Auxiliary diagnosis method and device based on medical big data

Abstract

The invention relates to an auxiliary diagnosis method and device based on medical big data, equipment and a medium. The method comprises the steps of obtaining an original medical record text and extracting medical information through a natural language processing technology to generate medical record data; performing clustering analysis on the medical record data to output a personalized feature combination; determining a target disease type and key medical features based on the personalized feature combination, and obtaining clinical quantized values of the key medical features; if the clinical quantized value of the key medical feature accords with a first preset threshold value, generating a trend prediction result by using a preset illness state dynamic time sequence model based on the key medical feature; calling a diagnosis rule base corresponding to the target disease type based on the trend prediction result to generate a diagnosis path; and carrying out association mapping on the key information of the diagnosis path and the path nodes to output a personalized medical service scheme. Through data driving and model analysis, intelligent and personalized generation of a disease diagnosis path is realized, and diagnosis efficiency and scheme pertinence are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical and health informatics, and in particular relates to an auxiliary diagnostic method and device based on medical big data. Background Technology

[0002] With the rapid development of healthcare informatics and artificial intelligence technologies, big data-based clinical decision support systems have emerged. These systems can assist doctors in disease diagnosis by analyzing massive amounts of medical record data. In traditional technologies, assisted diagnosis mainly relies on the following two methods: rule engine-driven systems, which match patient clinical indicators with a pre-set static diagnostic rule base to generate standardized treatment pathways; and data-driven predictive models, which use machine learning algorithms to predict the risk of specific diseases and output probabilistic results.

[0003] However, existing technologies have the following significant drawbacks: rigid feature combination problem, traditional clustering methods are unable to dynamically generate personalized feature combinations that adapt to the clinical characteristics of patients from multi-dimensional individual features, resulting in bias in the extraction of key medical features; lack of dynamic temporal modeling, most existing systems only rely on the current clinical quantitative values ​​for diagnosis, lacking the ability to model the temporal evolution of the disease, and cannot generate dynamic predictions when key feature values ​​are at the threshold; disconnect between path generation and execution, the path schemes output by the diagnostic rule base are not verified against the historical case path base, and no real-time mapping relationship between path nodes and clinical quantitative values ​​is established, resulting in insufficient executability of personalized service schemes. Summary of the Invention

[0004] Therefore, it is necessary to provide an auxiliary diagnostic method and device based on medical big data that can solve the above problems.

[0005] Firstly, this application provides an auxiliary diagnostic method based on medical big data, including:

[0006] The original medical record text is obtained, and medical information is extracted using natural language processing technology to generate medical record data;

[0007] Perform cluster analysis on medical record data and output personalized feature combinations;

[0008] Based on personalized feature combinations, the target disease type and key medical features are identified, and the clinical quantitative values ​​of the key medical features are obtained.

[0009] If the clinical quantitative value of the key medical feature meets the first preset threshold, then based on the key medical feature, a trend prediction result is generated using a preset dynamic time series model of the disease condition.

[0010] Based on the trend prediction results, the diagnostic rule base corresponding to the target disease type is called to generate a diagnostic path;

[0011] The key information of the diagnostic pathway is associated and mapped with the pathway nodes to output personalized medical service plans.

[0012] In one embodiment, cluster analysis is performed on medical record data to output personalized feature combinations, including:

[0013] The classification algorithm is used to perform multi-dimensional clustering of individual characteristics in medical record data to generate clinical similarity feature groups;

[0014] The weighted correlation degree between each clinical similarity feature group and the target disease is calculated based on a pre-set disease association model;

[0015] Based on clinical scenario adaptation parameters, a screening threshold is generated, and the weight correlation is screened to retain feature groups that exceed the screening threshold to form a primary personalized feature set.

[0016] The feature cross-validation algorithm eliminates redundant feature groups in the initial personalized feature set and outputs a personalized feature combination.

[0017] In one embodiment, based on personalized feature combinations, the target disease type and key medical features are determined, and the clinical quantitative values ​​of the key medical features are obtained, including:

[0018] Based on personalized feature combinations, the target disease type is determined through a disease classification model.

[0019] Calculate the association strength between each feature variable in the personalized feature combination and the individual's target disease type;

[0020] If the strength of the individual association exceeds the second preset threshold, then an independent clinical causal test based on a medical knowledge base or clinical guidelines is performed on the feature variables corresponding to the strength of the individual association.

[0021] If a feature variable passes the independent clinical causality test, it is identified as a key medical feature, and the clinical quantitative value of the key medical feature is extracted from the medical record data.

[0022] In one embodiment, based on the trend prediction results, a diagnostic rule base corresponding to the target disease type is invoked to generate a diagnostic path, including:

[0023] The trend prediction results are combined with the diagnostic rule base to perform path combination reasoning and generate a preliminary diagnostic path scheme.

[0024] Based on the preliminary diagnostic pathway scheme, similar case pathway sets are extracted from the historical case database corresponding to the target disease type and deviation detection is performed to obtain pathway deviation information;

[0025] The node sequence of the preliminary diagnostic path plan is optimized and adjusted based on the path deviation information to generate a diagnostic path.

[0026] In one embodiment, key information of the diagnostic pathway is mapped to pathway nodes to output a personalized medical service plan, including:

[0027] Based on the descriptive text of each node in the diagnostic path, the information extraction technology of the predefined rule engine is used to identify and extract predefined key medical entities and their clinical quantitative values, resulting in a set of medical entity-quantitative value key-value pairs.

[0028] Based on the set of medical entity-quantified value key-value pairs, each key-value pair is associated with its corresponding path node according to the node sequence relationship of the diagnostic path, forming a medical entity-quantified value mapping relationship;

[0029] Based on the mapping relationship between medical entities and quantitative values, clinical quantitative values ​​are configured into the corresponding fields of preset medical service templates to output personalized medical service plans.

[0030] In one embodiment, the weighted correlation between each clinical similarity feature group and the target disease is calculated using the following formula:

[0031]

[0032] Among them, F k Grouping based on the k-th clinical similarity feature, where D is the target disease, w i For feature f i The preset weighting coefficients, sim(f) i D) is a feature f i The similarity function with the target disease D, where m is the feature group F. k The number of features in;

[0033] The feature cross-validation algorithm eliminates redundant feature groups in the initial personalized feature set, outputting personalized feature combinations, including:

[0034] Calculate the mutual information values ​​between different feature groups in the primary personalized feature set to form a mutual information matrix MI = [mi ij ], where matrix element mi ij This represents the mutual information value between feature group i and feature group j;

[0035] Traverse the mutual information matrix MI, when mi ij If the value exceeds the preset third threshold, select one of feature group i and feature group j to retain;

[0036] The set of feature groups retained after traversal and filtering is output as a personalized feature combination.

[0037] In one embodiment, medical information is extracted using natural language processing technology to generate medical record data, including:

[0038] The original medical record text was segmented and standardized with medical terminology to obtain a standardized terminology sequence.

[0039] Semantic role labeling is performed on the standardized term sequence to obtain semantic labeling results;

[0040] The semantic annotation results are mapped to preset structured fields to generate medical record data.

[0041] Secondly, this application also provides an auxiliary diagnostic device based on medical big data, comprising:

[0042] The medical record data processing module is used to acquire the original medical record text and extract medical information through natural language processing technology to generate medical record data;

[0043] The feature clustering analysis module is used to perform clustering analysis on medical record data and output personalized feature combinations.

[0044] The key feature determination module is used to determine the target disease type and key medical features based on personalized feature combinations, and to obtain the clinical quantitative values ​​of the key medical features.

[0045] The dynamic time series prediction module is used to generate trend prediction results based on the key medical features and using a preset dynamic time series model of the disease if the clinical quantitative value of the key medical features meets the first preset threshold.

[0046] The diagnostic path generation module is used to generate a diagnostic path by calling the diagnostic rule library corresponding to the target disease type based on the trend prediction results.

[0047] The service solution output module is used to associate and map key information of the diagnostic path with path nodes, and output personalized medical service solutions.

[0048] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned auxiliary diagnostic method based on medical big data.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described auxiliary diagnostic method based on medical big data.

[0050] The aforementioned auxiliary diagnostic method, device, computer equipment, and storage medium based on medical big data acquire original medical record text and extract medical information using natural language processing technology to generate structured medical record data. It then dynamically outputs personalized feature combinations through cluster analysis of the medical record data. Based on these personalized feature combinations, it determines the target disease type and key medical features and obtains their clinical quantitative values. When the quantitative values ​​meet a first preset threshold, it triggers a preset dynamic temporal model of the disease condition to generate trend prediction results. Based on the trend prediction results, it calls the diagnostic rule base corresponding to the target disease type to generate a diagnostic path, and outputs a personalized medical service plan by associating and mapping the key information of the diagnostic path with the path nodes. This solves the problems of rigid feature combinations, lack of dynamic temporal modeling, and disconnect between path generation and execution in traditional technologies, significantly improving diagnostic efficiency and the feasibility of the plan. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of an auxiliary diagnostic method based on medical big data according to the present invention;

[0053] Figure 2 This is a structural diagram of an auxiliary diagnostic device based on medical big data according to the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] The hardware architecture for implementing this method may include: terminal devices (such as doctor workstations) responsible for collecting original medical record text input and displaying the final service plan; server clusters providing core computing capabilities such as natural language processing, cluster analysis, and time series modeling; and network infrastructure to enable secure data transmission. When clinical needs arise to improve diagnostic efficiency and personalization, the terminal transmits the original medical record text input by the doctor to the server via the network. The server performs medical record data processing, feature cluster analysis, and dynamic time series prediction, generates a diagnostic path, and returns it to the terminal for association mapping to form a visualized service plan.

[0056] In one embodiment, such as Figure 1As shown, an auxiliary diagnostic method based on medical big data is provided. This embodiment illustrates the application of this method to a terminal device. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0057] S01: Obtain the original medical record text and extract medical information through natural language processing technology to generate medical record data.

[0058] The process involves: acquiring raw medical record text (unstructured medical documents recorded by doctors, including patient symptoms, signs, medical history, and treatment processes, as the data input source); processing medical information (including disease names, symptom descriptions, laboratory indicators, and other key clinical entities) using natural language processing (a method that uses computer algorithms to parse and understand human language); performing word segmentation on the raw medical record text, breaking it down into discrete words or phrases, and combining this with standardized medical terminology integration, mapping non-standard terms to standardized terms through a pre-built medical terminology database to generate a standardized term sequence; performing semantic role annotation on the standardized term sequence, identifying the semantic role of each term (e.g., patient as subject, symptom as object, and onset time as timestamp) to obtain semantic annotation results; and mapping the semantic annotation results to pre-defined structured fields (e.g., patient basic information, diagnostic records, medication data, etc.) to generate structured medical record data, thereby improving data usability through automation technology.

[0059] S02 performs cluster analysis on medical record data and outputs personalized feature combinations.

[0060] The process involves using classification algorithms (such as K-means or hierarchical clustering) to perform multi-dimensional clustering of individual characteristics (such as laboratory indicators, symptom descriptions, and medical history features) in medical record data, generating feature groups with clinical similarity (e.g., classifying blood pressure fluctuation frequency and nocturnal palpitation into a cardiovascular-related feature group). Optionally, based on a preset disease association model, the weighted correlation degree between each feature group and the target disease can be calculated. Combined with clinical scenario adaptation parameters (e.g., increasing the weight of timeliness features in emergency scenarios), a screening threshold can be dynamically generated, retaining feature groups with weights exceeding the threshold to form a primary personalized feature set. The mutual information value between different feature groups in the primary set can be calculated. When the mutual information value is greater than a preset redundancy threshold (e.g., 0.8), a high-weight group can be selected and retained based on the correlation strength between the feature group and the target disease (e.g., retaining the glycated hemoglobin trend group while discarding the fasting blood glucose fluctuation group). Finally, the personalized feature combination after eliminating redundancy is output.

[0061] S03, based on personalized feature combinations, determines the target disease type and key medical features, and obtains the clinical quantitative values ​​of the key medical features.

[0062] This process involves inputting personalized feature combinations into a pre-trained disease classification model (such as a classifier based on random forests or deep neural networks). By analyzing the statistical distribution pattern of the feature combinations and their mapping relationship with disease labels, the model outputs the target disease type with the highest probability. The individual association strength between each feature variable and the target disease type is calculated (e.g., quantifying feature importance through information gain or chi-square tests). When the association strength exceeds a preset threshold, an independent clinical causality test based on a medical knowledge base is triggered. This test verifies the medical rationality of the feature variables and the disease by matching causal chain rules in clinical guidelines (e.g., the NCEP-ATPIII standard's causal definition of blood lipid indicators and coronary heart disease). Features that pass the test are identified as key medical features, and their corresponding real-time clinical quantitative values ​​can be directly extracted from structured medical record data fields. This dual-layer filtering mechanism of dynamic association strength calculation and clinical causality verification solves the feature extraction bias problem in traditional techniques, ensuring that key medical features are statistically significant (association strength threshold screening) and conforming to medical causal logic (clinical guideline verification), providing an accurate and reliable input benchmark for subsequent time-series predictions.

[0063] S04. If the clinical quantitative value of the key medical feature meets the first preset threshold, then based on the key medical feature, a trend prediction result is generated using a preset dynamic time series model of the disease condition.

[0064] The process involves real-time comparison of the clinically quantified values ​​of key medical features with a first preset threshold (a disease progression threshold defined by clinical guidelines). For example, if a patient's blood glucose levels are within the critical range of 6.8-7.2 mmol / L for three consecutive tests, the time-series modeling process is activated if the threshold is met. Based on historical sequence data of this key medical feature (e.g., blood glucose monitoring records from the past 30 days), a preset dynamic time-series model (trained using historical cases and a machine learning model capable of capturing feature evolution patterns, such as LSTM or Transformer time-series networks) is input for multi-step prediction. An encoder-decoder architecture learns the potential patterns of feature changes (e.g., the nonlinear relationship between blood glucose fluctuations and insulin sensitivity), outputting a trend prediction for a specific future time window (e.g., blood glucose levels will remain above 8.5 mmol / L for the next 72 hours). Through these steps, when the key feature value is at a clinically critical state, the time-series model simulates the disease progression trajectory, addressing the problem of incomplete disease progression prediction in traditional techniques.

[0065] S05: Based on the trend prediction results, call the diagnostic rule library corresponding to the target disease type to generate a diagnostic path.

[0066] The process involves combining trend prediction results with a diagnostic rule base (a pre-defined set of structured diagnostic logics categorized by disease, including diagnostic criteria, testing order, and treatment decision trees) to perform path combination reasoning. Predicted values ​​(e.g., progressive hyperglycemia) are mapped to conditional branches in the rule base (e.g., persistent hyperglycemia → initiating insulin therapy), generating a preliminary diagnostic path. Based on this preliminary path, a set of similar case paths (e.g., treatment records of 100 cases of worsening blood sugar) is extracted from a historical case database corresponding to the target disease type. A node sequence comparison algorithm is used for deviation detection (e.g., detecting a missing retinopathy screening step in the current plan), outputting path deviation information. The node sequence of the preliminary plan is dynamically optimized and adjusted based on the deviation information (e.g., inserting missing examination nodes and adjusting medication order), generating the final diagnostic path. By using trend prediction as input to trigger the generation of the basic path in the diagnostic rule base, and then verifying and correcting path deviations using historical case empirical data, this approach addresses the disconnect between path generation and clinical practice in traditional techniques, ensuring that the diagnostic path is both standardized and individually adaptable.

[0067] S06 maps key information from the diagnostic pathway to pathway nodes, and outputs personalized medical service plans.

[0068] This process involves using the descriptive text of each node in the diagnostic pathway (the smallest execution unit in the diagnostic pathway, such as enhanced CT scan or warfarin anticoagulation therapy) to identify and extract predefined key medical entities (such as drug names and test indicators) and their clinical quantitative values ​​(such as 5 mg / kg or 120 / 80 mmHg) based on the information extraction technology of a predefined rule engine (such as a BERT-based entity recognition model). This generates a set of medical entity-quantitative value key-value pairs. Based on the node sequence relationships in the diagnostic pathway (such as preoperative preparation, node priority, and surgical operation nodes), a graph neural network is used to establish a node topology mapping, dynamically associating each key-value pair with the corresponding path node, forming a structured medical entity-quantitative value mapping relationship. The clinical quantitative values ​​in the mapping relationship are automatically configured into the corresponding fields of a preset medical service template (an electronic medical record template conforming to the HL7 standard), outputting a personalized medical service plan containing time nodes, execution content, and quantitative parameters. Through the collaboration of the rule engine and the graph neural network, deep coupling between the diagnostic pathway and real-time clinical data is achieved, solving the problem of disconnect between pathway generation and execution in traditional technologies.

[0069] The aforementioned medical big data-based assisted diagnosis method acquires original medical record text and uses natural language processing (NLP) technology for word segmentation, standardization, and semantic annotation to generate structured medical record data. It then performs multi-dimensional clustering analysis on the medical record data, generates clinical similarity feature groups through classification algorithms, calculates weighted correlation degrees using a pre-set disease association model, and filters out redundancies to form personalized feature combinations. This addresses the problem of rigid feature combinations leading to biased key feature extraction in traditional techniques. Based on these personalized feature combinations, a disease classification model determines the target disease type, calculates the individual correlation strength between feature variables and the disease, and performs causal testing using a medical knowledge base. Key medical features and their clinical quantitative values ​​are extracted. When the quantitative values ​​meet a first pre-set threshold, a pre-set disease condition is considered. The dynamic time series model learns from historical sequence data of key features to generate trend prediction results for future time windows, overcoming the shortcomings of existing systems that lack the ability to model the time series of disease evolution and cannot dynamically predict critical states. Based on the trend prediction results and the execution path of the diagnostic rule base, a preliminary diagnostic path is generated through combined reasoning. Then, a set of similar paths is extracted from the historical case database for deviation detection. After optimizing and adjusting the node sequence, a diagnostic path is generated. The key medical entities and quantitative values ​​of the diagnostic path nodes are mapped to preset medical service templates to output personalized solutions. This solves the problem of insufficient executability caused by the lack of historical case verification and the disconnect between nodes and quantitative values ​​in traditional path generation. Through data-driven and model analysis, the intelligent and personalized generation of diagnostic paths is achieved, improving diagnostic efficiency.

[0070] In one embodiment, cluster analysis is performed on medical record data to output personalized feature combinations, including:

[0071] S11, using a classification algorithm to perform multi-dimensional clustering of individual characteristics in medical record data to generate clinical similarity feature groups;

[0072] S12, Calculate the weighted correlation degree between each clinical similarity feature group and the target disease based on the preset disease association model;

[0073] S13, Generate screening thresholds based on clinical scenario adaptation parameters, and screen the weight correlation, retaining feature groups that exceed the screening thresholds to form a primary personalized feature set;

[0074] S14 uses a feature cross-validation algorithm to eliminate redundant feature groups in the initial personalized feature set and outputs a personalized feature combination.

[0075] Specifically, classification algorithms (such as K-means or hierarchical clustering) can be used to perform multi-dimensional clustering of individual characteristics (including laboratory indicators, symptom descriptions, and medical history features) in medical record data. Grouping can be generated based on the similarity of features in a medical sense (such as jointly reflecting a specific pathological mechanism or clinical manifestation) (e.g., grouping the frequency of blood pressure fluctuations and the degree of nocturnal palpitations into a cardiovascular-related feature group); based on a pre-set disease association model, formulas can be used. Calculate the weighted correlation between each feature group and the target disease (where w) i The feature weight coefficients are predefined by the medical knowledge base; the sim function achieves similarity measurement through the disease-feature correlation matrix; combined with clinical scenario adaptation parameters (such as increasing the screening threshold weight of timeliness features in emergency scenarios), a screening threshold is dynamically generated, and feature groups with weights exceeding the threshold are retained to form a primary personalized feature set; the mutual information values ​​between different feature groups in the primary set are calculated to construct the mutual information matrix MI = [mi ij ], traverse the matrix and when mi ij When the redundancy threshold exceeds a preset threshold (e.g., 0.8), a high-weight group is selected based on the correlation strength between the feature group and the target disease (e.g., retaining the glycated hemoglobin trend group while discarding the fasting blood glucose fluctuation group), and the personalized feature combination after redundancy elimination is output. Multi-dimensional clustering overcomes the limitations of traditional single-feature analysis, achieving semantic-level grouping of clinical features; a weighted association model replaces the static rule base to solve the problem of rigid feature combinations; mutual information verification is introduced to eliminate redundant features, ensuring the independence and representativeness of key medical features (e.g., retaining feature groups with strong pathological mechanism specificity).

[0076] In one embodiment, based on personalized feature combinations, the target disease type and key medical features are determined, and the clinical quantitative values ​​of the key medical features are obtained, including:

[0077] S21, Based on personalized feature combinations, the target disease type is determined through a disease classification model;

[0078] S22, Calculate the individual association strength between each feature variable in the personalized feature combination and the target disease type;

[0079] S23, If the individual association strength exceeds the second preset threshold, then perform an independent clinical causal test based on a medical knowledge base or clinical guidelines on the feature variables corresponding to the individual association strength.

[0080] S24. If the feature variable passes the independent clinical causality test, it is determined to be a key medical feature, and the clinical quantitative value of the key medical feature is extracted from the medical record data.

[0081] For example, personalized feature combinations (feature subsets formed through cluster analysis and redundancy elimination) can be input into a pre-trained disease classification model (such as a classifier based on random forest or deep neural networks). By analyzing the mapping relationship between the statistical distribution pattern of the feature combinations and disease labels, the target disease type with the highest probability (such as type II diabetes or coronary artery disease) can be output. Information gain or chi-square test algorithms can be used to quantify the individual association strength between each feature variable and the target disease type (such as calculating the contribution of blood glucose fluctuation coefficient to diabetes). When the individual association strength exceeds a second preset threshold (such as information gain > 0.5), an independent clinical causal test based on a medical knowledge base (such as WHO disease classification standards) or clinical guidelines is triggered: matching pre-set causal chain rules (such as using NCEP-ATPIII standards to verify the medical rationality of blood lipid indicators and coronary heart disease), excluding statistically significant features that are not clinically logically valid; features that pass the test are determined to be key medical features (such as glycated hemoglobin), and their real-time clinical quantitative values ​​(such as HbA1c = 7.8%) are directly extracted from the corresponding fields of structured medical record data. Preliminary feature screening is achieved by using correlation strength thresholds to solve the problem of interference from low-correlation features in traditional techniques; clinical causality tests are introduced to ensure that key features conform to medical guidelines (such as avoiding misclassification of advanced age as a direct causal feature of diabetes); quantitative values ​​are directly extracted from standardized fields of medical record data to avoid errors from manual secondary data entry.

[0082] In one embodiment, based on the trend prediction results, a diagnostic rule base corresponding to the target disease type is invoked to generate a diagnostic path, including:

[0083] S31, Perform path combination reasoning on the trend prediction results and the diagnostic rule base to generate a preliminary diagnostic path scheme;

[0084] S32, Based on the preliminary diagnostic pathway scheme, extract similar case pathway sets from the historical case database corresponding to the target disease type and perform deviation detection to obtain pathway deviation information;

[0085] S33, Optimize and adjust the node sequence of the preliminary diagnostic path scheme based on the path deviation information to generate a diagnostic path.

[0086] Specifically, the trend prediction results (such as the deterioration trajectory of key medical features) can be dynamically matched with the diagnostic rule base (a pre-set set of structured diagnostic logic) corresponding to the target disease type. Through conditional branching reasoning of the rule engine (such as progressive hyperglycemia → initiation of insulin treatment path), a preliminary diagnostic path plan containing examination, medication, and monitoring nodes can be generated. Based on the preliminary plan, a set of similar case paths (such as the diagnosis and treatment records of 100 cases of worsening blood sugar) is extracted from the historical case database of the target disease type. A node sequence comparison algorithm (a dynamic time warping algorithm based on edit distance) is used to detect deviations, identify missing, redundant, or out-of-order nodes (such as finding that the current plan lacks a retinal disease screening link), and output path deviation information. According to the path deviation information, the node sequence of the preliminary plan is dynamically adjusted: missing nodes are inserted (such as adding fundus examination), redundant nodes are deleted (such as removing duplicate blood routine tests), and the execution order is optimized (such as moving the insulin dose adjustment node to the front), generating a final diagnostic path that adapts to the disease trend. By converting trend prediction results into trigger condition parameters in the rule base, real-time adaptation of disease progression and diagnostic pathways is achieved; deviation detection of historical case pathways (such as using the DTW algorithm to compare node sequence differences) ensures that the generated pathways have been validated in clinical practice; and dynamic addition, deletion, or reordering of pathway nodes based on deviation information (such as automatically shortening the detection node interval in emergency scenarios) improves the individual adaptability of the pathways.

[0087] In one embodiment, key information of the diagnostic pathway is mapped to pathway nodes to output a personalized medical service plan, including:

[0088] S41, based on the description text of each node in the diagnostic path, the information extraction technology of the predefined rule engine is used to identify and extract predefined key medical entities and their clinical quantitative values ​​to obtain a set of medical entity-quantitative value key-value pairs;

[0089] S42, based on the set of medical entity-quantified value key-value pairs, according to the node sequence relationship of the diagnostic path, each key-value pair is associated with its corresponding path node to form a medical entity-quantified value mapping relationship;

[0090] S43, based on the mapping relationship between medical entities and quantitative values, configures clinical quantitative values ​​into the corresponding fields of the preset medical service template and outputs personalized medical service plans.

[0091] For example, based on the descriptive text of each node in the diagnostic path (including semantic descriptions of node execution content such as "enhanced CT scan" or "warfarin anticoagulation therapy"), the information extraction technology of a predefined rule engine (based on a BERT-based entity recognition model) can be used to identify and extract predefined key medical entities (including drug names, test indicators, and other medical terms) and their clinical quantitative values ​​(such as contrast agent dosage of 5 mg / kg or blood pressure of 120 / 80 mmHg), generating a set of medical entity-quantitative value key-value pairs (the key being the entity identifier and the value being the real-time quantitative value). Based on this set of key-value pairs, according to the node sequence relationship of the diagnostic path (such as preoperative preparation nodes taking precedence over surgical operation nodes), a node topology mapping model can be established through a graph neural network to dynamically associate each key-value pair with its corresponding path node, forming a structured medical entity-quantitative value mapping relationship (such as mapping warfarin dosage = 5 mg to the anticoagulation therapy node). Based on this mapping relationship, the clinical quantitative value is automatically configured to a preset medical service. The system outputs personalized medical service plans (such as treatment plan documents arranged along a timeline) by including time series, execution content, and quantitative parameters in the corresponding fields (drug dosage field or detection threshold field) of the service template (such as electronic medical record templates conforming to the HL7 standard). High-precision entity recognition (such as extracting INR values ​​with a threshold of 2.5 from postoperative monitoring node text) is achieved through pre-trained models such as BERT, replacing traditional manual annotation and improving data extraction efficiency. A dynamic graph structure is constructed using node sequence relationships (treating nodes as graph vertices and sequence dependencies as edges) to achieve automatic association between key-value pairs and nodes, resolving adaptation biases in static mapping. The mapping relationships are injected into HL7 standard template fields (such as CDA documents), allowing the solution to be directly integrated into hospital information systems, improving executability. Through deep coupling of real-time data and path nodes, the system overcomes the limitations of traditional technologies where path solutions are detached from clinical quantitative data, combining individual adaptability (dynamic quantitative value binding) and clinical feasibility (standard template support), thus improving diagnostic efficiency.

[0092] In one embodiment, S51, the weighted correlation between each clinical similarity feature group and the target disease is calculated using the following formula:

[0093]

[0094] Among them, F k Grouping based on the k-th clinical similarity feature, where D is the target disease, w i For feature f i The preset weighting coefficients, sim(f) i D) is a feature f i The similarity function with the target disease D, where m is the feature group F. k The number of features in;

[0095] The feature cross-validation algorithm eliminates redundant feature groups in the initial personalized feature set, outputting personalized feature combinations, including:

[0096] S52, calculate the mutual information values ​​between different feature groups in the primary personalized feature set, forming the mutual information matrix MI = [mi ij ], where matrix element mi ij This represents the mutual information value between feature group i and feature group j;

[0097] S53, traverse the mutual information matrix MI, when mi ij If the value exceeds the preset third threshold, select one of feature group i and feature group j to retain;

[0098] S54 outputs the set of feature groups retained after traversing and filtering as a personalized feature combination.

[0099] For example, this can be achieved through the weighted correlation formula. Calculate the association strength between clinical similarity feature grouping and target disease, where F k Grouping for the k-th clinical similarity feature (e.g., a combination of cardiovascular-related symptoms), where D is the target disease (e.g., coronary artery disease), w i For feature f i The preset weighting coefficients (predefined by the medical knowledge base based on clinical importance), sim(f i D) is a feature f i The similarity function with the target disease D (calculated through the disease-feature correlation matrix, such as Euclidean distance or cosine similarity), where m is the feature group F. k The number of features in the dataset is used to quantify the overall association between feature groups and the disease through weighted aggregation (e.g., calculating the association value between the blood pressure fluctuation + dyslipidemia group and coronary heart disease); the mutual information value mi between different feature groups in the primary personalized feature set is calculated. ij This forms the mutual information matrix MI = mi ij ], (e.g., calculating the information overlap between the blood glucose fluctuation group and the insulin resistance group), traverse the matrix, when mi ij When the value exceeds a preset third threshold (e.g., 0.8), the highest-weighted group is selected based on the correlation coefficient W (e.g., retaining the glycated hemoglobin group while discarding the redundant fasting blood glucose group). The set of retained feature groups after screening is used as the final personalized feature combination (e.g., retaining 3 core features to form a diabetes-specific combination). This is achieved through the formula W(F... k D) To achieve numerical correlation assessment between feature grouping and disease, replacing subjective experience judgment; : Utilize mutual information matrix to identify and eliminate highly overlapping feature groups (such as when mi...). ij(When the value is >0.8, choose one of two) to ensure the independence and representativeness of feature combinations; in the screening of redundant feature groups, priority is given to retaining high-weight groups to strengthen key pathological features and solve the problem of rigid feature combinations in background technology.

[0100] In one embodiment, medical information is extracted using natural language processing technology to generate medical record data, including:

[0101] S61, perform word segmentation and medical terminology standardization integration on the original medical record text to obtain a standardized terminology sequence;

[0102] S62, perform semantic role labeling on the standardized term sequence to obtain the semantic labeling results;

[0103] S63 maps the semantic annotation results to preset structured fields to generate medical record data.

[0104] Specifically, the original medical record text (unstructured medical documents recorded by doctors, containing symptoms, signs, and treatment processes) is segmented. A pre-trained segmentation model (such as BERT-BiLSTM-CRF) is used to decompose the continuous text into discrete medical terminology units. Through medical terminology standardization and integration processing (based on UMLS or SNOMED CT terminology database), synonymous terms (such as myocardial infarction → myocardial infarction) are mapped to standard medical terms, generating a standardized terminology sequence. Semantic role labeling is performed on the standardized terminology sequence (using a semantic role labeling model based on an attention mechanism) to identify the semantic role of each term in the medical context (including the agent being the patient, the patient being the symptom, and the timestamp being the onset time, etc.), resulting in semantic labeling results containing role tags. The semantic labeling results are mapped to preset structured fields (achieved through a field mapping rule engine, such as mapping agent: patient + patient: chest pain + timestamp: 2023-05-01 to the chief complaint field patient 2023-05-01 chest pain), generating machine-readable structured medical record data. By using a professional terminology database, the semantics of clinical descriptions are normalized (e.g., palpitations → heart palpitations), thus resolving feature extraction biases caused by synonyms; semantic role labeling is used to capture the clinical relationships of medical entities (e.g., distinguishing between patient-reported chest pain and doctor-diagnosed chest pain), enhancing the semantic accuracy of data; and based on preset field rules, the automatic conversion of semantic tags to data fields is achieved (e.g., mapping the <time, symptom> combination to the present medical history field), replacing manual data entry and improving the machine readability of medical record data.

[0105] The aforementioned medical big data-based assisted diagnosis method acquires original medical record text and uses natural language processing (NLP) techniques to perform word segmentation, medical terminology standardization, and semantic role labeling to generate structured medical record data. The medical record data is then subjected to multi-dimensional clustering using a classification algorithm to generate clinical similarity feature groups. Based on a pre-set disease association model, the weighted correlation between each group and the target disease is calculated. Combined with clinical scenario adaptation parameters, a screening threshold is generated to form a primary personalized feature set. Redundant feature groups are eliminated by calculating a mutual information matrix, resulting in a personalized feature combination. Based on this combination, a disease classification model is used to determine the target disease type. When the individual correlation strength between the feature variable and the disease exceeds a threshold, an independent clinical causality test is performed based on a medical knowledge base to extract key medical features and their clinical quantitative values. When the quantitative values ​​meet a first pre-set threshold, a pre-set dynamic time-series model of the disease condition is used to further refine the key features. Historical sequence data is used to learn and generate trend prediction results for future time windows. Based on these results, combined with the diagnostic rule base, a preliminary diagnostic path is generated through inference. Similar path sets are extracted from the historical case database of the target disease type for deviation detection and optimization of node sequences to generate a diagnostic path. A rule engine is used to identify key medical entities and quantified values ​​from the diagnostic path node text. Based on the node sequence relationship, they are mapped to a preset medical service template to output a personalized solution. Multi-dimensional clustering and redundancy elimination solve the problem of key feature extraction deviation caused by the rigidity of traditional feature combination. Dynamic time series model compensates for the lack of disease evolution time series modeling capability in the existing system. Historical case deviation detection and real-time mapping of nodes and quantified values ​​solve the problem of disconnect between path generation and execution, realizing intelligent and personalized generation of diagnostic paths, improving diagnostic efficiency and the pertinence and feasibility of the solution.

[0106] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0107] Based on the same inventive concept, this application also provides a medical big data-based auxiliary diagnostic device for implementing the aforementioned medical big data-based auxiliary diagnostic method. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of a medical big data-based auxiliary diagnostic device provided below can be found in the limitations of the medical big data-based auxiliary diagnostic method described above, and will not be repeated here.

[0108] In one exemplary embodiment, such as Figure 2 As shown, an auxiliary diagnostic device based on medical big data is provided, including:

[0109] The medical record data processing module 101 is used to acquire the original medical record text and extract medical information through natural language processing technology to generate medical record data.

[0110] The feature clustering analysis module 102 is used to perform clustering analysis on medical record data and output personalized feature combinations.

[0111] The key feature determination module 103 is used to determine the target disease type and key medical features based on personalized feature combinations, and to obtain the clinical quantitative value of the key medical features.

[0112] The dynamic time series prediction module 104 is used to generate trend prediction results based on the key medical features and using a preset dynamic time series model of the disease if the clinical quantitative value of the key medical features meets the first preset threshold.

[0113] The diagnostic path generation module 105 is used to generate a diagnostic path by calling the diagnostic rule library corresponding to the target disease type based on the trend prediction results.

[0114] The service solution output module 106 is used to associate and map the key information of the diagnostic path with the path nodes, and output a personalized medical service solution.

[0115] In one embodiment, the feature clustering analysis module 102 is further configured to:

[0116] The classification algorithm is used to perform multi-dimensional clustering of individual characteristics in medical record data to generate clinical similarity feature groups;

[0117] The weighted correlation degree between each clinical similarity feature group and the target disease is calculated based on a pre-set disease association model;

[0118] Based on clinical scenario adaptation parameters, a screening threshold is generated, and the weight correlation is screened to retain feature groups that exceed the screening threshold to form a primary personalized feature set.

[0119] The feature cross-validation algorithm eliminates redundant feature groups in the initial personalized feature set and outputs a personalized feature combination.

[0120] In one embodiment, the key feature determination module 103 is further configured to:

[0121] Based on personalized feature combinations, the target disease type is determined through a disease classification model.

[0122] Calculate the association strength between each feature variable in the personalized feature combination and the individual's target disease type;

[0123] If the strength of the individual association exceeds the second preset threshold, then an independent clinical causal test based on a medical knowledge base or clinical guidelines is performed on the feature variables corresponding to the strength of the individual association.

[0124] If a feature variable passes the independent clinical causality test, it is identified as a key medical feature, and the clinical quantitative value of the key medical feature is extracted from the medical record data.

[0125] In one embodiment, the diagnostic path generation module 105 is further configured to:

[0126] The trend prediction results are combined with the diagnostic rule base to perform path combination reasoning and generate a preliminary diagnostic path scheme.

[0127] Based on the preliminary diagnostic pathway scheme, similar case pathway sets are extracted from the historical case database corresponding to the target disease type and deviation detection is performed to obtain pathway deviation information;

[0128] The node sequence of the preliminary diagnostic path plan is optimized and adjusted based on the path deviation information to generate a diagnostic path.

[0129] In one embodiment, the service scheme output module 106 is further configured to:

[0130] Based on the descriptive text of each node in the diagnostic path, the information extraction technology of the predefined rule engine is used to identify and extract predefined key medical entities and their clinical quantitative values, resulting in a set of medical entity-quantitative value key-value pairs.

[0131] Based on the set of medical entity-quantified value key-value pairs, each key-value pair is associated with its corresponding path node according to the node sequence relationship of the diagnostic path, forming a medical entity-quantified value mapping relationship;

[0132] Based on the mapping relationship between medical entities and quantitative values, clinical quantitative values ​​are configured into the corresponding fields of preset medical service templates to output personalized medical service plans.

[0133] In one embodiment, the feature clustering analysis module 102 is further configured to:

[0134] The weighted correlation between each clinical similarity feature group and the target disease is calculated using the following formula:

[0135]

[0136] Among them, F k Grouping based on the k-th clinical similarity feature, where D is the target disease, w i For feature f i The preset weighting coefficients, sim(f) i D) is a feature f i The similarity function with the target disease D, where m is the feature group F. k The number of features in;

[0137] The feature cross-validation algorithm eliminates redundant feature groups in the initial personalized feature set, outputting personalized feature combinations, including:

[0138] Calculate the mutual information values ​​between different feature groups in the primary personalized feature set to form a mutual information matrix MI = [mi ij ], where matrix element mi ij This represents the mutual information value between feature group i and feature group j;

[0139] Traverse the mutual information matrix MI, when mi ij If the value exceeds the preset third threshold, select one of feature group i and feature group j to retain;

[0140] The set of feature groups retained after traversal and filtering is output as a personalized feature combination.

[0141] In one embodiment, the medical record data processing module 101 is further configured to:

[0142] The original medical record text was segmented and standardized with medical terminology to obtain a standardized terminology sequence.

[0143] Semantic role labeling is performed on the standardized term sequence to obtain semantic labeling results;

[0144] The semantic annotation results are mapped to preset structured fields to generate medical record data.

[0145] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the aforementioned medical big data-based assisted diagnostic method.

[0146] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0147] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0148] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A diagnostic aid method based on medical big data, characterized in that, The method includes: The original medical record text is obtained, and medical information is extracted using natural language processing technology to generate medical record data; Cluster analysis is performed on the medical record data to output personalized feature combinations; Based on the personalized feature combination, the target disease type and key medical features are determined, and the clinical quantitative values ​​of the key medical features are obtained. If the clinical quantitative value of the key medical feature meets the first preset threshold, then based on the key medical feature, a trend prediction result is generated using a preset dynamic time series model of the disease condition. Based on the trend prediction results, the diagnostic rule base corresponding to the target disease type is invoked to generate a diagnostic path; The key information of the diagnostic path is associated and mapped with the path nodes to output a personalized medical service plan.

2. The method according to claim 1, characterized in that, The cluster analysis of the medical record data, outputting personalized feature combinations, includes: The individual characteristics in the medical record data are clustered in multiple dimensions using a classification algorithm to generate clinical similarity feature groups; The weighted correlation degree between each clinical similarity feature group and the target disease is calculated based on a preset disease association model; Based on clinical scenario adaptation parameters, a screening threshold is generated, and the weight correlation is screened to retain feature groups that exceed the screening threshold to form a primary personalized feature set. The feature cross-validation algorithm is used to eliminate redundant feature groups in the primary personalized feature set and output the personalized feature combination.

3. The method according to claim 1, characterized in that, The process of determining the target disease type and key medical features based on the personalized feature combination, and obtaining the clinical quantitative values ​​of the key medical features, includes: Based on the personalized feature combination, the target disease type is determined through a disease classification model; Calculate the strength of association between each feature variable in the personalized feature combination and the individual with the target disease type; If the individual association strength exceeds the second preset threshold, then an independent clinical causal test based on a medical knowledge base or clinical guidelines is performed on the feature variables corresponding to the individual association strength. If the feature variable passes the independent clinical causality test, it is determined to be a key medical feature, and the clinical quantitative value of the key medical feature is extracted from the medical record data.

4. The method according to claim 1, characterized in that, The step of generating a diagnostic path by calling the diagnostic rule base corresponding to the target disease type based on the trend prediction result includes: The trend prediction results are combined with the diagnostic rule base to perform path combination reasoning, generating a preliminary diagnostic path scheme; Based on the preliminary diagnostic path scheme, similar case path sets are extracted from the historical case database corresponding to the target disease type and deviation detection is performed to obtain path deviation information; The node sequence of the preliminary diagnostic path scheme is optimized and adjusted based on the path deviation information to generate the diagnostic path.

5. The method according to claim 1, characterized in that, The process of associating and mapping key information of the diagnostic path with path nodes to output a personalized medical service plan includes: Based on the description text of each node in the diagnostic path, the information extraction technology of the predefined rule engine is used to identify and extract predefined key medical entities and their clinical quantitative values ​​to obtain a set of medical entity-quantitative value key-value pairs. Based on the set of medical entity-quantified value key-value pairs, and according to the node sequence relationship of the diagnostic path, each key-value pair is associated with its corresponding path node to form a medical entity-quantified value mapping relationship. Based on the medical entity-quantitative value mapping relationship, the clinical quantitative value is configured into the corresponding field of the preset medical service template, and the personalized medical service plan is output.

6. The method according to claim 2, characterized in that, The weighted correlation between each clinical similarity feature group and the target disease is calculated using the following formula: Among them, F k Grouping based on the k-th clinical similarity feature, where D is the target disease, w i For feature f i The preset weighting coefficients, sim(f) i D) is a feature f i The similarity function with the target disease D, where m is the feature group F. k The number of features in; The step of eliminating redundant feature groups in the initial personalized feature set through a feature cross-validation algorithm and outputting the personalized feature combination includes: Calculate the mutual information values ​​between different feature groups in the primary personalized feature set to form a mutual information matrix MI = [mi ij ], where matrix element mi ij This represents the mutual information value between feature group i and feature group j; Traverse the mutual information matrix MI, when mi ij If the value exceeds the preset third threshold, select one of feature group i and feature group j to retain; The set of feature groups retained after traversal and filtering is output as the personalized feature combination.

7. The method according to claim 1, characterized in that, The process of extracting medical information and generating medical record data using natural language processing technology includes: The original medical record text is subjected to word segmentation and medical terminology standardization integration processing to obtain a standardized terminology sequence; Semantic role labeling is performed on the standardized term sequence to obtain semantic labeling results; The semantic annotation results are mapped to preset structured fields to generate the medical record data.

8. An auxiliary diagnostic device based on medical big data, characterized in that, The device includes: The medical record data processing module is used to acquire the original medical record text and extract medical information through natural language processing technology to generate medical record data; The feature clustering analysis module is used to perform clustering analysis on the medical record data and output personalized feature combinations; The key feature determination module is used to determine the target disease type and key medical features based on the personalized feature combination, and to obtain the clinical quantitative value of the key medical features. The dynamic time-series prediction module is used to generate a trend prediction result based on the key medical features and using a preset dynamic time-series model of the disease condition if the clinical quantitative value of the key medical feature meets the first preset threshold. The diagnostic path generation module is used to generate a diagnostic path by calling the diagnostic rule library corresponding to the target disease type based on the trend prediction results. The service solution output module is used to associate and map the key information of the diagnostic path with the path nodes, and output a personalized medical service solution.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images