Multi-agent and time sequence dynamic memory based diagnosis and treatment decision support method and system

Abstract

The present disclosure is suitable for the field of medical artificial intelligence, and provides a diagnosis and treatment decision support method and system based on multi-agent and time sequence dynamic memory, comprising: constructing a time sequence dynamic memory network, which comprises a short-term memory layer, a near-end memory layer and a long-term memory layer; according to the dynamic index change rate stored in the near-end memory layer, monitoring is performed through a memory-driven adaptive trigger engine, and when the change rate exceeds a dynamic optimization threshold, re-collection and processing of evaluation data are triggered; a multi-agent decision system is constructed based on the time sequence dynamic memory network, comprising a master agent and multiple professional agents, and decision suggestion information is generated through a retrieval enhancement generation mechanism; an evaluation weight is adjusted and an effect feedback matrix is updated through a bidirectional closed loop optimization engine; a causal chain is constructed and counterfactual reasoning is performed through a cross-modal causal reasoning engine, and intervention effect estimation data is output; and an ethical and explainability module is used for ethical verification and fairness detection, and a structured report is output.

Description

Technical Field

[0002] This application belongs to the field of medical artificial intelligence, and specifically relates to a diagnostic decision support method based on multi-agent and temporal dynamic memory. Background Technology

[0003] Coronary artery disease (CAD) is one of the leading causes of disease prevalence, disability, and mortality worldwide, with a prevalence rate as high as 27.8% in people over 60 years of age. Elderly CAD patients often suffer from multiple geriatric syndromes, including frailty, sarcopenia, cognitive impairment, fall risk, malnutrition, and polypharmacy. These syndromes are characterized by dynamic fluctuations, multifactorial coupling, and reversibility of interventions, posing significant challenges to the comprehensive management of elderly CAD patients. Comprehensive geriatric assessment is an effective means of screening for geriatric syndromes. Through multidisciplinary collaboration, it assesses the physical function, psychological and social support of the elderly, providing a basis for developing individualized comprehensive management plans. However, existing medical information systems for elderly CAD patients suffer from problems such as a singular assessment perspective, insufficient application of comprehensive geriatric assessment, limited intervention methods, reliance on individual physician experience for clinical decision-making, and fragmented multimodal data. In recent years, the rapid development of multimodal large language models has provided a new approach to solving these problems. These models can integrate and process heterogeneous data such as text, images, and signals, enabling a deeper understanding and complex reasoning about the patient's condition. However, existing medical AI models mostly remain at the level of "single assessment - static output" in terms of perceptual intelligence, lacking the ability to track the dynamic fluctuations of geriatric syndromes, and the data processing process is opaque, making it difficult to gain the trust of clinicians. In addition, the application of general multi-agent frameworks in the medical field is still in its early stages, lacking in-depth customization for geriatric comorbidity management scenarios.

[0004] In existing technologies, medical data processing systems mainly suffer from the following technical problems: First, it is impossible to dynamically track the dynamic fluctuations of geriatric syndromes. Because geriatric syndromes have the characteristic of fluctuating over time, the existing single-assessment model creates a monitoring blind spot, making it difficult to capture the trend of changes in the condition in a timely manner. Secondly, it is difficult to analyze the problem of multi-factor coupling. When elderly patients with coronary heart disease are complicated with multiple geriatric syndromes, the factors influence each other and are mutually causal. Existing technologies lack effective multimodal data fusion and causal reasoning mechanisms, resulting in fragmented data processing. Third, the data processing process is not transparent. Existing models mostly use black-box reasoning, lack interpretable output and uncertainty quantification, making it difficult to meet the requirements for traceability of decision-making basis in clinical applications.

[0005] The aforementioned problems prevent existing systems from making the leap from passive response to proactive monitoring, from single-point assessment to longitudinal tracking, and from statistical correlation to causal reasoning, which severely restricts the accuracy, interpretability, and safety of comprehensive management decisions for elderly patients with coronary heart disease and geriatric syndromes. Summary of the Invention

[0006] To address the aforementioned problems, this application provides a diagnostic decision support method based on multi-agent and temporal dynamic memory, comprising: A temporal dynamic memory network was constructed, which includes a short-term memory layer, a proximal memory layer, and a long-term memory layer, which are used to store the current session context, patient longitudinal trajectory data, and patient stable profile and historical intervention effect data, respectively. Based on the rate of change of dynamic indicators stored in the proximal memory layer, the system monitors the data through a memory-driven adaptive triggering engine. When the rate of change exceeds the dynamic optimization threshold, the system triggers the re-collection and processing of comprehensive elderly assessment data. Based on the temporal dynamic memory network and the evaluation data obtained after triggering, a multi-agent decision-making system is constructed. The multi-agent decision-making system includes a master agent and professional agents capable of risk warning, comprehensive elderly assessment, and prescription data generation. The master agent performs task planning and calls on professional agents to execute sub-tasks. Each agent obtains information from the temporal dynamic memory network and external knowledge base through a retrieval-enhanced generation mechanism and generates decision-making suggestions. Based on the risk level data output by the professional intelligent agent, the weight parameters of each dimension of the comprehensive elderly assessment are adjusted through a two-way closed-loop optimization engine, and the effect feedback matrix is ​​updated based on the effect data after prescription execution. Based on multimodal data and medical knowledge graphs, a causal chain is constructed through a cross-modal causal reasoning engine to identify key pathogenic factors. Counterfactual inference is then performed under the constraints of the medical knowledge graph to output predicted data on intervention effects. Based on the generated prescription data, ethical rules are verified and fairness is tested, and a structured report containing uncertainty quantification parameters is output.

[0007] Furthermore, the longitudinal trajectory data stored in the proximal memory layer includes indicator change curves and individual historical variability data; The long-term memory layer stores the effect feedback matrix, which records historical prescription data and its effect score data.

[0008] Furthermore, the dynamic optimization threshold is adjusted online using a reinforcement learning algorithm, and the reward function of reinforcement learning is constructed based on the weighted sum of the false negative rate and the false positive rate.

[0009] Furthermore, the multi-agent decision-making system configures an independent workspace for each agent, which includes configuration files for defining roles, collaborative relationships, invoking skills, storing memories, recording preferences, and identifying identities.

[0010] Furthermore, specialized intelligent agents are used to perform sub-tasks such as comprehensive elderly assessment data processing, image data analysis, risk warning data generation, prescription data generation, trajectory data analysis, causal reasoning data generation, and ethical verification data generation. Each professional intelligent agent corresponds to at least one skill, and the skills adopt a hierarchical loading mechanism.

[0011] Furthermore, adjusting the weighting parameters of each dimension of the comprehensive geriatric assessment includes: weighting the adjusted weighting parameters of each dimension based on the basic weighting parameters of that dimension, risk level data, and patient historical score variability data. The adjusted comprehensive assessment score data for the elderly is determined by the weighted sum of the original scores for each dimension and the adjusted weight parameters.

[0012] Furthermore, the update of the effect feedback matrix includes: determining the effect score data of the prescription on the patient population based on the ratio of the average improvement data to the standard deviation data.

[0013] Furthermore, the medical knowledge graph stores causal knowledge data in the form of triples and associates it with causal strength data; Counterfactual reasoning generates intervention hypothesis data and outputs predicted effect range data under the constraints of a medical knowledge graph.

[0014] Furthermore, ethical rule verification includes pre-filtering based on taboo rule data; Fairness testing includes detecting decision-making bias data across different groups; Uncertainty quantification includes outputting confidence interval data and evidence level data.

[0015] A diagnostic decision support system based on multi-agent and temporal dynamic memory includes: The temporal dynamic memory network construction module is used to construct a temporal dynamic memory network, which includes a short-term memory layer, a proximal memory layer, and a long-term memory layer, which are used to store the current session context, patient longitudinal trajectory data, and patient stable profile and historical intervention effect data, respectively. The memory-driven adaptive triggering engine is used to monitor the rate of change of dynamic indicators stored in the proximal memory layer. When the rate of change exceeds the dynamic optimization threshold, it triggers the re-collection and processing of comprehensive elderly assessment data. The multi-agent decision-making system construction module is used to construct a multi-agent decision-making system based on the temporal dynamic memory network and the evaluation data obtained after triggering. The multi-agent decision-making system includes a master agent and multiple professional agents. The master agent performs task planning and calls professional agents to execute sub-tasks. Each agent obtains information from the temporal dynamic memory network and external knowledge base through a retrieval enhancement generation mechanism and generates decision suggestion information. The two-way closed-loop optimization engine is used to adjust the weight parameters of each dimension of the comprehensive assessment of the elderly based on the risk level data output by the risk warning module, and to update the effect feedback matrix based on the effect data after the prescription is implemented. A cross-modal causal reasoning engine is used to construct causal chains based on multimodal data and medical knowledge graphs, identify key pathogenic factors, and perform counterfactual inference under the constraints of the medical knowledge graph to output intervention effect prediction data. The Ethics and Interpretability module is used to perform ethical rule verification and fairness checks on the generated prescription data, and output a structured report containing uncertainty quantification parameters.

[0016] A computer storage medium stores computer instructions, which, when executed by a processor, specifically perform the steps in any of the methods described above.

[0017] A computer program product includes computer instructions, which, when executed by a processor, specifically perform the steps in any of the methods described above.

[0018] Compared with the prior art, this application has the following advantages: This invention constructs a temporal dynamic memory network comprising a short-term memory layer, a proximal memory layer, and a long-term memory layer, and combines it with a memory-driven adaptive triggering engine to monitor the rate of change of dynamic indicators stored in the proximal memory layer in real time. When the rate of change exceeds the dynamic optimization threshold, it automatically triggers the re-collection and processing of geriatric comprehensive assessment data, effectively solving the monitoring blind spot problem caused by the inability to dynamically track the dynamic fluctuations of geriatric syndromes in existing technologies.

[0019] By constructing a multi-agent decision-making system, the master agent plans tasks and calls multiple specialized agents to execute sub-tasks such as comprehensive elderly assessment data processing, image data analysis, risk warning data generation, prescription data generation, trajectory data analysis, causal reasoning data generation, and ethical verification data generation. Each agent obtains information from the temporal dynamic memory network and external knowledge base through a retrieval-enhanced generation mechanism and generates decision-making suggestions. At the same time, a cross-modal causal reasoning engine constructs causal chains based on medical knowledge graphs, identifies key pathogenic factor data, and performs counterfactual inference to output intervention effect prediction data. This effectively solves the problems of difficult analysis of multi-factor coupling and fragmented data processing in existing technologies.

[0020] The two-way closed-loop optimization engine feeds back the risk level data output by the risk warning module to the elderly comprehensive assessment module to dynamically adjust the weight parameters of each assessment dimension. It also feeds back the effect data after prescription execution to the effect feedback matrix to optimize subsequent decisions. At the same time, the ethics and interpretability module verifies the ethical rules and fairness of the generated prescription data and outputs a structured report containing uncertainty quantification parameters. This solves the problems of opaque data processing and lack of interpretable output in the existing technology.

[0021] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the multi-agent decision-making process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system deployment and core processing principle provided in the embodiments of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The decision support method and system for the diagnosis and treatment of geriatric comorbidities based on a multi-agent decision-making framework and a temporal dynamic memory mechanism provided in this invention can be applied to geriatric medicine departments, cardiology departments, or primary healthcare institutions in general hospitals. It assists medical personnel in the comprehensive management of elderly patients with coronary heart disease complicated by frailty, sarcopenia, cognitive impairment, fall risk, malnutrition, and polypharmacy. In existing technologies, medical data processing systems mainly suffer from the following technical problems: First, it is impossible to dynamically track the dynamic fluctuations of geriatric syndromes. Because geriatric syndromes have the characteristic of fluctuating over time, the existing single-assessment model creates a monitoring blind spot, making it difficult to capture the trend of changes in the condition in a timely manner. Secondly, it is difficult to analyze the problem of multi-factor coupling. When elderly patients with coronary heart disease are complicated with multiple geriatric syndromes, the factors influence each other and are mutually causal. Existing technologies lack effective multimodal data fusion and causal reasoning mechanisms, resulting in fragmented data processing. Third, the data processing process is not transparent. Existing models mostly use black-box reasoning, lack interpretable output and uncertainty quantification, making it difficult to meet the requirements for traceability of decision-making basis in clinical applications.

[0026] To address the aforementioned issues, this invention constructs a comprehensive management decision support system that is interpretable, traceable, and sustainably optimized throughout the entire process from data acquisition to decision information output. This system achieves full-cycle data storage and longitudinal tracking by building a temporal dynamic memory network, proactive monitoring and dynamic evaluation through a memory-driven adaptive triggering engine, task decomposition and professional division of labor through a multi-agent decision-making system, causal chain construction and counterfactual inference through a cross-modal causal reasoning engine, dynamic weight adjustment and experience feedback through a bidirectional closed-loop optimization engine, and compliance verification and uncertainty output through an ethics and interpretability module.

[0027] The following is in conjunction with the appendix Figure 1-2 The present invention will be described in further detail below.

[0028] S1. Construction of Temporal Dynamic Memory Network This temporal dynamic memory network includes a short-term memory layer, a proximal memory layer, and a long-term memory layer, which are used to store the current session context, patient longitudinal trajectory data, and patient stability profile and historical intervention effect data, respectively.

[0029] The short-term memory layer employs a combination of a cached database and vector storage to store the complete dialogue history of the current interaction session, including user messages, agent thought processes, and tool call records. This history is automatically archived as a log file after the session ends. This layer provides immediate contextual information for current decisions, ensuring the continuity and consistency of those decisions.

[0030] The proximal memory layer is constructed using a time-series database to store longitudinal trajectory data from multiple patient visits. Specifically, this layer stores data including change curves of indicators such as frailty score, gait speed, grip strength, and weight, as well as individual historical variability. Individual historical variability is obtained by calculating the standard deviation of historical scores for each indicator and is used for baseline adjustment in subsequent risk weighting. A trend feature vector is constructed for each patient in the proximal memory layer to characterize the changing trends of each indicator over time.

[0031] The long-term memory layer uses a relational database to store stable patient profiles and historical intervention effect data. Stable patient profiles include relatively stable information such as age, medical history, allergy history, and long-term medication use. Historical intervention effect data exists in the form of an effect feedback matrix, recording historical prescriptions and their implementation effect scores. This matrix is ​​stored in the relational database to provide experience-based references for subsequent decision-making for similar patients.

[0032] Each memory layer follows the memory interface specification of the multi-agent framework. Each agent has independent memory access permissions in the work area and can read and write the corresponding memory layer through the standard interface.

[0033] S2, memory-driven adaptive triggering mechanism.

[0034] The memory-driven adaptive triggering engine monitors the rate of change of dynamic indicators stored in the proximal memory layer. When the rate of change exceeds the dynamic optimization threshold, it automatically triggers the re-collection and processing of comprehensive elderly assessment data.

[0035] The formula for calculating the rate of change of dynamic indicators is:

[0036] in, , They are respectively , The index value at any given time.

[0037] When the rate of change of any indicator exceeds a preset threshold (e.g., a decrease in walking speed >20% / month), a CGA reassessment is automatically triggered without the need for a doctor's instruction. When multiple indicators show a synergistic deterioration trend, a comprehensive early warning is triggered.

[0038] The dynamic optimization threshold is adjusted online using a reinforcement learning algorithm. Specifically, the Q-learning algorithm can be used to achieve adaptive optimization of the threshold. The reward function of reinforcement learning is constructed based on the weighted sum of the false negative rate and the false positive rate, and its expression is as follows: The false negative rate is the ratio of the number of non-triggered worsening events to the total number of worsening events, while the false positive rate is the ratio of the number of false triggers to the total number of triggers. This is a balancing coefficient used to adjust the weights of the false negative rate and false positive rate in the reward function. When the false negative rate is high, the system tends to lower the threshold to improve sensitivity; when the false positive rate is high, the system tends to raise the threshold to improve specificity.

[0039] Through this mechanism, the system can adaptively adjust the trigger threshold for different patients and at different stages, achieving personalized proactive monitoring. When the absolute value of the rate of change of any indicator exceeds the current dynamic optimization threshold, the system automatically triggers the re-collection and processing of comprehensive elderly assessment data without the need for manual instructions from doctors. When multiple indicators show a synergistic deterioration trend, the system triggers a comprehensive early warning, indicating that special attention is needed.

[0040] S3, Construction of a multi-agent decision-making system.

[0041] Based on the temporal dynamic memory network and the evaluation data obtained after triggering, a multi-agent decision-making system is constructed, which includes a master agent and multiple specialized agents.

[0042] Specifically, the master intelligent agent plans tasks and calls on specialized intelligent agents to execute sub-tasks, while other specialized intelligent agents obtain information from the temporal dynamic memory network and external knowledge base through a retrieval-enhanced generation mechanism and generate decision-making suggestions.

[0043] The multi-agent decision-making system configures an independent workspace for each agent, which includes the following types of configuration files: Role definition files are used to define the roles of intelligent agents and system instructions, such as "You are an expert in comprehensive geriatric assessment and must follow the guidelines for comprehensive geriatric assessment"; Collaboration relationship files are used to describe the collaborative relationships between multiple agents. For example, the master agent can call the elderly comprehensive assessment agent, the image analysis agent, etc. A whitelist file for tools that list permitted skills and external resource access services; a long-term memory file for storing long-term memory. User preference files are used to record user preferences and context; and identity files are used to identify the identity of intelligent agents.

[0044] Specialized intelligent agents are used to perform sub-tasks such as comprehensive elderly assessment data processing, image data analysis, risk warning data generation, prescription data generation, trajectory data analysis, causal reasoning data generation, and ethical verification data generation.

[0045] Each professional intelligent agent corresponds to at least one skill. The skill adopts a hierarchical loading mechanism, including an index layer for loading metadata and descriptive information, a definition layer for loading instructions and configuration documents, and an execution layer for loading runnable scripts or code modules.

[0046] Referring to Table 1, the specific implementation of skills includes, but is not limited to, the following types: The comprehensive geriatric assessment skill takes patient identification, scale data, and gait video as inputs and outputs a structured comprehensive geriatric assessment report. Its unique mechanism lies in automatically extracting gait parameters from the video. Image analysis skills, with coronary CTA images as input, output plaque list, stenosis degree and calcification integral, are implemented using deep learning segmentation and detection algorithms; The risk warning skill takes as input a comprehensive elderly assessment report, imaging features, and test data, and outputs the probability and level of risk events, integrating multimodal features for analysis; Trajectory analysis skills take patient identifiers and time ranges as inputs and output indicators change rates and trend types as outputs, employing time series analysis and anomaly detection techniques. The causal reasoning skill takes a multimodal feature set as input and outputs a causal chain graph and a ranking of key pathogenic factors, employing graph neural network and knowledge graph alignment technology. Counterfactual reasoning skills take the current state and hypothetical intervention as inputs and output the predicted effect range and causal rationality score, and conduct intervention simulation based on causal models; The effect feedback skill takes prescription identifiers and follow-up data as inputs and outputs effect scores and feedback matrix updates, and adopts a reinforcement learning reward function design. The prescription generation skill takes patient profiles, comprehensive geriatric assessment results, risk information, and contraindication rules as inputs and outputs five types of prescription texts, which are achieved through rule engines and knowledge base queries. The ethics verification skill takes a prescription draft as input and outputs compliance indicators and reasons for violations. It uses a rule engine to verify contraindications, over-treatment, and fairness.

[0047] The specific implementation steps of the intelligent agent retrieval enhancement generation mechanism are as follows: After receiving a task, the master agent plans the retrieval requirements, retrieves information sequentially from the near memory layer, long memory layer, and external knowledge base, and evaluates the sufficiency of the information after each retrieval. Based on the evaluation results, it optimizes the retrieval strategy, such as adjusting keywords, expanding the time range, or switching data sources. Finally, it integrates the retrieval results with the current session context to generate decision suggestions, and writes the new knowledge or optimization strategies generated during this reasoning process into the corresponding memory layer.

[0048] Standardized external resource access is achieved through a standardized external resource access server cluster, which unifies access to various external medical data sources and knowledge resources. In some embodiments, the following servers may be included: The hospital information system interface server connects the hospital's electronic medical records and laboratory system, providing methods for querying basic patient information and test results, and extending the interface to obtain the changing trends of test indicators; The image archiving and communication system interface server connects to the image system, provides methods such as image retrieval and thumbnail acquisition, and extends the interface for multi-time point image comparison and analysis. Wearable device interface server connects to smart devices, provides methods such as heart rate subscription and step count query, and extends the interface for subscribing to indicator change trends; The medical knowledge base interface server connects clinical guidelines, drug databases, causal graphs, and ethical rule bases, providing methods such as guideline retrieval and drug interaction queries, and extending the interfaces for causal association queries and ethical verification.

[0049] S4, bidirectional closed-loop optimization Through a two-way closed-loop optimization engine, the weight parameters of each dimension of the comprehensive elderly assessment are adjusted based on the risk level data output by the risk warning module, and the effect feedback matrix is ​​updated based on the effect data after prescription execution.

[0050] The risk level output by the risk warning module is mapped to the weighted coefficients of each assessment dimension of the comprehensive elderly assessment, while individual historical variability is introduced for adjustment.

[0051] The specific calculation method for weight adjustment is as follows: in, For the first Adjusted weights for each evaluation dimension; Basic weights; For the corresponding risk level; The variability of the patient's historical scores; , This is the adjustment coefficient; risk i This corresponds to the risk level.

[0052] The higher the risk level, the greater the increase in the weight of the corresponding dimension. The greater the variability of the patient's historical score, the more drastic the fluctuation of the indicator in that dimension. The weight is then reduced accordingly to reduce the impact of fluctuation on the assessment results.

[0053] The adjusted formula for calculating the comprehensive assessment score for the elderly is as follows:

[0054] Wherein, the original score of dimension i is the original score value of the i-th evaluation dimension; w i The new comprehensive assessment score for the elderly is obtained by weighting and summing the dimensions according to the corresponding adjusted weights.

[0055] The effect feedback matrix is ​​updated as follows: the effect score of the prescription on the patient group is set as the ratio of the average improvement to the standard deviation, and its expression is: in, This indicates the effect score of prescription P on patient group R.

[0056] The average improvement is the average of the improvement in each efficacy indicator after multiple administrations of the same prescription in the same patient population, and the standard deviation is the standard deviation of the improvement in each efficacy indicator. This matrix is ​​used to optimize prescription recommendations for similar patients and also serves as input for online fine-tuning of the risk warning model.

[0057] After the prescription is administered, follow-up data is collected and used to update the effect feedback matrix. At the same time, newly collected labeled data is used as incremental samples to fine-tune the risk warning model online, which can be achieved using XGBoost or other incremental learning algorithms.

[0058] S5, Cross-modal causal reasoning By using a cross-modal causal reasoning engine, causal chains are constructed based on multimodal data and medical knowledge graphs to identify key pathogenic factors. Counterfactual inference is then performed under the constraints of the medical knowledge graph to output data predicting the intervention effect.

[0059] The medical knowledge graph is built based on clinical guidelines and expert consensus, storing causal knowledge data in the form of triplets and associating causal strength data. Each triplet contains a cause, relationship, and consequence, such as "weakness, leading to increased risk of falls, intensity 0.7", "sarcopenia, leading to decreased walking speed, intensity 0.8", and "malnutrition, leading to weakness, intensity 0.6".

[0060] The causal chain construction employs a graph neural network to align multimodal features with nodes in a medical knowledge graph and calculates the joint probability of the causal chains. Specifically, multimodal features (such as the degree of stenosis in coronary CTA, gait video gait speed, serum albumin, etc.) are aligned with corresponding nodes in the knowledge graph. The graph neural network calculates the joint probability of each causal path and outputs a causal path that conforms to medical logic. A typical output example is: 50% stenosis of the left anterior descending artery in coronary CTA leads to myocardial ischemia, which in turn leads to decreased exercise tolerance; a gait video gait speed of 0.6 m / s suggests an increased likelihood of sarcopenia, which in turn leads to an increased risk of falls; a serum albumin level of 32 g / L suggests malnutrition, which in turn leads to the progression of weakness; the above paths ultimately lead to a comprehensive decision conclusion.

[0061] Counterfactual inference, constrained by a medical knowledge graph, generates intervention hypothesis data and outputs predicted effect interval data. The specific implementation of counterfactual inference is as follows: Under the constraints of knowledge graphs, reasonable intervention hypotheses are generated to avoid hypotheses that violate common medical sense. Based on the current state and the hypothetical intervention, intervention simulation is carried out through causal models, and the predicted effect range and causal rationality score are output. For example, for the counterfactual question "If nutritional support is strengthened, how much can the patient's risk of falling be reduced?", the system infers the current state as frailty score 4 and fall risk 30%, and the counterfactual hypothesis is that serum albumin is increased to 35g / L (within the clinically feasible range). The inference result is that the frailty score is expected to drop to level 3 and the fall risk is expected to drop to 18%, and the confidence interval is output as 12% to 24%.

[0062] S6. Ethics and Explainability.

[0063] The ethics and interpretability module verifies ethical rules and checks fairness based on the generated prescription data, and outputs a structured report containing uncertainty quantification parameters.

[0064] Ethical rule validation includes pre-filtering based on contraindication rule data. Contraindication rules are extracted from authoritative clinical guidelines, and the built-in contraindication rules include: (1) Automatically reduce the weight of high-intensity exercise prescriptions for severely frail patients (score ≥ 5); (2) Non-pharmacological interventions are recommended as a priority for patients using multiple medications (≥5 types of medications); (3) Avoid recommending high-risk movements in balance training for patients with a history of falls.

[0065] The rules engine performs pre-filtering before prescription generation to ensure that the output prescription meets clinical safety requirements.

[0066] Fairness testing includes detecting decision-making bias data across different groups. Specifically, it involves periodically testing for decision-making biases in different genders and age groups. For example, it checks whether the proportion of male patients recommended for resistance training is significantly higher than that of female patients, or whether elderly patients are over-recommended for conservative treatment. If biases are detected, feature weights are automatically adjusted or manual review is triggered to ensure the fairness of the decision-making process.

[0067] Uncertainty quantification includes output confidence interval data and evidence level data. Specifically, uncertainty quantification information is added to each recommendation in the output report. For example: it is recommended to conduct resistance training twice a week for 30 minutes each time; the basis is the effectiveness based on similar cases (sample size of 50 cases, age 70 to 75 years, frailty score of 3 to 4), the evidence level is B (expert consensus); the confidence interval is 75% to 85%.

[0068] In practice, the system workflow of this embodiment includes an initial diagnosis process, a follow-up diagnosis process, and a continuous learning process.

[0069] Initial diagnosis process: First, the system accesses the hospital's information system, image archiving and communication system, and wearable devices via a standardized external resource access protocol to obtain data such as the patient's electronic medical record, coronary CTA images, laboratory indicators, and recent gait videos. The main control agent calls upon the geriatric comprehensive assessment agent to perform the initial comprehensive assessment and generate a baseline geriatric comprehensive assessment report; simultaneously, it calls upon the medical image analysis agent to analyze the coronary CTA and output results such as plaque characteristics and stenosis degree. The initial assessment results are stored in the proximal memory layer as the starting point of the patient's longitudinal trajectory. The risk warning agent integrates multimodal information such as the geriatric comprehensive assessment results, image features, and laboratory data to output an initial risk level. The causal reasoning agent identifies current key pathogenic factors based on a medical knowledge graph. The prescription generation agent invokes ethical verification skills to generate an initial personalized prescription plan, along with uncertainty quantification information. The generated prescription and its expected effect prediction are written into the effect feedback matrix in the long-term memory layer.

[0070] Follow-up process: The master control agent loads the trajectory data of the patient's indicators from previous visits from the proximal memory layer, and uses trajectory analysis skills to identify the deterioration or improvement trends of the indicators. If any dynamic indicator is detected to be rapidly deteriorating (the rate of change exceeds the dynamic optimization threshold), the system automatically triggers a re-execution of the comprehensive geriatric assessment. During the comprehensive geriatric assessment, the specialized intelligent agent dynamically adjusts the weights of each assessment dimension based on the current risk level and the patient's individual historical variability. It uses causal reasoning skills, combining the latest collected data with a medical knowledge graph, to identify key pathogenic factors and causal chains leading to disease deterioration. The prescription generation specialized intelligent agent combines historical effect feedback from long-term memory to select the prescription template with the best effect on similar patients, and again uses ethical verification skills for compliance filtering. Optional counterfactual reasoning can be performed, providing a comparison of the predicted intervention effects of "option A versus option B" to assist doctors in decision-making. The prescription generated during this visit and the expected effects are written into the memory layer, awaiting verification of the actual intervention effect from the next follow-up data.

[0071] Continuous learning process: During the next visit, the system automatically compares the actual improvement effect with the expected effect, calls the effect feedback skill to calculate the actual effect score of this prescription, writes the actual effect score into the long-term memory layer, and updates the experience weight of the corresponding prescription in the patient population. Newly collected labeled data is used as incremental samples to fine-tune the risk warning model online. Through federated learning, the anonymized effect gradient or model update parameters are uploaded to the cloud-based federated learning training cluster to participate in multi-center model optimization; after cloud aggregation, the global model parameters are distributed to the edge devices of various medical institutions to achieve cross-institutional knowledge sharing and continuous model evolution.

[0072] As a specific embodiment, the overall architecture of the system of this invention includes two parts: a cloud center and a medical institution edge terminal. The cloud center comprises a federated learning training cluster, a multi-center experience aggregation service, a global knowledge base, a causal knowledge graph, and an ethical rule base. The cloud communicates with each edge terminal through a standardized external resource access protocol, and is responsible for model distribution, knowledge updates, and federated aggregation. The medical institution edge terminal deploys a multi-agent framework runtime environment, which specifically includes a temporal dynamic memory network construction module, a memory-driven adaptive triggering engine, a multi-agent decision center, a standardized external resource access server cluster, a bidirectional closed-loop optimization engine, a cross-modal causal reasoning engine, and an ethics and interpretability module. All patient privacy data at the edge terminal is processed locally; only the anonymization effect gradient is uploaded to participate in cloud-based federated learning.

[0073] The system comprises three modules: a temporal dynamic memory network construction module, a short-term memory layer, a proximal memory layer, and a long-term memory layer, which store the current session context, patient longitudinal trajectory data, patient stability profile, and historical intervention effect data, respectively; a memory-driven adaptive triggering engine, which monitors the rate of change of dynamic indicators stored in the proximal memory layer and triggers the re-collection and processing of elderly comprehensive assessment data when the rate of change exceeds the dynamic optimization threshold; and a multi-agent decision-making system construction module, which constructs a multi-agent decision-making system based on the temporal dynamic memory network and the assessment data acquired after triggering, including a master agent and multiple specialized agents. The master agent performs task planning and calls specialized agents to execute sub-tasks. Each agent obtains information from the temporal dynamic memory network and an external knowledge base through a retrieval-enhanced generation mechanism and generates decision-making suggestions. A bidirectional closed-loop optimization engine adjusts the weight parameters of each dimension of the elderly comprehensive assessment based on the risk level data output by the risk warning module and updates the effect feedback matrix based on the effect data after prescription execution. The cross-modal causal reasoning engine constructs causal chains based on multimodal data and a medical knowledge graph, identifies key pathogenic factors, and performs counterfactual reasoning under the constraints of the medical knowledge graph to output predicted intervention effects. The ethics and interpretability module verifies ethical rules and checks fairness based on the generated prescription data, outputting a structured report containing quantified uncertainty parameters. Through this cloud-edge collaborative architecture and the coordinated functions of each module, this embodiment enables dynamic tracking and intelligent decision support for the entire lifecycle of data from elderly patients with comorbidities.

[0074] The present invention also discloses a computer-storable medium storing computer instructions, which, when executed by a processor, specifically perform the steps of any of the methods described above.

[0075] The present invention also discloses a computer program product, including computer instructions, which, when executed by a processor, specifically perform the steps of any of the methods described above. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

[0076] Table 1

Claims

1. A diagnostic decision support method based on multi-agent and temporal dynamic memory, characterized in that, include: A temporal dynamic memory network is constructed, which includes a short-term memory layer, a proximal memory layer, and a long-term memory layer, which are used to store the current session context, patient longitudinal trajectory data, and patient stable profile and historical intervention effect data, respectively. Based on the rate of change of dynamic indicators stored in the proximal memory layer, the memory-driven adaptive triggering engine monitors the data. When the rate of change exceeds the dynamic optimization threshold, the data of the comprehensive assessment of the elderly is re-collected and processed. Based on the temporal dynamic memory network and the evaluation data obtained after triggering, a multi-agent decision-making system is constructed. The multi-agent decision-making system includes a master agent and professional agents capable of risk warning, comprehensive elderly assessment, and prescription data generation. The master agent performs task planning and calls on professional agents to execute sub-tasks. Each agent obtains information from the temporal dynamic memory network and external knowledge base through a retrieval-enhanced generation mechanism and generates decision-making suggestions. Based on the risk level data output by the professional intelligent agent, the weight parameters of each dimension of the comprehensive assessment of the elderly are adjusted through a two-way closed-loop optimization engine, and the effect feedback matrix is ​​updated based on the effect data after the prescription is executed. Based on multimodal data and medical knowledge graphs, a causal chain is constructed through a cross-modal causal reasoning engine to identify key pathogenic factors. Under the constraints of the medical knowledge graph, counterfactual inference is performed to output predicted data on intervention effects. Based on the generated prescription data, ethical rules are verified and fairness is tested, and a structured report containing uncertainty quantification parameters is output.

2. The method according to claim 1, characterized in that, The longitudinal trajectory data stored in the proximal memory layer includes indicator change curves and individual historical variability data; The long-term memory layer stores an effect feedback matrix that records historical prescription data and its effect score data.

3. The method according to claim 1, characterized in that, The dynamic optimization threshold is adjusted online using a reinforcement learning algorithm, and the reward function of the reinforcement learning is constructed based on the weighted sum of the false negative rate and the false positive rate.

4. The method according to claim 1, characterized in that, The multi-agent decision-making system configures an independent workspace for each agent. The workspace includes configuration files for defining roles, collaborative relationships, invoking skills, storing memories, recording preferences, and identifying identities.

5. The method according to claim 1, characterized in that, The specialized intelligent agent is used to perform sub-tasks such as comprehensive elderly assessment data processing, image data analysis, risk warning data generation, prescription data generation, trajectory data analysis, causal reasoning data generation, and ethical verification data generation. Each professional intelligent agent corresponds to at least one skill, and the skills are loaded in a hierarchical manner.

6. The method according to claim 1, characterized in that, The adjustment of the weight parameters of each dimension of the comprehensive geriatric assessment includes: weighting the adjusted weight parameters of the dimension based on the basic weight parameters of that dimension, risk level data, and patient historical score variability data. The adjusted comprehensive assessment score data for the elderly is determined by the weighted sum of the original scores for each dimension and the adjusted weight parameters.

7. The method according to claim 1, characterized in that, The update of the effect feedback matrix includes: determining the effect score data of the prescription on the patient group based on the ratio of the average improvement data to the standard deviation data.

8. The method according to claim 1, characterized in that, The medical knowledge graph stores causal knowledge data in the form of triples and associates it with causal strength data. The counterfactual inference, under the constraints of the medical knowledge graph, generates intervention hypothesis data and outputs predicted effect range data.

9. The method according to claim 1, characterized in that, The ethical rule verification includes pre-filtering based on taboo rule data; The fairness test includes detecting decision-making bias data among different groups; The uncertainty quantification includes outputting confidence interval data and evidence level data.

10. A diagnostic decision support system based on multi-agent and temporal dynamic memory, characterized in that, include: A temporal dynamic memory network construction module is used to construct a temporal dynamic memory network, which includes a short-term memory layer, a proximal memory layer, and a long-term memory layer, which are used to store the current session context, patient longitudinal trajectory data, and patient stable profile and historical intervention effect data, respectively. The memory-driven adaptive triggering engine is used to monitor the rate of change of dynamic indicators stored in the proximal memory layer. When the rate of change exceeds the dynamic optimization threshold, it triggers the re-collection and processing of comprehensive elderly assessment data. A multi-agent decision-making system construction module is used to construct a multi-agent decision-making system based on the temporal dynamic memory network and the evaluation data obtained after triggering. The multi-agent decision-making system includes a master agent and multiple professional agents. The master agent performs task planning and calls professional agents to execute sub-tasks. Each agent obtains information from the temporal dynamic memory network and external knowledge base through a retrieval enhancement generation mechanism and generates decision suggestion information. The two-way closed-loop optimization engine is used to adjust the weight parameters of each dimension of the comprehensive assessment of the elderly based on the risk level data output by the risk warning module, and to update the effect feedback matrix based on the effect data after the prescription is implemented. A cross-modal causal reasoning engine is used to construct causal chains based on multimodal data and medical knowledge graphs, identify key pathogenic factors, and perform counterfactual reasoning under the constraints of the medical knowledge graphs to output intervention effect prediction data. The Ethics and Interpretability module is used to perform ethical rule verification and fairness checks on the generated prescription data, and output a structured report containing uncertainty quantification parameters.

Images