Medical record data mining and potential risk prediction method and system based on deep learning

Abstract

The invention discloses a medical record data mining and potential risk prediction method and system based on deep learning, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining a first semantic vector set containing the current symptom multi-role chief complaint information of a target child, multi-modal physiological data, basic treatment information and initial cognitive level model parameters, a conflict vector set is generated through conflict recognition, initial cognitive level model parameters are adjusted in combination with multi-modal physiological data, individual cognitive level model parameters are obtained, and then a questioning text set is generated according to the conflict vector set and the individual cognitive level model parameters; after feedback is received, conflict resolution is carried out on the first semantic vector set to obtain a second semantic vector set, the second semantic vector set is fused with basic doctor seeing information to generate medical record combination feature data, and a potential disease risk prediction result of the target child is generated according to the medical record combination feature data; the method has the beneficial effects that the multi-role chief complaint information conflict in the child medical record can be effectively processed, and the potential disease risk prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a medical record data mining and potential risk prediction method and system based on deep learning. BACKGROUND

[0002] Under the background of intelligent development of medical data, medical record data mining technology, as an important means of medical auxiliary decision-making, has gradually become a core supporting tool in disease prediction, risk assessment and diagnosis and treatment path optimization, etc. Through deep learning modeling of structured data and unstructured data in electronic medical records, the patient's condition can be analyzed comprehensively and potential risks can be identified in advance.

[0003] In the prior art, various medical record data mining methods are often based on the direct description of the patient himself to develop modeling. In particular, in the handling of chief complaints, through entity recognition, keyword extraction, symptom classification, etc., subjective symptoms are converted into structured features for model training. Such methods have good adaptability in adult groups, can effectively capture the correlation between chief complaints and disease risks, and improve the prediction accuracy and model generalization ability. However, for children patients, especially for younger children, their language expression ability has not yet developed maturely, and they usually cannot accurately describe their own symptoms. The source of the chief complaint content in their medical records often involves multiple roles, such as the child himself, the parents and the medical staff. There may be semantic conflicts or expression biases between their expressions, and there are often a large number of descriptive sentences, analogical expressions or time ambiguity sentences. These non-standardized expressions make it difficult for conventional medical record data mining methods to identify their true medical meaning, which may lead to misunderstanding of chief complaint semantics or error in symptom classification, and may cause prediction bias, disease misjudgment or risk underestimation, and even delay treatment opportunities or increase the rate of misdiagnosis.

[0004] Therefore, a medical record data mining and potential risk prediction method and system based on deep learning are proposed. SUMMARY

[0005] In view of the above prior art situation, the present application is proposed. The embodiments of the present application provide a medical record data mining and potential risk prediction method and system based on deep learning, which can effectively handle the conflict of multi-role chief complaint information in children's medical records, generate questioning text conforming to the cognitive level of children and improve the prediction accuracy of potential disease risks.

[0006] According to an aspect of the present application, a medical record data mining and potential risk prediction method based on deep learning is provided, comprising: obtaining a first semantic vector set, multi-modal physiological data, basic consultation information and initial cognitive level model parameters of a target child, the first semantic vector set containing multi-role complaint information around the current symptoms of the target child; performing conflict recognition on the first semantic vector set to generate a conflict vector set with semantic conflicts; adjusting the initial cognitive level model parameters according to the conflict vector set and the multi-modal physiological data through a preset weight adjustment algorithm based on causal reasoning and reinforcement learning to obtain individual cognitive level model parameters; generating a set of follow-up texts in line with the cognitive level of the target child according to the conflict vector set and the individual cognitive level model parameters through a deep learning-based strategy generation network; receiving feedback of the target child on each follow-up text in the set of follow-up texts to obtain a set of feedback texts; performing conflict resolution on the first semantic vector set according to the set of feedback texts to obtain a second semantic vector set with semantic conflicts eliminated; performing feature fusion processing on the second semantic vector set and the basic consultation information to generate medical record combined feature data; and generating a potential disease risk prediction result of the target child according to the medical record combined feature data through a preset risk reasoning model.

[0007] According to another aspect of the present application, a medical record data mining and potential risk prediction system based on deep learning is provided, comprising: a data acquisition module for acquiring a first semantic vector set, multi-modal physiological data, basic consultation information and initial cognitive level model parameters of a target child, the first semantic vector set containing multi-role complaint information around the current symptoms of the target child; a conflict recognition module for performing conflict recognition on the first semantic vector set to generate a conflict vector set with semantic conflicts; a parameter adjustment module for adjusting the initial cognitive level model parameters according to the conflict vector set and the multi-modal physiological data through a preset weight adjustment algorithm based on causal reasoning and reinforcement learning to obtain individual cognitive level model parameters; a text generation module for generating a set of follow-up texts in line with the cognitive level of the target child according to the conflict vector set and the individual cognitive level model parameters through a deep learning-based strategy generation network; a feedback receiving module for receiving feedback of the target child on each follow-up text in the set of follow-up texts to obtain a set of feedback texts; a conflict resolution module for performing conflict resolution on the first semantic vector set according to the set of feedback texts to obtain a second semantic vector set with semantic conflicts eliminated; a feature fusion module for performing feature fusion processing on the second semantic vector set and the basic consultation information to generate medical record combined feature data; and a prediction module for generating a potential disease risk prediction result of the target child according to the medical record combined feature data through a preset risk reasoning model.

[0008] According to another aspect of the present application, an electronic device is provided, comprising a memory for storing computer executable instructions, and a processor for executing the computer executable instructions, which, when executed by the processor, implement the steps of the method as described above.

[0009] According to another aspect of the present application, a computer storage medium is provided, having stored thereon computer executable instructions, which, when executed by a processor, implement the steps of the method as described above.

[0010] Compared with the prior art, the medical record data mining and potential risk prediction method and system based on deep learning according to the embodiments of the present application can generate follow-up text conforming to the cognition of children through conflict identification and dynamic parameter adjustment, effectively solve the problems of multi-role complaint information conflict and insufficient expression ability in children's medical records by combining multi-modal data fusion and feedback resolution mechanism, and has the advantage of improving the accuracy of potential disease risk prediction. BRIEF DESCRIPTION OF DRAWINGS

[0011] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The drawings provided in the present application are used to provide further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0012] Figure 1 Flowchart of the medical record data mining and potential risk prediction method based on deep learning of the present application.

[0013] Figure 2 Block diagram of the medical record data mining and potential risk prediction system based on deep learning of the present application.

[0014] Figure 3 Block diagram of the electronic device of the present application. DETAILED DESCRIPTION

[0015] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0016] Exemplary method

[0017] Figure 1The method for medical record data mining and potential risk prediction based on deep learning according to the embodiment of the present application is illustrated, which comprises: obtaining a first semantic vector set, multi-modal physiological data, basic consultation information and initial cognitive level model parameters for a target child, the first semantic vector set containing multi-role complaint information around the current symptoms of the target child; performing conflict recognition on the first semantic vector set to generate a conflict vector set with semantic conflicts; adjusting the initial cognitive level model parameters according to the conflict vector set and the multi-modal physiological data through a preset weight adjustment algorithm based on causal reasoning and reinforcement learning to obtain individual cognitive level model parameters; generating a set of follow-up texts in line with the cognitive level of the target child through a deep learning-based strategy generation network according to the conflict vector set and the individual cognitive level model parameters; receiving feedback from the target child on each follow-up text in the set of follow-up texts to obtain a set of feedback texts; performing conflict resolution on the first semantic vector set according to the set of feedback texts to obtain a second semantic vector set with resolved semantic conflicts; performing feature fusion processing on the second semantic vector set and the basic consultation information to generate medical record combined feature data; and generating a potential disease risk prediction result of the target child through a preset risk reasoning model according to the medical record combined feature data.

[0018] The first semantic vector set refers to the vectorized representation set of multi-role complaint information around the current symptoms of the target child, which can be realized by using natural language processing technology to perform semantic coding on the descriptions of the child, parents and medical staff, and is used to integrate the differential expressions of different roles on the same symptom.

[0019] The conflict recognition refers to detecting semantic contradictions or expression biases between the complaint information of different roles in the first semantic vector set, which can be realized by using semantic similarity calculation or an adversarial neural network model, and is used to locate the semantic fragments with inconsistent expressions.

[0020] The weight adjustment algorithm refers to a method for dynamically adjusting the cognitive level model parameters by combining causal reasoning and reinforcement learning, which can be realized by generating a parameter enhancement amount through a strategy network and calculating an effectiveness weight by combining causal contribution intensity, and is used to optimize the adaptability of the model to the cognitive ability of the child according to the conflict features and physiological data.

[0021] The strategy generation network refers to a model for generating follow-up texts in line with the cognitive level of the child based on deep learning, which can be realized by using a sequence-to-sequence model to generate texts in combination with cognitive constraints, and is used to output interactive questions with simple grammar and adapted vocabulary to clarify semantic conflicts.

[0022] The set of feedback texts refers to a set of response information of the target child to the follow-up texts, which can be realized by collecting through a natural language interaction interface and converting into semantic vectors, and is used to provide supplementary information to correct the bias in the initial complaint.

[0023] wherein the conflict resolution refers to eliminating semantic contradictions in the first semantic vector set using feedback information, which can be specifically implemented by semantic matching degree calculation and context relationship reconstruction through a graph attention network, for generating a complaint expression with consistent semantics.

[0024] wherein the feature fusion processing refers to jointly representing the resolved semantic vector and the basic medical information, which can be specifically implemented by feature splicing or a cross-modal attention mechanism, for constructing a combined feature that comprehensively reflects the child's condition.

[0025] wherein the risk inference model refers to a deep learning model for predicting potential disease risks based on medical record combined features, which can be specifically implemented by modeling the association between symptoms and diseases using a graph neural network or a hierarchical attention mechanism, for outputting an interpretable risk assessment result.

[0026] The core innovation of the present application lies in constructing a child-specific medical record representation through multi-role complaint conflict identification and dynamic feedback mechanism, combined with interactive questioning based on causal reinforcement learning parameter adjustment and cognitive adaptation, effectively solving the semantic deviation problem caused by ambiguous complaint information and multi-source conflicts of young patients, and improving the accuracy of pediatric disease risk prediction.

[0027] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0028] First, a multi-modal data acquisition device is used to obtain a first semantic vector set, multi-modal physiological data, basic medical information, and initial cognitive level model parameters of the target child. The first semantic vector set is converted from the oral symptom descriptions of children, parents, and medical staff by a natural language processing model. The multi-modal physiological data includes indicators such as body temperature, heart rate, and blood pressure. The basic medical information includes age, gender, and past medical history. The initial cognitive level model parameters are set based on the standard cognitive development model for children of different age groups.

[0029] Next, a pre-trained semantic conflict detection model is used to analyze the first semantic vector set, identify conflicting or inconsistent semantic fragments, and generate a conflict vector set. For example, the conflict between the parent's description of "persistent fever" and the child's statement of "no discomfort" is identified.

[0030] Then, a weight adjustment algorithm based on causal reasoning and reinforcement learning is used to input the conflict vector set and multi-modal physiological data, and dynamically adjust the initial cognitive level model parameters. The algorithm continuously optimizes the parameter adjustment strategy through reinforcement learning, and simultaneously uses a causal reasoning model to evaluate the impact of parameter adjustment on conflict resolution, ultimately obtaining individual cognitive level model parameters that adapt to the target child.

[0031] The strategy generation network based on deep learning generates a series of targeted follow-up texts according to the conflict vector set and the individual cognitive level model parameters. These follow-up texts match the language complexity and content depth with the child's cognitive level, facilitating the child's understanding and answering.

[0032] The medical staff asks the generated follow-up questions to the target child and records the child's feedback answers to form a feedback text set. Using natural language processing technology, the conflict parts in the first semantic vector set are corrected and supplemented according to the feedback text set, obtaining a second semantic vector set that eliminates semantic conflicts.

[0033] The second semantic vector set is processed by feature fusion with the basic medical information, and a comprehensive medical record combined feature data is generated using a multi-modal feature fusion algorithm. Finally, through a pre-trained risk reasoning model such as a deep neural network or an ensemble learning model, the medical record combined feature data is analyzed to output the potential disease risk prediction result of the target child.

[0034] Through the above scheme, the semantic conflict problem existing in the child's complaint information is effectively solved, and the accuracy and consistency of the medical record data are improved. By dynamically adjusting the cognitive model parameters, the individual differences of children are adapted, making the follow-up and interaction process more in line with the cognitive level of children. Based on the comprehensive analysis of multi-role complaint information and multi-modal data, the comprehensiveness and accuracy of potential disease risk prediction are enhanced. In addition, the interactive conflict resolution mechanism improves the reliability of the complaint information, providing a more reliable data basis for subsequent risk prediction. Overall, the scheme of the present application significantly improves the effect of pediatric medical record data mining, providing technical support for early identification and prevention of children's disease risks.

[0035] In some of the above schemes of the present application, adjusting the initial cognitive level model parameters specifically includes: obtaining historical conflict resolution result data; constructing a cognitive level model parameter vector according to the initial cognitive level model parameters ; constructing a state feature vector according to the conflict vector set and the multi-modal physiological data , wherein the function represents a feature fusion mapping; performing action selection on the state feature vector through a policy network in the reinforcement learning algorithm to output a cognitive parameter enhancement amount , i.e. ; generating an effectiveness weight vector indicating the causal contribution intensity of each cognitive dimension adjustment action to the conflict resolution task through a causal reasoning model according to the state feature vector , the cognitive parameter enhancement amount and the historical conflict resolution result data ; according to the effectiveness weight vector updating the initial cognitive level model parameters: wherein, denotes element-wise multiplication, denotes individual cognitive level model parameters for the target child patient.

[0036] Specifically:

[0037] First, historical conflict resolution result data is obtained from the database, including past cases of successfully resolving semantic conflicts and their corresponding cognitive parameter adjustment records. Then, the initial cognitive level model parameters are converted into vector form where each component represents a parameter value of a cognitive dimension, such as language comprehension ability, logical reasoning ability, etc.

[0038] Next, the conflict vector set and the multi-modal physiological data are fused through a feature fusion function to obtain the state feature vector . For example, f can be a multi-layer perceptron network that takes and as input and outputs the fused feature vector .

[0039] Then, a pre-trained reinforcement learning policy network is used to generate cognitive parameter enhancement with as input. It can be a deep neural network that is trained multiple times to obtain the optimal strategy.

[0040] Further, a causal inference model is used to analyze the causal relationship between , and historical data to generate the effectiveness weight vector . The causal inference model can be based on Bayesian networks or structural equation models to evaluate the actual effect of adjusting each cognitive dimension.

[0041] Finally, the initial parameters are added to the weighted enhancement to obtain the updated individual cognitive level model parameters . This process can be iterated multiple times until the parameters converge or reach a preset number of iterations.

[0042] By the technical solution, the application realizes self-adaptive adjustment of a cognitive level model parameter based on multi-source data. Thus, individual cognitive characteristics of a child patient can be accurately captured, and the accuracy of semantic conflict identification and resolution is improved. Meanwhile, by introducing a causal reasoning mechanism, the explainability and reliability of model adjustment are enhanced, and the problem of model instability caused by blind adjustment is avoided. In addition, by using a reinforcement learning method, the system can continuously learn and optimize from historical experience, and the generalization ability and adaptability of the model are improved.

[0043] In some solutions of the application, the conflict resolution of the first semantic vector set according to the feedback text set specifically includes: calculating semantic matching degrees between the feedback text set and the conflict vector set, constructing a feedback consistency score vector , wherein represents a semantic consistency degree of the i th conflict semantic segment and the feedback text; the feedback consistency score vector is weighted and fused with the conflict vector set to obtain a conflict enhanced semantic vector set , wherein represents an element-wise product; the conflict enhanced semantic vector set is reconstructed according to new semantic information in the feedback text set by a preset semantic graph attention network model, to obtain a semantic vector set after conflict resolution ; and the semantic vector set after conflict resolution is spliced with non-conflict semantic vectors in the first semantic vector set to generate a second semantic vector set.

[0044] Specifically, the feedback text set and the conflict vector set calculate the semantic matching degrees by cosine similarity, generate the feedback consistency score vector, and screen out the conflict segments with strong relevance to the feedback; the conflict enhanced semantic vector set combines the score vector and the original conflict vector by element-wise product, and reduces the weight of the low-relevance conflict segment; the semantic graph attention network model constructs the conflict enhanced semantic vector and the embedding vector of the feedback text into a graph structure, extracts the semantic dependency relationship across texts by using the multi-head attention mechanism, and eliminates the conflict residues caused by ambiguous expression; the reconstructed semantic vector and the original non-conflict vector are aligned and spliced after the feature dimensions pass through the full connection layer, to form a complete and non-contradictory second semantic vector set. For example, when the conflict vector contains “fever lasts for three days” and the feedback text supplements “actual intermittent low fever for two days”, the semantic graph attention network corrects the conflict vector to “intermittent low fever for two days” by associating the time description and the body temperature fluctuation characteristics. The second semantic vector set generated in this way can accurately reflect the real symptoms of the patient, and improve the reliability of subsequent risk prediction.

[0045] In implementation, first, a BERT-based semantic similarity calculation model is used to calculate the cosine similarity between the feedback text and the conflict vector, obtaining a feedback consistency score vector A. And Element-wise multiplication obtains . Then, a graph attention network (GAT) model is used, taking each semantic vector in as a node in the graph, and constructing edge connections according to semantic correlation. The GAT model aggregates information for the nodes through a multi-head attention mechanism, while fusing new semantic information in the feedback text, to realize context reconstruction, obtaining . Finally, the is concatenated with the non-conflict semantic vector to obtain a second semantic vector set.

[0046] Through the above technical solutions, the present application effectively eliminates the semantic conflict of multi-role expression in the chief complaint of a child patient. By introducing the feedback consistency score, more reliable semantic information is retained. The graph attention network is used for context reconstruction, fully utilizing the supplementary information in the feedback text. The finally generated second semantic vector set eliminates the semantic conflict, improves the accuracy and consistency of the chief complaint information, and lays a foundation for subsequent medical record data mining and risk prediction.

[0047] In some of the above schemes of the present application, generating a set of follow-up texts that conform to the cognitive level of the target child specifically includes: according to the individual cognitive level model parameters, screening semantic expression templates corresponding to the cognitive level of the target child from a preset cognitive template set; according to the semantic conflict content in the conflict vector set, performing semantic filling on the semantic expression templates to obtain an initial set of follow-up candidate texts; according to the individual cognitive level model parameters, constraining the vocabulary depth, syntax structure and sentence length of each candidate text in the initial set of follow-up candidate texts to obtain a cognitive adaptation follow-up text that is semantically complete and concisely expressed; and through a preset style control generation network, generating a set of follow-up texts that conform to the language expression preferences and interaction habits of the target child according to the cognitive adaptation follow-up text.

[0048] Among them, the cognitive template set is divided into multiple cognitive levels according to age, and each level corresponds to different sentence length thresholds and vocabulary banks, for example, the 3-5 year old cognitive level uses a single sentence length of no more than 8 words and a vocabulary bank containing 500 basic words. In the semantic filling process, the symptom keywords in the conflict vector are mapped to the template placeholders, for example, "abdominal pain" is filled into the sentence "How do you feel uncomfortable in [part]?" The vocabulary depth constraint is realized by replacing professional terms with life-like expressions, for example, "abdominal tenderness" is converted to "Will pressing the stomach hurt?" The style control generation network uses an adversarial training method, inputs historical dialogue data to extract child language style features, and outputs texts that conform to the first person questioning habits.

[0049] Specifically, the cognitive level matching module selects a corresponding template according to the age parameter, for example, a 7-year-old child selects a template containing complex sentences but no professional terms. The semantic filling engine maps "intermittent fever" in the conflict vector to "Has there been any [symptom] in the body recently?" to generate the initial candidate text. The vocabulary restrictor detects terms such as "inflammatory response" in the candidate text and replaces them with child-understandable vocabulary such as "inflammation and swelling". The sentence length controller splits complex conditional sentences into multiple simple interrogative sentences, for example, "If the headache lasts more than two days and is accompanied by vomiting, do you need to see a doctor?" is split into "Has the headache lasted more than two days?" and "Does the headache make you want to vomit?" The style generator adds mood modifiers to the text, such as converting "Is there tinnitus?" to "Does your ear ring, huh?" to form a set of follow-up questions that conform to the child's expression habits.

[0050] Through the above technical solutions, the application can generate follow-up question texts that conform to the target child's cognitive level and individualized language expression habits, improving the child's understanding of the follow-up questions and the accuracy of the answers. This improves the quality of the collection of children's chief complaint information and provides more reliable semantic input for subsequent medical record data mining and potential risk prediction. Further, the scheme enhances the affinity of the doctor-patient interaction through cognitive adaptation and style control, which helps to alleviate the tension of children during the medical treatment process and improves the diagnosis and treatment experience.

[0051] In some of the above schemes of the application, generating the medical record combined feature data specifically includes: extracting a structured semantic label vector corresponding to the current chief complaint symptom according to the second semantic vector set through a preset symptom structured coding network; concatenating the structured semantic label vector and the corresponding field in the basic medical information into a chief complaint information fusion vector; performing feature normalization processing on the individual cognitive level model parameters and the multi-modal physiological data, and constructing a child individual feature vector; performing feature concatenation processing on the chief complaint information fusion vector and the child individual feature vector to generate medical record combined feature data representing the current medical situation of the target child.

[0052] The symptom structured coding network adopts a pre-trained multi-layer Transformer architecture, identifies the mapping relationship between the symptom entities in the second semantic vector set and the medical ontology library through an attention mechanism, and outputs a dimension-standardized semantic label vector. The field matching of the basic medical information is realized through a preset field mapping table, ensuring that the structured semantic label vector and the symptom description field in the medical record are aligned in the vector space. The feature normalization processing adopts the Z-score standardization method to eliminate the differences in dimension and numerical range between the cognitive parameters and the physiological signals. The feature concatenation processing realizes the expansion of the vector dimension and the integration of information through a fully connected layer.

[0053] Specifically, the symptom structured coding network first performs sequence labeling on the second semantic vector set to identify semantic fragments containing symptom descriptions, and then performs entity linking through a medical knowledge graph to generate a semantic label vector containing ICD codes and symptom severity. This vector is cross-modally spliced with the past medical history field in the basic consultation information to form a chief complaint information fusion vector containing a time dimension and symptom evolution. At the same time, the individual cognitive level model parameters are subjected to logarithmic transformation to eliminate skew distribution, and the multi-modal physiological data are subjected to sliding window mean processing to eliminate noise. After standardization, the two are spliced into a feature vector reflecting individual differences of children. Finally, through nonlinear transformation of the full connection layer, the chief complaint information and individual characteristics are deeply fused to generate multi-dimensional medical record combination feature data containing time sequence features, symptom features, and physiological features, which are used as high-quality input of the risk reasoning model.

[0054] Through the above technical solutions, the application realizes the structured coding of children's chief complaint information and the fusion of multi-source heterogeneous data, and improves the representation ability of medical record data. By introducing individual cognitive level and physiological indicators, the personalized expression of features is enhanced, providing more comprehensive and accurate input data for subsequent risk prediction tasks. This method effectively solves the problems of inaccurate and limited expression of children's chief complaint information, and improves the quality and effectiveness of medical record data mining.

[0055] In some of the above schemes of the application, there is a problem that irrelevant redundant descriptions in complex sentence expressions are not effectively identified and removed during the processing, which introduces interference information during conflict identification, affecting the accuracy of semantic conflict detection and the reliability of subsequent risk prediction. The application further proposes that before the conflict identification of the first semantic vector set, it further includes: performing phrase-level segmentation and structure reorganization on the complex sentence expressions in the first semantic vector set, and removing the redundant descriptions irrelevant to the chief complaint.

[0056] Specifically, the semantic fragment boundary recognition network based on the BiLSTM-CRF structure performs named entity recognition and boundary labeling on the sentence elements in the first semantic vector set to extract a candidate chief complaint fragment set; a syntactic dependency graph is constructed according to the semantic weight distribution of the candidate chief complaint fragments, and the syntactic connection paths between the candidate chief complaint fragments are identified through graph structure traversal; the candidate chief complaint fragments with relevance are merged according to their connectivity and semantic similarity in the syntactic dependency graph, and the unmerged fragments and fragments with a semantic similarity lower than a preset similarity threshold are removed to obtain a new first semantic vector set.

[0057] For example, when the input contains the complex sentence "The stomach aches in waves, like being pricked by a needle, but seems to be a little better after eating", the semantic segment boundary recognition network first splits out four candidate complaint segments: "stomach aches", "in waves", "like being pricked by a needle", and "seems to be a little better after eating". Through dependency syntax analysis, it is determined that "in waves" modifies "stomach aches", "like being pricked by a needle" is a metaphorical description related to "stomach aches", and "seems to be a little better after eating" is related to the change of pain. In the syntactic dependency relation graph, the former three form a connected subgraph, the semantic similarity is 0.85, and they are merged into a single vector "intermittent severe abdominal pain"; the similarity between "seems to be a little better after eating" and the complaint segment is 0.65, which is lower than the preset threshold 0.7, and it is removed as redundant description. The new semantic vector set generated in this way eliminates irrelevant interference, retains the core symptom description, and improves the accuracy of subsequent conflict recognition.

[0058] Through the above technical solutions, the application can effectively extract key complaint information in children's medical records, eliminate redundant and repetitive expressions, and improve the accuracy of subsequent semantic conflict recognition. Through the structured processing and association analysis of semantic segments, it can better understand and integrate descriptions from multiple roles, and overcome the semantic ambiguity problem caused by the insufficient expression ability of children. At the same time, the semantic relationship modeling based on the graph structure helps to capture the logical relationship between symptoms, and provides a more reliable semantic basis for subsequent risk assessment.

[0059] Exemplary system

[0060] Figure 2The system for medical record data mining and potential risk prediction based on deep learning according to the embodiment of the present application is illustrated, comprising: a data acquisition module, configured to acquire a first semantic vector set, multi-modal physiological data, basic consultation information and initial cognitive level model parameters for a target child, the first semantic vector set containing multi-role complaint information around the current symptoms of the target child; a conflict identification module, configured to perform conflict identification on the first semantic vector set to generate a conflict vector set with semantic conflicts; a parameter adjustment module, configured to adjust the initial cognitive level model parameters according to the conflict vector set and the multi-modal physiological data through a preset weight adjustment algorithm based on causal reasoning and reinforcement learning, to obtain individual cognitive level model parameters; a text generation module, configured to generate a set of follow-up texts in line with the cognitive level of the target child through a deep learning-based strategy generation network according to the conflict vector set and the individual cognitive level model parameters; a feedback receiving module, configured to receive feedback of the target child on each follow-up text in the set of follow-up texts to obtain a set of feedback texts; a conflict resolution module, configured to perform conflict resolution on the first semantic vector set according to the set of feedback texts to obtain a second semantic vector set with resolved semantic conflicts; a feature fusion module, configured to perform feature fusion processing on the second semantic vector set and the basic consultation information to generate medical record combined feature data; and a prediction module, configured to generate a potential disease risk prediction result of the target child through a preset risk reasoning model according to the medical record combined feature data.

[0061] In one example, adjusting the initial cognitive level model parameters comprises: acquiring historical conflict resolution result data; constructing a cognitive level model parameter vector according to the initial cognitive level model parameters ; constructing a state feature vector according to the conflict vector set and the multi-modal physiological data , wherein the function represents feature fusion mapping; performing action selection on the state feature vector through a policy network in the reinforcement learning algorithm to output a cognitive parameter enhancement amount , i.e. ; generating an effectiveness weight vector indicating the causal contribution intensity of each cognitive dimension adjustment action to the conflict resolution task through a causal reasoning model according to the state feature vector , the cognitive parameter enhancement amount and the historical conflict resolution result data; and updating the initial cognitive level model parameters according to the effectiveness weight vector : wherein represents element-wise multiplication, and represents the individual cognitive level model parameters for the target child patient.

[0062]

[0062] In one example, the conflict resolution module resolves the conflicts in the first set of semantic vectors according to the set of feedback texts includes: calculating semantic matching degrees between the set of feedback texts and the set of conflict vectors, constructing a feedback consistency score vector , wherein represents a semantic consistency degree between the ith conflict semantic fragment and the feedback text; performing weighted fusion on the feedback consistency score vector and the set of conflict vectors to obtain a set of conflict-enhanced semantic vectors , wherein represents an element-wise product; performing context reconstruction on the set of conflict-enhanced semantic vectors according to the newly added semantic information in the set of feedback texts by a preset semantic graph attention network model to obtain a set of conflict-resolved semantic vectors ; and splicing the set of conflict-resolved semantic vectors with non-conflict semantic vectors in the first set of semantic vectors to generate a second set of semantic vectors.

[0063] In one example, the text generation module generates a set of follow-up question texts that conform to the cognitive level of the target child includes: selecting a semantic expression template corresponding to the cognitive level of the target child from a set of preset cognitive templates according to the individual cognitive level model parameters; performing semantic filling on the semantic expression template according to the semantic conflict content in the set of conflict vectors to obtain an initial set of follow-up question candidate texts; constraining the vocabulary depth, syntax structure, and sentence length of each candidate text in the initial set of follow-up question candidate texts according to the individual cognitive level model parameters to obtain a cognitive-adapted follow-up question text that is semantically complete and concisely expressed; and generating a set of follow-up question texts that conform to the language expression preferences and interaction habits of the target child according to the cognitive-adapted follow-up question text by a preset style control generation network.

[0064] In one example, the feature fusion module generates the medical record combined feature data includes: extracting a structured semantic label vector corresponding to the current chief complaint symptom according to the second set of semantic vectors by a preset symptom structured coding network; splicing the structured semantic label vector and the corresponding field in the basic consultation information into a chief complaint information fusion vector; performing feature normalization processing on the individual cognitive level model parameters and the multi-modal physiological data, and constructing a child individual feature vector; and performing feature concatenation processing on the chief complaint information fusion vector and the child individual feature vector to generate medical record combined feature data representing the current consultation situation of the target child.

[0065] In one example, the conflict identification module further includes, before performing conflict identification on the first set of semantic vectors: performing phrase-level segmentation and structure reorganization on the complex sentence expressions in the first set of semantic vectors, and eliminating redundant descriptions irrelevant to the chief complaint.

[0066] In one example, the phrase-level segmentation and structure reorganization of the complex sentence expression in the first semantic vector set by the conflict identification module comprises: performing named entity recognition and boundary labeling on the sentence elements in the first semantic vector set by a semantic segment boundary identification network based on a BiLSTM-CRF structure, extracting a candidate chief complaint segment set; constructing a syntactic dependency relation graph according to the semantic weight distribution of the candidate chief complaint segments, and identifying the syntactic connection paths between the candidate chief complaint segments through graph structure traversal; merging the candidate chief complaint segments with relevance according to their connectivity and semantic similarity in the syntactic dependency relation graph, and eliminating the unmerged segments and the segments with a semantic similarity lower than a preset similarity threshold, to obtain a new first semantic vector set.

[0067] Example electronic device

[0068] Figure 3 An electronic device according to embodiments of the application is illustrated. The electronic device can be the mobile device itself, or a standalone device that can communicate with the mobile device to receive input signals captured therefrom and send selected target driving behaviors thereto.

[0069] Figure 3 A block diagram of an electronic device according to embodiments of the application is illustrated.

[0070] As Figure 3 illustrated, the electronic device includes one or more processors and a memory.

[0071] The processor can be a central processing unit (CPU) or other form of processing unit having data processing and / or instruction execution capabilities, and can control other components in the electronic device to perform desired functions.

[0072] The memory can include one or more computer program products that can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory, for example, can include random access memory (RAM), cache, and / or the like. The non-volatile memory, for example, can include read-only memory (ROM), hard disk, flash memory, and / or the like. One or more computer program instructions can be stored on the computer-readable storage media, and the processor can execute the program instructions to implement the driving behavior decision method of various embodiments of the application described above and / or other desired functions.

[0073] In one example, the electronic device can further include input and output devices, which are interconnected through a bus system and / or other forms of connection mechanisms (not shown).

[0074] Of course, to simplify, Figure 3Only some of the components of the electronic device related to the present application are shown, and components such as a bus, an input / output interface, and the like are omitted. In addition to this, the electronic device can include any other appropriate components according to the specific application.

[0075] Exemplary computer-readable medium

[0076] Embodiments of the present application can also be computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, cause the processor to perform the steps described in the above "Exemplary Method" section of the present specification for a driving behavior decision method according to various embodiments of the present application.

[0077] The computer readable storage medium can be any combination of one or more computer readable medium. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium can include, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination of the above. More specific examples (a non-exhaustive list) of the computer readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0078] The above describes the basic principles of the present application in combination with specific embodiments, but it should be noted that the advantages, advantages, effects and the like mentioned in the present application are only examples and are not limiting, and these advantages, advantages, effects and the like cannot be considered as the must-have of each embodiment of the present application. In addition, the above specific details are only for the purpose of example and for the purpose of understanding, and the above details do not limit the present application to the must-use of the above specific details to realize.

[0079] The block diagrams of the devices, apparatuses, equipment, systems involved in the present application are only illustrative examples and are not intended to require or imply the connection, arrangement, configuration shown in the block diagram. As those skilled in the art will recognize, these devices, apparatuses, equipment, systems can be connected, arranged, configured in any manner. Words such as "include", "contain", "have", and the like are open-ended words, mean "include but not limited to", and can be used interchangeably. The words "or" and "and" used herein mean the word "and / or", and can be used interchangeably unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably.

[0080] It is also important to note that the devices, apparatuses and methods described in the present application can be embodied in a variety of other forms, modi fications and alt ernatives, some of which have been discussed above and some of which are gathe red as wi ll be apparent to those reasonably skilled in the art. The described aspects and embodiments are to be considered in a descriptive sense only and not for purposes of limitation. Therefore, the scope of the present application is not to be determined strictly by the description in the specification but by the appended claims, and all changes that come within the meaning and range of equivalency of the claims are intended to be embraced therein.

[0081] The above description of disclosed aspects is intended to be illustrative, and not restrictive. Other aspects, including modifications and alternative forms, will be apparent upon reading the disclosure. The various aspects of the disclosure are defined in the appended claims, and the equivalents thereof.

[0082] The above description has been given by way of example and is not intended to limit the application. Although various example aspects and embodiments have been discussed above, those of ordinary skill in the art will appreciate certain variations, modifications, alterations, additions and subcombinations thereof.

Claims

1. A method for medical record data mining and potential risk prediction based on deep learning, characterized by: include: Obtaining a first semantic vector set for the target child, multimodal physiological data, basic medical information, and initial cognitive level model parameters, wherein the first semantic vector set includes multi-role chief complaint information surrounding the target child's current symptoms; Performing conflict identification on the first semantic vector set to generate a conflict vector set with semantic conflicts; By presetting a weight adjustment algorithm based on causal reasoning and reinforcement learning, adjusting the initial cognitive level model parameters according to the conflict vector set and the multimodal physiological data, to obtain individual cognitive level model parameters; Generate a set of follow-up texts that meet the target child's cognitive level based on the conflict vector set and the individual cognitive level model parameters through a deep learning-based strategy generation network; receiving feedback from the target child on each follow-up text in the follow-up text set to obtain a feedback text set; performing conflict resolution on the first semantic vector set according to the feedback text set to obtain a second semantic vector set that eliminates semantic conflicts; Performing feature fusion processing on the second semantic vector set and the basic medical information to generate medical record combined feature data; Through a preset risk inference model, a potential disease risk prediction result of the target child is generated based on the combined characteristic data of the medical records.

2. The method for medical record data mining and potential risk prediction based on deep learning according to claim 1, characterized in that: The adjusting of the initial cognitive level model parameters comprises: Obtain historical conflict resolution result data; Construct a cognitive level model parameter vector according to the initial cognitive level model parameters ; According to the conflict vector set and the multimodal physiological data , construct the state feature vector , where the function represents feature fusion mapping; Through the policy network in the reinforcement learning algorithm For the state feature vector Perform action selection and output cognitive parameter enhancement ,Right now ; Through the causal inference model, according to the state feature vector , cognitive parameter enhancement and historical conflict resolution result data to generate the effectiveness weight vector indicating the causal contribution strength of each cognitive dimension adjustment action to the conflict resolution task ; According to the validity weight vector Update the initial cognitive level model parameters: ,in, represents element-wise product, represents the individual cognitive level model parameters for the target child patient.

3. The method for medical record data mining and potential risk prediction based on deep learning according to claim 1 or 2, characterized in that: The performing conflict resolution on the first semantic vector set according to the feedback text set includes: Calculate the semantic matching degree between the feedback text set and the conflict vector set, and construct a feedback consistency score vector ,in , represents the semantic consistency between the i-th conflicting semantic segment and the feedback text; The feedback consistency score vector With the conflict vector set Perform weighted fusion to obtain the conflict enhanced semantic vector set ,in, represents element-wise product; Through the preset semantic graph attention network model, the conflict enhanced semantic vector set is enhanced according to the newly added semantic information in the feedback text set. Perform context reconstruction to obtain the semantic vector set after conflict resolution ; The semantic vector set after conflict resolution The second semantic vector set is generated by splicing the non-conflicting semantic vectors in the first semantic vector set.

4. The method for medical record data mining and potential risk prediction based on deep learning according to claim 1, characterized in that: Generating a set of follow-up texts that meet the cognitive level of the target children includes: According to the individual cognitive level model parameters, a semantic expression template corresponding to the target child's cognitive level is selected from a preset cognitive template set; Semantically filling the semantic expression template according to the semantic conflict content in the conflict vector set to obtain an initial question candidate text set; Constraining the vocabulary depth, grammatical structure, and sentence length of each candidate text in the initial set of candidate question texts according to the individual cognitive level model parameters to obtain a cognitively adapted question text that is semantically complete and concise in expression; Through a preset style control generation network, a set of follow-up texts that conform to the target children's language expression preferences and interaction habits are generated based on the cognitive adaptation follow-up texts.

5. The method for medical record data mining and potential risk prediction based on deep learning according to claim 1, characterized in that: Generating medical record combination feature data includes: Extracting a structured semantic label vector corresponding to the current chief complaint symptom according to the second semantic vector set through a preset symptom structured coding network; Splicing the structured semantic label vector and the corresponding field in the basic medical information into a chief complaint information fusion vector; Performing feature normalization processing on the individual cognitive level model parameters and the multimodal physiological data, and constructing an individual feature vector of the child; The chief complaint information fusion vector and the child's individual feature vector are subjected to feature cascade processing to generate the medical record combined feature data representing the current medical situation of the target child.

6. The method for medical record data mining and potential risk prediction based on deep learning according to claim 1, characterized in that: Before performing conflict identification on the first semantic vector set, the method further includes: The complex sentence expressions in the first semantic vector set are segmented and restructured at the phrase level to remove redundant descriptions that are not related to the main complaint.

7. The method for medical record data mining and potential risk prediction based on deep learning according to claim 6, characterized in that: The phrase-level segmentation and structural reorganization of the complex sentence expressions in the first semantic vector set includes: The semantic segment boundary recognition network based on BiLSTM-CRF structure is used to perform named entity recognition and boundary annotation on the sentence elements in the first semantic vector set to extract the candidate main complaint segment set; Constructing a syntactic dependency graph based on the semantic weight distribution of the candidate main complaint segments, and identifying syntactic connection paths between the candidate main complaint segments by traversing the graph structure; The relevant candidate main complaint segments are merged according to their connectivity and semantic similarity in the syntactic dependency graph, and the unmerged segments and segments with semantic similarity lower than the preset similarity threshold are eliminated to obtain a new first semantic vector set.

8. A deep learning-based medical record data mining and potential risk prediction system, characterized by: include: a data acquisition module, configured to acquire a first semantic vector set for a target child, multimodal physiological data, basic medical information, and initial cognitive level model parameters, wherein the first semantic vector set includes multi-role chief complaint information surrounding the target child's current symptoms; a conflict identification module, configured to perform conflict identification on the first semantic vector set and generate a conflict vector set having semantic conflicts; a parameter adjustment module, configured to adjust the initial cognitive level model parameters according to the conflict vector set and the multimodal physiological data by using a preset weight adjustment algorithm based on causal reasoning and reinforcement learning, thereby obtaining individual cognitive level model parameters; A text generation module, configured to generate a set of follow-up texts that meet the cognitive level of the target child based on the conflict vector set and the individual cognitive level model parameters using a deep learning-based strategy generation network; a feedback receiving module, configured to receive the target child's feedback on each follow-up text in the follow-up text set, and obtain a feedback text set; a conflict resolution module, configured to resolve conflicts on the first semantic vector set according to the feedback text set to obtain a second semantic vector set that eliminates semantic conflicts; a feature fusion module, configured to perform feature fusion processing on the second semantic vector set and the basic medical information to generate medical record combined feature data; The prediction module is used to generate a potential disease risk prediction result for the target child based on the combined characteristic data of the medical records through a preset risk inference model.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images