Elderly care-oriented medical and nursing combined knowledge service system construction method

Abstract

The invention relates to the technical field of medicine and artificial intelligence, and discloses an elderly care-oriented medical and nursing combined knowledge service system construction method, which comprises the following steps of collecting sentiment expression data of the elderly, applying a nonlinear time sequence analysis and deep learning model, extracting sentiment dynamic characteristics, and obtaining sentiment expression data of the elderly; recognizing an early signal with reduced emotion regulation ability; on the basis of the obtained emotion dynamic features, a memory enhanced attention network is combined with a psychological knowledge graph, an individualized emotion cognitive interaction model of the old people is constructed, a dynamic interaction mechanism between an emotion state and a cognitive mode is revealed, and key intervention nodes are recognized; based on an emotion cognition interaction model and a key node recognition algorithm, the system can accurately position key links of mutual influence of emotion and cognition, and targeted intervention is achieved; and a dual-channel collaborative intervention strategy adjusts an emotional state and reconstructs a cognitive mode through multi-agent reinforcement learning, and breaks through an emotional cognitive vicious circle.

Description

Technical Field

[0001] The present invention relates to the fields of medicine and artificial intelligence technology, and more specifically, to a method for constructing a medical and nursing integrated knowledge service system for elderly care. Background Art

[0002] With the accelerated aging of the population, mental health issues are becoming increasingly prominent among the elderly, especially those living alone. Mental illnesses such as depression and anxiety have become significant factors affecting their quality of life. Currently, managing mental health in the elderly faces multiple challenges: First, mental health issues in the elderly are often hidden, making traditional screening methods difficult to identify early. Second, resources for psychological interventions for the elderly are limited, making interventions ineffective.

[0003] The existing technologies have the following main deficiencies: First, they mainly rely on cross-sectional assessments of the immediate emotional state of the elderly, ignoring the dynamic characteristics of emotional changes, and making it difficult to distinguish between normal emotional fluctuations and pathological changes; second, they lack recognition of the unique emotional expression patterns of the elderly, making it difficult to capture implicit psychological needs; third, they often focus solely on emotional symptoms or cognitive distortions, lack a collaborative intervention mechanism, and are unable to effectively break the vicious cycle between emotion and cognition; fourth, they lack personalized adaptability and are unable to meet the individual needs of different elderly people.

[0004] Therefore, there is an urgent need for a knowledge service system that can achieve early, accurate identification and efficient intervention, and improve the efficiency and accuracy of elderly mental health management through multimodal data analysis, intelligent algorithms and comprehensive intervention strategies. Summary of the Invention

[0005] The present invention provides a method for constructing a medical-nursing integrated knowledge service system for elderly care, which solves the technical problems of constructing a medical-nursing integrated knowledge service system in related technologies.

[0006] The present invention provides a method for constructing a medical and nursing integrated knowledge service system for elderly care, comprising the following steps:

[0007] Collect emotional expression data from older adults, apply nonlinear time series analysis and deep learning models to extract dynamic emotional features and identify early signs of decreased emotional regulation ability;

[0008] Based on the acquired dynamic emotional features, we use a memory-enhanced attention network combined with a psychological knowledge graph to construct an individualized emotional-cognitive interaction model for the elderly, revealing the dynamic interaction mechanism between emotional states and cognitive patterns and identifying key intervention nodes.

[0009] Based on the constructed interaction model, a multi-agent reinforcement learning algorithm is applied to generate an emotion-cognition dual-channel collaborative intervention strategy to maximize the effect of precise intervention;

[0010] Implement the generated intervention strategies, establish a closed-loop intervention execution and feedback mechanism, achieve dynamic evaluation of intervention effects and strategy optimization, and improve the accuracy and sustainability of mental health management for the elderly;

[0011] Integrate the functional modules into the medical and nursing knowledge service system to achieve collaborative work with other health management modules and build a complete elderly mental health service ecosystem.

[0012] As a further optimization scheme of the present invention, the functional modules include an emotion dynamic analysis module, an emotion cognitive interaction modeling module, a dual-channel collaborative intervention module, and an intervention execution module.

[0013] As a further optimization solution of the present invention, the extraction of emotional dynamic features specifically includes:

[0014] Collect multimodal data of the elderly through sensor networks, including voice, facial expressions, activity patterns, and physiological signals, as the basic input for affective state analysis;

[0015] Applying a deep multimodal fusion algorithm to extract characteristic parameters of emotional expression, including emotional intensity, emotional diversity, emotional inertia, and emotional resilience, and constructing an emotional dynamic feature vector;

[0016] Using nonlinear time series analysis methods, we conduct complexity analysis on the time series of emotional dynamic feature vectors and quantify the complexity changes of the emotional dynamic system.

[0017] A hybrid architecture of hidden Markov model and long short-term memory network is constructed to capture the long-term evolution pattern of emotional state transition and identify abnormal transition patterns.

[0018] As a further optimization solution of the present invention, the emotional cognitive interaction model construction specifically includes:

[0019] Based on the dynamic characteristics of emotions and cognitive evaluation data, the emotional state vector and cognitive pattern vector are constructed;

[0020] A memory-enhanced attention network is used to construct a bidirectional influence relationship model between emotion and cognition, capturing the mutual influence relationship between the two and storing historical state information through the attention weight matrix.

[0021] Based on the individual data of the elderly, an optimization algorithm combining gradient descent and variational inference is used to train an individualized emotion-cognitive interaction model.

[0022] Applying causal inference methods, based on the trained interaction model, we identify individual-specific key nodes of emotional cognitive interaction as the key targets for the formulation of subsequent intervention strategies.

[0023] As a further optimization solution of the present invention, the dual-channel collaborative intervention strategy generation specifically includes:

[0024] Based on the emotion-cognition interaction model, define the intervention optimization objective function and constraints;

[0025] Construct a multi-agent system consisting of an emotional intervention agent and a cognitive intervention agent, sharing state information and coordinating action selection through a collaborative mechanism to maximize joint rewards;

[0026] A deep deterministic policy gradient algorithm is used to optimize the intervention strategy, and a collaborative decision-making mechanism is designed to coordinate the timing of emotional intervention and cognitive intervention to maximize the synergistic effect.

[0027] As a further optimization solution of the present invention, the intervention execution specifically includes:

[0028] Build a multimodal intervention execution system that integrates multiple interaction channels such as virtual assistants, smart home environments, and wearable devices to achieve seamless execution of emotional and cognitive dual-channel intervention strategies;

[0029] Establish a multi-dimensional intervention effect evaluation mechanism, including immediate emotional response assessment, cognitive change assessment, and long-term health status assessment;

[0030] Bayesian adjustment methods and incremental learning models are applied to optimize intervention strategy parameters in real time based on intervention effectiveness and acceptance.

[0031] As a further optimization solution of the present invention, the knowledge service system integration and application specifically includes:

[0032] Build a modular system architecture, encapsulate technical components into standardized service modules, and form a scalable medical and nursing knowledge service system;

[0033] Realize the collaborative working mechanism between mental health management and other health management modules to build a panoramic picture of the health status of the elderly;

[0034] Build a personalized interaction strategy generation algorithm based on the elderly’s personal interests, life experiences, and social preferences;

[0035] Build a dynamic update and optimization mechanism for the knowledge base, continuously accumulate experience from intervention practice, and improve overall service quality.

[0036] As a further optimization solution of the present invention, the nonlinear time series analysis method includes calculating the entropy increase rate, Lyapunov index and fractal dimension of the emotional trajectory to quantify the complexity changes of the emotional dynamic system.

[0037] As a further optimization scheme of the present invention, the memory-enhanced attention network includes three parts: an encoder, a memory module and a decoder, wherein the memory module maintains a historical state buffer, and the multi-head attention mechanism is used to calculate the correlation between the current state and the historical memory.

[0038] A system for constructing a medical-nursing integrated knowledge service system for elderly care, configured to execute the aforementioned method for constructing a medical-nursing integrated knowledge service system for elderly care, comprising:

[0039] The emotional dynamics analysis module is used to collect and analyze multimodal emotional data of the elderly, extract emotional dynamic features, identify early signs of decreased emotional regulation ability, and provide basic data support for subsequent interventions;

[0040] The emotion-cognition interaction modeling module is used to build an individualized emotion-cognition interaction model, reveal the dynamic relationship between emotional state and cognitive pattern, identify key intervention nodes, and provide a theoretical basis for formulating intervention strategies;

[0041] The dual-channel collaborative intervention module is used to generate emotion-cognition dual-channel collaborative intervention strategies based on multi-agent reinforcement learning, achieving the optimal coordination of emotion and cognitive interventions and maximizing the intervention effect;

[0042] A personalized intervention execution module is used to execute intervention strategies through multimodal interaction channels, evaluate intervention effects in real time, and dynamically adjust intervention parameters to ensure the accuracy and sustainability of interventions;

[0043] The knowledge service integration module is used to integrate various functional modules, manage knowledge base updates, coordinate connections with other health management systems, and provide a unified service interface;

[0044] Each module communicates through a standardized API and adopts an event-driven model to achieve loosely coupled integration.

[0045] The beneficial effects of the present invention are: based on the emotion-cognition interaction model and the key node identification algorithm, the present invention enables the Zhengegege system to accurately locate the key links in the mutual influence of emotion and cognition, and achieve targeted intervention; and the dual-channel collaborative intervention strategy uses multi-agent reinforcement learning to simultaneously regulate emotional states and reconstruct cognitive patterns, effectively breaking the vicious cycle of emotion and cognition, and greatly improving the effect-output ratio of unit intervention resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a method for constructing a medical and nursing integrated knowledge service system for elderly care according to the present invention;

[0047] Figure 2 It is a detailed flow chart of the emotional dynamic feature extraction and analysis of the present invention;

[0048] Figure 3 It is a detailed flow chart of the construction of the emotion cognition interaction model of the present invention;

[0049] Figure 4 It is a detailed flow chart of the generation of the dual-channel collaborative intervention strategy of the present invention;

[0050] Figure 5 It is a detailed flow chart of the personalized intervention execution and feedback adjustment of the present invention;

[0051] Figure 6 It is a detailed flow chart of the integration and application of the knowledge service system of the present invention. DETAILED DESCRIPTION

[0052] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0053] At least one embodiment of the present invention discloses a method for constructing a medical and nursing integrated knowledge service system for elderly care, such as Figures 1 to 6 As shown, the following steps are included:

[0054] Step 1: Collect emotional expression data from the elderly, apply nonlinear time series analysis and deep learning models to extract dynamic emotional features and identify early signs of decreased emotional regulation ability;

[0055] The specific steps include:

[0056] Step 1.1: Multimodal emotion data collection;

[0057] The sensor network collects multimodal data of the elderly, including voice, facial expressions, activity patterns, and physiological signals, which serve as the basic input for emotional state analysis. Specifically, the sensor network includes: an audio acquisition module, a video acquisition module, a wearable physiological signal monitoring module, and an environmental interaction behavior recording module. In addition, the system performs preliminary cleaning and standardization on the collected raw data to form a unified format of the original emotional expression data set. .

[0058] Step 1.2, emotional dynamic feature vector construction;

[0059] Based on the collected multimodal data, this application applies a deep multimodal fusion algorithm to extract the characteristic parameters of emotional expression, including emotional intensity , emotional diversity , emotional inertia and emotional resilience Etc., construct the emotional dynamic feature vector:

[0060]

[0061] in, Represents the emotional dynamic feature vector; Indicates the intensity of emotional expression; An index representing the diversity of emotion types; Indicates the persistence of an emotional state; Indicates the rate of recovery from negative affect to baseline state; Represents the transpose of a vector.

[0062] The system uses a cross-modal self-attention mechanism to integrate emotional features of different modalities to ensure the robustness of feature extraction and avoid the impact of information loss of a single modality on feature extraction.

[0063] Step 1.3, dynamic analysis of emotional complexity;

[0064] After obtaining the emotional dynamic feature vector, this application uses a nonlinear time series analysis method to analyze the emotional dynamic feature vector. Specifically, the entropy increase rate of the emotional trajectory is calculated , Lyapunov exponent and fractal dimension , quantify the complexity changes of emotional dynamic systems:

[0065]

[0066] in, Indicates time The amount of change in emotional complexity; Indicates time emotional complexity; Indicates time emotional complexity; Indicates the time window size, Represents the time step.

[0067] The complexity index calculation formula is as follows:

[0068] Entropy increase rate: ;

[0069] in, represents the entropy increase rate; Indicates when the limit as it approaches infinity; represents the normalization coefficient; represents the Shannon entropy function; 、 、 Respectively represent 、 、 Emotional state at a point in time, Indicates the length of the time series.

[0070] Lyapunov exponent: ;

[0071] in, represents the Lyapunov exponent; represents a continuous-time variable; Indicates the deviation from the initial state; Indicates time The deviation when Indicates when the limit as it approaches infinity; Indicates the limit when the initial deviation approaches 0; represents the time normalization coefficient; Represents the natural logarithm.

[0072] Fractal dimension The fractal characteristics of emotional trajectories are calculated using the box counting method. Therefore, the changing trends of these complexity indicators can effectively distinguish normal emotional fluctuations from pathological changes, providing a quantitative basis for early warning.

[0073] Step 1.4, emotional state transition pattern recognition;

[0074] Finally, this application constructs a hybrid architecture of Hidden Markov Model and Long Short-Term Memory (LSTM) network to capture the long-term evolution pattern of emotional state transition. In this hybrid architecture, Hidden Markov Model is used to represent the transition probability matrix between discrete emotional states. ,in Indicates emotional state Transfer to state Correspondingly, the LSTM network learns sequence patterns, predicts the future direction of the emotional state sequence, and identifies abnormal transition patterns. It can be seen that the model outputs early warning signals of decreased emotional regulation ability. , when the warning index exceeds the threshold When a person is asked to do something, the system flags potential mental health risks.

[0075] Specifically, the hybrid architecture is implemented as follows: First, the emotional state space is discretized into a finite number of state sets:

[0076]

[0077] in, Represents the emotional state space; 、 、 Respectively represent 、 、 an emotional state; Represents the total number of emotional states.

[0078] Each state represents a combination of emotional states (such as "mild anxiety", "moderate depression", etc.). Then, the hidden Markov model is based on the observed sequence of emotional features:

[0079]

[0080] in, represents the observed emotional feature sequence; 、 、 Respectively represent 、 、 observed sentiment features; Represents the total number of observed sentiment features.

[0081] Estimated state transition probability matrix and the observed probability distribution On this basis, the LSTM network further captures long-term dependencies, and its input is a sequence of sentiment feature vectors:

[0082]

[0083] in, 、 、 Respectively represent 、 、 sentiment feature vectors; Represents the total number of sentiment feature vectors.

[0084] The output is the probability distribution of the predicted emotional state at the next moment.

[0085] The workflow of this hybrid architecture in a specific application scenario is as follows: When an elderly person is continuously monitored by the system in their daily lives, the system extracts emotional features from multimodal data and generates a sequence of emotional feature vectors. This sequence is then input into a hybrid architecture of a hidden Markov model and LSTM to calculate the probability distribution of emotional state transitions and the predicted value of the future emotional state. If the system detects an unusual transition pattern, such as a sudden increase in the probability of transitioning from "mild anxiety" to "severe depression," or if the LSTM-predicted future emotional state shows a sustained deterioration, an early warning signal is generated. For example, if speech analysis of an elderly person living alone shows a decrease in emotional intensity and diversity over several consecutive days, and their activity patterns indicate a decrease in social interaction, the hybrid architecture can predict that the individual is at risk of developing depression. The system then triggers an early warning and intervenes.

[0086] Step 2: Based on the acquired emotional dynamic features, we use a memory-enhanced attention network combined with a psychological knowledge graph to construct an individualized emotional cognitive interaction model for the elderly, revealing the dynamic interaction mechanism between emotional state and cognitive pattern, and identifying key intervention nodes;

[0087] The specific steps include:

[0088] Step 2.1, emotional state and cognitive pattern representation;

[0089] Based on the emotional dynamic characteristics and cognitive evaluation data obtained in step 1, the emotional state vector is constructed and cognitive pattern vectors .

[0090] Emotional state vector: ;

[0091] in, represents the emotional state vector; 、 、 Respectively represent 、 、 the intensity of each emotional dimension; Indicates the total number of emotional dimensions; Represents the transpose of a vector.

[0092] Cognitive Pattern Vector: ;

[0093] in, represents the cognitive pattern vector; 、 、 Respectively represent 、 、 the degree of each cognitive dimension; represents the total number of cognitive dimensions; Represents the transpose of a vector.

[0094] Specifically, the affective state vector contains multidimensional affective features, while the cognitive pattern vector includes elements such as cognitive distortions, belief systems, and thought patterns. It should be noted that the system extracts cognitive pattern features from the elderly's language expressions using natural language processing technology and combines them with psychological scale assessment data to form a comprehensive cognitive pattern representation.

[0095] Step 2.2: Modeling of bidirectional influence relationships;

[0096] This application uses a memory-enhanced attention network to construct a bidirectional influence relationship model between emotion and cognition. This model captures the mutual influence relationship between the two:

[0097]

[0098]

[0099] in, express The emotional state vector at the moment; express The emotional state vector at the moment; express The cognitive pattern vector of the moment; Model parameters representing the emotional state transition function; a transition function representing the affective state; express The cognitive pattern vector of the moment; transfer functions representing cognitive patterns; Model parameters representing cognitive mode transfer function; Represents the time step.

[0100] It should be noted that the memory enhancement mechanism uses the attention weight matrix Store historical status information, formally expressed as:

[0101]

[0102] in, represents the attention weight matrix; represents the query matrix; represents the transpose of the key-value matrix; represents the hidden layer dimension; Represents the softmax activation function.

[0103] The model is able to learn long-term dependencies that reflect the long-term accumulated mutual influence between affective states and cognitive patterns.

[0104] The specific implementation of the memory-enhanced attention network is as follows: The network architecture consists of three parts: an encoder, a memory module, and a decoder. The encoder maps the emotional state vector and cognitive pattern vector to a latent space representation. The memory module maintains a historical state buffer:

[0105]

[0106] in, Represents a memory module; 、 、 Respectively represent 、 、 memory units; Indicates the total number of memory cells.

[0107] Each memory unit stores past state information. The multi-head attention mechanism is used to calculate the correlation between the current state and historical memory:

[0108]

[0109] in, Indicates the The output of an attention head; Represents the attention calculation function; 、 and Respectively represent The query, key, and value matrices of each attention head are obtained by linear transformation of the current state and history memory; represents the transpose of the bond matrix; represents the dimension of the key vector; Represents the softmax activation function.

[0110] The outputs of the multi-head attention are integrated through concatenation and linear transformation:

[0111]

[0112] in, represents the output of the multi-head attention mechanism, represents the query matrix, represents the bond matrix, represents the value matrix; 、 、 Respectively represent 、 、 The output of an attention head; represents the total number of attention heads; Represents a linear transformation matrix.

[0113] The decoder predicts the next moment’s emotional state and cognitive pattern based on the encoder output and attention-weighted historical memory.

[0114] In a specific application scenario, the model's workflow is as follows: When an elderly person is detected experiencing emotional fluctuations, the system collects their recent emotional state data and cognitive assessment data and inputs them into the memory-enhanced attention network. For example, an elderly person exhibits persistent sadness after the death of their spouse. The system captures a continuous increase in the sadness dimension of their emotional state vector, while "excessive negative thinking" and "diminished self-worth" in their cognitive pattern vector also gradually increase. Through historical data analysis, the memory-enhanced attention network discovers the interactive reinforcement process of this emotional cognitive pattern and predicts that without intervention, the elderly person may develop depression. The system identifies "excessive negative thinking" as a key intervention node because attention weight analysis shows that this cognitive pattern has the highest weight on sadness, and formulates targeted intervention strategies accordingly.

[0115] Step 2.3: Individualized model training and optimization;

[0116] Based on the individual data of the elderly, this application uses an optimization algorithm that combines gradient descent and variational inference to train an individualized emotion-cognitive interaction model. Specifically, the objective function is set to minimize the weighted sum of prediction error and model complexity:

[0117]

[0118]

[0119] in, represents the objective function; 、 and represent the weight coefficients of the emotional prediction error, cognitive prediction error, and regularization term respectively; represents the mean square error function; represents the predicted affective state; Indicates a true emotional state; Represents a cognitive model of prediction; Represents a true cognitive pattern; represents the regularization term; Represents model parameters.

[0120] In addition, the system dynamically adjusts the weight coefficients to balance the model's fit and generalization ability, and outputs individualized emotion cognition interaction model parameters. .

[0121] Step 2.4, identification of key intervention nodes;

[0122] Finally, this application applies causal inference methods to identify individual-specific emotional cognitive interaction key nodes based on the trained interaction model. It should be understood that a critical node is defined as the point where the impact of emotion on cognition or cognition on emotion exceeds a threshold:

[0123]

[0124] in, Represents a set of key nodes; Indicates the emotional dimensions; Indicates the cognitive dimensions; Indicates the strength of the influence of cognitive dimensions on affective states; Indicates the strength of the influence of the affective dimension on the cognitive pattern; represents the emotional impact threshold; represents the cognitive impact threshold; represents the index of the sentiment dimension, An index representing a cognitive dimension.

[0125] Therefore, the system quantifies the influence strength between each dimension by calculating the partial derivative matrix, and combines the sensitivity analysis method to determine the most influential key node set. , as the key target for the formulation of subsequent intervention strategies, the calculation formula of the partial derivative matrix is:

[0126]

[0127]

[0128] in, A matrix representing the impact of emotion on cognition; A matrix representing the impact of cognition on emotion; Indicates the strength of the influence of cognitive dimensions on affective states; Indicates the intensity of the influence of the emotional dimension on the cognitive pattern.

[0129] Step 3: Based on the constructed interaction model, a multi-agent reinforcement learning algorithm is applied to generate an emotion-cognition dual-channel collaborative intervention strategy to maximize the precise intervention effect;

[0130] The specific steps include:

[0131] Step 3.1: Definition of intervention objectives and constraints;

[0132] Based on the emotion-cognitive interaction model constructed in step 2, this implementation defines the intervention optimization objective function:

[0133]

[0134] in, represents the objective function; indicates intervention strategies; Indicates that in the strategy expectations under represents the time step; Indicates that from the time step arrive The sum of the discount rewards; Represents the discount factor Power, used to balance the importance of current and future rewards; represents the total number of time steps; express The state of the moment; express intervention actions taken at all times; represents the reward function.

[0135] Specifically, the status Contains the emotional state vector , cognitive pattern vector and the environmental context vector , formally expressed as:

[0136]

[0137] in, express The state vector at the moment; 、 、 Represent the emotional state vector, cognitive mode vector and environmental context vector at time t respectively; Represents the transpose operator.

[0138] The intervention action space is divided into the emotional intervention action subspace and the cognitive intervention action subspace; its expression is:

[0139]

[0140] in, represents the intervention action space; represents the emotion intervention action subspace; represents the cognitive intervention action subspace.

[0141] Reward Function It is designed based on factors such as the degree of improvement in mental health status, resource consumption and acceptance of the elderly.

[0142] Additionally, constraints include resource limitations and psychological burden limitations ;

[0143] in, Represents resource consumption function; Indicates the resource limit threshold; represents the psychological burden function; Indicates the psychological burden limit threshold; Indicates status; Indicates intervention action.

[0144] Step 3.2: Design a multi-agent reinforcement learning framework.

[0145] This embodiment constructs an emotional intervention agent and cognitive intervention agents A multi-agent system composed of each agent. Each agent optimizes its own Q-value function based on the Deep Q-Network (DQN) architecture.

[0146] The two agents share state information and coordinate action selection through a collaborative mechanism to maximize joint rewards. It should be noted that this multi-agent reinforcement learning framework can simultaneously optimize strategies for both emotional and cognitive intervention channels, achieving a synergistic intervention effect.

[0147] The specific implementation of the multi-agent reinforcement learning framework is as follows: the system adopts the multi-agent reinforcement learning (MARL) framework based on the attention mechanism, including the emotional intervention agent and cognitive intervention agents Two sub-agents and a central critic network (Critic) are used to evaluate the value of joint actions;

[0148] The policy network structure of each agent consists of the following components:

[0149] 1. State encoder: convert the state Mapping to latent representation ;

[0150] 2. Action Generator: Generates action probability distribution based on state representation ;

[0151] 3. Value Function Estimator: Estimating State Value ;

[0152] Collaboration between agents is achieved through an attention mechanism, where each agent pays attention to the state and actions of other agents:

[0153]

[0154] in, Representing an agent The context vector of Indicates the removal of the agent All other agents except sum; Representing an agent For intelligent agents The attention weight of Representing an agent Status indication.

[0155] The calculation formula is:

[0156]

[0157] in, Representing an agent For intelligent agents The attention weight of represents the exponential function; It is a similarity metric function used to calculate the correlation or similarity between the state representations of two agents; Representing an agent Status representation; Representing an agent Status representation; Indicates the removal of the agent All other agents except sum; Representing an agent Status indication.

[0158] The training process uses the Actor-Critic framework, and the policy gradient update of each agent is as follows:

[0159]

[0160] in, Indicates the parameters Gradient operator of ; Representing an agent The objective function of Indicates about status and actions expectations; Represents the natural logarithm function, used to calculate the policy function The log-likelihood of ; Representing an agent The policy function of Representing an agent actions; Indicates status; Representing an agent Strategy network parameters; represents the Q-value function of the central critic network; Indicates emotional intervention action; represents cognitive intervention actions; represents the parameters of the central critic network.

[0161] To handle non-stationary training environments, the system implements an experience replay buffer and target network, and adopts soft update mechanism:

[0162]

[0163] in, represents the target network parameters; represents the soft update coefficient; Indicates the current network parameters; Represents a parameter update operator; Indicates the complement of the soft update coefficient.

[0164] In actual application scenarios, the workflow of this multi-agent reinforcement learning framework is as follows: the system continuously monitors the emotional state and cognitive patterns of the elderly and initiates intervention when abnormal patterns are detected. For example, for an elderly person at risk of mild depression, and By analyzing their emotional state vectors and cognitive pattern vectors, we identified that “social isolation” and “catastrophic thinking” are key issues. Select "increase social participation activities" as the emotional intervention action, and "Cognitive restructuring training" was selected as the cognitive intervention action. The system, through a virtual assistant, recommended nearby community activities (emotional intervention) and provided targeted cognitive behavioral therapy exercises (cognitive intervention). These two intervention channels worked synergistically to improve the elderly person's psychological state. As the intervention process progressed, the system continuously collected feedback data, and the two agents continuously adjusted and optimized their respective intervention strategies to form a personalized intervention plan.

[0165] Step 3.3: Optimization of intervention strategies and collaborative decision-making;

[0166] This implementation utilizes the Deep Deterministic Policy Gradient (DDPG) algorithm to optimize the intervention strategy, including both the policy network and the value network. Experience replay and target network techniques are used to reduce the impact of sample correlation and non-stationarity during training. Furthermore, this implementation also incorporates a collaborative decision-making mechanism to coordinate the timing of emotional and cognitive interventions. Based on the interactive effects of the two-channel interventions:

[0167]

[0168] in, represents the interaction effect of the two-channel intervention; represents the intervention effect function; Indicates emotional intervention action; represents cognitive intervention actions; Indicates empty intervention action.

[0169] In addition, the system determines the optimal intervention sequence based on the strength of each connection and the location of key nodes in the emotion-cognitive interaction model.

[0170] Step 4: Implement the generated intervention strategy, establish a closed-loop intervention execution and feedback mechanism, achieve dynamic evaluation of intervention effects and strategy optimization, and improve the accuracy and sustainability of mental health management for the elderly;

[0171] In another embodiment of the present application, this step establishes a closed-loop intervention execution and feedback mechanism to achieve dynamic evaluation of intervention effects and strategy optimization, thereby improving the accuracy and sustainability of mental health management for the elderly. It should be understood that this step mainly includes the following sub-steps:

[0172] Step 4.1: Construction of multimodal intervention execution system;

[0173] This implementation builds a multimodal intervention execution system that integrates multiple interaction channels, including virtual assistants, smart home environments, and wearable devices, to achieve seamless execution of dual-channel intervention strategies for both emotion and cognition. Specifically, the system is based on a microservices architecture and includes the following core components: intervention strategy parsing service, multimodal interaction coordination service, response generation service, and status monitoring service. Intervention strategies are converted into specific execution instruction sets through the microservices API:

[0174]

[0175] in, 、 、 Respectively represent 、 、 An execution instruction is handed over to the corresponding execution terminal for completion; Indicates the total number of executed instructions.

[0176] The specific implementation of the multimodal intervention execution system is as follows: The system adopts a container-based microservice architecture to ensure the independence and scalability of each component. The core microservices include:

[0177] Policy Parsing Service: This service breaks down high-level intervention strategies into specific execution instruction sequences, addressing logical relationships such as execution order, priority, and preconditions. This service utilizes a combination of decision trees and a rules engine to ensure the correctness and consistency of instructions.

[0178] Interaction Coordination Service: Manages the coordination and switching between multiple interaction channels (such as voice assistants, smart home devices, and mobile applications). It uses a publish-subscribe model for event-driven communication and uses message queues to ensure the sequential processing of interaction events.

[0179] Response Generation Service: Generates tailored responses for different interaction channels. For voice assistants, a Natural Language Generation (NLG) model is used to dynamically generate personalized conversational content. For visual interfaces, a response generation mechanism that combines templates with dynamic content is employed.

[0180] Condition Monitoring Service: This service monitors multimodal sensor data in real time, assesses the immediate effects of current interventions, and detects triggers that may require policy adjustments. This service uses a stream processing architecture and supports Complex Event Processing (CEP) to identify important patterns.

[0181] In a specific application scenario, the system workflow is as follows: After the system generates the collaborative intervention strategy of "increasing social participation + cognitive reconstruction" according to step 3, the strategy analysis service converts it into a series of instructions:

[0182] (1) Push community activity suggestions through smart speakers;

[0183] (2) Arrange cognitive behavioral exercises on mobile applications;

[0184] (3) Adjust the intelligent lighting system to create a positive emotional environment.

[0185] Interaction coordination services ensure that each instruction is executed at the appropriate time and through the most appropriate device. For example, when an elderly person wakes up in the morning, a smart speaker plays a greeting and reminds them of the community gardening activities that day. At the same time, the smart lighting system adjusts to bright, warm tones to boost their mood. If a senior citizen is detected to be depressed, the response generation service dynamically creates conversation content that incorporates positive memories to guide them in emotional regulation. Throughout this process, the status monitoring service continuously evaluates the effectiveness of the intervention. If it finds that social activity suggestions are not being accepted, the system adjusts its strategy to recommend small-scale social activities within the family.

[0186] Step 4.2: Real-time evaluation of intervention effects;

[0187] This implementation establishes a multi-dimensional intervention effectiveness assessment mechanism, encompassing immediate emotional response assessment, cognitive change assessment, and long-term health status assessment. Specifically, the system uses multimodal sensor data to calculate emotional state change vectors and cognitive pattern change vectors in real time, comprehensively evaluating the intervention effectiveness index.

[0188] Emotional state change vector: ;

[0189] in, Represents the emotional state change vector; Indicates current emotional state; Indicates the emotional state at the previous moment.

[0190] Cognitive pattern change vector: ;

[0191] in, represents the cognitive pattern change vector; Indicates the current cognitive mode; Represents the cognitive pattern of the previous moment.

[0192] Intervention effect index: ;

[0193] in, represents the intervention effect index; and represent the weight coefficients of changes in affective states and cognitive patterns, respectively; and evaluation functions representing changes in affective states and cognitive patterns, respectively; Represents the emotional state change vector; Represents the cognitive pattern change vector.

[0194] In addition, the system combines the subjective feedback and objective physiological indicators of the elderly to construct an intervention acceptance index:

[0195]

[0196] in, express Intervention receptivity index at the moment; represents the comprehensive evaluation function; represents the subjective feedback feature set; Represents the objective indicator feature set.

[0197] The real-time evaluation of intervention effects is implemented as follows: the evaluation system adopts a hierarchical temporal evaluation architecture, including three levels: micro (seconds) reaction evaluation, meso (days) state change evaluation, and macro (weeks / months) trend evaluation.

[0198] Micro-evaluation relies primarily on real-time physiological signals and behavioral response data. A variational auto-encoder (VAE) is used to extract a low-dimensional representation of the emotional state and calculate the change in state over a continuous time period. The change in emotional state is quantified using the following formula:

[0199]

[0200] in, Indicates changes in affective states at the micro level; Indicates from arrive The sum operation of Indicates the length of the time window; represents the time weight coefficient, giving higher importance to recent changes; express emotional state at the moment; express Emotional state at the moment.

[0201] The meso-level assessment focuses on changes in the elderly's daily behavior patterns and cognitive performance. The system applies a sequential pattern mining algorithm to extract behavioral pattern changes from daily activity sequences and assesses cognitive function changes through regular lightweight cognitive tests. The changes in cognitive patterns at the meso-level are calculated as follows:

[0202]

[0203] in, It indicates changes in cognitive patterns at the meso-level; Indicates from arrive The sum operation of represents the total number of cognitive dimensions; Indicates the The weight coefficient of each cognitive dimension; Indicates the current moment The scores of cognitive dimensions; express Time ago The scores of cognitive dimensions; Indicates the evaluation time interval.

[0204] The macro-assessment comprehensively analyzes the trend changes of long-term health indicators, social activity participation and mental health scale scores, applies time series decomposition technology to separate trend, seasonal and random components, and focuses on the direction and slope of change of trend components.

[0205] In a specific application scenario, after implementing a one-month intervention program for an elderly individual, the real-time evaluation system can generate multi-level effectiveness reports: At the micro level, the individual's average physiological stress indicators (such as heart rate variability) improved during the virtual natural environment intervention; at the meso level, the individual's frequency of daily social interactions increased, as did their scores on cognitive flexibility tests; and at the macro level, a decrease in their Geriatric Depression Scale (GDS) score, indicating a significant improvement in depressive symptoms. The system also identified that the individual had low receptiveness to cognitive restructuring exercises but high participation in community activities, and adjusted the intervention plan accordingly, increasing the proportion of social interventions.

[0206] Step 4.3, Bayesian adjustment and incremental learning;

[0207] This implementation uses a Bayesian adjustment method and an incremental learning model to optimize intervention strategy parameters in real time based on intervention effectiveness and acceptance. Specifically, the system maintains a posterior distribution of intervention parameters, continuously updates the distribution based on observed data, and selects the optimal parameters:

[0208]

[0209] in, represents the optimal intervention parameter; In parameter space Find the utility function in Maximum parameter; Indicates that the observed data Under the condition, the utility function expected value.

[0210] In addition, a deep incremental learning method is used to update the intervention strategy network parameters, which can be formally expressed as:

[0211]

[0212] in, represents the new intervention strategy network parameters; represents the old intervention strategy network parameters; represents the learning rate; Represents performance indicator function Parameters gradient.

[0213] Therefore, the system can continuously accumulate intervention experience, optimize personalized intervention strategies, and form adaptive personalized intervention plans.

[0214] The specific implementation of Bayesian adjustment and incremental learning is as follows: The system uses a Bayesian optimization framework based on Gaussian processes, representing the intervention parameter space as a multidimensional continuous space. Bayesian optimization uses sampling strategies during the iterative update process to balance exploration and exploitation, effectively finding the optimal parameters with limited sample data.

[0215] The core of Bayesian optimization is to build a proxy model to estimate the true utility function. The system uses a Gaussian process with a radial basis function (RBF) kernel as the proxy model:

[0216]

[0217] in, Represents the kernel function of the Gaussian process; represents the intervention parameter vector to be evaluated, represents the historical intervention parameter vector; represents the covariance amplitude parameter; represents the natural exponential function; represents the length scale parameter; Represents the squared Euclidean distance between two parameter vectors.

[0218] To determine the next set of parameters to evaluate, the system uses the Upper Confidence Bound (UCB) to obtain the function:

[0219]

[0220] in, Represents the confidence upper bound acquisition function; Represents a Gaussian process in The mean prediction at ; represents the parameter that controls the exploration-exploitation balance; Represents a Gaussian process in The standard deviation of the prediction.

[0221] The incremental learning part uses the Elastic Weight Consolidation (EWC) algorithm to adapt to new data while retaining the learned knowledge. The network parameter update formula is:

[0222]

[0223] in, represents the total loss function; Represents the loss function of the current batch; represents the regularization coefficient; Indicates that for all parameters sum; represents the Fisher information matrix, which is used to measure the importance of parameters; Indicates the current parameters; represents the previous optimal parameter value.

[0224] In actual application scenarios, for an elderly user who is receiving system intervention for the first time, the system first provides an initial intervention strategy based on a general model. As the intervention progresses, the Bayesian optimization framework continuously updates the posterior distribution of the parameter space. For example, the system may find that the user is most effective with cognitive exercises performed in the evening, and the optimal intervention duration for music therapy is between 15 and 20 minutes (the mood improvement effect is most significant). At the same time, the incremental learning module adjusts the intervention content generation model based on the user's specific feedback, such as gradually adjusting the voice assistant's communication style from formal to humorous, because the data shows that the user has a more positive emotional response to humorous content. After 2-3 weeks of continuous optimization, the system has formed a highly personalized intervention plan, the intervention effect index has improved compared to the initial strategy, and user satisfaction has been improved.

[0225] Step 5: Integrate the functional module into the medical and nursing knowledge service system to achieve collaborative work with other health management modules and build a complete elderly mental health service ecosystem;

[0226] This step integrates the functional modules (technical components) such as the emotion dynamic analysis module, emotion cognitive interaction modeling module, dual-channel collaborative intervention module, and personalized intervention execution module from steps 1 to 4 into the medical and nursing knowledge service system, achieving collaborative work with other health management modules and building a complete elderly mental health service ecosystem. It mainly includes the following steps:

[0227] Step 5.1: Modular system architecture construction;

[0228] This application constructs a modular system architecture, encapsulates various technical components into standardized service modules, and forms an extensible medical and nursing knowledge service system. Specifically, the system adopts a microservice architecture, including the following core modules: emotional dynamic analysis service, emotional cognitive interaction modeling service, dual-channel collaborative intervention service, and personalized intervention execution service. It should be noted that each module communicates through a standardized API and adopts an event-driven model to achieve loosely coupled integration. In addition, the system builds a unified data storage layer, which includes a time series database and a knowledge graph database to achieve efficient data management and query. Therefore, the data flow can be expressed as:

[0229]

[0230] in, Represents the original data stream; Represents the emotional dynamic analysis module; Represents the emotion-cognition interaction modeling module; represents a dual-channel collaborative intervention module; represents the personalized intervention execution module; Represents the feedback data stream, which is used to continuously optimize the performance of each module.

[0231] The specific implementation of the modular system architecture is as follows: The system adopts a cloud-native architecture based on container orchestration to ensure the independent deployment, scaling, and maintenance of each service module. The core technical components include:

[0232] The service mesh layer uses service mesh technologies such as Istio to manage communication between microservices, enabling traffic control, fault recovery, and secure communication. Service discovery and registration are implemented through the Kubernetes service registry, ensuring dynamic connectivity between system components.

[0233] API Gateway Layer: This layer uses GraphQL as the API interface specification, providing a unified query language and flexible data acquisition capabilities. The API Gateway implements functions such as request routing, load balancing, authentication and authorization, and rate limiting.

[0234] Data persistence layer: adopts a multi-model database architecture, including:

[0235] Time Series Database (InfluxDB): stores time series data such as the emotional state and physiological indicators of the elderly;

[0236] Graph database (Neo4j): stores knowledge graph structures and expresses complex relationships between health concepts;

[0237] Document database (MongoDB): stores unstructured or semi-structured user profiles, intervention records, etc.

[0238] Distributed Cache (Redis): Improves read performance for frequently accessed data.

[0239] Message Queue System: Kafka is used to implement event-driven inter-module communication, supporting asynchronous processing of data streams and system decoupling. Event topics include:

[0240] Emotional state change event (emotion.state.change);

[0241] Intervention strategy generated event (intervention.strategy.created);

[0242] Intervention execution event (intervention.execution.completed);

[0243] User feedback event (user.feedback.received).

[0244] Data Pipeline: Spark Streaming is used to process real-time data streams, and ETL tools are used to process batch data to ensure efficient flow and conversion of data between modules.

[0245] In a specific application scenario, the modular architecture works as follows: When an elderly person begins using the system, they first register their personal information and authorize data access through the API gateway. Subsequently, various sensor devices begin collecting data and transmit it to the system via the Internet of Things (IoT) gateway. The Emotional Dynamic Analysis Service processes this raw data, extracts emotional features, stores the results in a time series database, and publishes emotional state change events via Kafka. The Emotional Cognitive Interaction Modeling Service subscribes to these events, combines historical data to build a personalized interaction model, identifies key intervention nodes, and stores the results in a graph database. When the system detects a potential risk, the Dual-Channel Collaborative Intervention Service generates an intervention strategy based on the current state and publishes an Intervention Strategy Generation event. The Personalized Intervention Execution Service receives this event, coordinates multiple terminal devices to execute the intervention, and collects feedback data, forming a closed-loop optimization process. For example, the system detects that the emotional variability of an elderly person living alone has decreased for three consecutive days, and immediately initiates emotional dynamic analysis. After confirming the warning, the collaborative intervention service generates a dual-channel intervention strategy of "enhancing social interaction + cognitive reconstruction". The personalized execution service recommends the elderly person to participate in community chess and card activities through the smart speaker based on his interests and hobbies (obtained from the user portrait) and the current time (9 o'clock in the morning), and pushes positive thinking exercises on his tablet device.

[0246] Step 5.2: Collaborative working mechanism implementation;

[0247] This application realizes the collaborative working mechanism of mental health management and other health management modules. It should be noted that the system uses knowledge graph technology to build a panoramic picture of the health status of the elderly, including multi-dimensional information such as physiological health, cognitive status and mental health. Specifically, through the ontology model Define the relationship between the dimensions to form the following representation:

[0248]

[0249] in, Represents the health ontology model; Represents a set of health concepts; Represents a set of relationships between concepts; Represents a collection of attributes; Represents a collection of instances.

[0250] Accordingly, based on this knowledge graph, the system realizes collaborative decision-making among modules such as mental health intervention, cognitive assessment, and chronic disease management to ensure that the intervention strategy is consistent with the overall health status of the elderly.

[0251] The specific implementation of the collaborative working mechanism is as follows: the system builds a multi-level health knowledge graph to support cross-domain health information integration and reasoning. The core components of the knowledge graph include:

[0252] Ontology layer: defines the core concepts and relationships in the field of medical care and health. The ontology model includes the following main concept sets:

[0253] Mental health concepts: emotional state, cognitive patterns, mental illness, intervention methods, etc.;

[0254] Physiological health concepts: vital signs, disease status, medication, functional status, etc.;

[0255] Cognitive function concepts: memory ability, executive function, language ability, spatial perception, etc.

[0256] Concepts of social factors: social networks, family support, community resources, cultural background, etc.

[0257] Relational model: defines the relationships between concepts, including:

[0258] Causation: represents a chain of causes and effects whereby one health state leads to another;

[0259] Influence relationship: indicates the degree of mutual influence between different factors;

[0260] Compositional relationship: represents the hierarchical structure of the whole and its parts;

[0261] Temporal relationship: represents the evolution pattern of health status over time.

[0262] Reasoning engine: performs health status reasoning and intervention decisions based on the knowledge graph, including:

[0263] Rule reasoning: Reasoning based on IF-THEN rules defined by experts;

[0264] Path analysis: Analyze paths in the knowledge graph to discover indirect connections;

[0265] Probabilistic reasoning: Dealing with uncertainty and probabilistic relationships in health knowledge.

[0266] Conflict detection and reconciliation mechanism: Identify and resolve intervention conflicts between different health management modules, including:

[0267] Drug-psychological intervention conflict detection;

[0268] Cognitive training-emotional management priority adjustment;

[0269] Coordination of physiological treatment-psychological intervention time.

[0270] In a specific application scenario, the operating process of this collaborative working mechanism is as follows: when the system provides services to an elderly person who suffers from both hypertension and mild depression, the knowledge graph captures the influence relationship between "hypertension medication" and "mood fluctuations." When the mental health module is ready to implement emotional intervention, the system discovers through graph query that the user is taking beta-blockers, which may cause depression as a side effect. The collaborative decision-making mechanism adjusts the emotional intervention strategy accordingly, increasing the proportion of cognitive reconstruction, and at the same time recommends that the medical team consider adjusting the medication regimen. In another scenario, when the system detects a decrease in the user's cognitive function test score, the knowledge graph analysis shows that the user's emotional state has fluctuated more and more recently, and their sleep quality has declined. Through the association path analysis in the graph, the system infers that emotional problems may be upstream factors of cognitive changes. Based on this, it prioritizes emotional stabilization intervention and then implements cognitive training to form a coordinated and synergistic intervention plan.

[0271] Step 5.3: Generate personalized interaction strategy;

[0272] This application builds a personalized interaction strategy generation algorithm based on the elderly's personal interests, life experiences, and social preferences. Specifically, the system maintains a user interest model and, through collaborative filtering and content analysis methods, predicts the elderly's preferences for different interactive content and generates the most suitable psychological support interaction strategy. It should be noted that the interactive content library includes three types of content: emotional companionship, cognitive stimulation, and social promotion:

[0273] Emotional companionship content: providing music, stories, nostalgic activities, etc. based on current emotional state and personal preferences;

[0274] Cognitive stimulation content: Provides adaptive cognitive training games and knowledge quizzes based on cognitive ability assessment results;

[0275] Social promotion content: Recommend social activities and group interaction opportunities based on social network analysis.

[0276] In addition, the system optimizes the content recommendation strategy through the multi-armed bandit algorithm to maximize the participation and satisfaction of the elderly.

[0277] The specific implementation of personalized interaction strategy generation is as follows: the system builds a multi-dimensional elderly user portrait and a dynamic interaction strategy generation framework, which includes the following core components:

[0278] Multi-dimensional user portrait model: Integrate multi-source data to build a comprehensive user portrait, including:

[0279] Basic attribute dimensions: age, gender, educational background, professional experience, etc.;

[0280] Interest preference dimension: cultural hobbies, activity preferences, topic interests, etc.;

[0281] Life experience dimension: important life events, memories of life stages, etc.;

[0282] Social network dimensions: family relationships, circle of friends, social habits, etc.;

[0283] Interaction habit dimensions: device preference, interaction duration, feedback method, etc.

[0284] Content representation and matching system: This system creates semantic representations for interactive content and matches them with user profiles. Technical implementations include:

[0285] Content vectorization: Use language models such as BERT to generate semantic vectors for text content;

[0286] Multimodal feature extraction: extracting feature representations for images, audio, and other content;

[0287] Similarity calculation: Use methods such as cosine similarity to calculate the matching degree between user interests and content.

[0288] Hybrid recommendation algorithm: Combines multiple recommendation methods to generate personalized interactive content:

[0289] Content-based recommendation: recommendations based on content semantic similarity;

[0290] Collaborative filtering: Recommendations based on preferences of similar user groups;

[0291] Knowledge graph-enhanced recommendation: using domain knowledge to guide the recommendation process;

[0292] Context-aware recommendations: Consider contextual factors such as time, location, and emotions.

[0293] Exploration-Exploitation Balance Strategy: Using the Contextual Multi-Armed Bandit (CMAB) algorithm to balance exploring new content and exploiting known preferences:

[0294] LinUCB algorithm implementation: linearly mapping context features to reward predictions;

[0295] Thompson sampling: handling uncertainty by sampling from the posterior distribution;

[0296] Layered exploration strategy: Explore at two levels: content category and specific content.

[0297] In a specific application scenario, the personalized interaction strategy works as follows: Through initial investigation and ongoing interactions, the system learns that a 78-year-old retired teacher has a particular interest in classical music, gardening, and history, uses a tablet computer in the mornings and evenings, and has volunteered at her hometown's history museum. When this user experiences low mood, the system, instead of providing generic emotional support content, recommends music therapy content based on their user profile, including Beethoven's Sixth Symphony ("Pastoral"), paired with a photo gallery showcasing the historical changes in their hometown, and recommends volunteer opportunities at a nearby botanical garden. The system records the user's response to each recommendation (such as dwell time, emotional changes, and direct feedback) to update the recommendation model. The system also regularly explores new content types. For example, if it detects that the user has browsed several articles about astronomy, it will attempt to recommend activities at a local planetarium, observing feedback and adjusting the interest weighting within the user profile. Through this continuous learning and adaptation process, the system gradually develops a highly personalized interaction strategy, improving the acceptance and effectiveness of psychological support.

[0298] Step 5.4: Dynamic update and optimization of knowledge base;

[0299] This application builds a dynamic knowledge base update and optimization mechanism to continuously accumulate experience from intervention practices and improve the overall service quality. Specifically, the system uses an incremental learning method to regularly update the model parameters based on newly collected intervention effect data:

[0300]

[0301] in, represents the updated model parameters; Represents the model parameters before updating; Indicates that based on new data The calculated parameter increment.

[0302] It should be understood that the system also maintains a case library that records typical intervention cases and their effects, supporting case-based reasoning decision-making. As can be seen, the knowledge base update process is supervised by the following quality control indicators: model performance indicator, intervention effect indicator, and user satisfaction indicator, ensuring that the update process can continuously improve system performance.

[0303] The specific implementation method of dynamic updating and optimization of the knowledge base is as follows: the system establishes a multi-level knowledge accumulation and optimization framework to achieve the continuous accumulation and application of experience knowledge:

[0304] Distributed knowledge base architecture: Building a hierarchical, multi-perspective knowledge management system:

[0305] Core knowledge layer: Contains proven expertise and best practices;

[0306] Empirical knowledge layer: empirical rules extracted from actual intervention cases;

[0307] Individual knowledge layer: personalized knowledge for specific user groups;

[0308] Provisional knowledge layer: newly discovered knowledge and hypotheses to be verified.

[0309] Case database construction and retrieval mechanism: systematically collect, organize and index typical intervention cases:

[0310] Case representation: Each case is structured into a “context-intervention-outcome” triad;

[0311] Similarity calculation: Calculate case similarity based on weighted feature matching;

[0312] Efficient retrieval: Using technologies such as Locality Sensitive Hashing (LSH) to achieve fast case retrieval;

[0313] Case adaptation: Adapting intervention strategies from historical cases to current situations.

[0314] Incremental learning and model update mechanism: The system improves each component model through continuous learning:

[0315] Parameter increment calculation: ;

[0316] in, represents the parameter increment calculated based on the new data; represents the learning rate; Indicates the parameters Find the gradient; represents the loss function; Indicates new data; Represents the model parameters before updating;

[0317] Catastrophic forgetting protection: prevent new knowledge from overwriting old knowledge through experience replay and elastic weight merging;

[0318] Model version control: maintain model version history, support rollback and A / B testing;

[0319] Federated learning framework: Aggregating learning results of multiple instances while protecting privacy.

[0320] Quality control and knowledge verification: Establish a multi-dimensional quality assessment system:

[0321] Model performance indicators : precision, recall, F1 score, etc.;

[0322] Intervention effect indicators : Mood improvement rate, cognitive improvement, etc.;

[0323] User satisfaction index : Frequency of use, subjective ratings, withdrawal rates, etc.;

[0324] Knowledge consistency check: Ensure the consistency of new knowledge with the existing knowledge system.

[0325] In a specific application scenario, the workflow for dynamically updating and optimizing the knowledge base is as follows: After serving hundreds of elderly users for several months, the system accumulated a large amount of intervention data. Through data analysis, the system discovered a significant pattern: for female users aged 70-75 with a teaching background, "interest-based social activities" were, on average, 32% more effective than "general social activities" interventions. This finding was recorded in the empirical knowledge layer, and the intervention strategy generation model was updated through incremental learning. The system also identified a particularly successful case: a single elderly person who joined a community choir through the system's recommendation and experienced significant improvement in their depressive symptoms. This case is recorded in detail in the case database, including user characteristics, intervention process, and effectiveness evaluation. When the system encounters a new user with similar characteristics, it adjusts the intervention strategy based on this successful case. Simultaneously, the system continuously monitors the performance indicators of all models. If the emotional dynamics analysis model's accuracy in identifying "hidden anxiety" declines, it triggers targeted data collection and model optimization. Through this cyclical knowledge accumulation and optimization process, the system's overall service quality has gradually improved, with both intervention effectiveness and user satisfaction indicators showing a steady upward trend.

[0326] The implementation method provided in this application realizes a medical and nursing integrated knowledge service system for elderly care through an innovative combination of key technologies such as emotion dynamic analysis, emotion cognitive interaction modeling, dual-channel collaborative intervention, and personalized intervention execution. It has the following significant technical effects:

[0327] Improved accuracy of early warning: By extracting dynamic emotional features through nonlinear time series analysis and a hybrid neural network architecture, the system of this application can capture emotional change patterns that are difficult to detect with traditional methods. It should be understood that compared to traditional methods based solely on cross-sectional assessments, this embodiment can identify deep dynamic features such as emotional complexity and transfer patterns, effectively identifying potential mental health risks in advance and providing a critical time window for early intervention.

[0328] Significantly Improved Intervention Efficiency: Based on an emotion-cognitive interaction model and a key node identification algorithm, the system in this application can accurately locate the key links in the mutual influence between emotion and cognition, enabling targeted intervention. Furthermore, the dual-channel collaborative intervention strategy, through multi-agent reinforcement learning, simultaneously regulates emotional states and restructures cognitive patterns, effectively breaking the vicious cycle of emotion and cognition and significantly improving the effect-output ratio per unit of intervention resources.

[0329] Comprehensive Enhancement of Personalization: This application's system utilizes multi-level personalization technologies, including individualized interactive model training, adaptive intervention scheduling, and dynamic strategy optimization, to achieve full personalization of the entire process, from psychological state assessment to intervention strategy generation. As a result, this personalized, precision service model can adapt to the psychological characteristics, lifestyle habits, and personal preferences of different elderly individuals, significantly improving the acceptance and compliance of intervention plans.

[0330] Qualitative improvement in user experience: Through implicit psychological assessment mechanisms and contextually adaptive intervention, the system in this application naturally integrates mental health management into the daily lives of the elderly, avoiding the labeling effect and sense of exclusion brought about by direct assessment and intervention in traditional methods. It should be noted that this "unnoticed" service approach has greatly increased the system's user stickiness and the frequency of continuous interaction, laying the foundation for long-term mental health management.

[0331] A virtuous cycle of knowledge accumulation: The modular architecture and dynamic update mechanism of the knowledge service system enable the system to continuously accumulate experience from practice, optimize model parameters and intervention strategies, and form a virtuous cycle of accumulated experience and knowledge. In addition, this evolutionary system design ensures continuous improvement in service quality and adaptability to emerging issues.

[0332] Collaborative management of overall health: Through the collaborative working mechanism realized by knowledge graph technology, the system of this application can organically combine mental health with other dimensions such as cognitive state and physical health, avoiding the limitations of single-dimensional management, providing comprehensive and integrated health management services for the elderly, and significantly improving the overall quality of life of the elderly.

[0333] To sum up, the implementation method of this application breaks through the technical barriers of traditional mental health management, realizes the deep integration of emotional dynamic analysis, emotional cognitive interaction modeling and intelligent intervention, provides a powerful elderly mental health support capability for the medical and nursing knowledge service system, and effectively responds to the increasingly severe elderly mental health challenges in the context of an aging population.

[0334] Application examples of this embodiment

[0335] To verify the feasibility and effectiveness of this implementation, we conducted a six-month real-world test at a senior care center in a certain city. This center housed 120 elderly people aged 65 and over, including 42 living alone, with an average age of 73.5. This section details the specific implementation process and effectiveness of this system in real-world applications.

[0336] 1. Application scenarios and system deployment:

[0337] The application scenario of this integrated medical and nursing knowledge service system in elderly care centers primarily involves early identification and ongoing intervention for the mental health of elderly people living alone. The system's hardware infrastructure includes: a smart terminal device (with a camera, microphone, and touchscreen) installed in each elderly person's home; a wearable physiological monitoring bracelet; a smart home environment control system; and a central server cluster. The system is deployed using a cloud-edge-end architecture, with core algorithms running on central servers, real-time interaction functions executed on edge devices, and data collection and initial processing performed on the end devices.

[0338] During the initial deployment of the system, baseline data was collected from each individual over a two-week period to establish a baseline model of their individualized emotional expression and cognitive patterns. With the individual's consent, the system then collected multimodal data, including speech, facial expressions, daily activities, and physiological indicators, for subsequent emotional dynamic feature extraction and analysis.

[0339] 2. Example of emotional dynamic feature extraction and analysis:

[0340] Taking a 73-year-old female elderly person living alone (number E037) as an example, the system collects her voice, facial expression and behavior data during daily interactions through smart terminal devices. Using a deep multimodal fusion algorithm, the system calculates the elderly person's emotional dynamic feature vector for 30 consecutive days, including the emotional intensity. (Range 0-1), emotional diversity (Range 0-5), emotional inertia (range 0-1) and emotional resilience (Range 0-1).

[0341] The collected emotional dynamic feature vector time series is subjected to nonlinear analysis, and the system calculates the entropy increase rate of the elderly person's emotional trajectory , Lyapunov exponent and fractal dimension . Compared with the reference values of healthy elderly population ( , , ), the emotional complexity index of the elderly man showed a higher entropy increase rate and a lower fractal dimension, indicating that the predictability and complexity of emotional changes decreased.

[0342] Through analysis using a hybrid Hidden Markov Model (HMM) and LSTM architecture, the system discovered that the elderly individual had recently experienced an unusual transition from mild anxiety to moderate depression. The probability of this transition increased from a baseline of 0.12 to 0.37, exceeding the warning threshold of 0.25. The system then generated a warning signal, flagging the individual as potentially at risk for depression, and initiated the subsequent process of building an emotional cognitive interaction model.

[0343] 3. Example of building an emotional cognitive interaction model:

[0344] In response to the warning signal of the elderly E037, the system further constructed an emotional cognitive interaction model. By analyzing the conversation between the elderly and the system through natural language processing technology and combining it with the cognitive assessment questionnaire, the system constructed an emotional state vector. and cognitive pattern vectors , corresponding to different emotional dimensions (sadness, anxiety, tiredness, loneliness, helplessness) and cognitive dimensions (catastrophic thinking, overgeneralization, negative screening, personalization, black and white).

[0345] By analyzing the elderly individual's emotional and cognitive interaction history over the past three months, the memory-enhancing attention network discovered that an increase in the emotional dimension of "loneliness" was often accompanied by an increase in the cognitive pattern of "overgeneralization," and this change exhibited a clear temporal dependency. By calculating the partial derivative matrix, the system discovered that the influence weight of "loneliness → overgeneralization" (0.72) was significantly higher than the influence weights between the other dimensions (average 0.31), identifying this node as a key intervention node.

[0346] The system also discovered that the elderly person's "negative screening" cognitive pattern had a strong feedback-enhancing effect on the emotional dimension of "sadness" (influence weight 0.68), creating a vicious cycle of emotional cognition. Based on the identification of these key nodes, the system determined the emotional and cognitive dimensions for key intervention.

[0347] 4. Example of dual-channel collaborative intervention strategy generation:

[0348] Aiming at the emotional cognitive interaction model of elderly person E037, the system applies multi-agent reinforcement learning algorithm to generate personalized intervention strategies. and cognitive intervention agents Based on the elderly person's status information (including emotional state, cognitive patterns and environmental factors), action recommendations for emotional intervention and cognitive intervention are generated respectively.

[0349] After multiple rounds of collaborative decision-making optimization, the system identified the following dual-channel intervention strategy: the emotional channel focused on "increasing social interaction" (weight 0.65) and "music therapy" (weight 0.35); the cognitive channel focused on "cognitive restructuring" (weight 0.70) and "mindfulness practice" (weight 0.30). System analysis found that the expected effect of this dual-channel collaborative intervention strategy (predicted value 0.78) was significantly higher than that of either emotional intervention alone (predicted value 0.42) or cognitive intervention alone (predicted value 0.51).

[0350] In terms of timing, the system decided to first implement emotional intervention to regulate the elderly person's loneliness, and then gradually introduce cognitive intervention to change his overgeneralized thinking patterns. This gradual intervention sequence design, based on the elderly person's receptive model, can maximize the overall effectiveness of the intervention.

[0351] 5. Examples of personalized intervention implementation and feedback adjustment:

[0352] In the actual intervention process, the system implemented the aforementioned intervention strategy for elderly patient E037 through a multimodal intervention execution system. Specifically, the system first recommended, through a smart terminal device, that she participate in a choir activity organized by the care center (to address social loneliness) and that she play her favorite classical music in the evening (based on her interest model) to regulate her mood.

[0353] After the initial emotional intervention was successful, the system began guiding the elderly woman through cognitive restructuring exercises through video conversations, challenging her overgeneralized thinking, such as "no one understands me," and helping her identify and adjust this irrational cognition. The system also pushed short mindfulness meditation audio clips to help the elderly woman focus on the present moment and reduce the distraction of negative thoughts.

[0354] The system monitored the intervention's effectiveness in real time and found that the elderly patient was highly receptive to social activities (85% participation rate) but less receptive to cognitive restructuring exercises (only 40% completion rate). Based on this feedback, the system optimized the intervention parameters using a Bayesian adjustment method, shifting the format of cognitive intervention from direct cognitive challenges to a more indirect story-based analogy. This adjustment increased the completion rate of cognitive exercises to 75%.

[0355] 6. Continuous monitoring and intervention effectiveness evaluation:

[0356] Elderly patient E037 underwent three months of continuous monitoring and intervention, with regular evaluations of intervention effectiveness. Micro-level assessments revealed an average 25% improvement in heart rate variability during music therapy, indicating a significant reduction in stress levels. Meso-level assessments revealed an increase in the frequency of daily social interactions from one to two times per week to four to five times per week, and a decrease in the "overgeneralization" score on a cognitive assessment test from 7.8 to 4.2 out of 10.

[0357] Macro-level assessments, measured using standard scales, showed a significant improvement in depressive symptoms, with the Geriatric Depression Scale (GDS) score decreasing from 14 (mild depression range) before the intervention to 8 (normal range). The Quality of Life Assessment Questionnaire score increased from 62 before the intervention to 79 (out of 100), indicating a significant improvement in the patient's overall life satisfaction.

[0358] The system identified that social activities had the most significant effect on improving the patient's mood. It stored this empirical knowledge in the individual knowledge layer to optimize future intervention strategies. The system also determined the optimal time periods for the patient to respond to intervention measures (9-11 a.m. and 4-6 p.m.) and adjusted the intervention schedule accordingly, further improving its effectiveness.

[0359] 7. Technical effect verification:

[0360] In the application test of the entire elderly care center, the medical and nursing integrated knowledge service system of this implementation method showed significant technical effects in terms of early warning accuracy and intervention efficiency.

[0361] The following is quantitative verification data for the effectiveness of core technologies:

[0362] Early warning accuracy verification results: The system's early identification results for elderly depression and anxiety were compared with the diagnoses of professional psychiatrists. Among 42 elderly people living alone, the system successfully warned 15 of them of potential mental health risks. Of these, 13 were subsequently diagnosed with mild depression or anxiety disorders by professional doctors within 2-3 months, with an early warning accuracy rate of 86.7%. The system's average warning time was 8.5 weeks earlier than traditional scale screening methods, providing a valuable window of opportunity for early intervention. Notably, among the five elderly people identified by the system as being in a subclinical state (not meeting diagnostic criteria but with obvious symptoms), timely intervention prevented further deterioration of their symptoms in four of them, demonstrating a significant preventive effect.

[0363] Results of the intervention efficiency and effectiveness verification: A comparative study showed significant differences in the improvement in mental health between elderly individuals who received the dual-channel collaborative intervention system (experimental group, n=20) and those who received only traditional psychological counseling (control group, n=22). The average GDS score in the experimental group decreased by 5.8 points, while the control group only decreased by 2.3 points. The average quality of life questionnaire score in the experimental group increased by 15.7 points, while the control group increased by 7.2 points. Furthermore, the intervention time required to achieve the same therapeutic effect in the experimental group was 41% shorter than that in the control group, indicating a significant improvement in intervention efficiency. The advantages of the dual-channel collaborative intervention were particularly evident for elderly individuals with both emotional problems and cognitive distortions, validating the effectiveness of the emotional-cognitive interaction modeling and dual-channel collaborative intervention strategy employed in this implementation.

[0364] These real-world application examples fully demonstrate the technical feasibility and effectiveness of this implementation for early warning and targeted intervention of elderly mental health. The system's modular design and adaptive optimization mechanisms ensure its broad applicability across diverse elderly populations, providing strong technical support for addressing the increasingly severe mental health challenges faced by the elderly in the context of an aging population.

[0365] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for constructing a medical and nursing integrated knowledge service system for elderly care, characterized by: The following steps are involved: Collect emotional expression data from older adults, apply nonlinear time series analysis and deep learning models to extract dynamic emotional features and identify early signs of decreased emotional regulation ability; Based on the acquired dynamic emotional features, we use a memory-enhanced attention network combined with a psychological knowledge graph to construct an individualized emotional-cognitive interaction model for the elderly, revealing the dynamic interaction mechanism between emotional states and cognitive patterns and identifying key intervention nodes. Based on the constructed interaction model, a multi-agent reinforcement learning algorithm is applied to generate an emotion-cognition dual-channel collaborative intervention strategy to maximize the effect of precise intervention; Implement the generated intervention strategies, establish a closed-loop intervention execution and feedback mechanism, achieve dynamic evaluation of intervention effects and strategy optimization, and improve the accuracy and sustainability of mental health management for the elderly; Integrate the functional modules into the medical and nursing knowledge service system to achieve collaborative work with other health management modules and build a complete elderly mental health service ecosystem.

2. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 1, characterized in that: Functional modules include emotion dynamic analysis module, emotion cognitive interaction modeling module, dual-channel collaborative intervention module, and intervention execution module.

3. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 1, characterized in that: The extraction of emotional dynamic features specifically includes: Collect multimodal data of the elderly through sensor networks, including voice, facial expressions, activity patterns, and physiological signals, as the basic input for affective state analysis; Applying a deep multimodal fusion algorithm to extract characteristic parameters of emotional expression, including emotional intensity, emotional diversity, emotional inertia, and emotional resilience, and constructing an emotional dynamic feature vector; Using nonlinear time series analysis methods, we conduct complexity analysis on the time series of emotional dynamic feature vectors and quantify the complexity changes of the emotional dynamic system. A hybrid architecture of hidden Markov model and long short-term memory network is constructed to capture the long-term evolution pattern of emotional state transition and identify abnormal transition patterns.

4. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 1, characterized in that: The emotional cognitive interaction model construction specifically includes: Based on the dynamic characteristics of emotions and cognitive evaluation data, the emotional state vector and cognitive pattern vector are constructed; A memory-enhanced attention network is used to construct a bidirectional influence relationship model between emotion and cognition, capturing the mutual influence relationship between the two and storing historical state information through the attention weight matrix. Based on the individual data of the elderly, an optimization algorithm combining gradient descent and variational inference is used to train an individualized emotion-cognitive interaction model. Applying causal inference methods, based on the trained interaction model, we identify individual-specific key nodes of emotional cognitive interaction as the key targets for the formulation of subsequent intervention strategies.

5. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 1, characterized in that: The dual-channel collaborative intervention strategy generation specifically includes: Based on the emotion-cognition interaction model, define the intervention optimization objective function and constraints; Construct a multi-agent system consisting of an emotional intervention agent and a cognitive intervention agent, sharing state information and coordinating action selection through a collaborative mechanism to maximize joint rewards; A deep deterministic policy gradient algorithm is used to optimize the intervention strategy, and a collaborative decision-making mechanism is designed to coordinate the timing of emotional intervention and cognitive intervention to maximize the synergistic effect.

6. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 1, characterized in that: The intervention implementation specifically includes: Build a multimodal intervention execution system, integrate multiple interactive channels, and achieve seamless implementation of emotion-cognition dual-channel intervention strategies; Establish a multi-dimensional intervention effect evaluation mechanism, including immediate emotional response assessment, cognitive change assessment, and long-term health status assessment; Bayesian adjustment methods and incremental learning models are applied to optimize intervention strategy parameters in real time based on intervention effectiveness and acceptance.

7. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 1, characterized in that: Knowledge service system integration and application specifically include: Build a modular system architecture, encapsulate technical components into standardized service modules, and form a scalable medical and nursing knowledge service system; Realize the collaborative working mechanism between mental health management and other health management modules to build a panoramic picture of the health status of the elderly; Build a personalized interaction strategy generation algorithm based on the elderly’s personal interests, life experiences, and social preferences; Build a dynamic update and optimization mechanism for the knowledge base, continuously accumulate experience from intervention practice, and improve overall service quality.

8. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 3 is characterized in that: The nonlinear time series analysis method includes calculating the entropy increase rate, Lyapunov index and fractal dimension of the emotional trajectory to quantify the complexity changes of the emotional dynamic system.

9. The method for constructing a medical and nursing integrated knowledge service system for elderly care according to claim 4, characterized in that: The memory-enhanced attention network consists of three parts: an encoder, a memory module and a decoder. The memory module maintains a historical state buffer, and the multi-head attention mechanism is used to calculate the correlation between the current state and historical memory.

10. A system for constructing a medical-nursing integrated knowledge service system for elderly care, for executing a method for constructing a medical-nursing integrated knowledge service system for elderly care according to any one of claims 1 to 9, characterized in that: include: The emotional dynamics analysis module is used to collect and analyze multimodal emotional data of the elderly, extract emotional dynamic features, identify early signs of decreased emotional regulation ability, and provide basic data support for subsequent interventions; The emotion-cognition interaction modeling module is used to build an individualized emotion-cognition interaction model, reveal the dynamic relationship between emotional state and cognitive pattern, identify key intervention nodes, and provide a theoretical basis for formulating intervention strategies; The dual-channel collaborative intervention module is used to generate emotion-cognition dual-channel collaborative intervention strategies based on multi-agent reinforcement learning, achieving the optimal coordination of emotion and cognitive interventions and maximizing the intervention effect; A personalized intervention execution module is used to execute intervention strategies through multimodal interaction channels, evaluate intervention effects in real time, and dynamically adjust intervention parameters to ensure the accuracy and sustainability of interventions; The knowledge service integration module is used to integrate various functional modules, manage knowledge base updates, coordinate connections with other health management systems, and provide a unified service interface; Each module communicates through a standardized API and adopts an event-driven model to achieve loosely coupled integration.

Images