absorb new knowledge at a relatively fast speed, while maintaining the stability and reliability of the database.

[0053] In some embodiments, the memory evolution engine 20 can be configured to clean up expired information in the short-term memory layer 31, transfer memory nodes in the medium-term memory layer 32 with weights greater than a first preset threshold to the long-term memory layer 33, and reduce the weights of memory nodes in the long-term memory layer 33 with weights lower than a second preset threshold.

[0054] In some embodiments, the memory evolution engine 20 can be configured to read the memory state of the bionic memory center 30 and update the memory strength.

[0055] The memory evolution engine 20 is the core algorithm unit of the intelligent assisted diagnosis and treatment system 1 based on the bionic memory evolution mechanism. It interacts with all three levels of the bionic memory center 30, and is responsible for reading the memory state and updating the memory strength. It is responsible for dynamically updating the knowledge nodes and their weights in the three memory layers, and is the core computing unit driving the adaptive evolution of the entire bionic memory center 30.

[0056] In some embodiments, the memory evolution engine 20 can be a weight update model based on Hebbian Theory. Specifically, Hebbian Theory is a fundamental axiom in neuroscience explaining synaptic plasticity, namely, "the strengthening of connections between co-activated neurons." In the embodiments of this application, this theory is used to guide the dynamic adjustment of memory node weights.

[0057] The memory evolution engine 20 can reduce the weight of memory nodes in the long-term memory layer 33 that have not been retrieved within a preset time and whose weight is lower than a second preset threshold. Here, retrieval can refer to the system's invocation of external knowledge graphs and vector libraries during the retrieval enhancement generation process. When a memory node has not been invoked, recalled, or used in reasoning by the model for a long period of time, it is considered "not retrieved," and its weight can be reduced and it can be cleaned up.

[0058] In some embodiments, the memory evolution engine 20 can be configured to collect feedback values, calculate weight coefficients based on preset weights or confidence levels, and update the weights of memory nodes in each memory layer based on the feedback values ​​and weight coefficients. This allows the doctor's feedback signals to be directly converted into memory strength adjustment amounts.

[0059] In some embodiments, based on the above Hebbi theory, the formula for updating the weights of memory nodes is as follows:

[0060]

[0061] Where w_old and w_new represent the weights of the memory node before and after the update, respectively, i.e., the memory strength. It is a quantitative feedback signal; for example, if the suggestion is adopted, f=1; if it is slightly modified, f=0.5; and if it is rejected, f=0. It is an adaptive learning rate. These are the weighting coefficients.

[0062] In some embodiments, the memory evolution engine 20 can update weights based on an exponential decay model. Specifically, the weights of memory nodes can be automatically reduced based on their most recent usage time, and if feedback is received, the weights can be increased or further decayed based on that feedback.

[0063] In some embodiments, the memory evolution engine 20 can update weights based on an adaptive gating mechanism. That is, it controls "how much old knowledge to retain" and "how much new feedback to receive" respectively. The gating values ​​can be generated by rules or a lightweight model without requiring complex training.

[0064] In some embodiments, the intelligent assisted diagnosis and treatment system 1 based on a biomimetic memory evolution mechanism may further include a case memory sub-database 50. The case memory sub-database 50 may include multiple case memory sub-databases, each case memory sub-database being bound to a corresponding patient.

[0065] In some embodiments, the case memory sub-graph 50 can construct a time sequence diagram from standardized knowledge units according to time and logical relationships. For example, standardized knowledge units could be jaundice, CT scan showing common bile duct dilation, and ERCP procedure. The standardized knowledge units are obtained from the input electronic medical record text through structured processing, including: entity recognition, such as symptoms, surgical names, examination results, medical terminology standardization (e.g., mapping to SNOMED CT (Medical Systems Nomenclature - Clinical Terminology) or an in-hospital terminology database), and timestamp extraction. This information can be automatically extracted by a clinical NLP (Neuro-Linguistic Programming) model or rule templates.

[0066] This application embodiment can timestamp each node based on time information in the medical record, such as admission date, examination date, surgery date, event sequence, and treatment process logic, and connect them sequentially in chronological order to form a directed time sequence graph. If there is a clear causal or dependency relationship between nodes, corresponding logical edges will also be added.

[0067] In some embodiments, the case memory subgraph library 50 can be linked to the final database of the long-term memory layer 33 by calculating semantic similarity. Specifically, when performing diagnostic reasoning, this embodiment can quickly retrieve the case memory subgraph most similar to the current patient from the final database, providing a reference for individualized decision-making. By constructing and retrieving case memory subgraphs, this embodiment allows the system to provide highly similar historical cases for the current patient as a reference, effectively assisting doctors in developing plans more tailored to the individual patient's situation, and enabling personalized diagnostic and treatment support.

[0068] In some embodiments, the case memory sub-database 50 can be linked to the final database based on a graph neural network. This enables the automatic learning of deep relationships between the case sub-database and the main atlas.

[0069] In some embodiments, the dynamic weighted scoring algorithm retrieves data from the final database based on semantic similarity, structural importance, and memory strength. Semantic similarity refers to the textual relevance between the query and the content of the memory node; structural importance refers to the centrality of a node in the knowledge graph calculated using algorithms such as PageRank; and memory strength is the current weight of the memory node. In some embodiments, the dynamic weighted scoring algorithm can use preset weight coefficients to comprehensively score semantic similarity, structural importance, and memory strength. Furthermore, fixed weighted ranking can be replaced with a reinforcement learning (RL) strategy, thereby directly learning "which retrieval methods are more likely to be adopted by doctors."

[0070] Semantic similarity ensures that the retrieved content is semantically relevant to the doctor's query, while structural importance prioritizes nodes with high centrality or criticality in the medical knowledge graph, such as typical diagnostic pathways or frequently used treatment rules. Memory strength reflects the stability and reliability of knowledge gained through long-term use, giving higher weight to knowledge nodes that have been repeatedly validated recently. Finally, based on weighted scores, a multi-level, precise retrieval is performed, from terminology and treatment paragraphs to complete case studies, and the retrieval results are provided to the inference model to generate final suggestions.

[0071] This ensures that the evidence returned by the system is not only relevant but also reliable and representative, which helps the reasoning model generate accurate final recommendations.

[0072] In some embodiments, the search results can be multi-granular medical knowledge units. Specifically, the search results can be a single medical term, such as common bile duct dilatation, or a segment of diagnostic and treatment text, such as post-ERCP management principles, or a complete case memory subgraph or a typical treatment pathway fragment. The dynamic search module 40 can automatically select the granularity based on the query intent. For example, symptom-related questions are more likely to retrieve single symptom nodes, while treatment decision-related questions will retrieve structured case patterns containing examination results, treatment plans, and prognostic information.

[0073] In some embodiments, the intelligent assisted diagnosis and treatment system 1 based on the biomimetic memory evolution mechanism may further include a privacy and compliance module (not shown in the figure), which is used to encrypt the database.

[0074] In some embodiments, the privacy and compliance module can be deployed on the hospital's internal network, isolated from the external network. When providing aggregated data or conducting federated learning, the privacy and compliance module can add calculated noise to the data to protect individual privacy. The privacy and compliance module can manage role-based access and record all user operation logs, thereby meeting medical data security and compliance requirements. This ensures the security of sensitive medical data processing, fully complies with the "Information Security Technology - Personal Information Security Specification" and relevant data regulations in the medical industry, and guarantees safe and compliant operation.

[0075] This application's embodiments combine a path-based final database with dynamic retrieval enhancement generation technology, making the generated final recommendations more consistent with clinical pathways. The completeness and logic of its reasoning chain are improved by more than 30%, thereby enhancing the coherence of the reasoning logic.

[0076] In this application, structured data for constructing an initial database is first obtained through the data processing module 10. Then, the initial database is preprocessed using the short-term memory layer 31 to obtain a complete database. Next, the complete database is processed using the medium-term memory layer 32 to form memory patterns. Finally, the memory evolution engine 20 transfers the weighted and updated memory patterns to the long-term memory layer 33, forming the final database for retrieval. In this scenario, by constructing a multi-level memory structure that simulates the memory consolidation process from the hippocampus to the neocortex and then to the cerebral cortex—that is, a bionic memory center 30—and by enabling plastic updates to the database through activation and feedback, the medical information in the database becomes more accurate and stable when the structured data is transferred and stored to higher-level memory layers. This leads to improvements in the accuracy, interpretability, and personalization of the final recommendations.

[0077] This application also relates to an intelligent assisted diagnosis and treatment method based on a biomimetic memory evolution mechanism. The method includes: performing structured processing on multimodal medical data to obtain structured data including text fragments, medical entities, and entity relationships, wherein the structured data is used to construct an initial database, the initial database including a vector library storing text fragments and a knowledge graph constructed based on medical entities and entity relationships; performing pattern completion and fast associative retrieval processing on the initial database to obtain a complete database; abstracting diagnosis and treatment paths from the entity relationships in the complete database, and clustering medical entities to generate concept nodes to obtain memory patterns; updating the weights of the memory patterns and transferring the updated memory patterns to a long-term memory layer 33 to form a final database stored in the long-term memory layer 33; receiving query information, performing a retrieval in the final database of the long-term memory layer 33 based on the query information using a dynamic weighted scoring algorithm, sending the retrieval results to the inference model, and generating final suggestions.

[0078] Those skilled in the art will understand that all or part of the features / steps of the above-described method embodiments can be implemented by methods, data processing systems, or computer programs. These features may be implemented without hardware, entirely in software, or in a combination of hardware and software. The aforementioned computer program may be stored in one or more computer-readable storage media. When the computer program is executed (e.g., by a processor), it performs the steps of the above-described embodiments of the intelligent assisted diagnosis and treatment method based on a biomimetic memory evolution mechanism.

[0079] The aforementioned storage media capable of storing program code include: static hard disks, solid-state hard disks, random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), optical storage devices, magnetic storage devices, flash memory, magnetic disks or optical disks, and / or combinations of the above devices, that is, they can be implemented by any type of volatile or non-volatile storage devices or combinations thereof.

[0080] This application also provides a processing device embodiment, including one or more processors and a memory; wherein the memory is used to store one or more computer programs, and the one or more processors are used to execute the one or more computer programs stored in the memory, so that the processors execute the features / steps of the above-described intelligent assisted diagnosis and treatment method embodiment based on biomimetic memory evolution mechanism.

[0081] The above description is merely a preferred embodiment of this application. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of this application. Furthermore, under the teachings of this application, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of this application. Therefore, this application is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this application.

Claims

1. An intelligent auxiliary diagnosis and treatment system based on a biomimetic memory evolution mechanism, characterized in that, The system comprises a data processing module, a memory evolution engine, a biomimetic memory hub and a dynamic retrieval module, the biomimetic memory hub comprises a short-term memory layer, an intermediate-term memory layer and a long-term memory layer; The data processing module is configured to perform structured processing on multi-modal medical data to obtain structured data comprising text segments, medical entities and entity relationships, wherein the structured data is used to construct an initial database, and the initial database comprises a vector library storing the text segments and a knowledge graph constructed based on the medical entities and the entity relationships; The short-term memory layer is configured to perform pattern completion and rapid associative retrieval processing on the initial database to obtain a complete database, the intermediate-term memory layer is configured to perform diagnosis and treatment path abstraction on the entity relationships in the complete database and cluster processing on the medical entities to generate concept nodes, thereby obtaining a memory pattern; The memory evolution engine is configured to update the weights of the memory pattern, and transfer the updated memory pattern to the long-term memory layer to form a final database stored in the long-term memory layer; The dynamic retrieval module is configured to receive query information, perform retrieval in the final database of the long-term memory layer using a dynamic weighted scoring algorithm, and send the retrieval result to an inference model to generate a final suggestion; The short-term memory layer is configured to simulate the pattern completion by activating nodes to diffuse in the neighborhood of the knowledge graph, and adjust the synaptic strength in real time according to the activation frequency of the input signal; The intermediate-term memory layer is configured to extract diagnosis and treatment paths reflecting clinical logic from the nodes of the knowledge graph based on the entity relationships, aggregate the medical entities into concept nodes based on the embedding features of the medical entities or the neighborhood similarity of the structure of the knowledge graph, and integrate the text segments, the diagnosis and treatment paths and the concept nodes to construct the memory pattern as an associated subgraph; The dynamic weighted scoring algorithm is based on semantic similarity, structural importance and memory strength to retrieve in the final database.

2. The intelligent diagnosis and treatment system based on the bionic memory evolution mechanism according to claim 1, characterized in that, The short-term memory layer is further configured to automatically combine the activated nodes and relationships to form a temporary active subgraph, and mark the temporary active subgraph as a to-be-consolidated state after repeated activation or feedback in the temporary active subgraph.

3. The intelligent diagnosis and treatment system based on the bionic memory evolution mechanism according to claim 1, characterized in that, The memory evolution engine is configured to transfer memory nodes with weights greater than a first preset threshold in the intermediate-term memory layer to the long-term memory layer, and reduce the weights of memory nodes with weights less than a second preset threshold in the long-term memory layer.

4. The intelligent diagnosis and treatment system based on the bionic memory evolution mechanism according to claim 1, characterized in that, The memory evolution engine is further configured to collect feedback values, calculate a weight coefficient according to a preset weight or credibility, and update the weights of the memory nodes of each memory layer according to the feedback values and the weight coefficient.

5. The intelligent diagnosis and treatment system based on the bionic memory evolution mechanism according to claim 1, characterized in that, The intelligent auxiliary diagnosis and treatment system based on the biomimetic memory evolution mechanism can further comprise a case memory subgraph library, which is linked to the final database of the long-term memory layer by calculating semantic similarity.

6. The intelligent diagnosis and treatment system based on the bionic memory evolution mechanism according to claim 1, characterized in that, The data processing module is further configured to perform time series calibration on the structured data to generate a time-stamped knowledge unit sequence.

7. The intelligent diagnosis and treatment system based on the bionic memory evolution mechanism according to claim 6, characterized in that, The short-time memory layer calculates a score of the knowledge unit sequence based on text similarity, graph structure similarity, time recency and credibility.

Images