Intelligent medical multi-round dialogue diagnosis reasoning method and system based on deep learning

Abstract

The invention provides an intelligent medical multi-round dialogue diagnosis reasoning method and system based on deep learning, and relates to the technical field of deep learning, and the method comprises the steps: constructing time sequence features through employing an attention mechanism for time sequence symptom description, and carrying out bidirectional sequence modeling to generate comprehensive symptom features; using the medical knowledge graph to detect logic contradictions and information loss to generate standardized features; calculating information gain of an inquiry direction based on a deep neural network to generate optimal inquiry content; and iteratively updating according to user feedback until the diagnosis information entropy is lower than a threshold value, and outputting a diagnosis result. According to the invention, the accuracy and efficiency of medical diagnosis are improved.

Description

Technical Field

[0001] This invention relates to deep learning technology, and more particularly to a method and system for intelligent medical multi-turn dialogue diagnostic reasoning based on deep learning. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning is increasingly being applied in the field of intelligent healthcare. As an important branch of AI in healthcare, intelligent medical multi-turn dialogue diagnostic systems aim to simulate the doctor-patient consultation process, collecting patient symptom information through multi-turn interactive dialogue and making disease diagnostic inferences. These systems can alleviate the problem of uneven distribution of medical resources and improve the accessibility and efficiency of medical services. However, existing intelligent medical multi-turn dialogue diagnostic systems have the following shortcomings and deficiencies:

[0003] Existing systems lack the ability to effectively model the temporal evolution of symptoms in multi-turn dialogues. Patients' symptoms often change over time, and there are close logical connections between symptoms before and after. However, existing systems usually process each turn of dialogue independently, failing to capture the evolution of symptoms and the correlation between them, resulting in a lack of temporal coherence in diagnostic reasoning.

[0004] Existing systems perform poorly when processing non-standardized symptom descriptions from patients. Patients often use vague or non-technical terms in their symptom descriptions, making it difficult for the system to accurately understand the patient's true intentions. Furthermore, the lack of effective medical knowledge constraints to detect logical inconsistencies and missing information in the descriptions affects diagnostic accuracy.

[0005] Existing systems lack dynamic optimization capabilities in their consultation strategies. These systems typically employ preset, fixed consultation paths or simple greedy algorithms to select the next question, failing to perform global optimization based on the uncertainty of currently known symptom information and the information gain of each candidate question. This results in low consultation efficiency, requiring more rounds to achieve an accurate diagnosis. Summary of the Invention

[0006] This invention provides a method and system for intelligent medical multi-turn dialogue diagnostic reasoning based on deep learning, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a deep learning-based intelligent medical multi-turn dialogue diagnostic reasoning method, comprising:

[0008] The symptom descriptions input by the user are stored in chronological order. An attention mechanism is used to calculate the correlation weight of the symptom descriptions in each round. Based on the correlation weight, the current symptom description is fused with historical dialogue information to construct a chronological feature that includes the symptom evolution process. The chronological feature is then bidirectionally sequence-modeled to identify the logical relationship between the symptom descriptions and generate a comprehensive symptom feature that takes into account historical evolution.

[0009] Based on a medical knowledge graph, the comprehensive symptom features are analyzed for correlation, and contradictions in medical professional logic and missing information in the description are detected to generate standardized symptom features.

[0010] The standardized symptom features are input into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round. Based on user feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold, and the diagnosis result is output.

[0011] In one alternative implementation,

[0012] The steps include: storing user-inputted symptom descriptions chronologically; calculating association weights based on the correlation between symptom descriptions from different rounds using an attention mechanism; fusing current symptom descriptions with historical dialogue information based on these association weights to construct chronological features that include the symptom evolution process; performing bidirectional sequence modeling on these chronological features to identify the logical relationships between symptom descriptions; and generating comprehensive symptom features that consider historical evolution.

[0013] The symptom descriptions input by the user in each round are segmented into word sequences, and these word sequences are used to construct a dialogue history matrix. Based on the dialogue history matrix, the word vector matrix of the current round and the word vector matrix of the previous round are extracted. A word-level attention score is calculated for each word in the current round and all words in the previous round. A round representation is obtained based on the word-level attention score, and the inter-round association weights are calculated based on the round representations. The inter-round association weights are obtained through a non-linear transformation between the hidden states of the current round and the hidden states of the previous rounds.

[0014] The word-level attention score and the inter-round association weight are weighted and combined to obtain the temporal fusion feature. The temporal fusion feature is processed forward and backward through a bidirectional gated loop unit to obtain the temporal feature. The internal logical structure of the symptom description is identified by analyzing the forward and backward dependencies of the symptom description.

[0015] Calculate the semantic similarity of temporal features between adjacent rounds. If the semantic similarity is lower than a preset threshold, correct the temporal features of the current round based on the temporal features of the historical rounds. Calculate the difference vector of temporal features between adjacent rounds.

[0016] The corrected temporal features, difference vectors, and internal logical structure features identified through bidirectional sequence modeling are fused to obtain comprehensive symptom features.

[0017] In one alternative implementation,

[0018] The temporal fusion features are processed forward and backward using a bidirectional gated loop unit to obtain temporal features. The steps of identifying the internal logical structure of symptom descriptions by analyzing the forward and backward dependencies of symptom descriptions include:

[0019] The temporal fusion features are input into a bidirectional gated recurrent unit to obtain the forward hidden state sequence and the backward hidden state sequence;

[0020] The temporal window size is calculated based on the forward and backward hidden state sequences. The temporal mean and standard deviation are calculated based on the hidden state subsequences corresponding to the temporal window size. Fine-grained dependency strength is calculated based on the temporal mean and standard deviation. Coarse-grained dependency strength is obtained by multiplying the original hidden states. Adaptive weights are calculated based on the temporal position difference and the hidden states. The fine-grained dependency strength and the coarse-grained dependency strength are weighted and combined to obtain multi-granularity dependency features.

[0021] The multi-granularity dependency features and hidden state sequences are input into a recursive logic decomposition network, which constructs logical relationship representations layer by layer through a multi-layer cascaded structure. The logical relationship representations are input into a generator to generate candidate logical relationships, and the logical rationality is verified by a discriminator to obtain optimized logical relationships. The optimized logical relationships are combined with the bidirectional hidden state sequences to obtain the symptom logical structure representation.

[0022] In one alternative implementation,

[0023] The steps for generating standardized symptom features by performing correlation analysis on the comprehensive symptom features based on a medical knowledge graph, detecting medical logical contradictions and missing information in the description, include:

[0024] Transform the comprehensive symptom characteristics into a standard symptom set;

[0025] Based on the symptom ontology and symptom relationships in the medical knowledge graph, an association strength matrix, a pathological conflict matrix, and a necessary co-occurrence matrix are constructed. The association strength matrix records the degree of medical association between standard symptoms, the pathological conflict matrix records the medical conflict relationship between standard symptoms, and the necessary co-occurrence matrix records the necessary co-occurrence relationship between standard symptoms.

[0026] For each pair of standard symptoms in the standard symptom set, the degree of association is obtained by querying the association strength matrix, and a symptom association graph is constructed based on the degree of association; conflict relationships are obtained by querying the pathological conflict matrix, and when a conflict relationship exists, the corresponding symptom pair is marked as a logical contradiction point; a set of necessary accompanying symptoms is obtained by querying the necessary accompanying matrix, and necessary accompanying symptoms that are not in the standard symptom set are marked as information missing points;

[0027] Based on the symptom association diagram, the logical contradictions, and the missing information points, the standard symptom set is optimized to generate standardized symptom features that include symptom associations.

[0028] In one alternative implementation,

[0029] The steps of inputting the standardized symptom features into a deep neural network, which calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round include:

[0030] Standardized symptom features are encoded into symptom semantic vectors using a medical pre-trained language model. The temporal position codes of the standardized symptom features in the course of disease are extracted, and the professional knowledge feature vectors corresponding to the standardized symptom features are obtained. The symptom semantic vectors, the temporal position codes, and the professional knowledge feature vectors are input into a deep neural network to obtain symptom feature representations.

[0031] The deep neural network constructs a state vector based on the symptom feature representation, applies a multi-head attention mechanism to the historical dialogue sequence to obtain a dialogue context vector, and fuses the state vector and the dialogue context vector to obtain the current diagnostic state; based on the current diagnostic state, it calculates the expected information gain of each candidate consultation direction in the candidate consultation direction set;

[0032] The user's historical feedback information is input into the feedback encoding layer to obtain the feedback feature vector. The feedback feature vector is combined with the expected information gain to obtain the consultation evaluation vector. The deep neural network optimizes and selects the candidate consultation direction set based on the consultation evaluation vector to generate the optimal consultation content.

[0033] In one alternative implementation,

[0034] The steps for calculating the expected information gain of each candidate consultation direction in the candidate consultation direction set based on the current diagnostic status include:

[0035] For each candidate consultation direction in the candidate consultation direction set, a consultation feature vector is constructed. The consultation feature vector includes a symptom correlation vector, a temporal feature vector, and a severity vector. The consultation feature vector is input into a variational autoencoder for state transition prediction. The consultation feature is compressed into a low-dimensional latent representation through an encoding network. The latent representation is randomly sampled and the state transition result is reconstructed through a decoding network. Based on the multiple sampling and reconstruction results, a state transition probability distribution under different feedback types is generated.

[0036] Based on the state transition probability distribution, the diagnostic convergence gain, symptom confirmation gain, and risk identification gain for each candidate consultation direction are calculated. The diagnostic convergence gain is calculated by the difference between the conditional entropy of the disease distribution in the current state and the conditional entropy of the disease distribution in the predicted state. The symptom confirmation gain is calculated by weighting the symptom confirmation probability and feedback contribution obtained by analyzing historical feedback patterns through a recurrent neural network. The risk identification gain is calculated by multiplying the clinical risk level and symptom urgency level assessed by the medical knowledge graph.

[0037] Based on Bayesian decision theory, the certainty of the current diagnosis, the need for verification, and the urgency are calculated. The diagnostic convergence gain, symptom confirmation gain, and risk identification gain are dynamically weighted and combined to obtain the expected information gain of the candidate consultation direction.

[0038] In one alternative implementation,

[0039] Based on user feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold. The steps for outputting the diagnosis result include:

[0040] Extract symptom entities, attributes, and relationships from user feedback to construct incremental symptom feature data; merge and update the incremental symptom feature data with the original standardized symptom features, and perform consistency verification using medical knowledge;

[0041] Based on the updated standardized symptom features, calculate the disease diagnosis probability distribution and its information entropy; when the information entropy is greater than a preset threshold, return to the consultation optimization process, and when it is less than or equal to the preset threshold, enter the diagnosis result generation process.

[0042] A symptom-disease association matrix is ​​constructed. The diagnostic contribution of each symptom is calculated based on its discriminative power, frequency of occurrence, and clinical relevance in disease diagnosis. Symptoms with a diagnostic contribution greater than a first preset threshold are identified as diagnostic criteria. Based on symptom combination rules in a medical knowledge graph, clinically relevant symptom combinations are identified, and the disease diagnosis probability corresponding to the symptom combination is calculated.

[0043] Candidate diseases are ranked according to the disease diagnosis probability of the symptom combinations, and the disease with the highest diagnosis probability is selected as the diagnosis conclusion. The symptom contribution of the diagnostic basis is normalized to obtain the symptom weight. The conditional probability distribution of the diagnosis conclusion is weighted and summed based on the symptom weight, and then corrected by combining the clinical relevance of the symptom combinations to obtain the confidence level. The diagnosis conclusion, the diagnostic basis, the clinical relevance of the symptom combinations, and the corresponding confidence level are organized to form a diagnosis report.

[0044] A second aspect of this invention provides a deep learning-based intelligent medical multi-turn dialogue diagnostic reasoning system, comprising:

[0045] The first unit is used to store the symptom descriptions input by the user in chronological order, calculate the correlation weight of the correlation between the symptom descriptions in each round using an attention mechanism, and construct a chronological feature containing the symptom evolution process by fusing the current symptom description with historical dialogue information based on the correlation weight. The chronological feature is then used to perform bidirectional sequence modeling to identify the logical relationship between the symptom descriptions before and after, and generate a comprehensive symptom feature that takes into account the historical evolution.

[0046] The second unit is used to perform correlation analysis on the comprehensive symptom features based on the medical knowledge graph, detect medical professional logical contradictions and information gaps in the description, and generate standardized symptom features.

[0047] The third unit is used to input the standardized symptom features into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round. Based on the user's feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold, and the diagnosis result is output.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] The intelligent medical multi-turn dialogue diagnostic reasoning method based on deep learning provided by this invention can effectively capture the temporal evolution of patient symptoms by constructing temporal features and using attention mechanisms, and comprehensively considers the logical connections between symptoms, thereby improving the accuracy and completeness of symptom understanding.

[0054] This invention combines medical knowledge graphs to standardize symptoms, automatically detects and corrects logical contradictions in descriptions, supplements missing information, and converts patients' colloquial descriptions into standardized medical terminology, providing a more reliable data foundation for subsequent diagnosis.

[0055] This invention employs a dynamic consultation strategy based on information entropy. By calculating the expected information gain for different consultation directions, it intelligently generates the most valuable questions for the next round, significantly reducing unnecessary consultation steps, improving diagnostic efficiency, and ensuring the reliability of diagnostic results. This significantly improves the user experience and clinical applicability of intelligent medical diagnostic systems. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the intelligent medical multi-turn dialogue diagnostic reasoning method based on deep learning, as described in an embodiment of the present invention.

[0057] Figure 2 This chart compares the performance improvements of different semantic models in intelligent consultation. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0060] Figure 1 This is a flowchart illustrating the deep learning-based intelligent medical multi-turn dialogue diagnostic reasoning method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] The symptom descriptions input by the user are stored in chronological order. An attention mechanism is used to calculate the correlation weight of the symptom descriptions in each round. Based on the correlation weight, the current symptom description is fused with historical dialogue information to construct a chronological feature that includes the symptom evolution process. The chronological feature is then bidirectionally sequence-modeled to identify the logical relationship between the symptom descriptions and generate a comprehensive symptom feature that takes into account historical evolution.

[0062] Based on a medical knowledge graph, the comprehensive symptom features are analyzed for correlation, and contradictions in medical professional logic and missing information in the description are detected to generate standardized symptom features.

[0063] The standardized symptom features are input into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round. Based on user feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold, and the diagnosis result is output.

[0064] In one optional implementation, the steps of storing user-inputted symptom descriptions chronologically, calculating the correlation weights of symptom descriptions in each round using an attention mechanism, fusing the current symptom description with historical dialogue information based on the correlation weights to construct a chronological feature that includes the symptom evolution process, performing bidirectional sequence modeling on the chronological feature to identify the logical relationships between symptom descriptions, and generating comprehensive symptom features that consider historical evolution include:

[0065] The symptom descriptions input by the user in each round are segmented into word sequences, and these word sequences are used to construct a dialogue history matrix. The dimensions of the dialogue history matrix include dialogue round dimension, word sequence dimension, and word vector dimension. Based on the dialogue history matrix, the word vector matrix of the current round and the word vector matrix of the previous round are extracted. A word-level attention score is calculated for each word in the current round and all words in the previous round. The word-level attention score is obtained by matrix multiplication of the word vector matrix of the current round and the word vector matrix of the previous round. A round representation is obtained based on the word-level attention score, and the inter-round association weight is calculated based on the round representation. The inter-round association weight is obtained by nonlinear transformation of the hidden state of the current round and the hidden state of the previous round.

[0066] The word-level attention score and the inter-round association weight are weighted and combined to obtain the temporal fusion feature. The temporal fusion feature is processed forward and backward through a bidirectional gated loop unit to obtain the temporal feature. The internal logical structure of the symptom description is identified by analyzing the forward and backward dependencies of the symptom description.

[0067] Calculate the semantic similarity of temporal features of adjacent rounds. If the semantic similarity is lower than a preset threshold, correct the temporal features of the current round based on the temporal features of the historical rounds. Calculate the difference vector of temporal features of adjacent rounds and analyze the symptom evolution trend based on the difference vector.

[0068] The corrected temporal features, difference vectors, and internal logical structure features identified through bidirectional sequence modeling are fused to obtain comprehensive symptom features.

[0069] For example, when a user inputs a description of symptoms, the input is segmented into words. For instance, the sentence "I started having a headache yesterday, and I have a slight fever today" can be segmented into a word sequence of ["I", "yesterday", "started", "headache", ",", "today", "slight", "fever"]. The word sequence from each round of dialogue is then used to construct a dialogue history matrix, which contains three dimensions: dialogue round dimension, word sequence dimension, and word vector dimension. Assuming a 300-dimensional word vector is used, and there have been 3 rounds of dialogue, with each round containing an average of 10 words, then the dimensions of the dialogue history matrix would be [3, 10, 300].

[0070] Based on the constructed dialogue history matrix, the word vector matrix for the current round and the word vector matrices for previous rounds are extracted. Taking the third round of dialogue as an example, the word vector matrix for the current round has dimensions [1, 10, 300], and the word vector matrix for previous rounds has dimensions [2, 10, 300]. Through matrix transformation operations, the word vector matrix for the current round is transformed into [10, 300], and the word vector matrix for previous rounds is transformed into [20, 300]. A word-level attention score is calculated for each word in the current round and all words in previous rounds, achieved through matrix multiplication of the word vector matrices for the current round and previous rounds. The result is an attention score matrix with dimensions [10, 20], where each row represents the attention distribution of a word in the current round and all words in previous rounds.

[0071] The attention score matrix is ​​normalized so that the sum of the attention scores in each row is 1 for weighted calculation. Based on the normalized attention score matrix and the word vectors from previous rounds, the context representation for each word in the current round is calculated. These context representations are then merged using average pooling to obtain a round-level representation with dimensions [1, 300].

[0072] The inter-round association weights are calculated based on round representation. Specifically, a non-linear transformation network is used, taking the hidden state of the current round and the hidden states of previous rounds as input, and outputting the inter-round association weights. This network contains two fully connected layers with a non-linear activation function in between. Assuming there are two rounds of dialogue in the past, the output round association weight dimension is [1,2], representing the degree of association between the current round and the previous two rounds of dialogue.

[0073] The word-level attention scores are grouped by round, with each group corresponding to a historical round. Then, the attention scores of each group are weighted using the corresponding round association weights. Finally, the weighted attention scores are multiplied by the historical word vectors to obtain the temporal fusion features.

[0074] For temporal fusion features, a bidirectional gated recurrent unit is used for forward and backward processing to capture the logical relationships between symptom descriptions. Taking the third round of dialogue as an example, forward processing processes information from round 1 to round 3 in chronological order, while backward processing processes information from round 3 to round 1 in reverse order. Forward processing identifies how symptoms evolve over time; backward processing reveals the causal relationship between current and historical symptoms. The combined results of bidirectional processing yield temporal features with dimensions [3, 300], where 3 represents the number of rounds and 300 represents the feature dimension.

[0075] To ensure consistency in symptom description, the semantic similarity of temporal features from adjacent rounds is calculated. Taking cosine similarity as an example, if the semantic similarity between round 2 and round 3 is 0.3, which is lower than the preset threshold of 0.5, it indicates that the description of the current round differs significantly from the previous round. In this case, the current round is corrected based on the temporal features of historical rounds. Specifically, the temporal features of historical rounds are weighted according to the inter-round association weights and fused with the temporal features of the current round to obtain the corrected temporal features.

[0076] The difference vector between the temporal features of adjacent rounds is also calculated to analyze the trend of symptom evolution. Taking the second and third rounds as an example, the difference vector is the temporal feature of the third round minus the temporal feature of the second round, resulting in a vector with dimensions [1, 300]. By analyzing the direction and magnitude of the difference vector, it can be determined whether the symptoms are aggravated, alleviated, or stabilized.

[0077] The corrected temporal features, difference vectors, and internal logical structure features identified through bidirectional sequence modeling are fused together. The fusion method involves feature concatenation followed by dimensionality reduction through a fully connected layer, ultimately yielding a comprehensive symptom feature vector with dimensions [1, 500]. This feature includes temporal evolution information of symptom description, internal logical relationships, and symptom change trends, providing more comprehensive information support for subsequent symptom identification and disease diagnosis.

[0078] This invention effectively captures the correlation and evolution of symptom descriptions at different time points by constructing a dialogue history matrix and combining it with an attention mechanism. Through dual weighting of word-level attention and turn-based association weights, it accurately identifies important information and temporal dependencies in symptom expressions. Simultaneously, the introduction of semantic similarity testing and temporal feature correction mechanisms enables timely detection and correction of deviations in symptom descriptions, ensuring the coherence and reliability of symptom information. The overall solution maintains the detailed integrity of the original symptom descriptions while achieving information fusion across time dimensions.

[0079] In one optional implementation, the temporal fusion features are processed forward and backward using a bidirectional gated loop unit to obtain temporal features. The step of identifying the internal logical structure of the symptom description by analyzing the forward and backward dependencies of the symptom description includes:

[0080] The temporal fusion features are input into a bidirectional gated recurrent unit to obtain the forward hidden state sequence and the backward hidden state sequence;

[0081] The temporal window size is calculated based on the forward and backward hidden state sequences. The temporal mean and standard deviation are calculated based on the hidden state subsequences corresponding to the temporal window size. Fine-grained dependency strength is calculated based on the temporal mean and standard deviation. Coarse-grained dependency strength is obtained by multiplying the original hidden states. Adaptive weights are calculated based on the temporal position difference and the hidden states. The fine-grained dependency strength and the coarse-grained dependency strength are weighted and combined to obtain multi-granularity dependency features.

[0082] The multi-granularity dependency features and hidden state sequences are input into a recursive logic decomposition network, which constructs logical relationship representations layer by layer through a multi-layer cascaded structure. The logical relationship representations are input into a generator to generate candidate logical relationships, and the logical rationality is verified by a discriminator to obtain optimized logical relationships. The optimized logical relationships are combined with the bidirectional hidden state sequences to obtain the symptom logical structure representation.

[0083] For example, the temporal fusion features are input into a bidirectional gated recurrent unit for processing. Specifically, the temporal fusion features are feature sequences obtained from the symptom description text after word embedding and context encoding, assuming their dimension is L×D, where L represents the sequence length and D represents the feature dimension. The bidirectional gated recurrent unit consists of two parts: a forward GRU and a backward GRU. The forward GRU starts from the beginning of the sequence and processes the features at each time step sequentially; the backward GRU starts from the end of the sequence and processes the features in reverse order.

[0084] A bidirectional GRU processes the input features at each time step t to generate corresponding forward and backward hidden states. For a symptom description text containing 100 words, such as "The patient has recently experienced persistent headaches, accompanied by a slight fever (temperature 37.5℃), and loss of appetite, with symptoms gradually worsening over the past three days," the processing yields a forward and backward hidden state sequence of 100 time steps, with each hidden state having a dimension of 128. Based on the obtained forward and backward hidden state sequences, the temporal window size is calculated. The temporal window size is adaptively determined based on the semantic complexity of the symptom description and can be obtained through trend analysis of the hidden state sequence. For example, for the above symptom description, a temporal window size of 5 is determined, meaning that each analysis considers 5 adjacent time steps.

[0085] Based on a defined temporal window size, a corresponding subsequence of hidden states is extracted from the complete hidden state sequence. Taking a temporal window size of 5 as an example, for position t, the hidden states from time steps t-2 to t+2 are extracted as a subsequence. For positions at sequence boundaries, a padding strategy is used to ensure consistent subsequence lengths. For each hidden state subsequence, the temporal mean is the average of all hidden states across all dimensions, forming a 128-dimensional vector; the temporal standard deviation is the standard deviation of the hidden state values ​​across all dimensions, also forming a 128-dimensional vector. These statistics reflect the stability and degree of change in local semantics.

[0086] Fine-grained dependency strength is calculated based on the temporal mean and standard deviation. This fine-grained dependency strength captures subtle dependencies in symptom descriptions and quantifies them by comparing the similarity of statistical features at different locations. For example, for the symptom descriptions "persistent headache" and "mild fever," the similarity between the temporal mean and standard deviation at their corresponding locations is calculated; high similarity indicates a close dependency. Simultaneously, coarse-grained dependency strength is obtained through the product of the original hidden states. Specifically, for positions i and j, the corresponding hidden state vectors are element-wise multiplied, and the resulting vectors are summed to obtain a scalar value as the coarse-grained dependency strength. This calculation method captures global dependencies in symptom descriptions.

[0087] Adaptive weights are calculated based on temporal position differences and hidden state information. The temporal position difference refers to the distance between two positions in the sequence; the greater the distance, the weaker the dependency between them. An adaptive weight is calculated for each pair of positions by combining the position difference and hidden state information. For example, smaller weights are assigned to "persistent headache" and "gradually worsening symptoms," which are far apart; while larger weights are assigned to adjacent "mild fever" and "body temperature 37.5℃."

[0088] By adaptively weighting and combining fine-grained and coarse-grained dependency strengths, multi-granularity dependency features are obtained. These features encompass both fine-grained local dependencies and coarse-grained global dependency patterns, comprehensively describing the various logical connections between symptoms.

[0089] The Recursive Logical Decomposition Network (RPDN) receives multi-granularity dependency features and hidden state sequences as input, and identifies logical relationships in symptom descriptions layer by layer through a multi-layered cascaded structure. The network employs a hierarchical design: the bottom layer processes basic semantic relationships between adjacent positions, the middle layer processes logical associations between semantic blocks, and the top layer integrates global information to construct complex logical structures. Each layer contains an attention aggregation unit, a logical relationship inference unit, and a residual connection module. The attention aggregation unit calculates the attention between positions based on multi-granularity dependency features and aggregates relevant information; the logical relationship inference unit maps the aggregated features to a logical relationship space; and the residual connections ensure that high-level features retain basic relationship information from lower layers. The network processing is recursive, with the output of each layer serving as supplementary input for the next layer. Each layer progressively refines the logical representation: the first layer identifies basic relationships such as parallelism and modification; the second layer identifies relationships with larger spans such as causality and transition; and the third layer infers complex logic such as conditional and progressive relationships. Through this hierarchical recursive structure, the network can comprehensively capture the multi-level logical structure in symptom descriptions, forming a complete logical relationship representation.

[0090] After the logical relation representation is generated, it is input into a generator to generate candidate logical relations. The generator proposes a complete logical relation graph based on the currently recognized logical structure. For example, it generates a candidate logical relation such as "headache and fever are parallel symptoms, body temperature values ​​are supplementary descriptions of fever, and the overall symptoms show a worsening trend." The generated candidate logical relations are then verified for logical rationality by a discriminator. The discriminator evaluates the semantic consistency, medical rationality, and logical self-consistency of the relations, filtering out irrational relations and retaining logical relations that conform to medical knowledge and language rules, thus obtaining optimized logical relations.

[0091] The optimized logical relationships are combined with the bidirectional hidden state sequence to obtain the final symptom logical structure representation. This representation includes both the semantic information of the symptoms themselves and the logical connections between symptoms, providing comprehensive structured information support for subsequent symptom analysis and disease diagnosis.

[0092] This invention, through the construction of multi-granularity dependent features, considers both local details and global structure, enabling a comprehensive characterization of the logical connections between symptoms. The introduction of a recursive logic decomposition network achieves a hierarchical expression of logical relationships, and the adversarial verification mechanism ensures the rationality of the logical reasoning results.

[0093] In one optional implementation, the steps of performing correlation analysis on the comprehensive symptom features based on a medical knowledge graph, detecting medical logical contradictions and missing information in the description, and generating standardized symptom features include:

[0094] Transform the comprehensive symptom characteristics into a standard symptom set;

[0095] Based on the symptom ontology and symptom relationships in the medical knowledge graph, an association strength matrix, a pathological conflict matrix, and a necessary co-occurrence matrix are constructed. The association strength matrix records the degree of medical association between standard symptoms, the pathological conflict matrix records the medical conflict relationship between standard symptoms, and the necessary co-occurrence matrix records the necessary co-occurrence relationship between standard symptoms.

[0096] For each pair of standard symptoms in the standard symptom set, the degree of association is obtained by querying the association strength matrix, and a symptom association graph is constructed based on the degree of association; conflict relationships are obtained by querying the pathological conflict matrix, and when a conflict relationship exists, the corresponding symptom pair is marked as a logical contradiction point; a set of necessary accompanying symptoms is obtained by querying the necessary accompanying matrix, and necessary accompanying symptoms that are not in the standard symptom set are marked as information missing points;

[0097] Based on the symptom association diagram, the logical contradictions, and the missing information points, the standard symptom set is optimized to generate standardized symptom features that include symptom associations.

[0098] For example, a medical knowledge graph is constructed through the fusion of multi-source data, extracting structured knowledge from medical literature, clinical guidelines, and medical textbooks; integrating standardized medical terminology sets (such as SNOMEDCT, ICD-10, and MeSH) to obtain standardized symptom, disease, and drug ontology; utilizing natural language processing technology to extract entity relationships from medical texts; and combining manual review and supplementation of complex medical relationships by clinical experts. The construction process employs an iterative update mechanism, regularly incorporating new medical discoveries and clinical practice experience. Through entity linking and relationship mapping technologies, the consistency and integrity of knowledge are ensured, forming a structured knowledge network containing multi-dimensional medical entities such as symptoms, diseases, and drugs, and their complex relationships.

[0099] The medical knowledge graph includes a symptom ontology, a disease ontology, a drug ontology, and the relationships between these entities. The symptom ontology contains information such as standard symptoms, their synonyms, and hierarchical relationships. It first receives a comprehensive description of the patient's symptoms, such as "persistent headache for two weeks, accompanied by dizziness, nausea, and recently, loss of appetite and fatigue."

[0100] The comprehensive symptom descriptions are converted into a standard symptom set through a medical terminology recognition module. This module uses a medical terminology recognition algorithm, combined with symptom ontology matching, to standardize non-standard expressions. For example, "persistent headache" is mapped to the standard symptom "headache," while "lasting for two weeks" is extracted as the time attribute of the symptom; "dizziness," "nausea," "loss of appetite," and "fatigue" are mapped to their corresponding standard symptoms. After conversion, the standard symptom set S = {headache, dizziness, nausea, loss of appetite, fatigue} is obtained.

[0101] Three key matrices are constructed based on a medical knowledge graph. The association strength matrix R records the degree of medical association between standard symptoms. Each element R[i,j] in the matrix represents the association strength between symptom i and symptom j, with a value range of [0,1]. The association strength is calculated based on the symptom co-occurrence frequency, disease association relationships, and expert knowledge in the medical knowledge graph. For example, the association strength R[headache,dizziness] = 0.75 between "headache" and "dizziness" indicates a high correlation; the association strength R[headache,dizziness] = 0.25 between "headache" and "loss of appetite" indicates a weaker association.

[0102] The pathological conflict matrix C records the medical conflict relationships between standard symptoms. Each element C[i,j] in the matrix is ​​a Boolean value, where 1 indicates that there is a pathological conflict between symptom i and symptom j, and 0 indicates that there is no conflict. For example, the conflict relationship between "high fever" and "normal body temperature" is C[high fever, normal body temperature] = 1, indicating that the two do not occur at the same time; however, there is no pathological conflict between the five symptoms in this example, so the corresponding C values ​​are all 0.

[0103] The necessary co-occurrence matrix M records the necessary co-occurrence relationships between standard symptoms. Each element M[i,j] in the matrix is ​​a Boolean value, where 1 indicates that symptom j is a necessary co-occurrence symptom of symptom i, and 0 indicates that there is no necessary co-occurrence relationship. For example, the necessary co-occurrence relationship between "severe headache with sudden vomiting" and "neck stiffness" is M[severe headache with sudden vomiting, neck stiffness] = 1, indicating that when the former occurs, the latter should be checked for. In this example, "headache" and "nausea" are detected to coexist, and the "headache" lasts for a long time. The necessary co-occurrence matrix shows M[headache lasting for two weeks with nausea, blurred vision] = 1, indicating that the symptom of "blurred vision" exists but is not described by the patient.

[0104] Association analysis is performed on the standard symptom set. A symptom association graph G is constructed by querying the association strength matrix. Nodes in the association graph represent symptoms, edges represent the association relationships between symptoms, and edge weights represent the association strength. In this example, the association graph shows that "headache" has a high association strength with "dizziness" and "nausea," forming a closely related symptom cluster; "loss of appetite" has a high association strength with "fatigue," forming another symptom cluster; the association strength between the two symptom clusters is relatively low.

[0105] Logical contradictions are detected by querying the pathological conflict matrix. Each pair of symptoms (Si, Sj) in the symptom set is iterated over, and C[Si, Sj] is checked to see if it is 1. If a symptom pair with C[Si, Sj] = 1 exists, it is marked as a logical contradiction. No logical contradiction was detected in this example.

[0106] Information gaps are detected by querying the necessary adjoint matrix. For each symptom Si in the symptom set, find all symptoms Sj where M[Si,Sj] = 1 and Sj is not in the symptom set, and mark these Sj as information gaps. In this example, "blurred vision" is marked as an information gap.

[0107] Based on symptom correlation diagrams, logical contradictions, and missing information points, the standard symptom set is optimized. For missing information points, a prompt message is generated, suggesting that the doctor further inquire whether the patient has "blurred vision" as a symptom. If the patient confirms the presence of this symptom, it is added to the standard symptom set. If logical contradictions exist, the doctor is prompted to reconfirm conflicting symptom information.

[0108] Standardized features include: a set of standard symptoms, the relationships between symptoms (represented by the strength of the relationships), potential missing information points, and supplementary information such as the temporal attributes and severity of each symptom. In this example, standardized symptom features include: headache (lasting for two weeks), dizziness, nausea, loss of appetite, fatigue, the relationships between them, and "blurred vision" as a missing information point.

[0109] This invention describes the strength of associations, pathological conflicts, and necessary accompaniments between symptoms. By constructing and analyzing symptom association diagrams, logical contradictions and missing information in symptom descriptions can be quickly identified, and corresponding optimization suggestions can be provided. This knowledge-driven, standardized processing approach ensures the standardization of symptom features while fully utilizing professional medical knowledge for constraints and guidance, significantly improving the quality and usability of symptom features.

[0110] In one optional implementation, the standardized symptom features are input into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, and comprehensively evaluates the value of each candidate question to generate the optimal consultation content for the next round. The steps include:

[0111] Standardized symptom features are encoded using a medical pre-trained language model to obtain symptom semantic vectors. The temporal position encoding of the standardized symptom features in the course of disease is extracted, and the professional knowledge feature vectors corresponding to the standardized symptom features are obtained. The professional knowledge feature vectors include pathological conflict degree, progression abnormality degree, information completeness, and medical importance weights. The symptom semantic vectors, the temporal position encodings, and the professional knowledge feature vectors are input into a deep neural network to obtain symptom feature representations.

[0112] The deep neural network constructs a state vector based on the symptom feature representation, applies a multi-head attention mechanism to the historical dialogue sequence to obtain a dialogue context vector, and fuses the state vector and the dialogue context vector to obtain the current diagnostic state; based on the current diagnostic state, it calculates the expected information gain of each candidate consultation direction in the candidate consultation direction set;

[0113] The user's historical feedback information is input into the feedback encoding layer to obtain the feedback feature vector. The feedback feature vector is combined with the expected information gain to obtain the consultation evaluation vector. The deep neural network optimizes and selects the candidate consultation direction set based on the consultation evaluation vector to generate the optimal consultation content.

[0114] For example, standardized symptom features are encoded into symptom semantic vectors using a medical pre-trained language model. Specifically, a medical domain pre-trained language model based on the Transformer architecture (such as Med-BERT or Clinical-BERT) is employed. This model is first pre-trained on a general corpus through masked language modeling and next-sentence prediction tasks, and then fine-tuned for domain adaptation using Chinese and English medical literature, clinical electronic medical records, and mainstream medical textbooks. The fine-tuning process employs a contrastive learning strategy to cluster representations of similar symptoms in the vector space. Upon inputting standardized symptom features (e.g., "persistent headache, accompanied by nausea"), the model generates context-sensitive representations for each lexical unit, which are then fused through an attention pooling mechanism to obtain a 768-dimensional vector representation of the overall symptom. This vector accurately captures the medical semantic information and potential correlations of the symptoms.

[0115] Standardized symptom features are extracted and encoded temporally within the disease progression. This temporal encoding represents the order in which symptoms appeared and their duration. For example, a symptom description like "headache started 3 days ago, fever appeared 2 days ago" is converted into a relative time code. Specifically, the temporal code for headache is [-3,0], representing the period from 3 days ago to the present; the temporal code for fever is [-2,0], representing the period from 2 days ago to the present. The temporal codes are represented by a 32-dimensional vector, containing information such as symptom start time, end time, and duration.

[0116] Obtain the professional knowledge feature vector corresponding to the standardized symptom features. The professional knowledge feature vector includes four dimensions: pathological conflict degree, progression anomaly degree, information completeness, and medical importance weight. Pathological conflict degree indicates the degree of conflict between the current symptom and other known symptoms in medical theory, with a value ranging from 0 to 1. For example, the pathological conflict degree between "hypertension" and "hypotension" is 0.95. Progression anomaly degree indicates the degree of deviation from the typical disease course during symptom development, with a value ranging from 0 to 1. For example, the progression anomaly degree for "slow wound healing in diabetic patients" is 0.2, while the progression anomaly degree for "rapid deterioration of a diabetic patient's wound within a week" is 0.8. Information completeness indicates the completeness of the symptom description, with a value ranging from 0 to 1. For example, the information completeness for describing only "headache" is 0.3, while the information completeness for describing "throbbing pain in the right temple, with a pain level of 7 out of 10, accompanied by photophobia and nausea" is 0.9. Medical importance weights represent the importance of symptoms in the diagnostic process, with values ​​ranging from 0 to 1. For example, in the diagnosis of heart disease, the importance weight of "chest pain" is 0.9, while the importance weight of "mild fatigue" is 0.2.

[0117] The symptom semantic vector, temporal location encoding, and professional knowledge feature vector are input into a deep neural network to obtain a symptom feature representation. In practical applications, the symptom semantic vector (768-dimensional), temporal location encoding (32-dimensional), and professional knowledge feature vector (4-dimensional) are first mapped to the same dimension (256-dimensional) through their respective linear projection layers, and then concatenated to obtain a comprehensive feature vector (768-dimensional). This comprehensive feature vector is then passed through two fully connected layers (768-dimensional to 512-dimensional, and 512-dimensional to 256-dimensional) and the ReLU activation function to finally output a 256-dimensional symptom feature representation.

[0118] The state vector integrates all known symptom information to represent the current diagnostic state. Specifically, an attention pooling mechanism is applied to all symptom feature representations to generate a fixed-dimensional (256-dimensional) state vector. This attention mechanism calculates the importance weight of each symptom; for example, for a set of respiratory symptoms, "difficulty breathing" receives a weight of 0.4, "cough" receives a weight of 0.3, and "low-grade fever" receives a weight of 0.1. The symptom features are then weighted and summed based on these weights to obtain the state vector reflecting the current diagnostic state.

[0119] A multi-head attention mechanism is applied to the historical dialogue sequence to obtain the dialogue context vector. The historical dialogue includes the user's input of symptom descriptions and previous consultation content. Each round of dialogue is encoded as a vector, and these vector sequences are processed using an 8-head attention mechanism, with each attention head having a dimension of 32, resulting in a 256-dimensional dialogue context vector. This vector captures key information and contextual dependencies in the dialogue history.

[0120] The current diagnostic state is obtained by fusing the state vector with the dialogue context vector. In the specific implementation, the two 256-dimensional vectors are fused through a gating mechanism. The importance weights of each dimension of the two vectors are calculated, and then a weighted combination is performed to obtain a 256-dimensional fused vector, which represents the current diagnostic state.

[0121] Calculate the expected information gain for each candidate consultation direction in the candidate consultation direction set based on the current diagnostic state. Maintain a set of 100 common consultation directions, such as "inquiry about digestive symptoms" and "inquiry about cardiovascular symptoms." For each candidate direction, calculate its expected information gain, representing the amount of new information that direction can provide for the diagnosis. During the calculation, the current diagnostic state vector and the vector representation of each candidate consultation direction are interactively calculated to obtain the corresponding expected information gain value.

[0122] User historical feedback information is input into the feedback encoding layer to obtain a feedback feature vector. User feedback includes the completeness and consistency of the user's answers to previous consultation questions. This feedback information is encoded into a 64-dimensional feedback feature vector, reflecting the user's level of cooperation and the quality of information provided.

[0123] The expected information gain value for each candidate consultation direction is adjusted, and the original information gain value is combined with the feedback feature vector through a fully connected network to generate an adjusted evaluation score. For example, if a user's answers to cardiovascular-related questions are incomplete or inconsistent, the evaluation score for the cardiovascular consultation direction will be lowered.

[0124] The deep neural network optimizes and selects the optimal consultation content based on the consultation evaluation vector from the set of candidate consultation directions. It selects the consultation direction with the highest evaluation score and chooses a specific question from a pre-defined question template library. For example, if the "Digestive Symptoms Inquiry" direction receives the highest evaluation score of 0.92, the specific question under that direction, "Please describe whether you have abdominal pain? If so, please describe the specific location, nature (e.g., cramping, dull pain), and whether it changes with eating, bowel movements, etc." as the content for the next round of consultation to obtain more valuable diagnostic information.

[0125] Figure 2This paper compares the performance improvements of different semantic models on intelligent consultation systems. The horizontal axis represents three evaluation scenarios: common disease diagnosis, complex symptom analysis, and disease identification efficiency; the vertical axis represents diagnostic accuracy (0-1.0). The figure compares the performance of three models: the general Transformer model (white bars), the Med-BERT medical pre-trained model (slanted bars), and the multimodal fusion intelligent consultation model of this invention (grid bars). The results show that the model of this invention achieves the best performance in all scenarios, especially in disease identification efficiency, reaching an accuracy of 0.85, which is significantly better than the general model (0.4) and the medical pre-trained model (0.6), fully demonstrating the significant contribution of the multimodal fusion method to improving the performance of intelligent consultation systems.

[0126] This invention achieves intelligent optimization of the consultation process through deep neural networks, fusing symptom semantics, temporal location, and professional knowledge features in a multimodal manner to construct a rich state representation. By combining multi-head attention mechanisms and dialogue context analysis, it accurately grasps the dynamic characteristics of the consultation process. Through comprehensive evaluation of the expected information gain of candidate consultation directions and optimization selection based on user feedback, it achieves adaptive adjustment of the consultation strategy, improving both consultation efficiency and quality.

[0127] In one optional implementation, the step of calculating the expected information gain for each candidate consultation direction in the candidate consultation direction set based on the current diagnostic state includes:

[0128] For each candidate consultation direction in the candidate consultation direction set, a consultation feature vector is constructed. The consultation feature vector includes a symptom correlation vector, a temporal feature vector, and a severity vector. The consultation feature vector is input into a variational autoencoder for state transition prediction. The consultation feature is compressed into a low-dimensional latent representation through an encoding network. The latent representation is randomly sampled and the state transition result is reconstructed through a decoding network. Based on the multiple sampling and reconstruction results, a state transition probability distribution under different feedback types is generated.

[0129] Based on the state transition probability distribution, the diagnostic convergence gain, symptom confirmation gain, and risk identification gain for each candidate consultation direction are calculated. The diagnostic convergence gain is calculated by the difference between the conditional entropy of the disease distribution in the current state and the conditional entropy of the disease distribution in the predicted state. The symptom confirmation gain is calculated by weighting the symptom confirmation probability and feedback contribution obtained by analyzing historical feedback patterns through a recurrent neural network. The risk identification gain is calculated by multiplying the clinical risk level and symptom urgency level assessed by the medical knowledge graph.

[0130] Based on Bayesian decision theory, the certainty of the current diagnosis, the need for verification, and the urgency are calculated. The diagnostic convergence gain, symptom confirmation gain, and risk identification gain are dynamically weighted and combined to obtain the expected information gain of the candidate consultation direction.

[0131] For example, a consultation feature vector is constructed for each candidate consultation direction in the candidate consultation direction set. The consultation feature vector consists of three parts: a symptom correlation vector, a temporal feature vector, and a severity vector. The symptom correlation vector represents the degree of association between the candidate consultation direction and the known symptoms. For example, for the candidate consultation direction of "headache", if the patient has reported "fever", the correlation between the two is 0.35. The temporal feature vector captures the time sequence information of the symptoms. For example, the feature value of "headache" appearing before "fever" is 0.7. The severity vector represents the severity level of the symptoms. For example, the severity of "mild headache" is 0.2, while that of "severe headache" is 0.8.

[0132] After constructing the consultation feature vector, it is input into a variational autoencoder for state transition prediction. The variational autoencoder consists of an encoder network and a decoder network. The encoder network consists of a three-layer fully connected neural network. The number of nodes in the input layer is consistent with the dimension of the consultation feature vector, the number of nodes in the intermediate layer is 64, and the output layer generates a 16-dimensional mean vector and standard deviation vector to represent the low-dimensional latent space. Reparameterization is used to randomly sample from the latent space to generate a latent representation. For example, for the candidate consultation direction "chest pain", the generated latent representation is [-0.21, 0.35, 0.02, ..., 0.18].

[0133] The decoding network also consists of three fully connected layers, which reduce the latent representation to state transition results. Multiple samplings (typically 50 times) are performed on the same consultation feature vector, generating multiple state transition results through the decoding network. Based on these results, the state transition probability distribution under different feedback types is calculated. For example, for the candidate consultation direction "dyspnea," the obtained state transition probability distribution is: positive feedback probability 0.45, negative feedback probability 0.30, and partially positive feedback probability 0.25.

[0134] After obtaining the state transition probability distribution, three gains are calculated for each candidate consultation direction: diagnostic convergence gain, symptom confirmation gain, and risk identification gain. When calculating the diagnostic convergence gain, the conditional entropy of the disease distribution in the current state is first calculated, i.e., the degree of uncertainty in the current diagnosis. First, based on the currently known set of symptoms, the posterior probability of each disease is calculated. For example, when the known symptoms are "fever 39℃" and "cough," the posterior probability of "pneumonia" is 0.45, "influenza" is 0.35, "bronchitis" is 0.15, and other diseases are 0.05. The conditional entropy calculation process is as follows: take the logarithm of the posterior probability of each disease, multiply it by its probability value, then sum over all diseases and take the negative value. Using the above example, the conditional entropy is 1.38.

[0135] Based on the state transition probability distribution, the conditional entropy of the disease distribution in the predicted state is calculated. The calculation of the conditional entropy in the predicted state needs to consider various feedback scenarios after the consultation. Taking the question "Are you having difficulty breathing?" as an example, there are three feedback options: affirmative ("Yes, I am having difficulty breathing"), negative ("I am not having difficulty breathing"), or partially affirmative ("I feel a little unwell occasionally"). For each feedback scenario, the updated conditional entropy of the disease distribution is calculated. Assume the conditional entropy drops to 0.85 with affirmative feedback, 1.26 with negative feedback, and 1.02 with partially affirmative feedback. Based on the previously predicted feedback probability distribution (0.45, 0.30, 0.25), the weighted average expected conditional entropy is calculated as 0.45 × 0.85 + 0.30 × 1.26 + 0.25 × 1.02 = 1.01. Therefore, the diagnostic convergence gain for this consultation direction is the difference between the current conditional entropy and the expected conditional entropy: 1.38 - 1.01 = 0.37.

[0136] The calculation of symptom confirmation gain relies on the analysis of historical feedback patterns by a recurrent neural network (RNN). A GRU (Gated Recurrent Unit) structure is used, with the input being the historical consultation sequence and its feedback results, and the output being the symptom confirmation probability for the current candidate consultation direction. The RNN analyzes historical feedback patterns using a bidirectional GRU structure, with the input layer receiving the encoding of historical consultation content and the corresponding feedback type. For example, the historical consultation "Is there phlegm in my cough?" and the feedback "There is a small amount of white phlegm" can be encoded as a combination of consultation and feedback vectors. The GRU hidden layer dimension is set to 128, capturing the temporal dependencies in the sequence through forward and backward propagation. Finally, the symptom confirmation probability for the candidate consultation direction is output through a fully connected layer and a softmax activation function.

[0137] The symptom confirmation gain is calculated by weighting the symptom confirmation probability and the feedback contribution. The feedback contribution is calculated using value scores from similar consultation scenarios in historical data. Historical cases similar to the current candidate consultation direction are retrieved from the consultation database, and their feedback contribution scores to the final diagnosis are extracted and normalized. For example, feedback on consultations regarding "abdominal pain" typically contributes more to the diagnosis of "appendicitis" (score 0.85) than to the diagnosis of "gastritis" (score 0.40). For instance, for the candidate consultation direction of "abdominal pain," with a symptom confirmation probability of 0.65 and a feedback contribution of 0.8, the symptom confirmation gain can be calculated as 0.52.

[0138] Risk identification gain is calculated using a medical knowledge graph. The knowledge graph retrieves the clinical risk level associated with candidate consultation directions, while simultaneously assessing the urgency of the symptoms. The risk identification gain is the product of these two factors. For example, if the clinical risk level of "sudden, severe chest pain" is 0.9 and the urgency level is 0.95, then its risk identification gain is 0.855.

[0139] Based on Bayesian decision theory, three key indicators for the current diagnosis are calculated: certainty, validation need, and urgency. Certainty reflects the reliability of the current diagnosis, ranging from 0 to 1; for example, the certainty of a preliminary diagnosis of "pneumonia" is 0.72. Validation need indicates the necessity of confirming key symptoms; for example, the validation need for a diagnosis of "myocardial infarction" is 0.85. Urgency indicates the time sensitivity of the diagnosis; for example, the urgency for a diagnosis of "stroke" is 0.93.

[0140] The diagnostic convergence gain, symptom confirmation gain, and risk identification gain are dynamically weighted and combined based on these three indicators. This dynamic weighting process is implemented using a soft maximum function. First, scaling parameters are applied to the degree of certainty, validation requirement, and urgency. For example, applying scaling parameters [2.0, 1.5, 1.8] to [0.3, 0.6, 0.8] yields [0.6, 0.9, 1.44]. Then, the exponent values ​​[1.82, 2.46, 4.22] are calculated and normalized to obtain weights [0.21, 0.29, 0.50]. Finally, these weights are multiplied by the diagnostic convergence gain, symptom confirmation gain, and risk identification gain respectively, and summed to obtain the overall expected information gain. When the degree of certainty is low, the weight of the diagnostic convergence gain is increased; when the validation requirement is high, the weight of the symptom confirmation gain is increased; and when the urgency is high, the weight of the risk identification gain is increased. For example, for a certain diagnostic state, the three indicators are [0.3, 0.6, 0.8], and the generated weights are [0.5, 0.2, 0.3]. If the three gains for a candidate consultation direction are [0.6, 0.4, 0.7], then its final expected information gain is 0.5×0.6+0.2×0.4+0.3×0.7=0.59.

[0141] This invention employs a variational autoencoder for state transition prediction, effectively simulating the results of different consultation feedback through random sampling and reconstruction mechanisms. Based on this, a multi-dimensional consultation value evaluation system is established, comprehensively considering diagnostic convergence gain, symptom confirmation gain, and risk identification gain. Dynamic weight adjustment using Bayesian decision theory ensures that the selection of consultation direction considers both diagnostic efficiency and medical safety, achieving intelligent optimization of the consultation strategy.

[0142] In one optional implementation, the standardized symptom features are updated based on user feedback on the consultation content, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold. The step of outputting the diagnosis result includes:

[0143] Extract symptom entities, attributes, and relationships from user feedback to construct incremental symptom feature data; merge and update the incremental symptom feature data with the original standardized symptom features, and perform consistency verification using medical knowledge;

[0144] Based on the updated standardized symptom features, calculate the disease diagnosis probability distribution and its information entropy; when the information entropy is greater than a preset threshold, return to the consultation optimization process, and when it is less than or equal to the preset threshold, enter the diagnosis result generation process.

[0145] A symptom-disease association matrix is ​​constructed. The diagnostic contribution of each symptom is calculated based on its discriminative power, frequency of occurrence, and clinical relevance in disease diagnosis. Symptoms with a diagnostic contribution greater than a first preset threshold are identified as diagnostic criteria. Based on symptom combination rules in a medical knowledge graph, clinically relevant symptom combinations are identified, and the disease diagnosis probability corresponding to the symptom combination is calculated.

[0146] Candidate diseases are ranked according to the disease diagnosis probability of the symptom combinations, and the disease with the highest diagnosis probability is selected as the diagnosis conclusion. The symptom contribution of the diagnostic basis is normalized to obtain the symptom weight. The conditional probability distribution of the diagnosis conclusion is weighted and summed based on the symptom weight, and then corrected by combining the clinical relevance of the symptom combinations to obtain the confidence level. The diagnosis conclusion, the diagnostic basis, the clinical relevance of the symptom combinations, and the corresponding confidence level are organized to form a diagnosis report.

[0147] For example, the system receives user feedback on their online consultations and extracts symptom entities, attributes, and relationships from the feedback using natural language processing (NLP). For instance, when a user reports "headache lasting 3 days, throbbing pain, accompanied by mild nausea," "headache" is identified as the symptom entity, "lasting 3 days" and "throbbing" as attributes, and "accompanied by mild nausea" indicates a co-occurrence relationship between "headache" and "nausea." A named entity recognition (NAME) model is used to process the user's text. This model, based on a deep learning architecture, combines a medical dictionary and a rule engine to accurately identify symptom entities. Attribute extraction is achieved through keyword matching and dependency parsing, while relationship extraction is based on co-occurrence rules and contextual semantic analysis. The extracted results constitute incremental symptom feature data, including symptom names, attribute values, and relationship triples.

[0148] The incremental symptom feature data is fused and updated with the original standardized symptom features. Each symptom entity is checked to see if it already exists in the original feature set; if it does, its attribute values ​​and relationships are updated; otherwise, it is added as a new symptom. For example, if the original feature set contains "headache (severity: moderate)" and the incremental data contains "headache (nature: pulsating)", the fused data becomes "headache (severity: moderate, nature: pulsating)". During the update process, consistency checks are performed using a medical knowledge base to detect attribute value conflicts or logical contradictions in relationships. For example, if "fever (temperature: 38.5℃)" and "normal body temperature" contradict each other, the more reliable information is retained based on time-series information or credibility, and the conflict resolution process is recorded.

[0149] Based on the updated standardized symptom features, a disease diagnosis probability distribution is calculated. A symptom-disease joint probability table is maintained, recording the probability of each symptom occurring in each disease. For the user's symptom set, the posterior probability of each candidate disease is calculated. For example, for the symptom set {headache (throbbing), nausea, photophobia}, the diagnosis distribution is: migraine (0.65), meningitis (0.25), intracranial hypertension (0.08), and others (0.02). Based on this probability distribution, information entropy is calculated; the lower the information entropy, the more certain the diagnosis. If the calculated information entropy value is 0.98, which is greater than the preset threshold of 0.8, the process returns to the consultation optimization process to continue asking questions; if it is less than or equal to the threshold, the process proceeds to the diagnosis result generation process.

[0150] In the diagnostic result generation process, a symptom-disease association matrix is ​​constructed, where the matrix element values ​​represent the degree to which a symptom contributes to the disease diagnosis. The diagnostic contribution of each symptom is evaluated from three dimensions: discriminative power, frequency of occurrence, and clinical relevance. Discriminative power is assessed by calculating the difference in the probability of a symptom occurring in the target disease compared to other diseases; frequency of occurrence is based on the incidence of the symptom in the target disease; and clinical relevance is based on the importance weight of the symptom in medical guidelines. These three indicators are combined to calculate the diagnostic contribution for each symptom. For example, for the diagnosis of migraine, "throbbing headache" has a contribution of 0.85, "nausea" 0.65, and "photophobia" 0.72. Symptoms with a contribution greater than a preset threshold of 0.6 are identified as diagnostic criteria.

[0151] This approach uses a medical knowledge graph to identify clinically relevant symptom combinations. The knowledge graph stores typical symptom combination patterns for diseases; for example, a typical combination for migraines is "throbbing headache + nausea / vomiting + photophobia / phonophobia." The system checks if these combinations exist in a user's symptom set; if so, it increases the probability of diagnosing the corresponding disease. The diagnostic probability of a disease corresponding to a symptom combination is calculated using symptom combination rules. For example, the symptom combination "throbbing headache + nausea + photophobia" has a diagnostic probability of 0.88 for migraines.

[0152] Candidate diseases are ranked according to the diagnostic probability of symptom combinations, and the disease with the highest diagnostic probability is selected as the diagnosis. In the example above, migraine (0.88) ranks first and is selected as the diagnosis. The symptom contribution of the diagnostic criteria is normalized to obtain symptom weights: throbbing headache (0.38), nausea (0.29), and photophobia (0.33). Based on these weights, the conditional probability distribution of the diagnosis is weighted and summed, and then adjusted according to the clinical relevance of the symptom combinations, resulting in a diagnostic confidence of 0.92.

[0153] Finally, the diagnostic conclusion (migraine), diagnostic criteria (throbbing headache, nausea, photophobia), clinical relevance of the symptom combination (classic migraine triad), and corresponding confidence level (0.92) are organized into a structured diagnostic report. The report also includes symptom weight distribution, differential diagnosis, and recommended further investigations, providing users and doctors with comprehensive diagnostic reference information.

[0154] This invention constructs a complete closed-loop diagnostic mechanism, achieving dynamic adjustment of the diagnostic process through continuous symptom feature updates and diagnostic probability optimization. In the diagnostic result generation stage, a reliable diagnostic basis system is established through symptom contribution analysis and clinical relevance assessment. Simultaneously, symptom combination rules and conditional probability weighting are introduced to quantify the reliability of diagnostic conclusions. The overall solution ensures both the rigor of the diagnostic process and provides clear diagnostic interpretations, significantly improving the reliability and interpretability of intelligent diagnosis.

[0155] A second aspect of this invention provides a deep learning-based intelligent medical multi-turn dialogue diagnostic reasoning system, comprising:

[0156] The first unit is used to store the symptom descriptions input by the user in chronological order, calculate the correlation weight of the correlation between the symptom descriptions in each round using an attention mechanism, and construct a chronological feature containing the symptom evolution process by fusing the current symptom description with historical dialogue information based on the correlation weight. The chronological feature is then used to perform bidirectional sequence modeling to identify the logical relationship between the symptom descriptions before and after, and generate a comprehensive symptom feature that takes into account the historical evolution.

[0157] The second unit is used to perform correlation analysis on the comprehensive symptom features based on the medical knowledge graph, detect medical professional logical contradictions and information gaps in the description, and generate standardized symptom features.

[0158] The third unit is used to input the standardized symptom features into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round. Based on the user's feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold, and the diagnosis result is output.

[0159] A third aspect of the present invention provides an electronic device, comprising:

[0160] processor;

[0161] Memory used to store processor-executable instructions;

[0162] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0163] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0164] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based intelligent medical multi-turn dialogue diagnostic reasoning method, characterized in that, include: The symptom descriptions input by the user are stored in chronological order. An attention mechanism is used to calculate the correlation weight of the symptom descriptions in each round. Based on the correlation weight, the current symptom description is fused with historical dialogue information to construct a chronological feature that includes the symptom evolution process. The chronological feature is then bidirectionally sequence-modeled to identify the logical relationship between the symptom descriptions and generate a comprehensive symptom feature that takes into account historical evolution. Based on a medical knowledge graph, the comprehensive symptom features are analyzed for correlation, and contradictions in medical professional logic and missing information in the description are detected to generate standardized symptom features. The standardized symptom features are input into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round. Based on user feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold, and the diagnosis result is output.

2. The method according to claim 1, characterized in that, The steps include: storing user-inputted symptom descriptions chronologically; calculating association weights based on the correlation between symptom descriptions from different rounds using an attention mechanism; fusing current symptom descriptions with historical dialogue information based on these association weights to construct chronological features that include the symptom evolution process; performing bidirectional sequence modeling on these chronological features to identify the logical relationships between symptom descriptions; and generating comprehensive symptom features that consider historical evolution. The symptom descriptions input by the user in each round are segmented into word sequences, and these word sequences are used to construct a dialogue history matrix. Based on the dialogue history matrix, the word vector matrix of the current round and the word vector matrix of the previous round are extracted. A word-level attention score is calculated for each word in the current round and all words in the previous round. A round representation is obtained based on the word-level attention score, and the inter-round association weights are calculated based on the round representations. The inter-round association weights are obtained through a non-linear transformation between the hidden states of the current round and the hidden states of the previous rounds. The word-level attention score and the inter-round association weight are weighted and combined to obtain the temporal fusion feature. The temporal fusion feature is processed forward and backward through a bidirectional gated loop unit to obtain the temporal feature. The internal logical structure of the symptom description is identified by analyzing the forward and backward dependencies of the symptom description. Calculate the semantic similarity of temporal features between adjacent rounds. If the semantic similarity is lower than a preset threshold, correct the temporal features of the current round based on the temporal features of the historical rounds. Calculate the difference vector of temporal features between adjacent rounds. The corrected temporal features, difference vectors, and internal logical structure features identified through bidirectional sequence modeling are fused to obtain comprehensive symptom features.

3. The method according to claim 2, characterized in that, The temporal fusion features are processed forward and backward using a bidirectional gated loop unit to obtain temporal features. The steps of identifying the internal logical structure of symptom descriptions by analyzing the forward and backward dependencies of symptom descriptions include: The temporal fusion features are input into a bidirectional gated recurrent unit to obtain the forward hidden state sequence and the backward hidden state sequence; The temporal window size is calculated based on the forward and backward hidden state sequences. The temporal mean and standard deviation are calculated based on the hidden state subsequences corresponding to the temporal window size. Fine-grained dependency strength is calculated based on the temporal mean and standard deviation. Coarse-grained dependency strength is obtained by multiplying the original hidden states. Adaptive weights are calculated based on the temporal position difference and the hidden states. The fine-grained dependency strength and the coarse-grained dependency strength are weighted and combined to obtain multi-granularity dependency features. The multi-granularity dependency features and hidden state sequences are input into a recursive logic decomposition network, which constructs logical relationship representations layer by layer through a multi-layer cascaded structure. The logical relationship representations are input into a generator to generate candidate logical relationships, and the logical rationality is verified by a discriminator to obtain optimized logical relationships. The optimized logical relationships are combined with the bidirectional hidden state sequences to obtain the symptom logical structure representation.

4. The method according to claim 1, characterized in that, The steps for generating standardized symptom features by performing correlation analysis on the comprehensive symptom features based on a medical knowledge graph, detecting medical logical contradictions and missing information in the description, include: Transform the comprehensive symptom characteristics into a standard symptom set; Based on the symptom ontology and symptom relationships in the medical knowledge graph, an association strength matrix, a pathological conflict matrix, and a necessary co-occurrence matrix are constructed. The association strength matrix records the degree of medical association between standard symptoms, the pathological conflict matrix records the medical conflict relationship between standard symptoms, and the necessary co-occurrence matrix records the necessary co-occurrence relationship between standard symptoms. For each pair of standard symptoms in the standard symptom set, the degree of association is obtained by querying the association strength matrix, and a symptom association graph is constructed based on the degree of association; conflict relationships are obtained by querying the pathological conflict matrix, and when a conflict relationship exists, the corresponding symptom pair is marked as a logical contradiction point; a set of necessary accompanying symptoms is obtained by querying the necessary accompanying matrix, and necessary accompanying symptoms that are not in the standard symptom set are marked as information missing points; Based on the symptom association diagram, the logical contradictions, and the missing information points, the standard symptom set is optimized to generate standardized symptom features that include symptom associations.

5. The method according to claim 1, characterized in that, The steps of inputting the standardized symptom features into a deep neural network, which calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round include: Standardized symptom features are encoded into symptom semantic vectors using a medical pre-trained language model. The temporal position codes of the standardized symptom features in the course of disease are extracted, and the professional knowledge feature vectors corresponding to the standardized symptom features are obtained. The symptom semantic vectors, the temporal position codes, and the professional knowledge feature vectors are input into a deep neural network to obtain symptom feature representations. The deep neural network constructs a state vector based on the symptom feature representation, applies a multi-head attention mechanism to the historical dialogue sequence to obtain a dialogue context vector, and fuses the state vector and the dialogue context vector to obtain the current diagnostic state; based on the current diagnostic state, it calculates the expected information gain of each candidate consultation direction in the candidate consultation direction set; The user's historical feedback information is input into the feedback encoding layer to obtain the feedback feature vector. The feedback feature vector is combined with the expected information gain to obtain the consultation evaluation vector. The deep neural network optimizes and selects the candidate consultation direction set based on the consultation evaluation vector to generate the optimal consultation content.

6. The method according to claim 5, characterized in that, The steps for calculating the expected information gain of each candidate consultation direction in the candidate consultation direction set based on the current diagnostic status include: For each candidate consultation direction in the candidate consultation direction set, a consultation feature vector is constructed. The consultation feature vector includes a symptom correlation vector, a temporal feature vector, and a severity vector. The consultation feature vector is input into a variational autoencoder for state transition prediction. The consultation feature is compressed into a low-dimensional latent representation through an encoding network. The latent representation is randomly sampled and the state transition result is reconstructed through a decoding network. Based on the multiple sampling and reconstruction results, a state transition probability distribution under different feedback types is generated. Based on the state transition probability distribution, the diagnostic convergence gain, symptom confirmation gain, and risk identification gain for each candidate consultation direction are calculated. The diagnostic convergence gain is calculated by the difference between the conditional entropy of the disease distribution in the current state and the conditional entropy of the disease distribution in the predicted state. The symptom confirmation gain is calculated by weighting the symptom confirmation probability and feedback contribution obtained by analyzing historical feedback patterns through a recurrent neural network. The risk identification gain is calculated by multiplying the clinical risk level and symptom urgency level assessed by the medical knowledge graph. Based on Bayesian decision theory, the certainty of the current diagnosis, the need for verification, and the urgency are calculated. The diagnostic convergence gain, symptom confirmation gain, and risk identification gain are dynamically weighted and combined to obtain the expected information gain of the candidate consultation direction.

7. The method according to claim 1, characterized in that, Based on user feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold. The steps for outputting the diagnosis result include: Extract symptom entities, attributes, and relationships from user feedback to construct incremental symptom feature data; merge and update the incremental symptom feature data with the original standardized symptom features, and perform consistency verification using medical knowledge; Based on the updated standardized symptom features, calculate the disease diagnosis probability distribution and its information entropy; when the information entropy is greater than a preset threshold, return to the consultation optimization process, and when it is less than or equal to the preset threshold, enter the diagnosis result generation process. A symptom-disease association matrix is ​​constructed. The diagnostic contribution of each symptom is calculated based on its discriminative power, frequency of occurrence, and clinical relevance in disease diagnosis. Symptoms with a diagnostic contribution greater than a first preset threshold are identified as diagnostic criteria. Based on symptom combination rules in a medical knowledge graph, clinically relevant symptom combinations are identified, and the disease diagnosis probability corresponding to the symptom combination is calculated. Candidate diseases are ranked according to the disease diagnosis probability of the symptom combinations, and the disease with the highest diagnosis probability is selected as the diagnosis conclusion. The symptom contribution of the diagnostic basis is normalized to obtain the symptom weight. The conditional probability distribution of the diagnosis conclusion is weighted and summed based on the symptom weight, and then corrected by combining the clinical relevance of the symptom combinations to obtain the confidence level. The diagnosis conclusion, the diagnostic basis, the clinical relevance of the symptom combinations, and the corresponding confidence level are organized to form a diagnosis report.

8. A deep learning-based intelligent medical multi-turn dialogue diagnostic reasoning system, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to store the symptom descriptions input by the user in chronological order, calculate the correlation weight of the correlation between the symptom descriptions in each round using an attention mechanism, and construct a chronological feature containing the symptom evolution process by fusing the current symptom description with historical dialogue information based on the correlation weight. The chronological feature is then used to perform bidirectional sequence modeling to identify the logical relationship between the symptom descriptions before and after, and generate a comprehensive symptom feature that takes into account the historical evolution. The second unit is used to perform correlation analysis on the comprehensive symptom features based on the medical knowledge graph, detect medical professional logical contradictions and information gaps in the description, and generate standardized symptom features. The third unit is used to input the standardized symptom features into a deep neural network. The deep neural network calculates the expected information gain for different consultation directions, combines user feedback and dialogue history, comprehensively evaluates the value of each candidate question, and generates the optimal consultation content for the next round. Based on the user's feedback on the consultation content, the standardized symptom features are updated, and the consultation optimization process is repeated until the information entropy of the disease diagnosis distribution corresponding to the standardized symptom features is lower than a preset threshold, and the diagnosis result is output.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images