A medical data intelligent analysis method and system based on authorized boundary constraint

Abstract

The present application relates to the technical field of natural language processing, and more particularly to a medical data intelligent analysis method and system based on authorized boundary constraint, comprising: parsing the input business requirement text and authorization information, generating a structured analysis triple containing an analysis target, a dimension set and a data constraint set by using a preset language model and a conditional random field layer; filtering and extracting the original analysis data set according to the authorization information and the data constraint set layer, and constructing a comprehensive cleaning loss function to perform adaptive closed-loop cleaning, and outputting the available analysis data set; calculating the score according to the analysis target to select the optimal model, using the Bayesian optimization algorithm to perform hyperparameter iteration optimization, and calculating the output model analysis result set; matching the report template, calling the content splicing operator and the knowledge base to cascade and automatically fill the analysis results, generating a structured analysis report and outputting. The present application ensures that the medical data calculation does not break the authorized red line, and improves the report accuracy.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for intelligent analysis of medical data based on authorization boundary constraints. Background Technology

[0002] With the rapid development of medical big data, leveraging data analytics to uncover the commercial value and clinical decision-making evidence within heterogeneous medical data has become a crucial means for medical institutions and related enterprises to enhance their core competitiveness. Currently, when processing users' natural language business needs, existing technologies typically employ general-purpose Natural Language Processing (NLP) models or large-scale language models for entity recognition and intent parsing. After initial parsing, the system matches extracted keywords with a backend database and filters the retrieved raw medical dataset using general blind data cleaning algorithms. Finally, data scientists or engineers manually select models and fine-tune hyperparameters based on their subjective experience from a model library. Ultimately, the nonlinear computational data output by the model is mechanically and inefficiently populated into a pre-defined chart generation tool, producing an analysis report for user reference.

[0003] When faced with high-threshold medical professional terminology, existing technologies lack specific knowledge support in the medical field, making it easy for general models to generate semantic ambiguity. Furthermore, colloquial abbreviations cannot be standardized and mapped in a standardized manner, resulting in the system's inability to accurately capture and break down the real business needs of front-end users. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a medical data intelligent analysis method and system based on authorization boundary constraints, aiming to improve the problem that the system cannot accurately capture and break down the real business needs of front-end users.

[0005] In a first aspect, the present invention provides the following technical solution: a method for intelligent analysis of medical data based on authorization boundary constraints, comprising: S1. Receive the input natural language business requirement text and the corresponding user data authorization information, and use a preset language model and conditional random field layer to perform structured parsing on the natural language business requirement text to generate a structured parsing triple containing the analysis target, the set of analysis dimensions and the set of data constraints. S2. Based on the user data authorization information and the data constraint set, filter and extract the original analysis dataset from the full medical dataset, construct a comprehensive cleaning loss function, and adaptively clean the original analysis dataset by minimizing the comprehensive cleaning loss function to output a usable analysis dataset. S3. Calculate the matching score between each model in the medical-specific model library and the analysis target, select the model with the highest score as the optimal analysis model, and use the Bayesian optimization algorithm to adaptively iteratively tune the hyperparameters of the optimal analysis model. Use the tuned optimal analysis model to calculate the available analysis dataset and output the model analysis result set. S4. Match the business scenario report template according to the analysis target, call the content splicing operator and the preset medical business knowledge base, automatically fill the model analysis result set into the business scenario report template, generate a structured analysis report and output it.

[0006] Preferably, in S1, the structured parsing of the natural language business requirement text includes: The preset language model is used to perform semantic encoding on natural language business requirement text to obtain context vector representations at the marker word level; The context vector representation is input into the conditional random field layer, and path optimization identification is performed in combination with the preset entity label transition probability matrix to obtain the optimal medical entity label sequence. The corresponding medical entities are captured based on the optimal medical entity label sequence, and the structured parsed triples are generated by assembling parameters according to the semantic relationships between the medical entities.

[0007] Preferably, in S2, the raw analysis dataset is extracted from the full medical dataset by filtering: Based on the user data authorization information, the full medical dataset is subjected to boundary filtering to generate a unique authorized dataset; Within the internal computing space of the authorized dataset, multidimensional feature retrieval and condition extraction are performed based on the data constraint set to obtain the original analysis dataset that meets business requirements without exceeding the computing boundaries.

[0008] Preferably, in S2, adaptive cleaning of the original analysis dataset includes: Based on the data defect type of the original analysis dataset, the weighting coefficients of each loss term in the comprehensive cleaning loss function are automatically initialized; Calculate the current cleaning loss value after each cleaning iteration; Determine whether the cleaning loss value is lower than a preset threshold; if the cleaning loss value is not lower than the preset threshold, adjust the weighting coefficient according to the cleaning loss value and re-clean until the cleaning loss value is lower than the preset threshold and terminate the iteration.

[0009] Preferably, in S2, the adaptive cleaning of the original analysis dataset further includes: Real-time monitoring of missing data types in the original analysis dataset; When the missing data is determined to be a key medical diagnosis result, a preset medical knowledge graph is invoked, and the key medical diagnosis result is filled in by reasoning through the logical reasoning chain of associating diseases with common medical knowledge.

[0010] Preferably, in S2, the adaptive cleaning of the original analysis dataset further includes: The original analysis dataset was filtered using the statistical principle of three standard deviations to identify undetermined outlier data. The pending abnormal data is compared and verified with the preset hard rules of reasonable value range in the medical field; The undetermined abnormal data is marked as an abnormal value and corrected or deleted only if it exceeds the hard rule of reasonable value range in the medical field.

[0011] Preferably, in S3, the matching score between each model in the medical-specific model library and the analysis target includes: Extract the feature vectors of each model in the medical-specific model library. The feature vectors include the applicable scenarios of the model, the model complexity, and historical performance characteristics. The feature vectors of each model are compared with the analysis target to obtain the comparison feature values; By using the feature weights obtained through prior training with historical analysis data, the feature values ​​of each comparison are weighted and summed to obtain the matching score of each model, and the model with the highest matching score is selected as the optimal analysis model.

[0012] Preferably, in S3, the adaptive iterative tuning of the hyperparameters of the optimal analysis model using the Bayesian optimization algorithm includes: Within the preset hyperparameter search space, the available analysis dataset is divided into a training set and a validation set; The performance score of the current hyperparameter combination is predicted using a Gaussian process prior model, and the optimal hyperparameter combination is obtained by maximizing the performance evaluation index of the optimal analysis model on the validation set during the iterative optimization process. The optimal analysis model with the best combination of hyperparameters is used to train and compute on the available analysis dataset, and outputs a multi-dimensional set of model analysis results.

[0013] Preferably, in S4, the process of automatically populating the business scenario report template with the called content concatenation operator and the preset medical business knowledge base, and the model analysis result set, includes: Parse the business scenario report template and extract the content placeholders and their corresponding placeholder types contained therein; By using the content splicing operator, the nonlinear calculation conclusions in the model analysis result set are concatenated with the industry common sense decision rules extracted from the medical business knowledge base to generate the target text block. Based on the placeholder type, the target text block and the data charts automatically drawn from the model analysis results are collaboratively assembled into the corresponding content placeholders to generate a structured analysis report and output it.

[0014] Secondly, the present invention provides the following technical solution: a medical data intelligent analysis system based on authorization boundary constraints, the system comprising: The structured parsing module is used to receive the input natural language business requirement text and the corresponding user data authorization information, and to perform structured parsing on the natural language business requirement text using a preset language model and a conditional random field layer to generate a structured parsing triple containing the analysis target, the set of analysis dimensions, and the set of data constraints. The data filtering and cleaning module is used to filter and extract the original analysis dataset from the full medical dataset based on the user data authorization information and the data constraint set, construct a comprehensive cleaning loss function, and adaptively clean the original analysis dataset by minimizing the comprehensive cleaning loss function to output a usable analysis dataset. The model selection and optimization module is used to calculate the matching score between each model in the medical-specific model library and the analysis target, select the model with the highest score as the optimal analysis model, and use the Bayesian optimization algorithm to adaptively and iteratively optimize the hyperparameters of the optimal analysis model. The optimized optimal analysis model is then used to calculate the available analysis dataset and output a set of model analysis results. The report population and output module is used to match a business scenario report template according to the analysis target, call the content splicing operator and the preset medical business knowledge base, automatically populate the model analysis result set into the business scenario report template, generate a structured analysis report and output it.

[0015] The present invention has the following beneficial effects: 1. In this invention, by constructing a medical-specific ontology knowledge base covering five core entities, and cascading a finely tuned semantic parsing model for feature parsing and variable slot filling, the end-to-end automatic decomposition and precise mapping of unstructured business requests to ternary standard structured parameters is achieved. This solves the problem that general large models lack medical scenario knowledge support, resulting in the inability to standardize and normalize the mapping of ambiguous terms, abbreviations, and colloquial texts, thus failing to accurately capture users' real business requests.

[0016] 2. In this invention, by embedding authorization information as a core parameter into the algorithm's computation flow, and cascading "statistical + medical rules" dual verification and knowledge graph reasoning cleaning operators within the closed authorization boundary, fully automatic adaptive iterative cleaning and loss closed-loop control of multi-source medical heterogeneous missing values ​​and outliers are achieved. This solves the problems of traditional data analysis being prone to exceeding authority due to the post-authorization verification, as well as the common blind cleaning algorithm causing the accidental deletion of outlier feature points with high medical value such as rare diseases, and the introduction of mean noise that destroys data fidelity.

[0017] 3. In this invention, by establishing a dedicated model library covering five major medical scenarios and using a cascaded matching degree feature weighting algorithm to locate the optimal model, and combined with a Bayesian optimization algorithm to perform parameter tuning driven by validation set performance indicators, end-to-end automated high-precision adaptation of medical analysis model selection and hyperparameter search space is achieved. This solves the problem that traditional medical analysis relies heavily on engineers' experience for manual selection, and that hyperparameters are prone to getting stuck in local optima or oscillations when facing highly nonlinear medical datasets, leading to biased final model calculation conclusions.

[0018] 4. In this invention, by matching a preset report template to perform streaming placeholder parsing, and calling the content splicing operator to perform context disambiguation and sentence restructuring of the nonlinear calculation conclusions and medical knowledge base rules, and finally using a graphics rendering engine to perform multimodal content collaborative assembly, a closed-loop automatic generation of high-value structured reports that combine data charts, scientific conclusions and actionable suggestions is achieved. This solves the problem that traditional analysis reports can only perform simple and mechanical data piling, resulting in a serious disconnect between the content and the user's decision-making needs, and users still need to spend a lot of manpower to perform secondary processing before they can be used for decision-making. Attached Figure Description

[0019] Figure 1 This is a flowchart of a medical data intelligent analysis method based on authorization boundary constraints proposed in this invention; Figure 2 This is a flowchart illustrating the data filtering, extraction, adaptive cleaning, and refinement process of a medical data intelligent analysis method based on authorization boundary constraints proposed in this invention. Figure 3 This is an architecture diagram of a medical data intelligent analysis system based on authorization boundary constraints proposed in this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: In a first embodiment of the present invention, the present invention provides a method for intelligent analysis of medical data based on authorization boundary constraints, such as... Figures 1-2 As shown, it includes the following steps: S1. Receive the input natural language business requirement text and the corresponding user data authorization information, and use a preset language model and conditional random field layer to perform structured parsing on the natural language business requirement text to generate a structured parsing triple containing the analysis target, the set of analysis dimensions and the set of data constraints. Furthermore, in S1, the structured parsing of the natural language business requirement text includes: The preset language model is used to perform semantic encoding on natural language business requirement text to obtain context vector representations at the marker word level; The context vector representation is input into the conditional random field layer, and path optimization identification is performed in combination with the preset entity label transition probability matrix to obtain the optimal medical entity label sequence. The corresponding medical entities are captured based on the optimal medical entity label sequence, and the structured parsed triples are generated by assembling parameters according to the semantic relationships between the medical entities.

[0022] Specifically, after receiving the input natural language business requirement text and corresponding user data authorization information, the computer device first invokes a pre-built medical-specific ontology knowledge base. This knowledge base serves as a pre-defined rule support and includes disease ontology, drug ontology, regional ontology, time ontology, and business scenario ontology. When the data processing stream enters the semantic encoding stage, the computer device uses a medical semantic encoding model obtained by fine-tuning the MedBERT model using a corpus of millions of medical domain data as the preset language model.

[0023] After inputting natural language business requirement text into the medical semantic encoding model, word embeddings are calculated using its internal multi-layer bidirectional Transformer architecture. In each encoding layer, a multi-head self-attention mechanism transforms the token sequence into a query matrix, a key matrix, and a value matrix. Global contextual semantic weights are obtained by calculating the dot product of the query matrix and the key matrix, and then multiplied by the value matrix to capture the complex nested relationships of medical entities and long-distance semantic dependencies in the medical text. Finally, the model outputs a fixed-dimensional continuous dense vector for each token at the output layer, thus obtaining a token-level context vector representation containing specific semantic features of the medical scenario.

[0024] Subsequently, the aforementioned context vector representation is seamlessly passed to the Conditional Random Field (CRF) layer as a feature vector. The CRF layer internally maintains a pre-defined entity label transition probability matrix learned through training on historical medical text samples. In the specific sequence labeling calculation, the computer inputs the context vector representation into a linear mapping layer to calculate the emission score of the current labeled word as belonging to a specific medical ontology label such as disease, drug, region, time, or business scenario.

[0025] The Conditional Random Field (CRF) layer performs global probabilistic causal constraint calculations within the full path search space by combining the emission score with the state transition probabilities in the entity tag transition probability matrix. The conditional probability calculation formula constructed by the CRF layer is as follows: ; in, Indicates a given sequence of tokens Candidate label sequence under the condition The conditional probability; This represents the total number of marker words. In this embodiment, the maximum text segmentation length is set to a fixed constant of 128 marker words. Indicates the internal position number of the current marker word; Indicates the first Predicted entity labels corresponding to each location; Indicates the first Predicted entity labels corresponding to each location; This indicates that the entity label transition probability matrix represents the transition from label... Transfer to label The transition probability value; The first term, calculated based on the context vector representation, represents the... Output label at each position The launch score; It represents the normalized partition function for all possible candidate label sequences.

[0026] The computer device uses the Viterbi algorithm for dynamic programming optimization. Starting from the first labeled word position, the algorithm accumulates the emission score of the current state and the transition score from the previous state to the current state layer by layer, and records the local optimal path pointer. When it reaches the last labeled word position, it backtracks to the optimal path pointer, thereby obtaining the optimal medical entity label sequence in the entire unstructured text.

[0027] After obtaining the optimal medical entity label sequence, the data processing flow enters the entity capture and normalized triple assembly stage. The computer device extracts continuous marker word segments from the text based on the boundary markers in the optimal medical entity label sequence, thereby accurately capturing the corresponding medical entities. At this point, the computer device, in conjunction with the aforementioned medical-specific ontology knowledge base, performs synonym normalization mapping and classification on the captured medical entities, mapping identified non-standard disease abbreviations to standard disease ontology codes, and mapping colloquial time descriptions to standard timestamp ranges.

[0028] After standardization, the computer equipment fills multi-dimensional variable slots based on the semantic relationships of these entities in the business scenario. Specifically, the normalized disease ontology entity is filled into the variable slot of the analysis target as the core analytical requirement; the geographic ontology entity, time ontology entity, and drug ontology entity are merged into the variable slot of the analysis dimension set as discrete element sets; and the hard rules of specific conditions that limit the scope of analysis are converted into structured Boolean logic or interval key-value pairs and placed into the variable slot of the data constraint set. Through this topological assembly, a structured parsed triple consisting of the analysis target, the analysis dimension set, and the data constraint set is finally generated, providing accurate and structured control data for downstream data filtering and cleaning processes based on authorization boundaries.

[0029] S2. Based on the user data authorization information and the data constraint set, filter and extract the original analysis dataset from the full medical dataset, construct a comprehensive cleaning loss function, and adaptively clean the original analysis dataset by minimizing the comprehensive cleaning loss function to output a usable analysis dataset. Furthermore, in S2, filtering and extracting the raw analysis dataset from the full medical dataset includes: Based on the user data authorization information, the full medical dataset is subjected to boundary filtering to generate a unique authorized dataset; Within the internal computing space of the authorized dataset, multidimensional feature retrieval and condition extraction are performed based on the data constraint set to obtain the original analysis dataset that meets business requirements without exceeding the computing boundaries.

[0030] Furthermore, in S2, the adaptive cleaning of the original analysis dataset includes: Based on the data defect type of the original analysis dataset, the weighting coefficients of each loss term in the comprehensive cleaning loss function are automatically initialized; Calculate the current cleaning loss value after each cleaning iteration; Determine whether the cleaning loss value is lower than a preset threshold; if the cleaning loss value is not lower than the preset threshold, adjust the weighting coefficient according to the cleaning loss value and re-clean until the cleaning loss value is lower than the preset threshold and terminate the iteration.

[0031] Furthermore, in S2, the adaptive cleaning of the original analysis dataset also includes: Real-time monitoring of missing data types in the original analysis dataset; When the missing data is determined to be a key medical diagnosis result, a preset medical knowledge graph is invoked, and the key medical diagnosis result is filled in by reasoning through the logical reasoning chain of associating diseases with common medical knowledge.

[0032] Furthermore, in S2, the adaptive cleaning of the original analysis dataset also includes: The original analysis dataset was filtered using the statistical principle of three standard deviations to identify undetermined outlier data. The pending abnormal data is compared and verified with the preset hard rules of reasonable value range in the medical field; The undetermined abnormal data is marked as an abnormal value and corrected or deleted only if it exceeds the hard rule of reasonable value range in the medical field.

[0033] Specifically, after receiving the structured parsed triplet, the computer device automatically initiates a data filtering and extraction process based on authorization boundaries. The entire data processing flow first enters a two-stage hierarchical extraction phase. The computer device reads the input user data authorization information, extracts the list of purchased medical data products, the authorization time range, and the authorized disease range as hard calculation boundaries, and performs global pointer pruning on the full medical dataset to generate a dedicated authorization dataset.

[0034] Subsequently, the computer device parses the set of data constraints in the structured parsing triples, transforming them into multidimensional feature retrieval conditions composed of specific disease labels and numerical ranges. The computer device then injects these multidimensional feature retrieval conditions into the internal computational space of the authorized dataset for bounded retrieval and condition extraction, ultimately obtaining the original analysis dataset completely confined to the scope of the user data authorization information, thus preventing unauthorized calculations from the outset.

[0035] After acquiring the original analysis dataset, the data processing flow enters the adaptive cleaning and loss control stage. The computer scans the original analysis dataset to be cleaned and constructs a comprehensive cleaning loss function based on the identified data defect types. This comprehensive cleaning loss function is constructed by a weighted sum of missing value loss, outlier loss, and redundant data loss, and its algebraic calculation formula is as follows: ; in, This represents the total loss value of the comprehensive cleaning loss function; This represents the missing value loss, used to characterize the degree of data loss. This represents the outlier loss used to characterize the degree of data anomaly. This represents the loss of redundant data used to characterize the degree of data duplication; These represent the weighting coefficients for the corresponding loss terms. The computer equipment automatically initializes the weighting coefficients for each loss term based on the data defect types in the original analysis dataset.

[0036] After each data cleaning iteration is completed, the computer equipment calculates the current cleaning loss value in real time. And compare it with the preset convergence threshold. If If the value is not lower than the preset threshold, it indicates that the cleaning quality has not yet met the standard. The computer equipment automatically adjusts the weighting coefficients based on the loss gradient and the data characteristics of the current dataset. The value can be adjusted, for example, by increasing the value when the weight of missing data is high. The data stream is then redirected back to the cleaning iteration operator for further cleaning, until... The iteration terminates when the value falls below a preset threshold, and a usable analysis dataset is output.

[0037] In the iterative process of the aforementioned adaptive cleaning, the loss due to missing values ​​is addressed. During the calculation and imputation process, the computer equipment maintains a parallel monitoring stream, continuously monitoring the missing data types in the original analysis dataset. When it is determined that the missing data is a general attribute missing (categorical or numerical), the mode or median for the same age group is used for routine imputation.

[0038] When the missing data is determined to be a key medical diagnosis result, the computer device invokes a preset medical knowledge graph. The computer device performs multi-hop reasoning along the logical reasoning chain of related diseases and medical common sense in the medical knowledge graph, combines the characteristics of the patient's test results and past medical history, and derives a high-confidence diagnostic conclusion to fill in the key medical diagnosis result, and calculates the current missing value loss accordingly.

[0039] Meanwhile, regarding outlier loss The computer equipment performs dual-verification control logic for judgment and correction. First, the computer equipment filters the original analysis dataset using the statistical principle of three standard deviations, identifying undetermined outlier data that exceeds statistical boundaries.

[0040] Next, the computer device invokes a preset hard rule for reasonable value ranges in the medical field and compares the pending outlier data with this hard rule for verification. Only when the pending outlier data simultaneously exceeds three standard deviations in statistics and exceeds the hard rule for reasonable value ranges in the medical field is the computer device marked as an outlier and performs correction or deletion processing. This process accurately calculates outlier loss until the entire adaptive cleaning process converges and outputs a high-quality, usable analytical dataset.

[0041] S3. Calculate the matching score between each model in the medical-specific model library and the analysis target, select the model with the highest score as the optimal analysis model, and use the Bayesian optimization algorithm to adaptively iteratively tune the hyperparameters of the optimal analysis model. Use the tuned optimal analysis model to calculate the available analysis dataset and output the model analysis result set. Furthermore, in S3, the matching score between each model in the medical-specific model library and the analysis target includes: Extract the feature vectors of each model in the medical-specific model library. The feature vectors include the applicable scenarios of the model, the model complexity, and historical performance characteristics. The feature vectors of each model are compared with the analysis target to obtain the comparison feature values; By using the feature weights obtained through prior training with historical analysis data, the feature values ​​of each comparison are weighted and summed to obtain the matching score of each model, and the model with the highest matching score is selected as the optimal analysis model.

[0042] Furthermore, in S3, the adaptive iterative tuning of the hyperparameters of the optimal analysis model using the Bayesian optimization algorithm includes: Within the preset hyperparameter search space, the available analysis dataset is divided into a training set and a validation set; The performance score of the current hyperparameter combination is predicted using a Gaussian process prior model, and the optimal hyperparameter combination is obtained by maximizing the performance evaluation index of the optimal analysis model on the validation set during the iterative optimization process. The optimal analysis model with the best combination of hyperparameters is used to train and compute on the available analysis dataset, and outputs a multi-dimensional set of model analysis results.

[0043] Specifically, during the model selection phase, the computer equipment first analyzes the candidate models in the medical-specific model library and extracts the feature vectors of each model. These feature vectors include the model's applicable scenarios, model complexity, and historical performance characteristics. In the specific feature construction method, the computer equipment transforms the model's applicable scenarios, model complexity, and historical performance characteristics into discrete text labels or numerical ranges.

[0044] Subsequently, the computer device performs a multi-dimensional comparison between the feature vectors of each model and the analysis target obtained from the current structured parsing, acquiring the comparison feature values ​​for each item. To eliminate the blindness of manually assigned weights, the computer device calls the feature weights pre-trained using historical analysis data to perform a weighted summation calculation on the comparison feature values ​​for each item. The matching score calculation formula for each candidate model is as follows: ; in, Indicates the first medical-specific model in the library Matching score of each candidate model; This represents the total number of aligned feature terms in the feature vector; Indicates the sequence number of the currently compared feature; Indicates the first Feature vectors of candidate models; Indicate the current analysis objective; Indicates the first The candidate model in the th Alignment feature values ​​on each alignment feature term; Indicates the first Each matching feature term corresponds to a preset feature weight. The computer device iterates through all candidate models and selects the matching score. The highest-performing model is taken as the optimal analysis model.

[0045] After determining the optimal analysis model, the data processing flow enters the hyperparameter adaptive iterative tuning stage based on the Bayesian optimization algorithm. The computer equipment first divides the available analysis dataset output from the upstream into a training set and a validation set according to a fixed ratio within a preset hyperparameter search space.

[0046] During the iterative optimization process, the computer equipment utilizes a Gaussian process prior model to predict the performance score of the current hyperparameter combination on the validation set. To ensure the optimization process has a clear technical convergence objective, the iterative optimization aims to maximize the performance evaluation index of the optimal analysis model on the validation set, and involves multiple rounds of adaptive iteration. The performance evaluation index is dynamically specified according to business requirements; the F1 score is used in classification tasks, and the mean squared error is used in regression tasks.

[0047] In each iteration, the Bayesian optimization algorithm continuously updates the posterior distribution of the Gaussian process prior model by aiming to improve the acquisition function to balance exploration and utilization in the hyperparameter space. When the recursive calculation reaches a preset number of iterations, or when the continuous gain of the performance index falls below a preset termination threshold, the adaptive tuning process terminates, thereby obtaining the optimal combination of hyperparameters.

[0048] Finally, the computer device activates the optimal analysis model with the best combination of hyperparameters. This optimal analysis model reads the available analysis dataset for final training and computation, outputting a multi-dimensional set of model analysis results including nonlinear calculation conclusions, classification probabilities, and evolution trend predictions, and seamlessly transferring this set to the downstream structured report generation process.

[0049] S4. Match the business scenario report template according to the analysis target, call the content splicing operator and the preset medical business knowledge base, automatically fill the model analysis result set into the business scenario report template, generate a structured analysis report and output it.

[0050] Furthermore, in S4, the process of automatically populating the business scenario report template with the called content concatenation operator and the preset medical business knowledge base includes: Parse the business scenario report template and extract the content placeholders and their corresponding placeholder types contained therein; By using the content splicing operator, the nonlinear calculation conclusions in the model analysis result set are concatenated with the industry common sense decision rules extracted from the medical business knowledge base to generate the target text block. Based on the placeholder type, the target text block and the data charts automatically drawn from the model analysis results are collaboratively assembled into the corresponding content placeholders to generate a structured analysis report and output it.

[0051] Specifically, after receiving the set of model analysis results, the computer device automatically triggers a report generation pipeline. First, the computer device calls a pre-built template parsing engine to load a business scenario report template that perfectly matches the analysis target. During the parsing process, the computer device uses a streaming lexical analyzer to scan the underlying markup language file of the business scenario report template, accurately identifying and extracting each content placeholder and its corresponding placeholder type.

[0052] The placeholder types are strictly distinguished at the underlying level into text placeholders, chart placeholders, and indicator placeholders. Each content placeholder carries a unique character encoding identifier and scope restriction label, which serve as format control parameters for subsequent data injection and dynamic rendering, and are stored in a temporary key-value tree in memory.

[0053] Subsequently, the data processing flow enters the heterogeneous content cross-cascading stage based on semantic alignment. The computer device invokes the content splicing operator to extract nonlinear calculation conclusions from the model analysis result set, and simultaneously retrieves matching industry common sense decision rules from a preset medical business knowledge base according to the analysis objective.

[0054] The content concatenation operator does not perform blind, mechanical string splicing. Instead, it first extracts semantic feature vectors from nonlinear calculation conclusions and industry-standard decision-making rules. Then, it uses transition probability matrices and relevance scores to perform context disambiguation and sentence restructuring, thereby generating the target text block that combines data accuracy with medical logic support. The algebraic expression of the content concatenation operator is: ; in, This refers to the generated target text block; This represents the text of the nonlinear calculation conclusions extracted from the set of model analysis results and smoothed. This indicates the industry common sense decision-making rule text retrieved and matched from a pre-defined medical business knowledge base; This indicates a text concatenation operation with logical context disambiguation performed by the content concatenation operator.

[0055] Finally, the computer device initiates a multimodal collaborative assembly mechanism, invoking the underlying graphics and text rendering engine. The computer device reads the key-value tree temporarily stored in memory and, based on the previously extracted placeholder types, collaboratively assembles the target text block and the data chart automatically drawn from the model analysis result set into the corresponding content placeholders in the business scenario report template.

[0056] In the specific collaborative assembly and operation logic, the computer device injects the target text block into the text placeholder in the form of streaming characters through the text filling operator, and automatically triggers dynamic line height and font style alignment; at the same time, the rendering engine automatically draws the time series data, survival curves or confusion matrix data in the model analysis results into data charts in vector format, and binds their binary object handles to the corresponding chart placeholders.

[0057] The computer device then activates the layout manager to redraw and perform overflow checks on the filled document. Only when all content placeholders are correctly filled and pass the dual consistency checks of preset styles and medical common sense rules, the computer device finally converts the business scenario report template into a structured PDF or electronic medical record report format for output, completing the closed loop of intelligent medical data analysis based on authorization boundary constraints.

[0058] Example 2: Pharmaceutical companies and medical institutions often need to conduct academic promotion and market insight analysis for specific diseases or innovative drugs. However, front-end business personnel typically lack professional backgrounds in advanced data modeling and statistical cleaning. When proposing complex multidimensional cross-analysis requests, their colloquial abbreviations can easily lead to semantic ambiguity in the system and pose data compliance risks of penetrating and exceeding authorization compliance red lines during frequent cross-database association searches. To address these issues, this invention provides a medical data intelligent analysis system based on authorization boundary constraints, the structure of which is as follows: Figure 3 As shown. The specific implementation process of this system is as follows: The structured parsing module is used to receive the input natural language business requirement text and the corresponding user data authorization information, and to perform structured parsing on the natural language business requirement text using a preset language model and a conditional random field layer to generate a structured parsing triple containing the analysis target, the set of analysis dimensions, and the set of data constraints. The data filtering and cleaning module is used to filter and extract the original analysis dataset from the full medical dataset based on the user data authorization information and the data constraint set, construct a comprehensive cleaning loss function, and adaptively clean the original analysis dataset by minimizing the comprehensive cleaning loss function to output a usable analysis dataset. The model selection and optimization module is used to calculate the matching score between each model in the medical-specific model library and the analysis target, select the model with the highest score as the optimal analysis model, and use the Bayesian optimization algorithm to adaptively and iteratively optimize the hyperparameters of the optimal analysis model. The optimized optimal analysis model is then used to calculate the available analysis dataset and output a set of model analysis results. The report population and output module is used to match a business scenario report template according to the analysis target, call the content splicing operator and the preset medical business knowledge base, automatically populate the model analysis result set into the business scenario report template, generate a structured analysis report and output it.

[0059] Specifically, the structured parsing module receives input natural language business request text (such as "query the clinical adverse reaction trend of amoxicillin in East China for the past three years") and corresponding user data authorization information. This module utilizes a pre-built medical-specific ontology knowledge base containing five major entity categories as pre-defined rules, and cascades a finely tuned semantic parsing model (MedBERT) with a conditional random field layer to perform feature parsing and variable slot filling on the input text. The algorithm automatically decomposes the user's colloquial business request and generates structured parsing triples containing the analysis objective, a set of analysis dimensions, and a set of data constraints, achieving "asking questions in business language, and the algorithm automatically understanding."

[0060] The data filtering and cleaning module reads user data authorization information and the aforementioned structured parsed triples. To embed compliance constraints upfront, this module performs a two-stage hierarchical processing: first, it trims the entire medical dataset based on the user data authorization information, isolating a closed, dedicated authorization dataset; then, within this dedicated space, it performs conditional extraction based on the data constraint set to obtain the original analysis dataset that does not violate compliance red lines. Subsequently, the module constructs a comprehensive cleaning loss function, performs double verification using "statistics (three standard deviations) + hard rules for value selection in the medical field," and calls upon a medical knowledge graph to perform multi-hop path logical reasoning to fill in key missing items. High-quality data cleaning is automatically completed without manual intervention until the loss value is below a preset threshold, outputting a usable analysis dataset.

[0061] The model selection and optimization module is then implemented. This module pre-configures a medical-specific model library for five major scenarios: market insight, academic promotion, product positioning, clinical decision-making, and policy analysis. The module first extracts the feature vectors (scenario, complexity, historical performance) of each candidate model. It then performs a weighted summation of the feature weights obtained from training on historical data, automatically selecting the model with the highest score as the optimal analysis model, without the need for data scientists. After determining the model, the module divides the training and validation sets within a pre-defined hyperparameter search space and activates the Bayesian optimization algorithm. Using a Gaussian process prior model and a data acquisition function, it performs closed-loop iterative optimization with the goal of maximizing performance metrics (such as the F1 score) on the validation set to obtain the optimal hyperparameter combination. Finally, it uses the optimized model to calculate and output a multi-dimensional set of model analysis results.

[0062] The report population module loads a matching business scenario report template and parses out text, chart, and indicator placeholders using streaming lexical analysis. This module then calls a content concatenation operator to perform semantic alignment-based context disambiguation and sentence restructuring cascade processing on the non-linear calculation conclusions output by the model and industry common-sense decision-making rules from the medical business knowledge base, generating the target text block. Finally, the module activates a graphics and text rendering engine for multimodal collaborative assembly, injecting the target text block into the text placeholders and automatically rendering time-series data into vector charts bound to the chart placeholders. After page redrawing and overflow validation, a structured analysis report containing data charts, statistical conclusions, and actionable business suggestions is automatically generated and output. Users can directly use this report for strategic decision-making without further processing.

[0063] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A medical data intelligent analysis method based on authorization boundary constraints, characterized in that, include: S1. Receive the input natural language business requirement text and the corresponding user data authorization information, and use a preset language model and conditional random field layer to perform structured parsing on the natural language business requirement text to generate a structured parsing triple containing the analysis target, the set of analysis dimensions and the set of data constraints. S2. Based on the user data authorization information and the data constraint set, filter and extract the original analysis dataset from the full medical dataset, construct a comprehensive cleaning loss function, and adaptively clean the original analysis dataset by minimizing the comprehensive cleaning loss function to output a usable analysis dataset. S3. Calculate the matching score between each model in the medical-specific model library and the analysis target, select the model with the highest score as the optimal analysis model, and use the Bayesian optimization algorithm to adaptively iteratively tune the hyperparameters of the optimal analysis model. Use the tuned optimal analysis model to calculate the available analysis dataset and output the model analysis result set. S4. Match the business scenario report template according to the analysis target, call the content splicing operator and the preset medical business knowledge base, automatically fill the model analysis result set into the business scenario report template, generate a structured analysis report and output it.

2. The intelligent medical data analysis method based on authorization boundary constraints according to claim 1, characterized in that, In S1, the structured parsing of the natural language business requirement text includes: The preset language model is used to perform semantic encoding on natural language business requirement text to obtain context vector representations at the marker word level; The context vector representation is input into the conditional random field layer, and path optimization identification is performed in combination with the preset entity label transition probability matrix to obtain the optimal medical entity label sequence. The corresponding medical entities are captured based on the optimal medical entity label sequence, and the structured parsed triples are generated by assembling parameters according to the semantic relationships between the medical entities.

3. The intelligent medical data analysis method based on authorization boundary constraints according to claim 1, characterized in that, In S2, the raw analysis dataset is extracted by filtering from the full medical dataset, including: Based on the user data authorization information, the full medical dataset is subjected to boundary filtering to generate a unique authorized dataset; Within the internal computing space of the authorized dataset, multidimensional feature retrieval and condition extraction are performed based on the data constraint set to obtain the original analysis dataset that meets business requirements without exceeding the computing boundaries.

4. The intelligent medical data analysis method based on authorization boundary constraints according to claim 1, characterized in that, In S2, adaptive cleaning of the original analysis dataset includes: Based on the data defect type of the original analysis dataset, the weighting coefficients of each loss term in the comprehensive cleaning loss function are automatically initialized; Calculate the current cleaning loss value after each cleaning iteration; Determine whether the cleaning loss value is lower than a preset threshold; if the cleaning loss value is not lower than the preset threshold, adjust the weighting coefficient according to the cleaning loss value and re-clean until the cleaning loss value is lower than the preset threshold and terminate the iteration.

5. The intelligent medical data analysis method based on authorization boundary constraints according to claim 1, characterized in that, In S2, the adaptive cleaning of the original analysis dataset also includes: Real-time monitoring of missing data types in the original analysis dataset; When the missing data is determined to be a key medical diagnosis result, a preset medical knowledge graph is invoked, and the key medical diagnosis result is filled in by reasoning through the logical reasoning chain of associating diseases with common medical knowledge.

6. The intelligent medical data analysis method based on authorization boundary constraints according to claim 1, characterized in that, In S2, the adaptive cleaning of the original analysis dataset also includes: The original analysis dataset was filtered using the statistical principle of three standard deviations to identify undetermined outlier data. The pending abnormal data is compared and verified with the preset hard rules of reasonable value range in the medical field; The undetermined abnormal data is marked as an abnormal value and corrected or deleted only if it exceeds the hard rule of reasonable value range in the medical field.

7. The intelligent medical data analysis method based on authorization boundary constraints according to claim 1, characterized in that, In S3, the matching score between each model in the medical-specific model library and the analysis target includes: Extract the feature vectors of each model in the medical-specific model library. The feature vectors include the applicable scenarios of the model, the model complexity, and historical performance characteristics. The feature vectors of each model are compared with the analysis target to obtain the comparison feature values; By using the feature weights obtained through prior training with historical analysis data, the feature values ​​of each comparison are weighted and summed to obtain the matching score of each model, and the model with the highest matching score is selected as the optimal analysis model.

8. The intelligent analysis method for medical data based on authorization boundary constraints according to claim 1, characterized in that, In S3, the adaptive iterative tuning of the hyperparameters of the optimal analysis model using the Bayesian optimization algorithm includes: Within the preset hyperparameter search space, the available analysis dataset is divided into a training set and a validation set; The performance score of the current hyperparameter combination is predicted using a Gaussian process prior model, and the optimal hyperparameter combination is obtained by maximizing the performance evaluation index of the optimal analysis model on the validation set during the iterative optimization process. The optimal analysis model with the best combination of hyperparameters is used to train and compute on the available analysis dataset, and outputs a multi-dimensional set of model analysis results.

9. The intelligent analysis method for medical data based on authorization boundary constraints according to claim 1, characterized in that, In S4, the process of automatically populating the business scenario report template with the content concatenation operator and the preset medical business knowledge base includes: Parse the business scenario report template and extract the content placeholders and their corresponding placeholder types contained therein; By using the content splicing operator, the nonlinear calculation conclusions in the model analysis result set are concatenated with the industry common sense decision rules extracted from the medical business knowledge base to generate the target text block. Based on the placeholder type, the target text block and the data charts automatically drawn from the model analysis results are collaboratively assembled into the corresponding content placeholders to generate a structured analysis report and output it.

10. A medical data intelligent analysis system based on authorization boundary constraints, characterized in that, The system for a medical data intelligent analysis method based on authorization boundary constraints as described in any one of claims 1-9, the system comprising: The structured parsing module is used to receive the input natural language business requirement text and the corresponding user data authorization information, and to perform structured parsing on the natural language business requirement text using a preset language model and a conditional random field layer to generate a structured parsing triple containing the analysis target, the set of analysis dimensions, and the set of data constraints. The data filtering and cleaning module is used to filter and extract the original analysis dataset from the full medical dataset based on the user data authorization information and the data constraint set, construct a comprehensive cleaning loss function, and adaptively clean the original analysis dataset by minimizing the comprehensive cleaning loss function to output a usable analysis dataset. The model selection and optimization module is used to calculate the matching score between each model in the medical-specific model library and the analysis target, select the model with the highest score as the optimal analysis model, and use the Bayesian optimization algorithm to adaptively and iteratively optimize the hyperparameters of the optimal analysis model. The optimized optimal analysis model is then used to calculate the available analysis dataset and output a set of model analysis results. The report population and output module is used to match a business scenario report template according to the analysis target, call the content splicing operator and the preset medical business knowledge base, automatically populate the model analysis result set into the business scenario report template, generate a structured analysis report and output it.

Images