Man-machine collaborative intelligent health consultant interaction method and system

Abstract

The invention discloses a man-machine collaborative intelligent health consultant interaction method and system, and relates to the technical field of digital services in the field of medicine health, and the method comprises the following steps: building a multi-modal data collection channel, collecting the multi-source interaction data of a user in real time, designing a hierarchical feature extraction network for feature extraction, and carrying out feature extraction; performing cross-modal attention weighting on the extracted features by using a cross-modal attention fusion mechanism to obtain a context representation vector, and storing the context representation vector in a time sequence context buffer area; through dynamic context perception and drift detection, fine changes in the user interaction process are accurately captured, personalized interaction of the user is achieved in real time, personalized accurate services can be provided according to the real-time state and requirements of the user, deep personalized experience is achieved, the satisfaction degree of the user and the sales strength of medicine products are improved, and the user experience is improved. And realizing intelligent cross-domain reasoning by fusing knowledge maps of knowledge of a medical domain, a commercial domain and a user behavior domain and utilizing a GNN reasoning path.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital services in the medical and health field, and in particular to a human-computer collaborative intelligent health consultant interaction method and system. BACKGROUND

[0002] With the continuous development of the digital transformation of the medical and health industry, intelligent health consultant interaction systems are increasingly widely used in retail pharmacies and clinic pharmacies, becoming a key tool for improving service efficiency and quality. Currently, most intelligent health consultant systems mainly rely on rule engines or basic machine learning models to provide health consultation and product recommendation services for users.

[0003] However, the existing intelligent health consultant interaction system has two significant defects in actual application: for example, in terms of personalized interaction, the existing intelligent health consultant interaction system mostly uses static user portraits and fixed interaction templates, but the user's health status, emotions, and environment change constantly during the interaction process, and the system cannot perceive these dynamic changes in real time and respond, resulting in outdated or non-compliant recommended content, and fixed interaction templates cannot be flexibly adjusted according to the user's specific performance and feedback during the interaction process and lack of pertinence, reducing user experience and the effectiveness of personalized services; in terms of knowledge processing, the existing intelligent health consultant interaction system is usually limited to a single domain knowledge base of medical knowledge or drug information, lacking cross-domain correlation analysis capability, but it is well known that health consultation involves medical knowledge, drug information, user behavior data, marketing activities, and other fields, and a single domain knowledge base cannot meet the comprehensive and integrated consultation needs, making it difficult for the system to effectively integrate and reason multiple sources of information, resulting in one-sidedness of the recommended results, and the system cannot provide a comprehensive health consultation and product recommendation solution for users, thus the human-computer collaborative intelligent health consultant interaction platform has the problems of insufficient personalized interaction and limited knowledge processing.

[0004] The existing technology has the following deficiencies: in addition, the existing intelligent health consultant interaction system lacks context awareness capability and is difficult to capture subtle changes in the user interaction process, resulting in the inability to provide accurate services according to the user's real-time feedback and situation; the system's knowledge representation is rigid and cannot achieve the organic integration and associated use of multi-domain knowledge such as medicine, business, and user behavior, limiting the system's knowledge utilization efficiency and recommendation accuracy, and when providing health consultation and product recommendations, the system makes a single decision based on pre-set rules or simple machine learning models, lacks human-computer collaborative optimization based on real-time context, and cannot fully combine human expert experience and system algorithm advantages to make decisions, resulting in less flexible and accurate decision results, thus severely restricting the further development and application of intelligent health consultant interaction systems in the medical and health industry.

[0005] The above information disclosed in the background section is only for enhancing the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art that is already known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide a human-computer collaborative intelligent health consultant interaction method and system, which solves the problems in the background art by dynamic context perception and drift detection, knowledge graph fusion of multi-field knowledge and GNN reasoning, and intelligent task allocation based on uncertainty quantification.

[0007] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a human-computer collaborative intelligent health consultant interaction method, comprising the following steps: S1, a multi-modal data acquisition channel is established, multi-source interaction data of a user is collected in real time, a hierarchical feature extraction network is designed for feature extraction, a cross-modal attention fusion mechanism is used to perform cross-modal attention weighting on the extracted features, a context representation vector is obtained, and stored in a time sequence context buffer; S2, an online adaptive clustering algorithm is used to cluster the context representation vector according to the user's personalized behavior, the number and position of the clustering center are dynamically adjusted, an isolation forest algorithm is used to detect the deviation degree of the new context vector from the historical mode, a drift strength is output, a drift type classifier is established to identify the drift type of the user's behavior, and a sensitivity adaptive mechanism is set to trigger an abnormal signal; S3, a three-domain fusion knowledge graph is constructed, graph anchor positioning is performed in combination with user queries, a dynamic activation mechanism of graph neural network GNN is designed, reasoning is performed according to the drift type and the abnormal signal, a multi-hop reasoning path is generated, the path confidence is calculated in combination with the attention weight, a reasoning result sorting algorithm is self-defined, the reasoning results are weighted and scored to obtain the optimal recommendation result; S4, an uncertainty quantification model is established, the task complexity is calculated in combination with the recommendation result confidence and the drift strength, a context multi-armed bandit algorithm is used for human-computer task dynamic allocation, a personalized interaction content is generated by using a conditional variational autoencoder, feedback information of the user is collected, and the system model parameters are updated through online learning of an incremental learning engine.

[0008] Optionally, the multi-modal data acquisition channel is built by using a multi-source sensor interface and an interaction log recorder, including a text channel, a speech channel, a visual channel and a behavior sequence channel. When the user initiates an interaction session with the intelligent health consultant interaction platform, the multi-source interaction data of the user is collected in real time and in parallel through the established multi-modal data acquisition channel, including text semantics, speech information, visual content and behavior sequence raw data. The collected raw data is pre-processed by cleaning, denoising and standardizing; The pre-processed data is temporarily stored in a data buffer, and a unified timestamp is added to each data to form the user's multi-source interaction data.

[0009] Optionally, the design steps of the hierarchical feature extraction network are as follows: Pre-trained neural network models including BERT model, Wav2Vec2 model, ResNet model and LSTM model are deployed by using text encoder, speech processor, image recognizer and behavior analyzer to extract hierarchical features from multi-modal data; The BERT model is used to process the pre-processed text semantics, and the BERT model learns the semantic information of language through pre-training, which can convert the input text semantics into semantic feature vectors; The Wav2Vec2 model is used to extract features from pre-processed speech information, including using ASR capability to obtain transcribed text and extracting speech feature vectors that can represent linguistic features; The ResNet model is used to extract features from pre-processed visual content, and the last classification layer of the ResNet model is removed, and the output of the last pooling layer is taken as the image feature vector of the visual content; The LSTM model is used to process the pre-processed behavior sequence, and the hidden state of the last time step of the LSTM model is taken as the behavior feature vector of the entire behavior sequence; The feature vectors are processed by feature dimension reduction and alignment using a fully connected layer, and the semantic feature vector, speech feature vector, image feature vector and behavior feature vector are projected into a unified and 256-dimensional common subspace.

[0010] Optionally, the context representation vector obtaining step is as follows: Each modal feature vector after projection, alignment and unification is regarded as a group of vector sequences; For each modal feature, the attention score between each pair of modal feature vectors is calculated, and the Softmax function is used for normalization to obtain a group of weight matrices; The cross-modal attention fusion mechanism is used to weight according to the weight matrix calculated in the last step, calculate the importance of each feature vector in different modalities, and perform residual connection with the modal feature itself to obtain the enhanced modal feature vector; After cross-modal attention weighting, all enhanced modal feature vectors are concatenated and weighted and summed again for combination, and a final fully connected layer is used for fine-tuning to output a unified dimension context representation vector; The generated context representation vector is stored in a time context buffer together with the current session ID and timestamp.

[0011] Optionally, the online adaptive clustering K-means++ algorithm performs the steps of user personalized behavior clustering as follows: Randomly select a context representation vector from the received set of context representation vectors as the first cluster center according to the K-means++ algorithm; Calculate the shortest distance of each context representation vector to the selected cluster center using the Euclidean distance, and select the context representation vector with the maximum distance probability as the new cluster center; Repeat the cluster center selection process until K cluster centers are selected; Calculate the distance of each new context representation vector to the selected K cluster centers, and assign it to the cluster where the nearest cluster center is located online; Update the corresponding cluster center of the cluster using the exponentially weighted moving average method; Set a cluster structure threshold, and analyze the distance of the new context representation vector to all cluster centers with the cluster structure threshold.

[0012] Optionally, the output step of the drift intensity is as follows: Read the user's historical context representation vectors stored in the time context buffer, and train the isolation forest model using the user's historical context representation vectors; Input the new context vector into the trained isolation forest model, and traverse from the root node to the leaf node in each isolation tree to calculate the path length of the new context vector in each tree; Average the path lengths of all new context vectors in each tree to obtain the average path length of the new context vector in the forest, wherein the average path length calculation formula is , wherein represents the average path length, represents the total number of isolation trees in the forest, represents the path length of the new context vector on each tree; Calculate the normalized anomaly score according to the average path length, and map it to the drift intensity, wherein the normalized anomaly score calculation formula is , and , wherein represents the normalized anomaly score, represents the normalization factor, which is the average path length of a binary search tree of samples, represents the harmonic number; The mapping expression of the drift intensity is , wherein is expressed as a drift intensity, is expressed as a minimum and maximum normalized boundary of an anomaly score.

[0013] Optionally, the identification step of the drift type is as follows: The context representation vector is decomposed into emotion features, intent features and domain feature groups according to semantic dimensions; The type-specific drift score is calculated by weighted absolute difference summation for each drift type, quantifying the degree of deviation of each type of the corresponding feature group of the historical clustering center; The type-specific drift score is converted into the probability of each drift type using a softmax function, and the main drift type is determined by probabilistic type; Based on the preset threshold of each drift type, the multi-type drift situation is identified, and when the probabilities of multiple types exceed the threshold, it is marked as a composite type drift.

[0014] Optionally, the step of setting the dynamic activation mechanism of the graph neural network GNN to generate the multi-hop reasoning path is as follows: Based on the three-domain fusion knowledge graph, the entity linking technology is used to map the user query to a corresponding set of entities in the knowledge graph as the starting graph anchor point set for reasoning; According to the drift type and the abnormal signal, the dynamic activation mechanism of the graph neural network GNN is designed, and the activation function parameters of the graph neural network GNN are dynamically adjusted; Using GNN for multi-hop reasoning, starting from the starting graph anchor point set, through the message passing mechanism of GNN, iteratively aggregating neighbor information and updating node representation, while recording the activated nodes and edges at each hop, forming a reasoning path, and generating a multi-hop reasoning path; The range limit of the path length is set to 2-4 hops as a screening condition to filter out reasonable candidate reasoning paths from all collected multi-hop reasoning paths.

[0015] Optionally, the step of establishing an uncertainty quantification model to calculate the task complexity is as follows: Combining the recommendation result confidence and the drift intensity, an uncertainty quantification model is established to analyze the uncertainty components of each recommendation result, including cognitive uncertainty of the recommendation result, accidental uncertainty of data noise, and inherent uncertainty of the uncertainty quantification model. The cognitive uncertainty is a measure of the lack of knowledge of the uncertainty quantification model for the recommendation result, and the expression of the cognitive uncertainty is , wherein is expressed as quantified cognitive uncertainty; The accidental uncertainty is a measure of the noise level of the context data of the input user behavior, and the expression of the accidental uncertainty is In the formula, This can be expressed as quantified random uncertainty. This can be represented as the historical volatility of user behavior; Inherent uncertainty measures the reliability of the uncertainty quantification model itself; therefore, the expression for inherent uncertainty is: In the formula, This can be expressed as the inherent uncertainty in quantification. Represented as inherent coefficients, This represents the time interval for updating the uncertainty quantification model. A multi-factor weighted model is used to comprehensively calculate the quantified uncertainty, resulting in the task complexity. The formula for calculating task complexity is as follows: ,and In the formula, This is expressed as task complexity. Represented as the weight of quantified cognitive uncertainty, Weights are represented as inherent uncertainties due to chance. Weights are represented as quantified inherent uncertainties. Tasks are categorized into operation levels based on complexity, including when... At that time, the operation level is low complexity, which is a routine user consultation and is handled autonomously by AI; when At that time, the operation level was medium complexity, with AI as the primary driver and physician supervision. when At that time, the operation level was characterized by high complexity, with human-led and AI-assisted operation.

[0016] A human-computer collaborative intelligent health consultant interaction system includes a multimodal context perception module: it collects multi-source interaction data of users in real time using a collection channel, designs a hierarchical feature extraction network to extract feature data, uses a cross-modal attention fusion mechanism to perform cross-modal attention weighting, obtains a context representation vector, and stores it in a temporal context buffer; Context drift detection module: Reads and analyzes context representation vectors, uses online adaptive clustering K-means++ algorithm to cluster context representation vectors for user-personalized behavior, uses isolation forest algorithm to detect user behavior drift, outputs drift intensity, identifies drift type, sets sensitivity adaptive mechanism, triggers abnormal signal when drift anomaly occurs, and transmits drift type and abnormal signal to dynamic knowledge reasoning module; Dynamic knowledge reasoning module: Based on the constructed three-domain fusion knowledge graph, the module locates graph anchor points by combining user queries, designs a dynamic activation mechanism for graph neural networks (GNN), performs reasoning based on drift type and abnormal signals, generates multi-hop reasoning paths, calculates path confidence, uses a custom reasoning result ranking algorithm to weight and score the reasoning results, obtains the optimal recommendation result, and transmits it to the human-machine collaborative decision-making and feedback optimization module. Human-Machine Collaborative Decision-Making and Feedback Optimization Module: Establish an uncertainty quantification model, calculate task complexity based on recommendation result confidence and drift intensity, learn the optimal task allocation strategy using a contextual multi-armed gambling machine algorithm, dynamically allocate human and machine tasks, generate personalized interactive content matching user context using a conditional variational autoencoder, collect user feedback information, and continuously optimize system performance by learning and updating system model parameters online through an incremental learning engine.

[0017] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention accurately captures subtle changes in user interaction through dynamic context awareness and drift detection, enabling real-time personalized user interaction. It provides precise, personalized services based on the user's real-time status and needs, achieving a deeply personalized experience, improving user satisfaction and boosting pharmaceutical product sales. By integrating knowledge graphs from the medical, business, and user behavior domains and utilizing GNN inference paths, it achieves intelligent cross-domain reasoning, not only improving recommendation accuracy but also providing explainable and reasonable inference paths. Furthermore, by establishing an uncertainty quantification model to quantify cognitive uncertainty, accidental uncertainty, and inherent uncertainty for intelligent task allocation, it optimizes human resource utilization while ensuring the quality of handling complex problems, achieving efficient human-machine collaboration and continuous system evolution. This improves processing efficiency and problem-solving quality, promoting the further development and application of the pharmaceutical and healthcare industry. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the human-computer collaborative intelligent health consultant interaction method of the present invention.

[0020] Figure 2 This is a flowchart of step S1 in the intelligent health consultant interaction method of the present invention.

[0021] Figure 3 This is a block diagram of the intelligent health consultant interaction system for human-machine collaboration of the present invention. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0023] Example 1 This invention provides, for example Figures 1-2 The human-computer collaborative intelligent health advisor interaction method shown includes the following steps: S1. Establish a multimodal data acquisition channel through multi-source sensor interfaces and interactive log recorders to collect multi-source interactive data of users in real time. Design a hierarchical feature extraction network to extract features from the multi-source interactive data. Use a cross-modal attention fusion mechanism to perform cross-modal attention weighting on the extracted features to obtain a machine-understandable and uniform-dimensional context representation vector, which is stored in a temporal context buffer. The hierarchical feature extraction network includes a BERT model to process text semantics, a Wav2Vec2 model to process speech features, a ResNet model to process visual content, and an LSTM model to process behavior sequences. Specifically, the steps for collecting multi-source interaction data from users are as follows: A multimodal data acquisition channel was built using a multi-source sensor interface and an interactive logger, including a text channel, a voice channel, a visual channel, and a behavior sequence channel. The text channel captures the text query entered by the user through the input box or chat interface of the application. The voice channel captures the user's voice input through the device's microphone, which is transcribed into text in real time, while retaining the original audio waveform data for analyzing the linguistic features of intonation, speech rate, and energy in the speech. The visual channel uses the device's camera to capture static images or video clips related to the user's health status, such as photos of tongue coating, skin symptoms, and the color and shape of stool and urine. The behavior sequence channel continuously records the user's behavior event sequence with high-frequency timestamps through an interaction log recorder, including mouse movement trajectory, click events of recommended medicines, page dwell time, deletion and modification behavior during input, scrolling speed of searched products, etc. When a user initiates an interactive session with the intelligent health consultant platform, the platform collects multi-source interactive data from the user in real time and in parallel through the established multimodal data acquisition channel. This includes raw data such as text semantics, voice information, visual content, and behavioral sequences. The collected raw data undergoes preprocessing including cleaning, noise reduction, and standardization. This includes word segmentation, stop word removal, and spell correction for text semantics; noise reduction, silence removal, and volume normalization for speech information; size normalization, center cropping, and color standardization for visual content; and integration of discrete events in the behavioral sequence into structured logs based on session ID and time sequence. The preprocessed data is temporarily stored in a data buffer, and each data entry is given a unique timestamp, forming multi-source interaction data of the user. This ensures that the multimodal data is aligned on the timeline, preparing for subsequent fusion processing. The expression for the multi-source interaction data is as follows: In the formula, Represented as multi-source interactive data, This is represented as the preprocessed text semantics. This represents the preprocessed speech information. This represents the preprocessed visual content. Represented as a preprocessed sequence of behaviors. It is represented as a uniform timestamp.

[0024] Specifically, the design steps for the hierarchical feature extraction network are as follows: Pre-trained neural network models, including BERT, Wav2Vec2, ResNet, and LSTM models, are deployed using text encoders, speech processors, image recognizers, and behavior analyzers to perform hierarchical feature extraction on multimodal data. The BERT model is used to process the semantics of preprocessed text to obtain sentence-level embeddings. The BERT model learns semantic information through pre-training, converting the input text semantics into semantic feature vectors. The expression for the semantic feature vector is: In the formula, This is represented as a semantic feature vector extracted from the preprocessed text semantics. This represents the symbol used for feature extraction processing using the BERT model. These are the trainable parameters representing the weights and biases of the BERT model. This represents the preprocessed text semantics input to the BERT model. This represents a label indicating the hidden layer state of the corresponding BERT model; The Wav2Vec2 model is used to extract features from the preprocessed speech information. This includes using ASR capabilities to obtain the transcribed text and extracting speech feature vectors that can characterize linguistic features, such as pitch, speech rate, and energy. The expression for the speech feature vector is as follows: In the formula, This is represented as a speech feature vector extracted from the preprocessed speech information. This represents the symbol used for feature extraction processing using the Wav2Vec2 model. This represents the preprocessed speech information input to the Wav2Vec2 model. These are represented as the trainable parameters of the Wav2Vec2 model; The ResNet model is used to extract features from the preprocessed visual content. By introducing residual blocks, the vanishing gradient problem in deep neural networks is solved. The last classification layer of the ResNet model is removed, and the output of the last pooling layer is taken as the image feature vector of the visual content. The expression for the image feature vector is: In the formula, This is represented as an image feature vector extracted from the preprocessed visual content. This is represented by the symbol for feature extraction processing using the ResNet model. This represents the preprocessed visual content input to the ResNet model. These are the trainable parameters of the ResNet model; The LSTM model is used to process the preprocessed behavior sequence, effectively capturing the temporal dependencies between user behaviors. The hidden state of the last time step of the LSTM model is used as the behavior feature vector of the entire behavior sequence, where the expression for the behavior feature vector is: In the formula, This is represented as a behavioral feature vector extracted from the preprocessed behavioral sequence. This represents the symbol used for feature extraction processing using the LSTM model. This represents the preprocessed sequence of behaviors input into the LSTM model. These are the trainable parameters of the LSTM model. By using fully connected layers to perform feature dimensionality reduction and alignment on the individual feature vectors, the semantic, speech, image, and behavioral feature vectors are projected into a unified 256-dimensional common subspace for subsequent cross-modal feature vector fusion. The expression for feature dimensionality reduction and alignment is as follows: , , , In the formula, It is represented as a unified 256-dimensional feature vector after semantic feature vector projection alignment. This is represented as the weight matrix of the fully connected layer used for dimensionality reduction in the BERT model. This is represented as the bias term of the fully connected layer used for dimensionality reduction in the BERT model. This is represented as a unified 256-dimensional feature vector after projection alignment of the speech feature vectors. This is represented as the weight matrix of the fully connected layer used for dimensionality reduction in the Wav2Vec2 model. This is represented as the bias term of the fully connected layer used for dimensionality reduction of the Wav2Vec2 model. It is represented as a unified 256-dimensional feature vector after the image feature vector is projected and aligned. This is represented as the weight matrix of the fully connected layer used for dimensionality reduction in the ResNet model. This is represented as the bias term of the fully connected layer used for dimensionality reduction in the ResNet model. It is represented as a unified 256-dimensional feature vector after the behavioral feature vector is projected and aligned. This is represented as the weight matrix of the fully connected layer used for dimensionality reduction in the LSTM model. This represents the bias term of the fully connected layer used for dimensionality reduction of the LSTM model.

[0025] Specifically, the steps to obtain the context representation vector are as follows: The modal feature vectors after projection alignment and unification are considered as a set of vector sequences, expressed as follows: ,and , respectively correspond , correspond , correspond , correspond ; For each modal feature, the attention score between each pair of modal feature vectors is calculated and normalized using the Softmax function to obtain a set of weight matrices. These matrices indicate the importance of each feature vector in different modalities during feature vector fusion. The expression for this set of weight matrices is as follows: ,and In the formula, Represented as modality For modes Attention weights Represented as the Softmax function, This is represented as a learnable weight matrix for generating query vectors. Represented as the first after projection alignment and unification Feature vectors of each modality The learnable weight matrix is ​​represented as the key vector. This is represented as the projection alignment unified result, different from the first. Feature vectors of each modality This is represented as transpose. This is expressed as the dimension of the key vector. Represented as the mode after cross-modal information enhancement The new feature vector, The learnable weight matrix is ​​represented as a value vector. This means that cross-modal attention computation involves feature vectors from two different modalities; The cross-modal attention fusion mechanism is used to weight the weight matrix calculated in the previous step, calculate the importance of each feature vector in different modalities, and perform residual connection with its own modal features to obtain the enhanced modal feature vector; After cross-modal attention weighting, all enhanced modal feature vectors are concatenated, weighted, and summed again for combination. This result is then fine-tuned through a final fully connected layer, outputting a unified-dimensional context representation vector. The expression for this unified-dimensional context representation vector is: In the formula, Represented as a context representation vector of uniform dimension, This is represented as a normalization operation for a fully connected layer. Represented as the weights for fusion in the fully connected layer. Represented as Concatenating new feature vectors of each modality This represents the bias used for fusion of fully connected layers; The generated context representation vector is stored in the timing context buffer along with the current session ID and timestamp.

[0026] S2. The online adaptive clustering K-means++ algorithm is used to cluster the context representation vector for user-personalized behavior, dynamically adjusting the number and position of cluster centers. The isolation forest algorithm is used to detect the degree of deviation between the new context vector and the historical pattern, output the drift intensity, establish a drift type classifier to identify the drift types of user behavior such as emotion drift, intent drift, and domain drift, and set a sensitivity adaptive mechanism to trigger abnormal signals. Specifically, the steps of the online adaptive clustering K-means++ algorithm for clustering user-personalized behavior are as follows: Randomly select a context representation vector from the received set of context representation vectors using the K-means++ algorithm as the first cluster center; The Euclidean distance is used to calculate the shortest distance from each context representation vector to the selected cluster center, and the context representation vector with the highest distance probability is selected as the new cluster center. Repeat the cluster center selection process until K cluster centers are selected; Calculate the distance between each new context representation vector and the selected K cluster centers, and assign it online to the cluster containing the nearest cluster center. The distance calculation formula is as follows: In the formula, Represented as the th in the new context representation vector The element and the first The distance between cluster centers This is represented as the total number of elements in the context representation vector. The first one, represented as the new context representation vector One element, Represented as the first The first cluster center One element; The expression for online cluster center allocation is: In the formula, This indicates online allocation of cluster centers. This is represented as the function that takes the closest distance; The cluster centers of the given clusters are updated using the exponentially weighted moving average method, where the update expression for the cluster centers is: ,and In the formula, This is represented as the updated cluster center. Represented as the learning rate; Set a clustering structure threshold and analyze the distance between the new context representation vector and all cluster centers. When the distance between the context representation vector and all cluster centers exceeds the clustering structure threshold, and the number of currently selected cluster centers is less than the maximum allowed number of clusters, then use the old cluster merging method or the sparse cluster deletion method to create a new cluster, which is used to dynamically adjust the number and position of cluster centers until the iteration converges.

[0027] Specifically, the steps for outputting the drift intensity are as follows: Read the user's historical context representation vector stored in the time-series context buffer, and train the isolation forest model using the user's historical context representation vector. The isolation forest model consists of several isolation trees, and each isolation tree uses random subsamples. The new context vector is input into the trained isolation forest model. In each isolation tree, the path length of the new context vector in each tree is calculated by traversing from the root node to the leaf node. The average path length of each new context vector in each tree is calculated by averaging the path lengths of all new context vectors in the forest. The formula for calculating the average path length is as follows: In the formula, Expressed as average path length, This represents the total number of isolation trees in the forest. This is represented as the path length of the new context vector on each tree; The standardized outlier score is calculated based on the average path length and mapped to drift intensity. The formula for calculating the standardized outlier score is as follows: ,and In the formula, Represented as standardized outlier scores, Represented as a standardized factor, which is The average path length of the binary search tree for each sample. Represented as harmonic numbers; The mapping expression for drift intensity is: In the formula, Expressed as drift intensity, These are represented as the minimum and maximum normalized boundaries of the outlier scores.

[0028] Specifically, the steps for identifying drift types are as follows: The context representation vector is decomposed into emotion features, intent features, and domain features according to the semantic dimension. The emotion features are the emotion vector parts extracted from the text and speech. Intent features are the intent vector portions extracted from text and behavioral sequences; Domain features are the domain topic vectors extracted from text and images; For each drift type, a type-specific drift score is calculated by summing the weighted absolute differences, quantifying the degree of shift of each type in the corresponding feature group of the historical cluster centers. The formula for calculating the type-specific drift score is as follows: ,and In the formula, Represented as a specific score for a particular drift type. This is represented as decomposing all feature dimensions in the feature group from the context representation vector. Summation, Represented as the i-th The importance weight of feature dimensions to the current drift type Represented as a context representation vector In the Values ​​in each feature dimension Represented as the first The feature dimension at the th feature dimension The values ​​of each reference cluster center state This is represented as an emotion type-specific drift score. Represented as intent type-specific drift score, Represented as domain type-specific drift score; The softmax function is used to convert type-specific drift scores into probabilities for each drift type. This probabilistic type determination identifies the dominant drift type, facilitating comparison and decision-making. The expression for probabilistic type determination is as follows: In the formula, Represented as a representation vector in a given context At that time, the probability of belonging to a specific drift type, This is expressed as a temperature parameter that controls the sharpness of the probability distribution. Represented as an exponential function, This is represented as the summation of all drift types across all emotions, intentions, and domains; Multiple types of drift are identified based on preset thresholds for each drift type. When the probability of multiple types exceeds the threshold, it is marked as a composite type drift. The expression for identifying multiple types of drift is as follows: hour, In the formula, This represents the preset threshold for each drift type. This is represented as a composite type of drift. Represented as the union symbol, This indicates that the current drift type is added to the composite drift set.

[0029] Specifically, the steps for setting up the sensitivity adaptation mechanism are as follows: Based on the drift intensity distribution of historical data, an initial sensitivity threshold is set and calibrated as follows. ; Each time a user behavior drift is detected, the drift intensity and drift type are recorded; The number of times the user drifted within the past 100 fixed interaction operations was counted, and the mean and variance of the user's historical drift intensity were calculated, where the mean and variance were denoted as . ; An adaptive threshold is calculated based on the user baseline. A sliding window mechanism is used to dynamically update the mean and variance statistics to smooth the sensitivity threshold, resulting in a user-adaptive sensitivity threshold. The formula for calculating the adaptive threshold is as follows: In the formula, Represented as an adaptive threshold calculated based on the user baseline. This is expressed as a sensitivity coefficient; The update expressions for the mean and variance statistics are as follows: ,and In the formula, This represents a dynamic update of the mean statistic. Represented as the forgetting factor, This is represented as a dynamic update of the variance statistic. Expressed as drift intensity; The formula for calculating the user-adaptive sensitivity threshold is as follows: In the formula, This is represented as the user-adaptive sensitivity threshold; An anomaly signal is triggered by comparing the current drift intensity with the user's adaptive sensitivity threshold, and the drift intensity and drift type information of the user behavior are output.

[0030] S3. Construct a three-domain fusion knowledge graph based on the medical domain, business domain, and user behavior domain. Combine user queries to locate graph anchor points. Design a dynamic activation mechanism for graph neural networks (GNNs). Perform inference based on drift type and abnormal signals to generate multi-hop inference paths. Combine attention weights to calculate path confidence. Define a custom inference result ranking algorithm. After comprehensively weighting and scoring the inference results based on path length, confidence, and user preferences, obtain the optimal recommendation result. Specifically, the construction steps of the three-domain integrated knowledge graph are as follows: Information including entities, attributes, and relationships is collected from the medical domain, commercial domain, and user behavior domain respectively. The collected information is then cleaned, deduplicated, and normalized to obtain preprocessed information. In the medical domain, an ontology is constructed based on the standard medical terminology system of SNOMEDCT or ICD-11, which includes entity types such as disease, symptoms, drugs, treatment plan, side effects, and contraindications, as well as semantic relationships such as "treatment", "cause", and "contraindication". Design business entities in the business domain, including entities such as products, inventory, promotions, suppliers, and pricing strategies, as well as relationships such as "belongs to category", "is currently on promotion", and "inventory status". Construct a user behavior ontology in the user behavior domain, which includes entities such as user profile, historical interactions, preference tags, spending power, and health goals, as well as relationships such as "purchased", "preferences", and "following". Identify entities in each domain and align entities using a knowledge fusion algorithm. Merge entities representing the same object in different domains and establish cross-domain relationships, including medical-commercial connections, medical-behavioral connections, and medical-behavioral connections. For medical-commercial connections, establish equivalence relationships between drug entities and product entities, and "applicable scenario" relationships between disease entities and related health products. For medical-behavioral connections, establish the "matching degree" relationship between users' health goals and recommended drugs, and the "association" relationship between users' historical consultations and disease entities; For medical-behavioral connections, establish a "preference" relationship between user preferences and product categories, and a "fit" relationship between user spending power and price ranges; Entity linking technology is used to extract relationships between entities from preprocessed information, and the knowledge graphs of the three domains are merged to construct a three-domain fused knowledge graph. The merged three-domain knowledge graph is stored in the Neo4j graph database, where nodes represent entities and edges represent relationships. A periodic update mechanism is set up to update the three-domain knowledge graph based on preprocessed information such as new medical knowledge, business activity data, and user behavior data.

[0031] Specifically, the steps for setting up the dynamic activation mechanism of a graph neural network (GNN) to generate multi-hop inference paths are as follows: Based on a three-domain fusion knowledge graph, entity linking technology is used to map user queries to a set of corresponding entities in the knowledge graph, which serves as the starting graph anchor point set for reasoning. Based on the drift type and anomalous signals, a dynamic activation mechanism for a graph neural network (GNN) is designed, and the activation function parameters of the GNN are dynamically adjusted. The expression for the activation function parameters of the GNN is as follows: In the formula, This represents the attention weight of node v to its neighbor u. This is represented by the softmax activation function. Represented as the ReLU activation function, This is represented as the parameter vector of the attention mechanism. Let the learnable weight matrix be represented as that of neighbor u. Let it be represented as the feature vector of neighbor u. This is represented as the learnable weight matrix for node v. This is represented as the feature vector of node v. Represented as a vector concatenation symbol, This is represented as transpose; Multi-hop inference is performed using a Generative Neural Network (GNN). Starting from the initial set of graph anchors, the GNN iteratively aggregates neighbor information and updates node representations via its message passing mechanism. Simultaneously, at each hop, the activated nodes and edges are recorded to form an inference path, generating a multi-hop inference path. The expression for updating nodes using the GNN is: In the formula, This is represented as the feature vector of node v in layer l. Represented as nodes The neighborhood group, This is represented as the learnable weight matrix of node v in layer l. This is represented as the feature vector of neighbor u in layer l-1; The path length is limited to 2-4 hops as a filtering condition. Reasonable candidate inference paths are selected from all collected multi-hop inference paths. The expression for selecting reasonable candidate inference paths is as follows: ,and In the formula, This represents the reasoning path used to select reasonable candidates. Represented as each reasoning path, This is represented as the path activation threshold. Represented as the selected first One reasoning path, This represents the total number of selected reasoning paths, and .

[0032] Specifically, the steps for calculating path confidence are as follows: For each selected inference path, extract path features for each node and relation, calculating the mean attention weight, path length, relation reliability, and semantic consistency. The expression for the mean attention weight is: ,and In the formula, Represented as the mean of attention weights, Represented as path length, Represented as the first The reasoning path to the first Attention weights on the edges of the reasoning path This can be represented as the summation of attention weights over all inference path edges; The expression for path length is: In the formula, Represented as path length, Represented as the number of hops in the path; The expression for the reliability of a relationship is: In the formula, This represents the overall reliability of the relationship in the reasoning path. It represents the chain multiplication symbol for multiplying all relations. Represented as relationships in a three-domain integrated knowledge graph , Represented as relationships derived from the three-domain fusion knowledge graph Prior reliability, Represented as from the first The reasoning path is through the relationship Connect to the Triples for a reasoning path; The expression for the semantically consistent path features is: In the formula, Represented as the semantic consistency score of the reasoning path. Represented as a cosine similarity function, Represented as embedded in the first Vector representation of the starting entity of each reasoning path. Represented as embedded in the first Vector representation of the entity terminating a reasoning path; The structural confidence score is calculated based on path topology features, the semantic confidence score is calculated based on embedding similarity, and the contextual fit score is calculated based on the matching of path and user context. The formula for calculating the structural confidence score is as follows: In the formula, This is represented as structural confidence. It is represented as a natural exponential function. It is expressed as the length attenuation coefficient, and ; The formula for calculating semantic confidence is: In the formula, Represented as semantic confidence; The formula for calculating context fit is: In the formula, This is represented as the path's fit with the user context. This can be represented as the overall vector representation of the path obtained by aggregating the node representations within the path. It is represented as a vector representation of the user context obtained from the user query; Structural confidence, semantic confidence, and contextual fit are used as confidence components. A weighted geometric mean fusion algorithm is used to calculate the overall confidence score. The confidence score is then normalized using the softmax function across all reasonable candidate inference paths for the same query. The formula for calculating the overall confidence score is as follows: ,and In the formula, This represents the overall confidence level of the reasoning path. The weights are represented as structural confidence scores. The weights are represented as semantic confidence. The weight is represented as the context fit.

[0033] Specifically, the steps for generating the optimal recommendation result are as follows: Collect the confidence scores of all generated inference paths, and use the terminal entities of all generated inference paths as candidate recommendation results; Taking into account path confidence, user preferences, commercial value, and health fit, a scoring function is designed for each candidate recommendation result, and the weighted scoring fusion process is dynamically adjusted according to the drift type to obtain a comprehensive score; Based on the comprehensive scores, the results are sorted in descending order, and the reasoning result with the highest comprehensive score is selected as the optimal recommendation result.

[0034] S4. Establish an uncertainty quantification model, calculate task complexity by combining recommendation result confidence and drift intensity, learn the optimal task allocation strategy using the contextual multi-armed gambling machine algorithm, and then dynamically allocate human and machine tasks. Use conditional variational autoencoders to generate personalized interactive content that matches the user context, collect user feedback information, and update the system model parameters online through incremental learning engine.

[0035] Specifically, the steps for establishing an uncertainty quantification model to calculate task complexity are as follows: By combining the confidence level and drift intensity of the recommendation results, an uncertainty quantification model is established to analyze the uncertainty components of each recommendation result, including cognitive uncertainty, random uncertainty from data noise, and inherent uncertainty of the uncertainty quantification model. Cognitive uncertainty measures the degree of knowledge lack of the uncertainty quantification model regarding the recommendation results; therefore, the expression for cognitive uncertainty is: In the formula, This can be expressed as quantified cognitive uncertainty. Random uncertainty measures the noise level of contextual data inputting user behavior; therefore, the expression for random uncertainty is: In the formula, This can be expressed as quantified random uncertainty. This can be represented as the historical volatility of user behavior; Inherent uncertainty measures the reliability of the uncertainty quantification model itself; therefore, the expression for inherent uncertainty is: In the formula, This can be expressed as the inherent uncertainty in quantification. Represented as inherent coefficients, This represents the time interval for updating the uncertainty quantification model. A multi-factor weighted model is used to comprehensively calculate the quantified uncertainty, resulting in the task complexity. The formula for calculating task complexity is as follows: ,and In the formula, This is expressed as task complexity. Represented as the weight of quantified cognitive uncertainty, Weights are represented as inherent uncertainties due to chance. Weights are represented as quantified inherent uncertainties. Tasks are categorized into operation levels based on complexity, including when... At that time, the operation level is low complexity, which is a routine user consultation and is handled autonomously by AI; when At that time, the operation level was medium complexity, with AI as the primary driver and physician supervision. when At that time, the operation level was characterized by high complexity, with human-led and AI-assisted operation.

[0036] Specifically, the steps for dynamic allocation of human and machine tasks are as follows: Define the action space, including four types of actions: actions processed entirely by AI, actions processed entirely by human physicians, actions processed by AI first and then by human physicians, and actions processed by human physicians first and then by AI assistance. The features of the current human-machine task are constructed into a context vector, including task complexity, drift intensity, user identity, current busy status of the human agent, and time pressure status feature information. Based on the feature information of human-machine task allocation, a contextual multi-armed gambling machine model is established, and the number of arms and the parameters of the reward function of the contextual multi-armed gambling machine algorithm are initialized. By continuously learning using the contextual multi-armed gambling machine algorithm, the maintenance value of each state-action is estimated, the optimal task allocation strategy is found, and the human-machine task is dynamically allocated based on the optimal task allocation strategy learned by the contextual multi-armed gambling machine model.

[0037] Specifically, the steps for generating personalized interactive content are as follows: The design of the Conditional Variational Autoencoder (CVAE) model includes an encoder that encodes user contextual representation information containing user profile, current emotion, historical interaction and preset response template as latent variables, and a decoder that generates personalized interactive content based on the latent variables and conditional information such as task complexity and human-computer task allocation. The CVAE model is trained using historical data to learn the latent relationship between user context and interactive content. Samples are taken from the latent variable distribution obtained from the encoding, input into the decoder to generate multiple candidate responses, and the candidate responses are sorted to select the optimal response. The feature vector of the user's current context is input into the trained CVAE model to generate personalized interactive content that matches the user's context.

[0038] Example 2 This invention provides, for example Figure 3 The human-computer collaborative intelligent health consultant interaction system shown includes a multimodal context perception module: it collects multi-source interaction data of users in real time using the acquisition channel, designs a hierarchical feature extraction network to extract feature data, uses a cross-modal attention fusion mechanism to perform cross-modal attention weighting, obtains a context representation vector and stores it in a temporal context buffer; Context drift detection module: Reads and analyzes context representation vectors, uses online adaptive clustering K-means++ algorithm to cluster context representation vectors for user-personalized behavior, uses isolation forest algorithm to detect user behavior drift, outputs drift intensity, identifies drift type, sets sensitivity adaptive mechanism, triggers abnormal signal when drift anomaly occurs, and transmits drift type and abnormal signal to dynamic knowledge reasoning module; Dynamic knowledge reasoning module: Based on the constructed three-domain fusion knowledge graph, the module locates graph anchor points by combining user queries, designs a dynamic activation mechanism for graph neural networks (GNN), performs reasoning based on drift type and abnormal signals, generates multi-hop reasoning paths, calculates path confidence, uses a custom reasoning result ranking algorithm to weight and score the reasoning results, obtains the optimal recommendation result, and transmits it to the human-machine collaborative decision-making and feedback optimization module. Human-Machine Collaborative Decision-Making and Feedback Optimization Module: Establish an uncertainty quantification model, calculate task complexity based on recommendation result confidence and drift intensity, learn the optimal task allocation strategy using a contextual multi-armed gambling machine algorithm, dynamically allocate human and machine tasks, generate personalized interactive content matching user context using a conditional variational autoencoder, collect user feedback information, and continuously optimize system performance by learning and updating system model parameters online through an incremental learning engine.

[0039] Specifically, the hierarchical feature extraction network is designed to consist of a text encoder, a speech processor, an image recognizer, and a behavior analyzer.

[0040] The present invention provides a human-computer collaborative intelligent health consultant interaction system, which is implemented through the above-mentioned human-computer collaborative intelligent health consultant interaction method. For details of the specific method and process of the human-computer collaborative intelligent health consultant interaction system, please refer to the above-mentioned embodiment of the human-computer collaborative intelligent health consultant interaction method, which will not be repeated here.

[0041] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0042] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0043] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0044] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A human-in-the-loop intelligent health advisor interaction method, characterized in that, Comprise the following steps: S1, establish multi-modal data acquisition channel, real-time acquisition of multi-source interaction data of users, design hierarchical feature extraction network for feature extraction, use cross-modal attention fusion mechanism to cross-modal attention weighting on the extracted features, obtain context representation vector, and store in the time sequence context buffer; S2, using online adaptive clustering algorithm for user personalized behavior clustering of context representation vector, dynamically adjusting the number and position of clustering center, using isolation forest algorithm to detect the deviation degree of new context vector and historical mode, outputting drift intensity, establishing drift type classifier to identify the drift type of user behavior, setting sensitivity adaptive mechanism to trigger abnormal signal; S3, construct three-domain fusion knowledge graph, combine user query to locate graph anchor point, design dynamic activation mechanism of graph neural network GNN, infer according to drift type and abnormal signal, generate multi-hop inference path, calculate path confidence combining attention weight, self define a kind of inference result sorting algorithm, get the optimal recommendation result after weighted scoring of inference result; S4, establish an uncertainty quantification model, combine the recommendation result confidence and the drift intensity to calculate the task complexity, use context multi-armed bandit algorithm for human-machine task dynamic allocation, use conditional variational autoencoder to generate personalized interaction content, collect user feedback information, and update the system model parameters through incremental learning engine.

2. The human-robot collaborative intelligent health advisor interaction method of claim 1, wherein, The multi-modal data acquisition channel is built by using multi-source sensor interface and interaction log recorder, including text channel, speech channel, visual channel and behavior sequence channel; When the user initiates an interaction session with the intelligent health consultant interaction platform, the multi-modal data acquisition channel is established, and the multi-source interaction data of the user is collected in real time and synchronously, including text semantics, speech information, visual content and behavior sequence raw data; The collected raw data is preprocessed by cleaning, denoising and standardization; The preprocessed data is temporarily stored in the data buffer, and each data is marked with a unified timestamp to form the user's multi-source interaction data.

3. The human-assisted intelligent health advisor interaction method of claim 2, wherein, The design steps of the hierarchical feature extraction network are as follows: Deploy pre-trained neural network models including text encoder, speech processor, image recognizer and behavior analyzer, including BERT model, Wav2Vec2 model, ResNet model and LSTM model, to extract features at different levels for multi-modal data; The BERT model is used to process the preprocessed text semantics, and the BERT model learns the semantic information of the language through pre-training, which can convert the input text semantics into semantic feature vectors; The Wav2Vec2 model is used to extract features from preprocessed speech information, including using ASR capability to obtain transcribed text and extracting speech feature vectors that can represent linguistic features; The ResNet model is used to extract features from preprocessed visual content, removing the classification layer of the ResNet model, and taking the output of the last pooling layer as the image feature vector of the visual content; The behavior sequence after behavior preprocessing is processed using an LSTM model, and the hidden state of the last time step of the LSTM model is taken as the behavior feature vector of the entire behavior sequence; Each feature vector is processed by feature dimension reduction and alignment using a fully connected layer, and the semantic feature vector, the speech feature vector, the image feature vector, and the behavior feature vector are projected into a unified and 256-dimensional common subspace.

4. The human-assisted intelligent health advisor interaction method of claim 3, wherein, The context representation vector is obtained as follows: The modality feature vectors after projection, alignment, and unification are regarded as a group of vector sequences. For each modality feature, the attention score between each pair of modality feature vectors is calculated, and the attention score is normalized by a Softmax function to obtain a group of weight matrices. The importance of each feature vector in different modalities is calculated by weighting according to the weight matrix calculated in the previous step using a cross-modal attention fusion mechanism, and a residual connection is performed with the modality feature itself to obtain an enhanced modality feature vector. After cross-modal attention weighting, all enhanced modality feature vectors are concatenated and then weighted and summed for combination, and a final fully connected layer is used for fine-tuning to output a unified dimension context representation vector. The generated context representation vector is stored in the temporal context buffer together with the current session ID and timestamp.

5. The human-assisted intelligent health advisor interaction method of claim 4, wherein, The steps of the online adaptive clustering K-means++ algorithm for user personalized behavior clustering are as follows: A context representation vector is randomly selected from the received set of context representation vectors as the first cluster center according to the K-means++ algorithm. The shortest distance of each context representation vector to the selected cluster center is calculated using the Euclidean distance, and the context representation vector with the maximum distance probability is selected as the new cluster center. The cluster center selection process is repeated until K cluster centers are selected. The distance of each new context representation vector to the selected K cluster centers is calculated, and the new context representation vector is assigned to the cluster center with the closest distance. The corresponding cluster center of the cluster where the new context representation vector is located is updated using the exponentially weighted moving average method. A clustering structure threshold is set, and the distance of the new context representation vector to all cluster centers is analyzed with the clustering structure threshold.

6. The human-assisted intelligent health advisor interaction method of claim 5, wherein, The output steps of the drift intensity are as follows: The user's historical context representation vectors stored in the temporal context buffer are read, and the isolation forest model is trained using the user's historical context representation vectors. The new context vector is input into the trained isolation forest model, and the path length of the new context vector in each tree is calculated by traversing from the root node to the leaf node in each isolation tree. The average path length of the new context vector in the forest is obtained by averaging the path lengths of the new context vector in each tree, wherein the calculation formula of the average path length is , wherein represents the average path length, represents the total number of isolated trees in the forest, represents the path length of the new context vector on each tree. The normalized anomaly score is calculated according to the average path length, and is mapped to the drift intensity, wherein the calculation formula of the normalized anomaly score is , and , wherein, is expressed as a normalized anomaly score, is expressed as a normalized factor, and is the average path length of the binary search tree of the sample, is expressed as a harmonic number; The mapping expression for the drift strength is where, is expressed as the drift strength, is expressed as the minimum and maximum normalized boundaries of the anomaly score.

7. The human-collaborative intelligent health advisor interaction method of claim 6, wherein, The identification steps of the drift type are as follows: The context representation vector is decomposed into emotion features, intent features, and domain features according to the semantic dimension; For each type of drift, the type-specific drift score is calculated by summing the weighted absolute differences, and the degree of deviation of each type of feature group of the historical cluster center is quantified. The type-specific drift score is converted into the probability of each drift type using the softmax function, and the main drift type is determined by probabilistic type. The multiple types of drift conditions are identified based on preset threshold values of each drift type, and when the probabilities of multiple types exceed the threshold values, the multiple types of drift are marked as composite type drift.

8. The human-assisted intelligent health advisor interaction method of claim 7, wherein, The steps of setting the dynamic activation mechanism of the graph neural network GNN to generate the multi-hop reasoning path are as follows: Based on the three-domain fusion knowledge graph, the entity linking technology is used to map the user query to a set of corresponding entities in the knowledge graph as the starting graph anchor point set for reasoning; According to the drift type and the abnormal signal, the dynamic activation mechanism of the graph neural network GNN is designed, and the activation function parameters of the graph neural network GNN are dynamically adjusted; Using GNN for multi-hop reasoning, starting from the starting graph anchor point set, through the message passing mechanism of GNN, iteratively aggregating neighbor information and updating node representation, while recording the activated nodes and edges at each hop to form a reasoning path, and generating a multi-hop reasoning path; The range limit of the path length is set to 2-4 hops as a screening condition to filter out reasonable candidate reasoning paths from all collected multi-hop reasoning paths.

9. The human-collaborative intelligent health advisor interaction method of claim 7, wherein, The steps of establishing the uncertainty quantification model to calculate the task complexity are as follows: Combining the recommendation result confidence and the drift strength, an uncertainty quantification model is established to analyze the uncertainty components of each recommendation result, including the cognitive uncertainty of the recommendation result, the accidental uncertainty of the data noise, and the inherent uncertainty of the uncertainty quantification model, wherein the cognitive uncertainty is used to measure the knowledge deficiency of the uncertainty quantification model to the recommendation result, and the expression of the cognitive uncertainty is , wherein is expressed as the quantified cognitive uncertainty; The accidental uncertainty is a measure of the noise level of the contextual data of the input user behavior, and the expression of the accidental uncertainty is wherein is expressed as a quantized accidental uncertainty, is expressed as a historical volatility of the user behavior; The inherent uncertainty is to measure the reliability of the uncertainty quantification model itself, and the expression of the inherent uncertainty is , wherein, is the inherent uncertainty quantified, is the inherent coefficient, is the uncertainty quantification model update time interval; The quantified uncertainties are synthesized by using a multi-factor weighting model to obtain the task complexity, wherein a calculation formula of the task complexity is , and , wherein, is the task complexity, is the weight of the quantified cognitive uncertainty, is the weight of the accidental inherent uncertainty, is the weight of the quantified inherent uncertainty. According to the task complexity, the operation level is divided into operation levels, including when the operation level is low complexity of user routine consultation and AI autonomous processing; When the operation level is AI-dominant and physician-supervised medium complexity; When the operational level is human-led and AI-assisted high complexity.

10. A human-computer collaborative intelligent health consultant interaction system, which is realized by the human-computer collaborative intelligent health consultant interaction method in any one of claims 1-8, characterized in that, It includes a multi-modal context perception module: real-time acquisition of multi-source interaction data of users through acquisition channels, design of a hierarchical feature extraction network to extract feature data, cross-modal attention weighting using a cross-modal attention fusion mechanism, obtaining context representation vectors and storing them in a time-series context buffer; The context drift detection module reads and analyzes the context representation vectors, uses the online adaptive clustering K-means++ algorithm to cluster the context representation vectors according to user personalized behavior, uses the isolation forest algorithm to detect user behavior drift, outputs the drift strength, identifies the drift type, sets a sensitivity adaptive mechanism, triggers an abnormal signal when a drift anomaly occurs, and transmits the drift type and abnormal signal to the dynamic knowledge reasoning module; The dynamic knowledge reasoning module: based on the constructed three-domain fusion knowledge graph, combined with user query for graph anchor point positioning, designed dynamic activation mechanism of graph neural network GNN, according to the drift type and the abnormal signal for reasoning, generating multi-hop reasoning path, calculating path confidence, using self-defined reasoning result sorting algorithm for weighted scoring of reasoning result, getting the optimal recommendation result, and transmitting to the human-machine collaborative decision and feedback optimization module; The human-machine collaborative decision and feedback optimization module: establishes an uncertainty quantification model, calculates the task complexity according to the recommendation result confidence and the drift strength, uses the context multi-armed bandit algorithm to learn the optimal task allocation strategy, performs human-machine task dynamic allocation, generates personalized interactive content matched with the user's context using the conditional variational autoencoder, collects user feedback information, and updates the system model parameters through the incremental learning engine for online learning, continuously optimizing system performance.

Images