Medical record analysis and intelligent inquiry method and device

Abstract

The invention relates to the field of intelligent inquiry, in particular to a medical record analysis and intelligent inquiry method and device. The method comprises the following steps: acquiring an electronic medical record file of a patient, performing multi-dimensional implicit symptom description analysis, and constructing multi-dimensional medical record analysis information; time sequence matching mapping is carried out based on the multi-dimensional medical record analysis information, dynamic evolution visualization is carried out, and a time sequence evolution medical record portrait is constructed; performing multi-scale wavelet transform decomposition on the sequential evolution medical record portrait, performing symptom evolution rule mining, and identifying a symptom evolution key trigger point and a turning point; performing potential pathology prediction according to the symptom evolution key trigger point and the turning point to generate a potential pathology candidate set; and generating intelligent inquiry questions based on the potential pathology candidate set, performing priority adjustment, and constructing an intelligent inquiry mechanism. Based on accurate medical record analysis of the patient, the targeted inquiry problem is generated, the high-quality automatic inquiry problem is generated, and the preliminary diagnosis efficiency and the preliminary pathological analysis accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical consultation, and more particularly to a method and apparatus for medical record analysis and intelligent medical consultation. Background Technology

[0002] With the rapid development of artificial intelligence and medical information technology, medical services are gradually transforming towards intelligence and precision. Especially with the increasing prevalence of Electronic Medical Records (EMRs), the scale and complexity of medical data are constantly increasing. How to efficiently and accurately analyze medical record information and automatically generate clinically valuable diagnostic questions based on this has become one of the key technologies for improving the quality and practicality of intelligent consultation systems. Traditional medical consultations rely on doctors' clinical experience and on-site communication. However, driven by the rapid development of new medical service models such as telemedicine and intelligent assisted diagnosis, automated medical record analysis and intelligent diagnostic question generation technologies are increasingly demonstrating their importance and urgent need.

[0003] In practical applications, medical records typically contain rich but unstructured textual content, such as chief complaints, present illness, past medical history, examination reports, and medical orders. This information often uses diverse terminology, has complex semantics, and lacks a unified standard, posing a significant challenge to automated parsing. Furthermore, the focus of consultations differs significantly across departments and disease areas. Extracting key information from known medical records and intelligently generating targeted consultation questions based on specific disease contexts is one of the technical challenges in achieving a high-quality human-computer interactive medical consultation system.

[0004] Existing medical record parsing methods largely rely on rule-based text processing techniques or traditional natural language processing models. These methods are prone to misjudgment or omission when dealing with complex semantics, polysemous words, and medical terminology, failing to meet the accuracy and generalization requirements of real-world medical scenarios. Furthermore, most current methods for generating consultation questions still rely on template matching or manual design, lacking a deep understanding of the contextual medical record content. The generated questions often suffer from repetition, redundancy, or insufficient clinical significance, making it difficult to truly assist doctors or replace the role of initial manual diagnosis. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method and apparatus for medical record analysis and intelligent consultation, thereby resolving at least one of the aforementioned technical problems.

[0006] To achieve the above objectives, the present invention provides a method for medical record analysis and intelligent consultation, comprising the following steps: Step S1: Obtain the patient's electronic medical record, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record analysis information; Step S2: Perform time-series matching and mapping based on multi-dimensional medical record parsing information, and visualize the dynamic evolution to construct a time-series evolution medical record profile; Step S3: Perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and mine the symptom evolution pattern to identify key trigger points and turning points in symptom evolution. Step S4: Based on the key trigger points and turning points in symptom evolution, perform potential pathology prediction and generate a potential pathology candidate set; Step S5: Generate intelligent consultation questions based on the potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

[0007] This specification provides a medical record analysis and intelligent consultation device for performing the medical record analysis and intelligent consultation method described above, comprising: The medical record parsing module is used to obtain patients' electronic medical records, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record parsing information. The dynamic evolution module is used to perform time-series matching and mapping based on multi-dimensional medical record parsing information, and to visualize the dynamic evolution to build a time-series evolution medical record profile. The symptom evolution mining module is used to perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and to mine the symptom evolution pattern, identifying key trigger points and turning points in symptom evolution. The potential pathology prediction module is used to predict potential pathologies based on key trigger points and turning points in symptom evolution, and generate a potential pathology candidate set. The intelligent consultation module is used to generate intelligent consultation questions based on a potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

[0008] The beneficial effects of this invention are specifically as follows: It enables comprehensive mining and fusion of structured and unstructured medical record data (such as chief complaints, examinations, and medical history), improving the completeness and multi-faceted understanding of medical record information. Utilizing natural language processing and semantic analysis, it extracts implicit symptoms, ambiguous descriptions, and clues to disease evolution, laying a high-quality data foundation for subsequent modeling. Constructing structured multidimensional medical record information facilitates subsequent model processing, visualization, and decision support. By linking various patient symptoms, signs, and examination results along a timeline, it forms a visualized evolutionary trajectory, helping doctors or systems understand the disease progression. Dynamic visualization enhances doctors' intuitive judgment and provides a clear time-series data structure for AI modeling. Building medical record profiles helps discover potential patterns, supporting personalized health management and similar case reasoning. Utilizing the time-frequency localization characteristics of wavelet transform, it effectively uncovers deep-seated patterns such as the periodicity, abrupt changes, and trend changes in symptom changes. It accurately identifies key trigger points (such as the first appearance or sudden change of symptoms) and turning points (the dividing line between worsening and remission of the condition), providing key node information for pathological prediction. It supports dynamic modeling of complex diseases, enhancing the system's ability to perceive and warn of potential risks. By introducing temporal context and symptom evolution paths, it improves the accuracy and reliability of pathological predictions, outperforming single-point static prediction methods. Candidate set construction narrows the scope of intelligent consultation, reduces the exploration of irrelevant paths, and improves system decision-making efficiency. It supports concurrent prediction of multiple pathologies, which helps in the early identification of comorbidities or complications. Based on candidate pathologies and key nodes, the system can automatically generate targeted and hierarchical consultation questions, improving user interaction efficiency and diagnostic accuracy. Prioritization ensures that the system starts asking the most critical questions, avoiding redundant questions and optimizing the user experience. The constructed intelligent consultation mechanism has self-learning and feedback optimization capabilities, supporting applications in multiple scenarios such as telemedicine, assisted diagnosis, and intelligent triage. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the steps of a medical record analysis and intelligent consultation method according to the present invention; Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1. Figure 3 This is a flowchart illustrating the detailed implementation steps of step S2. Detailed Implementation

[0010] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0011] This application provides a method and apparatus for medical record analysis and intelligent consultation. The executing entities of the method and apparatus include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio-visual management system, an information management system, and a cloud-based data management system.

[0012] Please see Figures 1 to 3 This invention provides a method for medical record analysis and intelligent consultation, including the following steps: Step S1: Obtain the patient's electronic medical record, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record analysis information; Step S2: Perform time-series matching and mapping based on multi-dimensional medical record parsing information, and visualize the dynamic evolution to construct a time-series evolution medical record profile; Step S3: Perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and mine the symptom evolution pattern to identify key trigger points and turning points in symptom evolution. Step S4: Based on the key trigger points and turning points in symptom evolution, perform potential pathology prediction and generate a potential pathology candidate set; Step S5: Generate intelligent consultation questions based on the potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

[0013] In the embodiments of the present invention, see Figure 1 The diagram below illustrates the steps of a medical record analysis and intelligent consultation method according to the present invention. In this example, the steps of the medical record analysis and intelligent consultation method include: Step S1: Obtain the patient's electronic medical record, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record analysis information; In this embodiment, the patient's complete electronic medical record is acquired, including outpatient medical records, inpatient records, laboratory reports, imaging data, and medication lists, among other comprehensive medical data. Deep semantic parsing technology based on the BERT-Clinical pre-trained model is used, with text segmentation granularity set to character level and word vector dimension set to 768 dimensions, to perform fine-grained semantic encoding on the medical record text. A named entity recognition algorithm is used to extract explicit symptom entities, with an accuracy threshold set above 85%. Simultaneously, a Latent Dirichlet Allocation (LDA) topic model is used to mine implicit symptom descriptions, with 20 topics and 1000 iterations. A symptom-disease association matrix is ​​constructed. The dataset has a dimension of 500×300, covering the association between 500 common symptoms and 300 diseases. TF-IDF weighting is used to quantify symptom importance, with a weight threshold of 0.3 or higher for inclusion in the analysis. Dependency parsing is used to identify causal, concurrent, and temporal relationships among symptoms, with a syntactic analysis depth of 3 layers. Semantic disambiguation is performed using a medical ontology database, achieving an accuracy rate of over 92%. Finally, a multi-dimensional medical record parsing dataset is generated, including explicit symptom features, implicit symptom inferences, symptom severity scores, and symptom duration estimates. The dataset contains a comprehensive symptom profile with an average of 150 feature dimensions per patient.

[0014] Step S2: Perform time-series matching and mapping based on multi-dimensional medical record parsing information, and visualize the dynamic evolution to construct a time-series evolution medical record profile; In this embodiment, timestamp standardization is performed based on multi-dimensional medical record parsing information, converting time information of different formats into a unified Unix timestamp format with hourly accuracy. A time series alignment algorithm is constructed, and the Dynamic Time Warping (DTW) method is used to handle irregular sampling problems in time series. The DTW window constraint parameter is set to 10% of the sequence length. Sliding time window technology is used for symptom evolution analysis, with a window size of 7 days and a sliding step of 1 day to ensure continuous monitoring of symptom changes. A symptom intensity quantitative scoring system is established, using a 0-10 scoring standard, where 0 points indicate symptom disappearance and 10 points indicate the most severe symptoms. Cubic spline interpolation is used to... The method of value interpolation fills in missing values ​​in time series data, with interpolation accuracy controlled within 5%. Kalman filtering is used to filter noise in the time series data, with process noise variance set to 0.1 and observation noise variance set to 0.01. A multi-layer time series mapping network is constructed with 5 layers, each with 256, 128, 64, 32, and 16 nodes, respectively, using the ReLU activation function. Time series clustering analysis is used to identify periodic patterns of symptom onset, with 8 main clusters. Finally, a dynamically evolving time series medical record profile is generated, containing a visual representation of four core dimensions: symptom development trajectory, symptom interaction relationships, treatment response patterns, and disease progression prediction curves.

[0015] Step S3: Perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and mine the symptom evolution pattern to identify key trigger points and turning points in symptom evolution. In this embodiment, the temporal evolution of medical records is decomposed using multi-scale wavelet transform. The Daubechies-4 wavelet is selected as the mother wavelet function, as it possesses good time-frequency localization characteristics. The decomposition is set to six levels, each corresponding to a different time scale: Level 1 corresponds to short-term fluctuations of 1-2 days, Level 2 to subacute changes of 3-7 days, Level 3 to medium-term trends of 1-2 weeks, Level 4 to long-term evolution of 1 month, Level 5 to seasonal changes of 3 months, and Level 6 to annual periodicity of 1 year. Statistical analysis is performed on the wavelet coefficients of each level, calculating their mean, variance, skewness, kurtosis, and other statistical characteristics to construct a 16-dimensional statistical feature vector. Wavelet packet energy analysis is used to quantify the energy distribution of symptom evolution across different frequency bands. An energy distribution threshold of 5% or more of the total energy is considered to indicate symptom evolution. Clinical significance: Wavelet coherence analysis identifies phase relationships between different symptoms, with a coherence coefficient greater than 0.7 indicating a strong correlation; continuous wavelet transform is used to detect abrupt changes in symptom evolution, with the abrupt change detection threshold set at three times the standard deviation of the local variance; a wavelet ridge extraction algorithm is constructed to identify the dominant frequency components of symptom evolution, with a frequency resolution set at 0.01Hz; multi-resolution analysis technology is used to separate the trend, periodic, and noise terms in the symptom evolution signal, with a signal-to-noise ratio requirement of over 20dB; the wavelet modulus maxima method is used to accurately locate key trigger points and turning points in symptom evolution, with a positioning accuracy within 6 hours; ultimately, an average of 8-12 key time points are identified per patient, including the initial symptom onset, deterioration turning point, treatment response point, and symptom relief point, providing accurate temporal feature input for subsequent pathological prediction.

[0016] Step S4: Based on the key trigger points and turning points in symptom evolution, perform potential pathology prediction and generate a potential pathology candidate set; In this embodiment, a pathological prediction feature engineering is constructed based on key trigger points and turning points in symptom evolution, extracting 84-dimensional temporal feature parameters, including the time interval between symptom onset, symptom duration, rate of change in symptom severity, and symptom combination patterns. A potential pathological prediction model is constructed using the Gradient Boosting Decision Tree (GBDT) algorithm, with a tree depth of 8, a learning rate of 0.1, a subsampling rate of 0.8, and a feature subsampling rate of 0.6. A medical knowledge graph mapping mechanism is established, containing 15,000 disease entities, 50,000 symptom entities, and 120,000 entity relationships. The TransE algorithm is used for knowledge representation learning, with an embedding dimension of 200. A symptom-disease association probability matrix is ​​constructed, and the point mutual information (PMI) method is used to calculate the association strength between symptoms and diseases. A PMI threshold of 2.0 or higher indicates a strong association. A Bayesian algorithm is then used to... A Yesian network is used for disease probability inference. The network contains 200 nodes and 500 directed edges, and a variational inference algorithm is used for probability calculation, with a convergence precision of 0.001. Multiple prediction model results are fused using an ensemble learning method, including random forest, support vector machine, and neural network, with weights set to 0.4, 0.3, and 0.3 respectively. Uncertainty quantification techniques are used to evaluate the confidence of the prediction results, employing the Monte Carlo dropout method with 1000 sampling iterations. A disease prediction probability threshold of 15% or higher is set for inclusion in the candidate set, while limiting the candidate set size to no more than 20 diseases, arranged in descending order of prediction probability. Finally, a potential pathology candidate set is generated, containing detailed information such as disease name, prediction probability, confidence interval, list of supporting symptoms, and key differential diagnostic points, providing precise diagnostic guidance for intelligent consultation.

[0017] Step S5: Generate intelligent consultation questions based on the potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

[0018] In this embodiment, a decision tree-based consultation strategy is constructed based on a potential pathological candidate set. Information gain ratio is used as the node splitting criterion, with a maximum tree depth of 10 layers and a minimum leaf node sample size of 5. The diagnostic value weight of each candidate disease is calculated, comprehensively considering disease probability, disease severity, and treatment urgency, with weight coefficients set to 0.5, 0.3, and 0.2, respectively. An analytic hierarchy process (AHP) is used to construct a question priority evaluation matrix, with the matrix dimension being the square of the number of candidate diseases, and the consistency ratio controlled within 0.1. A consultation question template library is established, containing 1500 standardized question templates covering four categories: symptom inquiry, medical history investigation, physical examination, and auxiliary examinations. Natural language generation technology is used to convert medical terminology into patient-understandable expressions, achieving a readability score at the 8th grade reading level. A dynamic consultation path planning algorithm is constructed, employing an A* search strategy for optimization. To improve consultation efficiency, a heuristic function is used, representing the ratio of the reduction in diagnostic uncertainty to the consultation time cost. A multi-turn dialogue context management mechanism is established, maintaining a 128-dimensional dialogue state vector containing information such as questions asked, patient responses, and diagnostic probability updates. A reinforcement learning algorithm is used to continuously optimize the consultation strategy, employing the Q-learning method with a learning rate of 0.01 and an initial exploration rate of 0.1 gradually decreasing to 0.01. A consultation quality evaluation index system is constructed, including three dimensions: diagnostic accuracy, consultation rounds, and patient satisfaction, with a target accuracy of over 85% and consultation rounds limited to 15 rounds. Finally, an adaptive intelligent consultation mechanism is built, which adjusts the question generation strategy in real time based on patient responses, achieving personalized and efficient intelligent consultation interaction. After the consultation, a structured diagnostic report is generated, including main diagnostic suggestions, a differential diagnosis list, and further examination suggestions.

[0019] In this embodiment, see Figure 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Obtain the patient's electronic medical record, which includes records of all medical visits, laboratory and examination reports, medication history, and auxiliary examination data; Data integrity is identified in the patient's sub-medical record files, data with missing structures is marked and adaptively removed to obtain anomaly-removed electronic medical records. Natural language processing is performed on the abnormal electronic medical records to identify symptom entities and obtain symptom entity information. Inter-entity relationships are extracted from symptom entity information to obtain semantic association features between entities; Based on the semantic association features between entities, semantic disambiguation and standard medical terminology conversion are performed to construct a personalized symptom knowledge graph; We analyze the implicit symptom descriptions of personalized symptom knowledge graphs from multiple dimensions to construct multi-dimensional medical record analysis information.

[0020] In this embodiment, after obtaining authorization from the hospital and the patient, previous medical records, laboratory reports, medication history, and auxiliary examination data are integrated from multiple information systems to form a unified data source. Data sources include Hospital Information System (HIS), Laboratory Information System (LIS, PACS), Drug Management System, and Follow-up Record Database. Standardized extraction of multi-source information is achieved through data interfaces based on the HL7 protocol and the FHIR (Fast Healthcare Interoperability Resources) standard. To achieve cross-system consistency, unique encrypted patient identifiers are used for matching and deduplication of each data source, ensuring accurate merging of records for the same patient across different systems. The collected data includes structured fields (such as laboratory indicators, examination items, and drug dosages) and unstructured text fields (such as chief complaint, present illness history, medical records, and medical orders). In the preprocessing stage, text data undergoes unified character encoding (UTF-8), abnormal symbols and noise removal, invalid or duplicate fields removal, and preliminary word segmentation and sentence segmentation. For numerical data, interval standardization methods are used to uniformly convert different units and dimensions. After collecting and preprocessing electronic medical record data, each sub-record needs to undergo structural integrity identification and anomaly removal. The main purpose is to eliminate low-quality medical records due to missing fields, non-standard formatting, or incomplete content, thereby improving the accuracy of subsequent analysis. Based on the "Basic Specifications for Electronic Medical Records (Trial)" and standard clinical medical record writing practices, structural field integrity detection rules are established, including key sections such as chief complaint, present illness, past medical history, physical examination, auxiliary examinations, diagnostic conclusions, and medication information. For text fields such as chief complaint and present illness, minimum character limits and mandatory field requirements are set; for laboratory reports, thresholds for the missing percentage of key indicators are set. When the missing percentage exceeds a set value (e.g., 30%), the record is marked as incomplete. For text content, a deep learning-based binary classification model is introduced to determine the overall structural integrity of the medical record text. A pre-trained language model is used to encode the text, and a classification layer outputs labels indicating whether the record is complete or not. The discrimination results are combined with the rule detection results to finally generate missing field labeling information. For medical records with a low percentage of missing fields, only the missing parts are removed; for medical records with a percentage of missing fields exceeding the threshold, the entire record is removed. After this process, the structural consistency and field completeness of the input dataset are effectively guaranteed.

[0021] This process employs a multi-layered NLP workflow to extract structured symptom entities from raw, unstructured medical record text. First, the text undergoes standardized word segmentation, part-of-speech tagging, and stop word filtering to provide a clear text structure for subsequent entity recognition. In the word segmentation stage, a BiLSTM-CRF model incorporating a medical terminology dictionary (derived from the National Medical Terminology Collection and Disease Classification Standards) is used to effectively identify professional symptom vocabulary and clinical expressions. Subsequently, a pre-trained Chinese medical BERT model is used to represent the text's features, and Symptom Named Entity Recognition (Symptom-NER) is performed based on sequence labeling. The labeling system uses the BIO scheme, categorizing symptom-related segments in the text into B-Symptom, I-Symptom, and O categories. The model identifies symptom phrases based on contextual information, including common symptoms, compound symptom descriptions, and implicit symptom expressions. The text is transformed into a structured set of symptom entity information, with each entity containing a symptom name, location of occurrence, contextual description, category label, and confidence score, laying the foundation for subsequent relation extraction and knowledge graph construction. This system utilizes relation extraction techniques to identify semantic connections between entities, including symptoms and locations, symptoms and properties, symptoms and time, and accompanying relationships between symptoms. Dependency parsing is performed on the text to obtain the grammatical dependency structures within sentences. By identifying subject-verb-object, modifier-head, and verb-complement dependencies, syntactic features are provided for subsequent relation extraction. Next, a bidirectional attention mechanism combined with a hybrid structure of Transformer and BiLSTM is employed to deeply represent the contextual information of entities and classify entity pairs into relations. Relation categories cover multiple dimensions, including symptom localization, property modification, time description, and accompanying symptoms. The output is represented in triplet form: (entity 1, relation, entity 2). During relation extraction, classification thresholds and confidence filtering mechanisms are set to ensure high accuracy of the extracted semantic relations. Finally, the system organizes all entity relations into a semantic association feature matrix, reflecting the multi-level semantic connections between entities.

[0022] This method achieves entity uniqueness and standardization by matching context vector representations with standard terminology databases. First, a pre-trained language model encodes the entity context, calculating the semantic similarity between the entity and standard terminology databases (including SNOMEDCT and MeSH-CN), using cosine similarity as the matching metric. A similarity threshold, such as 0.85, is set; entities exceeding this threshold are directly mapped to their corresponding standard terms. For entities below the threshold but with near-synonymous relationships, supplementary matching is performed using edit distance calculation and cluster analysis. After mapping, each entity is assigned a unique standard term code, achieving semantic consistency. Based on this, a personalized symptom knowledge graph is constructed and stored in a graph database (such as Neo4j). Node types in the graph include symptom, location, time, and nature, while edge types correspond to the semantic relationships extracted in the previous stage. Each node records attributes such as frequency of occurrence, contextual confidence, and timestamp. Graph mining techniques, such as Graph Neural Networks (GNNs), are used to propagate and aggregate features from the knowledge graph, identifying symptom information that is not directly mentioned but has strong semantic connections. For example, by observing the propagation of symptoms through multi-hop relationships such as location, nature, and time, implicit potential symptom clues can be inferred. The analytical dimensions include a temporal dimension (such as the time progression of symptom onset and disease evolution), a spatial dimension (the involved anatomical locations and spread), a semantic dimension (the nature and severity of symptoms), and a relational dimension (the mutual accompaniment and logical relationships between symptoms). After multi-dimensional analysis, medical record information is transformed into a structured output format, including entity lists, multi-dimensional feature descriptions, and potential diagnostic clues, providing a high-precision data foundation and semantic support for intelligent consultation, assisted diagnosis, and clinical decision-making.

[0023] In this embodiment, see Figure 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: The system identifies time-series nodes in the patient's electronic medical record, extracts the date of consultation, examination time, and medication cycle, and generates multiple key time nodes. Based on the analysis of multiple key time points, including the onset time of the disease, the period of symptom exacerbation, the period of remission, and the timing of key medical interventions, a standardized medical record timeline is constructed. Based on the standardized medical record timeline, multi-dimensional medical record parsing information is matched and mapped in a time sequence, and dynamic evolution visualization is performed to construct a time-series evolution medical record profile.

[0024] In this embodiment, standardized time elements are extracted from medical records, examination reports, medication orders, and medical progress records. These elements include the date of consultation, the specific time of various examinations, and the start and end times and cycles of medication. A regularized time recognition module is used to uniformly parse the time information appearing in the text, including both absolute time (e.g., "March 15, 2023") and relative time (e.g., "third day after admission," "one week ago"). For absolute time, it is directly converted to the standard ISO 8601 time format; for relative time, the date of consultation or admission is used as a baseline, and the corresponding absolute time point is calculated through rule matching and contextual reasoning. During the time information extraction process, various time expressions are also normalized; for example, "2023 / 03 / 15," "03-15-2023," and "March 15th" are uniformly converted to the same standard format. In structured medication information, the start time, discontinuation time, and duration of treatment are extracted from medical records based on prescriptions. Combined with the report time and test time fields from examination reports, an ordered set of multiple key time nodes is generated. Each node corresponds to a key event in the medical record (such as the first visit, first imaging examination, important laboratory test, medication start, symptom change record, etc.). All time nodes are sorted chronologically, and a baseline for the timeline is established, typically choosing the first visit time or the first recorded symptom onset time as the zero point. Subsequently, by calculating the time intervals between different nodes, the onset phase of the disease, the symptom exacerbation phase, the remission phase, and important medical intervention time points are identified. For example, analyzing the time expression of the chief complaint and present illness history can determine the first appearance time of symptoms; comparing the timestamps of changes in symptom description can determine the stage of symptom exacerbation; and using the test and medication record timestamps, key intervention events, such as the first use of antibiotics and the date of the first surgery, are marked. During the construction of the timeline, a time conflict detection and interpolation inference mechanism is employed to address situations where time discontinuities or conflicts exist. Abnormal time points are automatically adjusted or supplemented through contextual logic and conventional medical temporal relationships to ensure the integrity and logical consistency of the timeline. The resulting standardized medical record timeline uses a unified time scale as its benchmark, covering the entire process of disease occurrence, progression, intervention, and remission. Standardized labels and attributes (such as event type, symptom status, and intervention measures) are assigned to various event nodes on the timeline, laying the foundation for subsequent temporal mapping of multidimensional information.

[0025] All symptom entities, test results, medication information, and auxiliary examination results are matched with timeline nodes according to their occurrence time. For entities with precisely labeled times, a one-to-one mapping is performed directly; for entities lacking explicit time labels, the most likely corresponding time interval is inferred through a contextual time reasoning module. For example, "fever appeared on the second day after surgery" is mapped to the second time point after the surgery. After matching, the corresponding multidimensional information for each time node is integrated, including symptom changes, examination indicators, medication status, and medical interventions, forming a time-series hierarchical information structure. Subsequently, based on this structure, time series analysis and dynamic graph visualization technology are used to generate a time-series evolutionary medical record profile. In the profile, the timeline serves as the horizontal main line, and multidimensional symptom and event information is dynamically expanded as vertical layers. Through node size, color, connections, and dynamic changes, the entire process of disease occurrence, development, intervention, and relief is intuitively displayed. This temporal evolution profile can reflect the order in which symptoms appear, the correspondence between symptoms and examinations / medications, and the evolutionary patterns of key nodes in the disease development process, providing structured and visualized time-series support for intelligent consultation, clinical auxiliary diagnosis, and disease tracking.

[0026] In this embodiment, step S3 includes the following steps: Pathological trend analysis was performed on the temporal evolution of medical records at different time points, and trend features of medical records at multiple time points were extracted. Based on the trend characteristics of the medical records, the severity, duration, and combination patterns of accompanying symptoms are calculated to obtain a symptom co-occurrence matrix; Spatiotemporal correlation analysis of symptoms was performed on the symptom co-occurrence matrix to construct a multi-layer symptom node network; Multi-scale wavelet transform decomposition is performed on the multi-layer symptom node network, and the symptom evolution law is mined to identify key trigger points and turning points in symptom evolution.

[0027] In this embodiment, by longitudinally comparing and extracting features from medical record profiles at multiple key time points, the trend characteristics of symptoms, test indicators, and treatment interventions over time are identified. Specifically, the various symptoms, examinations, and medication events mapped on the timeline are first processed into time series, transforming information such as symptom frequency, changes in test values, and medication intensity into a feature sequence that changes over time. For example, for laboratory indicators, specific values ​​at each time point are extracted, and the rate of change, direction of increase / decrease, and stability are calculated; for symptom entities, the number of occurrences, duration, and new occurrence / remission ratios in different time periods are calculated; for treatment information, the start and end times and intensity changes of medication are extracted and time-aligned with symptom changes. Subsequently, a sliding window method is used to segment and statistically analyze the time series, extracting trend features within each time window, such as symptom upward trends, worsening trends of test indicators, and treatment intervention response trends. The width of the sliding window is adaptively set according to the disease type and record density; for example, a 3–5 day window can be selected for acute diseases, and a 1–2 week window can be selected for chronic diseases. Trend characteristics are quantified using indicators such as numerical change rate, trend slope, and temporal correlation, forming a set of pathological trend characteristics at multiple time points. The performance of each symptom in the time series is quantitatively described, including frequency of occurrence, duration of a single occurrence, overall disease duration, variability in occurrence, and correspondence with examination results. Severity calculation integrates symptom frequency, duration, and correlation with key examination indicators. For example, for symptoms like "dyspnea," the temporal changes in blood oxygen saturation can be combined to calculate the synchronous correlation coefficient between symptom and indicator changes, thereby quantifying the severity level. Subsequently, for multiple symptoms appearing within the same time window, an accompanying combination pattern is constructed to identify which symptoms have co-occurrence characteristics over time. By binarizing the occurrence of symptoms within the sliding time window, the co-occurrence frequency and proportion between different symptoms are calculated, constructing a symptom co-occurrence matrix. The elements of the matrix represent the frequency or weighted intensity of two symptoms occurring simultaneously within the same time period. To enhance robustness to sparse data, methods such as Pearson correlation coefficient, Jaccard similarity, or mutual information can be used to further weight co-occurrence relationships, so that the co-occurrence relationships between symptoms not only reflect the number of times they occur simultaneously, but also reflect the statistical correlation between symptoms.

[0028] The symptom co-occurrence matrix is ​​viewed as an adjacency matrix of a weighted undirected graph, where nodes represent symptoms and edge weights represent the temporal co-occurrence strength of two symptoms. Spatial dimension information is added, using the anatomical location or organ system corresponding to the symptoms as node attributes, thus simultaneously reflecting the temporal associations and spatial distribution characteristics of symptoms within the same network. To enhance the hierarchical nature of the analysis, the symptom network is divided into multiple layers: the bottom layer is the temporal co-occurrence layer, representing the pattern of symptoms co-occurring over time; the middle layer is the spatial anatomy layer, representing the distribution relationship of symptoms among different organ systems; and the top layer is the clinical semantic layer, representing the pathological or etiological correlations between symptoms. During network construction, a threshold filtering mechanism is used to remove edges with low weights and high random co-occurrence probabilities; for example, edges with weights less than twice the standard deviation of the overall mean can be deleted to highlight significant symptom associations. Through this multi-layered network structure, symptom association patterns in the temporal, spatial, and semantic dimensions can be simultaneously represented in a single graph, providing a fundamental graph structure support for subsequent dynamic evolutionary pattern mining. The time series data for each symptom node (e.g., frequency of occurrence, co-occurrence intensity curves) is input into the wavelet decomposition module. The Continuous Wavelet Transform (CWT) method is used to decompose the original time series into time-frequency components at different scales. Morlet wavelets or Mexican Hat wavelets are commonly used when selecting wavelet basis functions to achieve high resolution in both the time and frequency domains. Multi-scale decomposition effectively distinguishes between short-term sudden changes, periodic fluctuations, and long-term trends in symptoms. Local extremum detection and energy density analysis are used in the decomposition results to identify moments of significant change at specific time scales; these moments correspond to trigger points or turning points in symptom evolution. By comparing these key moments across different symptom nodes, the chronological order and causal relationships between symptoms can be further revealed, thereby uncovering potential patterns in disease progression.

[0029] In this embodiment, step S4 includes the following steps: Based on the key trigger points and turning points of symptom evolution, the frequency and temporal changes of symptoms are evaluated to obtain the temporal evaluation value of symptom evolution; Based on a pre-defined medical symptom spectrum, deep semantic matching is performed on the symptom evolution time sequence assessment values, and similarity is calculated to obtain a pathological similarity index. Based on the pathological similarity index, the correlation analysis of symptom combinations is performed, and potential pathological predictions are made to generate a potential pathological candidate set.

[0030] In this embodiment, the frequency, duration, and co-occurrence intensity of each symptom on the time axis are matched with identified trigger points and turning points to construct a complete symptom temporal evolution curve. For each symptom node, a set of time series features are defined, including: ① rate of change in attack frequency; ② rate of increase or decrease in symptom occurrence before and after the trigger point; ③ amplitude of fluctuation in duration; ④ change in co-occurrence intensity with other symptoms before and after the turning point; ⑤ evolutionary energy distribution at different time scales (derived from wavelet decomposition results). During the calculation process, sliding statistics are performed on a fixed-length time window (e.g., 5 days or one week) before and after the trigger point to calculate the difference, acceleration, and trend slope of symptom frequency, and the results are normalized to make the temporal features of different symptoms comparable. Subsequently, the multidimensional temporal feature vector of each symptom is comprehensively scored to form a symptom evolution temporal evaluation value. The evaluation value not only reflects the evolution intensity of the symptom in the overall time dimension but also reflects its change characteristics before and after key events, which can be used for subsequent matching with standard medical symptom spectra and calculation of pathological similarity. After obtaining the symptom evolution time-series assessment values, these assessment results need to be deeply semantically matched with a pre-defined medical symptom spectrum to calculate the pathological similarity index. The medical symptom spectrum is a standardized symptom time-series template library constructed based on standard clinical guidelines, disease knowledge bases (such as ICD-10 and SNOMED CT), and real clinical data, covering the symptom occurrence order, evolution trend, and temporal characteristics of various typical diseases. A unified semantic vectorization representation is applied to the symptom evolution time-series assessment value vector and the medical symptom spectrum template, encoding the time-series features and symptom semantic labels together into a high-dimensional vector. For example, pre-trained language models such as BERT can be used to encode symptom semantics, and temporal features can be represented using a Temporal CNN or bidirectional LSTM, ultimately concatenating them to form a composite feature representation vector. Subsequently, cosine similarity, Euclidean distance, or Dynamic Time Warping (DTW) methods are used to calculate the similarity between the patient's symptom evolution vector and the medical symptom spectrum vector. To improve matching accuracy, a symptom weight coefficient is introduced, giving higher matching weights to key symptoms (such as initial symptoms or high-risk symptoms). The calculation results are output as a pathological similarity index, which is used to quantitatively describe the degree of similarity between the patient's current symptom evolution pattern and the symptom spectrum of various standard diseases. The pathological similarity index can be set from 0 to 1. The higher the value, the closer it is to the typical evolution process of a certain disease, providing a reliable basis for potential pathological prediction.

[0031] Based on the similarity between the symptom spectrum of each disease and the patient's symptom evolution assessment value, diseases with similarity scores higher than a preset threshold (e.g., 0.75) are selected as an initial candidate set. For multiple candidate diseases, their corresponding symptom combination patterns are further analyzed. By constructing a symptom combination association graph, the symptom overlap and combination structure similarity between candidate diseases are quantified. Specifically, the core symptom set of each candidate disease is used as a node set, and the Jaccard similarity coefficient, mutual information, or Graph Edit Distance between the patient's actual symptom set and the disease's core symptom set are calculated to obtain the matching degree at the symptom combination level. At the same time, the order of symptom combinations in the time dimension is sorted and matched to identify patterns that are highly consistent with the disease spectrum in the symptom evolution order. When multiple diseases have similar symptom patterns, clustering or ranking algorithms are used to classify candidate diseases, prioritizing the output of disease types with high pathological similarity and high symptom combination matching degree. Finally, a potential pathological candidate set is generated, and a similarity score and symptom matching description are attached to each candidate pathology for diagnostic suggestions and subsequent clinical decision support in the intelligent consultation system.

[0032] In this embodiment, the specific steps of step S5 are as follows: The pathological diagnostic value is obtained by performing a step-by-step pathological diagnostic value analysis on the potential pathological candidate set; Based on the pathological diagnostic value, intelligent consultation questions are generated, resulting in multiple pathological consultation questions; Contextual questioning logic analysis was performed on multiple pathology questions, and priority adjustments were made to obtain a progressive question sequence; Personalized consultation style adaptation is performed based on a progressive question sequence to build an intelligent consultation mechanism; Real-time consultation and interaction are conducted based on an intelligent consultation mechanism, and data is uploaded to the cloud instantly.

[0033] In this embodiment, the standard symptom spectrum corresponding to the candidate pathology is extracted and compared with the patient's current symptom time-series assessment value to calculate the symptom evolution consistency score. Subsequently, the core diagnostic symptoms of the candidate pathology are matched with the patient's actual presentation, calculating coverage and matching degree. For example, binary matching is used for essential symptoms, and weighted matching is used for non-essential but important symptoms. For candidate pathologies involving examination and test results, a consistency score with typical pathological features is calculated through matching analysis of key experimental indicators and imaging features. Finally, information such as symptom similarity, evolutionary trend similarity, examination consistency, and feature weights are integrated, and the pathological diagnostic value is obtained through weighted summation or a logistic regression-based scoring model. The higher the score, the closer the candidate pathology is to the patient's current overall clinical presentation, and the higher its diagnostic value during the consultation process. Key symptoms, signs, and examination items that have not been fully validated are extracted from the standard diagnostic process of the candidate pathology. For example, if the core diagnostic criteria for a certain pathology include specific accompanying symptoms, but these are not recorded in existing medical records or have low confidence, the system uses these symptoms as the consultation target. Subsequently, based on a medical consultation template library and a semantic generation model (such as a Seq2Seq or T5 pre-trained model), targeted questions are automatically generated, including closed-ended questions (such as "Have you recently experienced nighttime chest pain?") and open-ended questions (such as "Please describe your recent discomfort"). During the generation process, entity recognition and grammatical rules ensure the accuracy and medical rigor of the question content, while a language model is used to adjust the sentence structure to make the questions more suitable for human-computer interaction scenarios.

[0034] All generated questions undergo semantic parsing and logical classification, categorizing them into five main types: symptom confirmation, disease progression, accompanying symptoms, supplementary examinations, and diagnosis exclusion. Subsequently, a dependency graph is constructed using contextual logical dependency analysis. For example, symptom confirmation questions are prioritized, accompanying symptom questions are asked after core symptom confirmation, and exclusion questions typically appear later in the consultation process. Logical relationship analysis is implemented using dependency syntax trees and semantic relation extraction techniques to extract "precedence-follow-up" relationships and semantic dependencies. Then, questions are initially prioritized based on pathological diagnostic value, with higher-value pathologies having higher priority. Within the same pathology, the order is further adjusted based on question category and logical relationships, forming a progressive sequence from core symptom confirmation → disease progression refinement → accompanying symptom expansion → supplementary examinations → exclusion verification. Conflicting or duplicate questions are merged or deleted using a redundancy detection module to reduce consultation interference and improve interaction efficiency. The consultation style is adaptively selected using the patient's basic information, past consultation records, and language features. The consultation style is divided into several modes, such as concise, detailed explanation, reassuring, and guiding, each differing in question expression, question-and-answer rhythm, and interaction depth. During the construction of the consultation mechanism, the system automatically adapts the expression of progressive question sequences, such as converting medical terminology into easily understandable language or breaking down complex questions into multiple short sentences to improve patient comprehension. Simultaneously, the consultation rhythm is adjusted in real-time based on the patient's response speed, keyword usage, and semantic structure to ensure smooth interaction. For specific diseases or elderly patient scenarios, a voice interaction assistance mechanism can be introduced to provide auditory feedback and guidance. The progressive question sequence is presented to the patient step-by-step according to a predetermined logic, with interaction methods including text input, speech recognition, and multimodal input, supporting both mobile and PC applications. The patient's answers are processed by the natural language understanding module, extracting symptom entities, time information, and negative expressions, and updating the structured medical record data in real-time. When the answer information involves existing symptom nodes, its confidence level is dynamically adjusted; when new symptom information appears, the symptom set is automatically expanded and time is annotated. Subsequently, structured information, raw semantic data, and interaction logs generated during the consultation process are uploaded to the cloud server in real time via a cloud data interface, facilitating subsequent disease evolution modeling and cross-institutional information sharing. The upload employs encrypted transmission and access control to ensure data security and privacy compliance. The entire process achieves a closed loop from candidate pathological diagnostic value analysis to question generation, logical organization, personalized presentation, and real-time information updates, providing an efficient, accurate, and dynamic information collection and processing mechanism for the intelligent consultation system.

[0035] In this embodiment, the specific steps for real-time consultation interaction based on the intelligent consultation mechanism and instant cloud upload are as follows: Real-time consultation and interaction are conducted based on an intelligent consultation mechanism to obtain adaptive consultation and interaction results. Based on the adaptive consultation interaction results, multi-model fusion diagnosis is performed, and a preliminary diagnostic analysis report is output. The preliminary diagnostic analysis report is uploaded to the cloud in real time, enabling doctors to access and evaluate it in real time.

[0036] In this embodiment, through multiple rounds of interaction with the patient, the system dynamically collects information on symptoms, disease course, medication, and related examinations. The collected results undergo real-time semantic understanding and adaptive adjustment to form an adaptive consultation interaction result. Specifically, the system first initiates a consultation with the patient one question at a time according to a progressive question sequence, supporting text, voice, and multimodal input to improve the flexibility of information collection. The patient's answers are processed by the natural language understanding module, including word segmentation, named entity recognition, temporal expression extraction, negation recognition, and semantic relationship analysis, converting unstructured answers into standardized structured medical record information. During this process, a dynamic confidence update mechanism adjusts the confidence of each symptom node or examination information in real time; for example, it increases the confidence of repeatedly confirmed information and decreases the confidence of ambiguous or contradictory information. If the patient's answer involves unpredictable symptoms or logical jumps, the system dynamically inserts supplementary questions based on context and semantic reasoning mechanisms to achieve adaptive adjustment, thereby ensuring the integrity and consistency of the information. Simultaneously, the system statistically analyzes patient response time, semantic complexity, and keyword density during the interaction process. These parameters are used to dynamically optimize the difficulty and pace of subsequent questions, ensuring both medical rigor and a positive interactive experience throughout the consultation. Combining the advantages of multiple medical intelligent models, model integration and fusion techniques are employed to improve diagnostic accuracy and robustness. First, the symptoms, time, disease course, and examination data in the interaction results are standardized and matched with a medical knowledge base and a pathology standard database to form feature vectors for model input. Subsequently, multiple heterogeneous diagnostic models are used for parallel data analysis. These models include rule-based diagnostic engines (such as knowledge graph reasoning based on ICD-10 and clinical guidelines), deep learning-based disease classification models (such as multi-layer neural networks or Transformer structures), and probabilistic reasoning models based on Bayesian networks. Each model outputs a corresponding list of disease candidates and their confidence scores. To fully leverage the complementarity of each model, multi-model fusion strategies are introduced, such as weighted averaging, Stacking ensemble, or Dempster-Shafer evidence theory, to fuse the outputs of each model and generate more stable and medically consistent preliminary diagnostic results. During the fusion process, specific weights are assigned. For example, knowledge graph models are given high weights to ensure medical reliability, deep learning models are given relatively high weights to improve the ability to identify complex symptoms, and Bayesian models are given moderate weights to handle uncertainty. The fused output includes a disease candidate set, confidence intervals, key symptom matching descriptions, and potential examination suggestions, ultimately forming a structured preliminary diagnostic analysis report with high interpretability and clinical application value.

[0037] The report content is structured in a hierarchical manner, including basic patient information, symptom time-series analysis results, model fusion diagnostic output, confidence scores, and suggested examinations. It is encapsulated using standard medical data exchange formats (such as the FHIR standard) to ensure compatibility with Hospital Information Systems (HIS) and Electronic Medical Records (EMR) systems. Subsequently, the report is uploaded to a cloud server via a secure cloud communication protocol (such as HTTPS with two-way authentication or a VPN tunnel), with data encryption and access control to ensure patient privacy and medical data security. After upload, doctors can access the report in real time through a dedicated cloud-based doctor's workbench to view the symptom evolution process, multi-model fusion diagnostic results, and system suggestions. The doctor's interface supports interactive browsing; for example, clicking on a symptom node displays the corresponding original medical history and confidence change curve, while clicking on a disease candidate shows the contribution and inference path of different models to that candidate. Doctors can also annotate, correct, or supplement the report; corrections are synchronized back to the system database in real time, forming a two-way feedback mechanism that helps in continuous model optimization. Through this mechanism, the intelligent consultation system not only automates and standardizes the diagnostic process, but also provides doctors with efficient auxiliary decision-making tools, achieving a deep integration of intelligent consultation and clinical practice.

[0038] In this embodiment, a medical record analysis and intelligent consultation device is provided for performing the medical record analysis and intelligent consultation method described above, including: The medical record parsing module is used to obtain patients' electronic medical records, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record parsing information. The dynamic evolution module is used to perform time-series matching and mapping based on multi-dimensional medical record parsing information, and to visualize the dynamic evolution to build a time-series evolution medical record profile. The symptom evolution mining module is used to perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and to mine the symptom evolution pattern, identifying key trigger points and turning points in symptom evolution. The potential pathology prediction module is used to predict potential pathologies based on key trigger points and turning points in symptom evolution, and generate a potential pathology candidate set. The intelligent consultation module is used to generate intelligent consultation questions based on a potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

[0039] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0040] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for medical record analysis and intelligent consultation, characterized in that, Includes the following steps: Step S1: Obtain the patient's electronic medical record, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record analysis information; Step S2: Perform time-series matching and mapping based on multi-dimensional medical record parsing information, and visualize the dynamic evolution to construct a time-series evolution medical record profile; Step S3: Perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and mine the symptom evolution pattern to identify key trigger points and turning points in symptom evolution. Step S4: Based on the key trigger points and turning points in symptom evolution, perform potential pathology prediction and generate a potential pathology candidate set; Step S5: Generate intelligent consultation questions based on the potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

2. The medical record analysis and intelligent consultation method according to claim 1, characterized in that, The specific steps of step S1 are as follows: Obtain the patient's electronic medical record, which includes records of all medical visits, laboratory and examination reports, medication history, and auxiliary examination data; Data integrity is identified in the patient's sub-medical record files, data with missing structures is marked and adaptively removed to obtain anomaly-removed electronic medical records. Natural language processing is performed on the abnormal electronic medical records to identify symptom entities and obtain symptom entity information. Inter-entity relationships are extracted from symptom entity information to obtain semantic association features between entities; Based on the semantic association features between entities, semantic disambiguation and standard medical terminology conversion are performed to construct a personalized symptom knowledge graph; We analyze the implicit symptom descriptions of personalized symptom knowledge graphs from multiple dimensions to construct multi-dimensional medical record analysis information.

3. The medical record analysis and intelligent consultation method according to claim 1, characterized in that, The specific steps of step S2 are as follows: The system identifies time-series nodes in the patient's electronic medical record, extracts the date of consultation, examination time, and medication cycle, and generates multiple key time nodes. Based on the analysis of multiple key time points, including the onset time of the disease, the period of symptom exacerbation, the period of remission, and the timing of key medical interventions, a standardized medical record timeline is constructed. Based on the standardized medical record timeline, multi-dimensional medical record parsing information is matched and mapped in a time sequence, and dynamic evolution visualization is performed to construct a time-series evolution medical record profile.

4. The medical record analysis and intelligent consultation method according to claim 1, characterized in that, Step S3 is as follows: Pathological trend analysis was performed on the temporal evolution of medical records at different time points, and trend features of medical records at multiple time points were extracted. Based on the trend characteristics of the medical records, the severity, duration, and combination patterns of accompanying symptoms are calculated to obtain a symptom co-occurrence matrix; Spatiotemporal correlation analysis of symptoms was performed on the symptom co-occurrence matrix to construct a multi-layer symptom node network; Multi-scale wavelet transform decomposition is performed on the multi-layer symptom node network, and the symptom evolution law is mined to identify key trigger points and turning points in symptom evolution.

5. The medical record analysis and intelligent consultation method according to claim 1, characterized in that, The specific steps of step S4 are as follows: Based on the key trigger points and turning points of symptom evolution, the frequency and temporal changes of symptoms are evaluated to obtain the temporal evaluation value of symptom evolution; Based on a pre-defined medical symptom spectrum, deep semantic matching is performed on the symptom evolution time sequence assessment values, and similarity is calculated to obtain a pathological similarity index. Based on the pathological similarity index, the correlation analysis of symptom combinations is performed, and potential pathological predictions are made to generate a potential pathological candidate set.

6. The medical record analysis and intelligent consultation method according to claim 1, characterized in that, The specific steps of step S5 are as follows: The pathological diagnostic value is obtained by performing a step-by-step pathological diagnostic value analysis on the potential pathological candidate set; Based on the pathological diagnostic value, intelligent consultation questions are generated, resulting in multiple pathological consultation questions; Contextual questioning logic analysis was performed on multiple pathology questions, and priority adjustments were made to obtain a progressive question sequence; Personalized consultation style adaptation is performed based on a progressive question sequence to build an intelligent consultation mechanism; Real-time consultation and interaction are conducted based on an intelligent consultation mechanism, and data is uploaded to the cloud instantly.

7. The medical record analysis and intelligent consultation method according to claim 6, characterized in that, The specific steps for real-time consultation interaction based on the intelligent consultation mechanism and instant cloud upload are as follows: Real-time consultation and interaction are conducted based on an intelligent consultation mechanism to obtain adaptive consultation and interaction results. Based on the adaptive consultation interaction results, multi-model fusion diagnosis is performed, and a preliminary diagnostic analysis report is output. The preliminary diagnostic analysis report is uploaded to the cloud in real time, enabling doctors to access and evaluate it in real time.

8. A medical record analysis and intelligent consultation device, characterized in that, For performing the medical record analysis and intelligent consultation method as described in claim 1, including: The medical record parsing module is used to obtain patients' electronic medical records, perform multi-dimensional analysis of implicit symptom descriptions, and construct multi-dimensional medical record parsing information. The dynamic evolution module is used to perform time-series matching and mapping based on multi-dimensional medical record parsing information, and to visualize the dynamic evolution to build a time-series evolution medical record profile. The symptom evolution mining module is used to perform multi-scale wavelet transform decomposition on the temporal evolution medical record profile, and to mine the symptom evolution pattern, identifying key trigger points and turning points in symptom evolution. The potential pathology prediction module is used to predict potential pathologies based on key trigger points and turning points in symptom evolution, and generate a potential pathology candidate set. The intelligent consultation module is used to generate intelligent consultation questions based on a potential pathology candidate set, adjust priorities, and build an intelligent consultation mechanism.

Images