Intelligent physical examination data quality control analysis method and system

Abstract

The invention discloses an intelligent physical examination data quality control analysis method and system, and relates to the technical field of medical quality control. The problems that existing physical examination data are diversified in source, complex in format, single in quality control means and the like are solved. Structured data, medical images and texts are integrated, and standardized processing and data mapping are achieved through a unified platform. Static and dynamic rule libraries are constructed, and the threshold is dynamically adjusted in combination with individual features, so that personalized risk assessment is realized. Structured data anomaly detection is performed by adopting machine learning, an image quality problem is identified by utilizing a convolutional neural network, and text anomaly is processed and analyzed through a natural language, so that the anomaly detection accuracy is improved. The system realizes multi-dimensional quality control and intelligent early warning, enhances the integrity and credibility of physical examination data, supports precise health management and disease early warning, improves the quality control efficiency, meets the intelligent analysis requirements of large-scale multi-modal physical examination data, and promotes the application of intelligent health management.

Description

Technical Field

[0001] The present invention relates to the field of medical quality control technology, and specifically to an intelligent health examination data quality control analysis method and system. Background Art

[0002] Current health checkup data quality control relies primarily on manual data entry and logical verification based on preset rules. This covers structured vital signs and laboratory test data, but lacks the ability to process unstructured data such as medical images and medical interview texts. Due to the diverse sources, complex and dispersed formats, and inconsistent data standards of health checkup data, data entry errors, missing information, and data inconsistencies, traditional static rules struggle to dynamically adapt to individual differences and data changes, and their limited anomaly detection methods result in low quality control efficiency and poor accuracy. This makes it difficult to meet the needs of intelligent analysis of large-scale, multimodal health checkup data, severely restricting the scientific and practical nature of health management and disease early warning.

[0003] The core problem facing the existing health examination data quality control is that the sources of health examination data are diverse and the formats are not uniform. Traditional static rules and single anomaly detection methods are difficult to meet the needs of data fusion, dynamic adjustment and personalized risk assessment, resulting in low quality control efficiency and insufficient accuracy, which seriously affects the reliability and effectiveness of health management. This problem makes it difficult to achieve comprehensive and accurate quality control of health examination data, increases the risk of erroneous data, affects subsequent health assessments and disease warnings, limits the promotion and application of intelligent health management systems, and reduces the overall level and efficiency of medical services. Summary of the Invention

[0004] In response to the problems mentioned in the above background technology, an intelligent health examination data quality control and analysis method and system are proposed.

[0005] The purpose of the present invention can be achieved through the following technical solutions: A smart health examination data quality control and analysis method, comprising:

[0006] S001: Collect multi-source physical examination data to obtain initial data; use a standardized data processing platform to perform field standardization, unit unification, and project mapping to obtain standardized data;

[0007] S002: Constructing a static rule base and a dynamic rule base to perform intelligent physical examination analysis on the standard data;

[0008] The static rule base is constructed by extracting original data diagnosis labels from a multi-source heterogeneous clinical database, modeling and evaluating the relationship between variables and diseases through statistical methods to obtain feature importance ranking; the feature importance ranking is subjected to clinical guideline standards and expert recommendations to obtain static rule items, and the static rule items are tested to obtain static rule entry items; and the obtained static rule entry items are uploaded to the static rule base;

[0009] Build a dynamic rule library, collect basic patient information, extract unstructured medical records through natural language processing, and construct structured individual feature vectors. For each chronic disease, establish a mapping table to provide the starting value for continuous dynamic calculation. Build a dynamic threshold adjustment model to obtain a dynamic offset driven by individual features. Replace the original threshold in the static rule with the dynamically adjusted final discrimination threshold to form an individualized rule entity. Obtain individualized rule entities through the application of a rule engine for item-by-item quality control. Compare the final discrimination threshold with the actual detection value. When the final discrimination threshold is less than or equal to the actual detection value, trigger the decision path output.

[0010] S003: Obtaining physical examination data for abnormality detection and generating structured prompt information;

[0011] S004: Summarize the abnormality mark output; obtain the unified summary of three types of abnormality marks: structured, image, and text, and generate a physical examination abnormality report.

[0012] As a preferred embodiment of the present invention, the specific process of testing the static rule items includes process testing and correlation testing;

[0013] The process inspection is to standardize the format of each static rule. When the rule expression meets the parsing requirements of the rule engine, the variables involved in the rule items are mapped one by one with the data source fields to identify field spelling differences, unit inconsistencies or missing variables. If the variables referenced in the rule are named FBG in the original data, the standardized replacement will be completed through the field mapping table. If there are no field spelling differences, unit inconsistencies or missing variables, the boundary value input test of the key variables in the rule is performed and simulation is performed. The missing test rule can be correctly triggered. If it can be correctly triggered, the rule base is preloaded and simulation reasoning is performed one by one. Whether each rule is triggered, the triggering frequency and output status are recorded to check for invalid rules that are never triggered.

[0014] The correlation detection is used to detect abnormal correlations between variables in the physical examination data; based on the correlation modeling results stored in the static rule base, a set of statistically verified highly correlated variable pairs is obtained; by constructing a standard correlation model, for subsequent comparison of actual physical examination data: using a regression model, a function expression between variables is established to obtain the correlation between variables;

[0015] Obtain the correlation matching judgment of real-time physical examination data and compare the deviation between the actual value and the predicted value. The process is as follows: Where δ is the deviation between the actual value and the predicted value; if δ exceeds the set threshold, it is judged as a correlation anomaly;

[0016] When it is detected that the correlation corresponding to a variable deviates seriously from the model expectation, it is marked as a correlation anomaly item, and the anomaly level is set according to the degree of deviation; according to the degree of deviation, the anomaly level is set, and a quality control report prompt is further generated.

[0017] As a preferred embodiment of the present invention, the process of obtaining the physical examination data for abnormality detection and generating structured prompt information is as follows:

[0018] The anomaly detection includes three data dimensions, namely, anomaly detection of structured data, anomaly detection of medical imaging data, and anomaly detection of data;

[0019] Detect anomalies in structured data and obtain physical examination indicators of corresponding items;

[0020] Anomaly detection in medical imaging data triggers anomaly marking;

[0021] Generate structured prompt information for data anomaly detection.

[0022] As a preferred embodiment of the present invention, the specific process of detecting anomalies in the structured data to obtain corresponding physical examination indicators is as follows:

[0023] Generate the corresponding vector feature P1 based on the patient portrait, calculate the degree of abnormality through reconstruction error, and mark the sample as abnormal if the reconstruction loss is greater than the preset threshold; specifically:

[0024] For each physical examination indicator, the abnormality score is calculated using the formula:

[0025] Score = f 型 (x 标 , P1), where x 标 is a structured indicator, P1 is the individual characteristic of the patient, and f 型 is a trained detection model; the static rule base or the dynamic rule base gives a threshold T for comparison and judgment. When the score is greater than T, the corresponding physical examination indicator is marked as abnormal.

[0026] As a preferred embodiment of the present invention, the specific process of triggering an abnormality mark for the abnormality detection of the medical imaging data is as follows:

[0027] A set of image quality recognition models based on convolutional neural networks (CNN) are pre-trained; when image data is received, the evaluation process is started: the pre-processed image is input into the CNN model, the deep feature vector is extracted, and the probability score of multiple abnormal dimensions is output through the classification head: P IMG ={p 模糊 ,p 失踪 ,p 闭塞}, where p 模糊 is the image blur probability, p失踪 is the target area missing probability and p 闭塞 To block the interference probability; then set the quality threshold according to each abnormal category, which is configured by the doctor or system administrator in the rule base; if an abnormal dimension is greater than the set quality threshold of the corresponding category, the abnormal flag is triggered.

[0028] As a preferred embodiment of the present invention, the specific process of generating structured prompt information based on the abnormality detection of the data is as follows:

[0029] Perform basic language processing on the raw text data; use a rule-based segmentation algorithm combined with a semantic stop word list to segment the patient's self-report text into text sentences, and use named entity recognition to extract text entities; then normalize the identified entities to construct a structured entity table; use the extracted entities as graph nodes, establish semantic edges between entities, and form a semantic graph for individual patients;

[0030] Based on the obtained semantic graph, semantic analysis is performed to determine whether there are conflicting or inconsistent logical relationships between texts; this includes contradiction detection and semantic similarity calculation;

[0031] Among them, contradiction detection uses the natural language inference model to enter two sentences to determine whether their relationship is entailment, neutrality or contradiction;

[0032] Semantic similarity calculation, converting text pairs into sentence vectors Consistency assessment is performed using cosine similarity: If the score is lower than the preset consistency threshold, it is determined to be a semantic conflict.

[0033] Another aspect of the present invention provides an intelligent health checkup data quality control and analysis system, comprising: a data access module, a data standardization module, a rule engine module, an anomaly detection module and a quality control report module;

[0034] The data access module uniformly collects multi-source physical examination data and transfers the original data to the data standardization module;

[0035] The data standardization module maps fields, unifies units, and standardizes terminology for data from different sources to form a unified data format; the cleaned data flows into the rule judgment module and the anomaly identification module;

[0036] The rule engine module makes preliminary judgments on standardized data based on static and dynamic rules; it transmits the potential problem items identified to the anomaly identification module for further verification or supplementation;

[0037] The anomaly detection module deeply mines potential anomalies for different data types;

[0038] The quality control report module integrates rule judgment and abnormality detection results to output physical examination abnormality reports.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention integrates structured data (such as blood pressure and blood sugar), medical imaging data (such as ultrasound and CT scans), and text data (such as medical records and chief complaints). Through a unified data processing platform, it achieves field standardization, unit unification, and project mapping. This multi-source data fusion and standardization ensures data accuracy and consistency, providing a reliable data foundation for subsequent quality control analysis.

[0041] 2. The present invention constructs a static rule base for standardized risk identification and early warning judgment of common diseases, which has the characteristics of high stability and high versatility. At the same time, a dynamic rule base is established to dynamically adjust the thresholds in the static rules, thereby generating individualized auxiliary decision paths. This combination of static and dynamic rule bases enhances the system's clinical identification accuracy and personalized risk assessment capabilities for different individuals.

[0042] 3. This invention performs anomaly detection on structured data, medical imaging data, and text data. It uses machine learning models (such as AutoEncoder) for unsupervised anomaly detection, employs convolutional neural network (CNN) models to identify image quality issues, and uses natural language processing (NLP) techniques to analyze semantic conflicts and inconsistent descriptions in text data. This multi-dimensional anomaly detection and intelligent early warning mechanism improves quality control efficiency and accuracy, ensuring the logical integrity and usability of medical examination data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0044] Figure 1 A diagram showing the steps of the method of the present invention;

[0045] Figure 2 This is a principle block diagram of the present invention. DETAILED DESCRIPTION

[0046] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] It should be understood that the terms “include” and “comprising” used in the specification and claims of the present disclosure indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0048] It should also be understood that the terminology used in this disclosure is for the purpose of describing specific embodiments only and is not intended to limit the disclosure. As used in this disclosure and the claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should be further understood that the term "and / or" as used in this disclosure and the claims refers to any and all possible combinations of one or more of the associated listed items, including and including these combinations.

[0049] See also Figure 1 As shown, the present invention provides a smart health examination data quality control and analysis method, comprising:

[0050] S001: Collect multi-source physical examination data, recorded as initial data, including: structured data, image data and text data;

[0051] Structured data includes basic physical examination data such as blood pressure and blood sugar, which comes from physical examination equipment (such as electronic blood pressure monitors and blood glucose meters). Image data includes medical images such as ultrasound and CT scans, which are transmitted in DICOM format and uploaded to imaging systems or cloud platforms. Using relevant APIs, image files are transferred to standardized data processing platforms for subsequent processing. Text data includes medical records, chief complaints, and other text data, which comes from questionnaire systems or doctors' handwritten notes. Natural language processing (NLP) technology is used to parse, clean, and convert them into structured data.

[0052] The data is uniformly entered into a standardized data processing platform for field standardization, unit unification, project mapping, and other processing to obtain standard data;

[0053] S002: Build a static rule base and a dynamic rule base to perform intelligent data inspection, specifically:

[0054] In a preferred embodiment of the present invention, a static rule base is first constructed; the rule base is used to realize standardized risk identification and early warning judgment of common diseases, and has the characteristics of high stability and high versatility;

[0055] The process of building a static rule base includes but is not limited to the following steps:

[0056] S211: Extracting raw data from multi-source heterogeneous clinical databases. In one specific example, 20,000 diabetes-related cases covering a three-year period were extracted from a tertiary hospital database, encompassing over 200 clinical variables and final diagnosis labels.

[0057] S212: To ensure the accuracy and transferability of rule generation, the original data is processed, including missing value filling, outlier processing, and feature construction;

[0058] S213: Model and evaluate the relationship between variables and diseases using a variety of statistical methods, including:

[0059] For the relationship between categorical variables and disease labels, the chi-square test was used to assess its significance;

[0060] For the correlation between continuous variables (such as blood glucose and blood pressure) and diseases, Spearman rank correlation analysis was used;

[0061] Introducing the Mutual Information method to identify nonlinear relationships;

[0062] Then set the p-value threshold (e.g., p < 0.01) and the correlation coefficient threshold (e.g., r > 0.6) to screen key variables; perform machine learning analysis on the screened key variables to obtain feature importance ranking;

[0063] During the actual implementation process, it was found that HbA1c, fasting blood glucose, BMI, and family history were highly correlated with the diagnosis of diabetes, among which the Pearson correlation coefficient of HbA1c reached 0.86 and the p value was less than 0.001.

[0064] S214: Based on the above analysis results, combined with clinical guideline standards and expert recommendations, static rule items are generated; the generated static rule items are tested to obtain static rule library items; finally, the obtained static rule library items are uploaded to the static rule library as a basis for subsequent dynamic rule adjustments;

[0065] The specific process of testing the static rule items includes process testing and correlation testing;

[0066] The process is verified by normalizing the format of each static rule. When the rule expression meets the parsing requirements of the rule engine, the variables involved in the rule items are mapped one by one to the data source fields to identify problems such as field spelling differences, inconsistent units, or missing variables. If the variable referenced in the rule (such as "fasting blood glucose") is named "FBG" in the original data, it will be standardized and replaced through the field mapping table to ensure semantic consistency. If there are no problems such as field spelling differences, inconsistent units, or missing variables, the key variables in the rule are tested for boundary values ​​(such as maximum, minimum, and missing) and simulated. For example, HbA1c=0, HbA1c=15, HbA1c, missing test rules are tested to see if they can be correctly triggered to avoid data anomalies causing rule execution failure. If they can be correctly triggered, the rule library is preloaded and simulation reasoning is performed one by one. The triggering, triggering frequency, and output of each rule are recorded to identify invalid rules that "never trigger". This step not only verifies the correctness of execution but also evaluates the coverage and practical value of the rules.

[0067] The correlation test is mainly used to detect abnormal correlations between variables in physical examination data. That is, although some variables individually meet the reference range, they show statistical inconsistency or potential data anomalies when combined, thereby identifying "hidden errors". The detection process is as follows:

[0068] Based on the correlation modeling results stored in the static rule base, a set of statistically verified highly correlated variable pairs (or multivariate combinations) is obtained, such as [ALT, AST], with a correlation coefficient of r = 0.85, [HbA1c, FBG], with r = 0.86, and [SBP, DBP] (systolic blood pressure, diastolic blood pressure). Note: These variable pairs should exhibit a linear or monotonic relationship in normal samples.

[0069] By constructing a standard correlation model, we can prepare for the comparison of the actual physical examination data in the future: using regression model (linear / logistic regression) to establish the functional expression between variables;

[0070] Example expression:

[0071] HbA1c_expected=a×FBG+b

[0072]

[0073] Obtain correlation matching judgment of real-time physical examination data. For each piece of physical examination data, extract the actual value corresponding to the target variable, substitute it into the corresponding reference model, calculate the predicted value or reference interval, and compare the deviation between the actual value and the predicted value. The process is as follows:

[0074] Where δ is the deviation between the actual value and the predicted value; if δ exceeds the set threshold (such as 20%), it is determined to be a correlation anomaly;

[0075] When the correlation corresponding to a variable is detected to be seriously deviated from the model expectation, it is marked as a "correlation abnormality item" and the abnormality level (mild / moderate / severe) is set according to the degree of deviation. The abnormality level is set according to the degree of deviation, and a quality control report prompt is further generated. For example, if the HbA1c and fasting blood glucose values ​​do not match, there may be a test error or sample contamination, and retesting is recommended.

[0076] Finally, the relevant abnormal records are written into the quality control data table of the current physical examination and input into the next stage as part of the comprehensive credibility score. Combined with the expert knowledge system, automated treatment suggestions are output, such as prompting re-examination, suggesting attention to a certain indicator, and statistically analyzing rare but clinically acceptable individual differences.

[0077] In a preferred embodiment of the present invention, in order to improve the system's clinical recognition accuracy and personalized risk assessment capabilities for different individuals, a dynamic rule base is further proposed to dynamically adjust the thresholds in static rules, thereby generating personalized auxiliary decision paths, specifically:

[0078] S221: Collect basic patient information, extract unstructured medical records through natural language processing, and construct structured individual feature vectors. The collected features include but are not limited to: age, gender, weight, height, body mass index (BMI), confirmed disease types (such as diabetes and hypertension), and the severity level of each disease (e.g., level 0 to level III);

[0079] In practical applications, the patient profile is represented as a vector: p = [age, gender, weight, height, body mass index, disease type, severity]

[0080] S222: For each chronic disease (such as diabetes, hypertension, etc.), a mapping table is established to record the basic judgment thresholds corresponding to different severity levels, which serve as the starting values ​​for subsequent dynamic calculations;

[0081] For example, the HbA1c indicator for diabetes is defined as follows:

[0082] Severity Level Judgment threshold 0 6.0％ Level 1 6.2％ Level 2 6.5％ Level 3 6.8％

[0083] The mapping table can be obtained through retrospective clinical data statistics, or it can be set by expert group review.

[0084] S223: Construct a dynamic threshold adjustment model. The basic form of the model is:

[0085] T_final = T_base + Δ, where: T_final is the final discrimination threshold after dynamic adjustment; T_base is the static basic threshold given in the mapping table; Δ is the dynamic offset driven by individual features;

[0086] The dynamic offset driven by the individual feature is calculated by Δ=α×f1+β×f2+γ×f3+υ×f4; wherein α, β, γ and υ are weighted coefficients obtained through historical data training; functions f1 to f4 represent mapping functions for standardizing the original features; f1 is a segmented processing of age, for example, ≥65 years old is considered old age, and the mapping is +0.2; f2 is an adjustment item for female gender, in which the adjustment item is -0.05; f3 is an offset of -0.1 if the weight is below the normal range (such as weight <50 kg); f4 is based on the disease level, and the corresponding offset value is +0.1, such as level II;

[0087] S224: replacing the original threshold in the static rule with the dynamically adjusted final discrimination threshold to form an individualized rule entity;

[0088] S231: By applying the rule engine to obtain individualized rule entities for item-by-item quality control, the final discrimination threshold is compared with the actual test value (such as the current HbA1c). When the final discrimination threshold is less than or equal to the actual test value, the decision path output is triggered; for example, it may indicate the clinical risk level (such as "moderate risk"), recommend measures (such as "OGTT test recommended"), etc. Conversely, when the final discrimination threshold is greater than the actual test value, it is determined that the patient's condition is well controlled and can continue to be observed, and regular review is recommended;

[0089] S232: Performing a physical examination based on a static rule base and a dynamic rule base;

[0090] S003: Obtain physical examination data for anomaly detection; the anomaly detection includes but is not limited to three types of data dimensions:

[0091] S301: Perform anomaly detection on structured data, including blood pressure, blood sugar, heart rate, blood lipids, liver and kidney function indicators, etc.

[0092] Based on the patient portrait, the corresponding vector feature P1 is generated. Then, a machine learning model is introduced for unsupervised anomaly detection. The AutoEncoder model is used to calculate the degree of anomaly through reconstruction error. If the reconstruction loss is greater than the preset threshold, the sample is marked as abnormal. Specifically:

[0093] For each physical examination indicator, the abnormality score is calculated using the formula:

[0094] Score = f 型 (x 标 , P1), where x 标is a structured indicator, P1 is the individual characteristic of the patient, and f 型 is a trained detection model; the static rule base or dynamic rule base then gives a threshold T for comparison and judgment. When the score is greater than T, the corresponding physical examination indicator is marked as abnormal;

[0095] For example, if the blood glucose value is detected to be 14.2mmol / L and the reconstruction error is greater than 0.5, the system will mark it as "abnormally high blood glucose" and prompt "recheck HbA1c is recommended." If the systolic blood pressure is 78mmHg, combined with individual characteristics P (such as advanced age, low weight) and is judged as an abnormal point by the AutoEncoder model, a "low blood pressure risk warning" will be prompted;

[0096] S302: Perform anomaly detection on medical imaging data, mainly targeting various medical images such as ultrasound images, CT images, and MRI images acquired during routine physical examinations;

[0097] In the image anomaly detection module, a set of image quality recognition models based on convolutional neural networks (CNN) are pre-trained to detect a variety of anomalies, including but not limited to the following categories:

[0098] Blurred image detection (whole or partial blurring of the image due to patient positioning, probe jitter, or abnormal device focus), target area missing detection (the image fails to contain the complete organ area required for diagnosis (e.g., the complete liver, kidney, or heart structure is not displayed)), and image occlusion detection (unexpected object interference in the image, such as the technician's fingers, operating instruments, or probe ghosting, etc.);

[0099] When the image data is received, the evaluation process is started: after image preprocessing, the preprocessed image is input into the CNN model, the deep feature vector is extracted, and the probability score of multiple abnormal dimensions is output through the classification head: P IMG ={p 模糊 ,p 失踪 ,p 闭塞}, where p 模糊 is the image blur probability, p 失踪 is the target area missing probability and p 闭塞 The probability of occlusion interference is calculated. A quality threshold (e.g., 0.6) is set for each abnormality category and configured in the rule base by the doctor or system administrator. If a certain abnormal dimension is greater than the set quality threshold for the corresponding category, an abnormality mark is triggered. Note: Once an abnormal image is detected, immediate processing is performed: abnormality annotation: the image is marked as "pending review" in the physical examination system; abnormality type is prompted: the specific abnormality category is fed back to the imaging technician's workstation, for example: "The image is blurred, please re-acquire the liver B-ultrasound image";

[0100] S303: Perform data anomaly detection to identify potential issues such as semantic conflicts, inconsistent descriptions, and grammatical errors in texts such as physical examination questionnaires, medical history self-reports, and initial diagnosis records, ensuring the logical integrity and availability of information in electronic health records;

[0101] First, perform basic language processing on the original text data; use a rule-based segmentation algorithm combined with a semantic stop word list to segment the patient's self-report text into sentences; then use a Jieba word segmenter to segment words and construct a bag-of-words model;

[0102] Use the named entity recognition model to obtain text sentences in the bag-of-words model for entity extraction, including disease names, symptom descriptions, medicines, lifestyles, etc.

[0103] Next, the identified entities are normalized. For example, “high blood sugar,” “high blood sugar,” and “abnormal sugar value” are all grouped into the “high blood sugar” concept node.

[0104] Finally, construct a structured entity table, such as: {"medical history": ["deny diabetes"], "current medication": ["daily use of insulin"]};

[0105] The extracted entities are then used as graph nodes to establish semantic edges between entities (e.g., “deny” → “diabetes” constitutes a negation relationship), forming a semantic graph for individual patients.

[0106] Based on the obtained semantic graph, semantic analysis is performed to determine whether there are conflicting or inconsistent logical relationships between texts; this includes contradiction detection and semantic similarity calculation;

[0107] Among them, contradiction detection uses the Natural Language Inference (NLI) model to input two sentences (or two entity sentence pairs) to determine whether their relationship is "implication, neutrality, or contradiction";

[0108] Example input: Sentence pair 1:

[0109] Q1: “Denial of history of diabetes”

[0110] Q2: “Daily use of insulin”

[0111] Output label: contradiction

[0112] Among them, semantic similarity calculation converts text pairs into sentence vectors Consistency assessment is performed using cosine similarity: If the score is lower than the preset consistency threshold (e.g., 0.3), it is determined to be a semantic conflict or description inconsistency; the consistency threshold can be dynamically adjusted based on the individualized rule base;

[0113] After the model detects a contradiction, it generates a structured prompt message, which includes the text location of the abnormal sentence and the targeted contradictory content.

[0114] S004: Summarize the abnormality mark output, obtain a unified summary of the three types of abnormality marks: structured, image, and text, and generate a physical examination abnormality report.

[0115] See also Figure 2 As shown, the present invention provides an intelligent health examination data quality control and analysis system, including: a data access module, a data standardization module, a rule engine module, an anomaly detection module and a quality control report module;

[0116] The data access module uniformly collects multi-source physical examination data, including structured data (blood routine, urine test, etc.), image data (CT, B-ultrasound, etc.) and text data (medical history, chief complaint, etc.); and transmits the raw data to the data standardization module;

[0117] The data standardization module performs field mapping, unit unification, and terminology standardization (such as connecting to Loinc / SNOMED) on data from different sources to form a unified data format; the cleaned data flows into the rule judgment module and the anomaly identification module;

[0118] The rule engine module makes preliminary judgments on standardized data based on static and dynamic rules; it transmits the potential problem items identified to the anomaly identification module for further verification or supplementation;

[0119] The anomaly detection module deeply mines potential anomalies for different data types:

[0120] Structured anomaly detection is based on unsupervised models to detect outliers (such as isolation forest and AutoEncoder);

[0121] Image anomaly detection uses CNN to detect image quality or missing key parts;

[0122] Text anomaly detection uses NLP to detect semantic contradictions;

[0123] Summarize abnormal results and rule judgment results into the quality control report module;

[0124] The quality control report module integrates rule judgment and abnormality detection results to output physical examination abnormality reports.

[0125] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for quality control and analysis of intelligent health examination data, characterized in that: include: S001: Collect multi-source physical examination data to obtain initial data; Through the standardized data processing platform, field standardization, unit unification, and project mapping are carried out to obtain standard data; S002: Constructing a static rule base and a dynamic rule base to perform intelligent physical examination analysis on the standard data; The static rule base is constructed to extract original data diagnosis labels from multi-source heterogeneous clinical databases, and the relationship between variables and diseases is modeled and evaluated by statistical methods to obtain feature importance ranking; The static rule items are obtained by ranking the feature importance according to the clinical guideline standards and expert recommendations, and the static rule items are tested to obtain the static rule library items; and the obtained static rule library items are uploaded to the static rule library; Build a dynamic rule library, collect basic patient information, extract unstructured medical records through natural language processing, and construct structured individual feature vectors. For each chronic disease, establish a mapping table to provide the starting value for continuous dynamic calculation. Build a dynamic threshold adjustment model to obtain dynamic offsets driven by individual features. The final discrimination threshold after dynamic adjustment replaces the original threshold in the static rule to form an individualized rule entity; By applying the rule engine to obtain individual rule entities for item-by-item quality control, the final discrimination threshold is compared with the actual detection value. When the final discrimination threshold is less than or equal to the actual detection value, the decision path output is triggered; S003: Obtaining physical examination data for abnormality detection and generating structured prompt information; S004: Summarize the abnormal mark output; Obtain a unified summary of three types of abnormal labels: structured, image, and text, and generate a physical examination abnormality report.

2. The method for quality control and analysis of intelligent health examination data according to claim 1, characterized in that: The specific process of testing the static rule items includes process testing and correlation testing; The process inspection is to standardize the format of each static rule. When the rule expression meets the parsing requirements of the rule engine, the variables involved in the rule items are mapped one by one with the data source fields to identify field spelling differences, unit inconsistencies or missing variables. If the variables referenced in the rule are named FBG in the original data, the standardized replacement will be completed through the field mapping table. If there are no field spelling differences, unit inconsistencies or missing variables, the boundary value input test of the key variables in the rule is performed and simulation is performed. The missing test rule can be correctly triggered. If it can be correctly triggered, the rule base is preloaded and simulation reasoning is performed one by one. Whether each rule is triggered, the triggering frequency and output status are recorded to check for invalid rules that are never triggered. The correlation detection is used to detect abnormal correlations between variables in the physical examination data; based on the correlation modeling results stored in the static rule base, a set of statistically verified highly correlated variable pairs is obtained; by constructing a standard correlation model, for subsequent comparison of actual physical examination data: using a regression model, a function expression between variables is established to obtain the correlation between variables; Obtain the correlation matching judgment of real-time physical examination data and compare the deviation between the actual value and the predicted value. The process is as follows: Where δ is the deviation between the actual value and the predicted value; if δ exceeds the set threshold, it is judged as a correlation anomaly; When it is detected that the correlation corresponding to a variable deviates seriously from the model expectation, it is marked as a correlation anomaly item, and the anomaly level is set according to the degree of deviation; according to the degree of deviation, the anomaly level is set, and a quality control report prompt is further generated.

3. The method for quality control and analysis of intelligent health examination data according to claim 1, characterized in that: The process of obtaining the physical examination data for abnormality detection and generating structured prompt information is as follows: The anomaly detection includes three data dimensions, namely, anomaly detection of structured data, anomaly detection of medical imaging data, and anomaly detection of data; Detect anomalies in structured data and obtain physical examination indicators of corresponding items; Anomaly detection in medical imaging data triggers anomaly marking; Generate structured prompt information for data anomaly detection.

4. The method for quality control and analysis of intelligent health examination data according to claim 1, characterized in that: The specific process of detecting anomalies in the structured data and obtaining corresponding physical examination indicators is as follows: Generate the corresponding vector feature P1 based on the patient portrait, calculate the degree of abnormality through reconstruction error, and mark the sample as abnormal if the reconstruction loss is greater than the preset threshold; specifically: For each physical examination indicator, the abnormality score is calculated using the formula: Score = f 型 (x 标 , P1), where x 标 is a structured indicator, P1 is the individual characteristic of the patient, and f 型 is a trained detection model; the static rule base or the dynamic rule base gives a threshold T for comparison and judgment. When the score is greater than T, the corresponding physical examination indicator is marked as abnormal.

5. The method for quality control and analysis of intelligent health examination data according to claim 1, characterized in that: The specific process of triggering abnormality marking for abnormality detection of the medical imaging data is as follows: A set of image quality recognition models based on convolutional neural networks (CNN) are pre-trained; when image data is received, the evaluation process is started: the pre-processed image is input into the CNN model, the deep feature vector is extracted, and the probability score of multiple abnormal dimensions is output through the classification head: P IMG ={p 模糊 ,p 失踪 ,p 闭塞 }, where p 模糊 is the image blur probability, p 失踪 is the target area missing probability and p 闭塞 To block the interference probability; then set the quality threshold according to each abnormal category, which is configured by the doctor or system administrator in the rule base; if an abnormal dimension is greater than the set quality threshold of the corresponding category, the abnormal flag is triggered.

6. The method for quality control and analysis of intelligent health examination data according to claim 1, characterized in that: The specific process of generating structured prompt information based on anomaly detection of the data is as follows: Perform basic language processing on the raw text data; use a rule-based segmentation algorithm combined with a semantic stop word list to segment the patient's self-report text into text sentences, and use named entity recognition to extract text entities; then normalize the identified entities to construct a structured entity table; use the extracted entities as graph nodes, establish semantic edges between entities, and form a semantic graph for individual patients; Based on the obtained semantic graph, semantic analysis is performed to determine whether there are conflicting or inconsistent logical relationships between texts; this includes contradiction detection and semantic similarity calculation; Among them, contradiction detection uses the natural language inference model to enter two sentences to determine whether their relationship is entailment, neutrality or contradiction; Semantic similarity calculation, converting text pairs into sentence vectors Consistency assessment is performed using cosine similarity: If the score is lower than the preset consistency threshold, it is determined to be a semantic conflict.

7. An intelligent health examination data quality control and analysis system, characterized in that: Applicable to implementing any one of claims 1-6 of the intelligent health examination data quality control analysis method, the system comprises: a data access module, a data standardization module, a rule engine module, an anomaly detection module and a quality control report module; The data access module uniformly collects multi-source physical examination data and transfers the original data to the data standardization module; The data standardization module maps fields, unifies units, and standardizes terminology for data from different sources to form a unified data format; the cleaned data flows into the rule judgment module and the anomaly identification module; The rule engine module makes preliminary judgments on standardized data based on static and dynamic rules; it transmits the potential problem items identified to the anomaly identification module for further verification or supplementation; The anomaly detection module deeply mines potential anomalies for different data types; The quality control report module integrates rule judgment and abnormality detection results to output physical examination abnormality reports.

Images