A method for adjusting priority of physical examination report conclusion based on user medical history

Abstract

This invention relates to the field of report optimization technology, and in particular to a method for prioritizing medical examination report conclusions based on user medical history. The method includes: collecting multi-source data; preprocessing the multi-source data to obtain a structured multi-source dataset; based on the structured multi-source dataset, mining the correlation between health data and diseases through association analysis, constructing a health trend prediction model using a deep Q-network, and obtaining multi-dimensional diagnostic analysis results; combining the multi-dimensional diagnostic analysis results with a priority calculation formula to determine the priority of diagnostic conclusions; calculating a comprehensive risk score through data fusion scoring, classifying risk levels according to the comprehensive risk score, and matching corresponding phased intervention strategies; inputting the multi-dimensional diagnostic analysis results, intervention strategy execution feedback data, and user health preferences into a comprehensive diagnostic decision model, and combining the diagnostic conclusion priorities to obtain differentiated health association strategies. This solution accurately identifies disease severity through dynamic priority adjustment.

Description

Technical Field

[0001] This invention relates to the field of report generation technology, and in particular to a method for prioritizing the conclusions of a physical examination report based on the user's medical history. Background Technology

[0002] In the current healthcare field, disease risk assessment and diagnostic decisions often rely on single data sources or traditional experience-based judgments, resulting in incomplete data coverage and limited analytical dimensions. For example, relying solely on clinical diagnostic data is insufficient to fully correlate individual baseline characteristics and potential influencing factors such as health behaviors, leading to delays in identifying early risks of chronic diseases. Simultaneously, the collection of health behavior data (such as diet and exercise) often relies on questionnaires and uses simplistic methods for handling missing values ​​(such as direct mean filling), failing to accurately reflect individual differences and further impacting the accuracy of diagnostic analysis. Furthermore, existing systems lack the capacity to process unstructured data (such as imaging reports and medical record texts), making it difficult to efficiently extract key medical information. This results in the ineffective integration of multi-source data, hindering the comprehensiveness and accuracy of diagnostic decisions.

[0003] In terms of applying diagnostic conclusions and health management interventions, traditional models lack dynamic priority adjustment mechanisms and the ability to adapt to personalized strategies. On the one hand, diagnostic conclusions are often ranked according to fixed standards, without dynamic optimization based on disease severity, intervention timeliness, and user needs, which may result in high-priority health issues not being addressed first. On the other hand, intervention strategies often adopt a "one-size-fits-all" approach, failing to customize differentiated plans based on individual risk levels, disease correlations, and user health preferences. Furthermore, there is a lack of tracking and feedback on intervention effects and strategy iteration, making it difficult to form a closed loop of health management. This fails to meet the personalized health management needs of different populations and is also difficult to adapt to the dynamic changes in conditions in medical settings. Summary of the Invention

[0004] This invention improves treatment efficiency by dynamically adjusting priorities to accurately identify the severity of diseases.

[0005] The technical solution proposed in this invention is: a method for adjusting the priority of physical examination report conclusions based on user medical history, the method comprising:

[0006] Collect multi-source data and preprocess the multi-source data to obtain a structured multi-source dataset;

[0007] Based on structured multi-source datasets, the correlation between health data and diseases is explored through association analysis, preliminary screening is carried out using pattern recognition classification, and a health trend prediction model is built by combining deep Q-networks to obtain multi-dimensional diagnostic analysis results.

[0008] By combining the results of multi-dimensional diagnostic analysis, a hierarchical weighting model is used, along with a priority calculation formula, to determine the priority of diagnostic conclusions.

[0009] Input structured multi-source datasets and multi-dimensional diagnostic analysis results, calculate a comprehensive risk score through data fusion scoring, and classify risk levels according to the comprehensive risk score to match corresponding phased intervention strategies;

[0010] By inputting multi-dimensional diagnostic analysis results, intervention strategy execution feedback data, and user health preferences into a comprehensive diagnostic decision-making model, and combining the priority of diagnostic conclusions, personalized control coefficients adapted to different health scenarios are calculated to obtain differentiated health-related strategies.

[0011] Preferably, the specific process for obtaining the structured multi-source dataset is as follows:

[0012] Individual basic characteristic data, health behavior data, clinical diagnostic data, and genetic and medical data were collected separately. Records with high data completeness were prioritized for retention. Numerical data with a missing rate exceeding 30% were marked as missing. Categorical data with a large proportion were classified separately, and outliers were identified and processed. For unstructured data, key information was extracted using natural language processing technology and transformed into structured data. Non-missing numerical data were processed using standardization methods to make the data meet the input requirements of subsequent analysis models. Numerical questionnaire data was not standardized, and subsequent analysis adopted targeted strategies to generate structured multi-source datasets.

[0013] Preferably, the specific process for mining the correlation between the health data and the disease is as follows:

[0014] From structured multi-source datasets, data dimensions closely related to diseases are selected, and data with medical relevance are grouped into transaction units to form transaction sets. Using the Apriori algorithm, with minimum support and minimum confidence parameters set, the transaction sets are mined to identify association rules whose support and confidence meet threshold requirements, thus preliminarily identifying potential associations between health data and diseases. Disease diagnoses and examination results from electronic medical records, as well as family medical histories collected from questionnaires, are transformed into nodes and edges of a knowledge graph. The weights of the edges are adjusted based on the frequency of associations in clinical data and medical literature. The knowledge graph is trained using a graph convolutional network to update node feature representations, mining direct, indirect, and hidden associations between health data and diseases. The results of association rule mining are verified and supplemented to clarify the association between health data and diseases.

[0015] Preferably, the specific process for obtaining the multi-dimensional diagnostic analysis results is as follows:

[0016] This process involves selecting disease-related data dimensions from structured multi-source datasets and combining them into transaction units. The Apriori algorithm is used to mine association rules that meet thresholds, and a knowledge graph is constructed. Graph convolutional networks are then used to mine direct, indirect, and hidden associations, clarifying the relationship between health data and disease. Based on the structured multi-source datasets and the association results, support vector machines (SVMs) and random forest classification models are trained. The SVMs handle data noise and non-linear distributions by optimizing penalty parameters and kernel functions, while the random forest classification models avoid overfitting by randomly selecting features and setting tree parameters. Together, they output preliminary disease screening results. Core health indicators are selected from the structured data to construct state vectors. A health management-related action space and hierarchical reward function are defined, and a deep Q-network model is trained. The training process is optimized through experience replay and dynamic parameter tuning, outputting disease risk, progression trends, and suggestions for optimized health behaviors. Finally, the potential disease risks obtained from association mining, the preliminary diagnostic results from pattern recognition classification, and the health trends and behavioral suggestions output by the deep Q-network are summarized to form a multi-dimensional diagnostic analysis result.

[0017] Preferably, the specific process for obtaining the priority of the diagnostic conclusion is as follows:

[0018] Based on multi-dimensional diagnostic analysis results, the severity of the disease is assessed with reference to the disease severity and criticality grading standards. Intervention timeliness levels are classified in conjunction with clinical guidelines. The rate of disease progression is quantified by changes in core physiological parameters, forming basic assessment data. A team of medical experts manually sets and regularly updates the weights of common disease characteristics based on the latest treatment guidelines. Simultaneously, they collaborate with rare disease treatment institutions to optimize the rare disease weight calculation model using professional data, thus correcting the basic assessment weights. The front-end interface collects user preferences regarding disease types, treatment expectations, and recovery periods. These user choices are mapped to three weight adjustment factors: disease type, treatment expectation, and recovery period. After normalization, personalized adjustment vectors are generated. The underlying basic assessment results, mid-level optimized weights, and top-level adjustment factors are integrated, substituted into the priority calculation formula, and combined with the disease progression rate adjustment coefficient to obtain a diagnostic conclusion priority score, determining the priority ranking.

[0019] Preferably, the specific process for determining the staged intervention strategy is as follows:

[0020] This process involves collecting structured multi-source datasets and multi-dimensional diagnostic analysis results. Based on the input data, a risk baseline score and current score are calculated through weighted summation. These are then combined with the health indicator change trend coefficients output by a deep Q-network to obtain a comprehensive risk score. Risk thresholds are set to classify users corresponding to the comprehensive risk score into three risk levels: high, medium, and low. Starting with core symptoms, a knowledge graph-based graph analysis algorithm is used to mine direct, indirect, and hidden risk associations between symptoms, constructing a risk hierarchy. For high-risk users, chronic disease management pathways and strict health management plans are matched; for medium-risk users, personalized health management plans and regular monitoring plans are customized; and for low-risk users, basic monitoring plans and health prevention recommendations are developed, forming a phased intervention strategy. Based on the results of the intervention strategy implementation and changes in user health indicators, the calculation parameters of the comprehensive risk score and the risk association mining model are adjusted, and the phased intervention strategy is updated.

[0021] Preferably, the specific process for obtaining the differentiated health association strategy is as follows:

[0022] The process involves collecting high-precision diagnostic conclusions, system optimization data, and user tags to obtain integrated data. Using this integrated data as input, a collaborative strategy model is constructed, embedding health scenario classification logic within the model. For different health scenarios, the model determines core health goals based on the priority of diagnostic conclusions, adjusts strategy adaptability parameters with reference to system optimization data, matches personalized needs based on user tags, and calculates personalized control coefficients suitable for each scenario. Based on these personalized control coefficients, basic intervention strategies are adjusted, intervention measures are strengthened for high-priority health issues, and the feasibility of strategies is optimized based on user preferences, generating differentiated health management strategies suitable for different scenarios. These differentiated strategies are then applied to actual health management, collecting user feedback and data on changes in health indicators to feed back into the collaborative strategy model and update the personalized control coefficient calculation logic.

[0023] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the aforementioned method for prioritizing the conclusions of a physical examination report based on a user's medical history.

[0024] The beneficial effects of this invention are:

[0025] 1. A differentiated preprocessing strategy was designed for multi-source data including individual basic characteristics, health behaviors, clinical diagnoses, and genetic medical data. This includes filling missing health behavior data with random errors by combining statistical data from the same region / age group; using disease-specific thresholds to identify outliers in clinical diagnostic data; and converting unstructured text into structured information through named entity recognition. Simultaneously, standardization processing was performed on numerical data classification (Z-score standardization for non-missing data and separate adaptation analysis strategies for "unknown" labeled data). This mechanism effectively solves the problems of "mean filling ignoring individual differences," "lack of medical adaptability in outlier identification," and "difficulty in utilizing unstructured data" in traditional data processing. It ensures that structured multi-source datasets retain both data integrity and medical rationality while accurately reflecting individual health characteristics, providing high-quality data support for subsequent intelligent diagnostic analysis and significantly reducing the risk of diagnostic bias due to data quality issues. Furthermore, by prioritizing the retention of records with high completeness during data cleaning and separately classifying data with a large proportion of "unknown" entries, data utilization efficiency is further improved, avoiding information loss caused by data discarding or simple classification in traditional processing. This allows multi-source data to truly achieve "aggregation that is usable and effective." The system also standardizes numerical data classification (Z-score standardization for non-missing data and separate adaptation analysis strategies for "unknown" labeled data). This mechanism effectively solves the problems of "mean filling ignoring individual differences," "lack of medical applicability in outlier judgment," and "difficulty in utilizing unstructured data" in traditional data processing. It ensures that structured multi-source datasets retain both data integrity and medical rationality while accurately reflecting individual health characteristics, providing high-quality data support for subsequent intelligent diagnostic analysis and significantly reducing the risk of diagnostic bias due to data quality issues.

[0026] 2. The multi-dimensional diagnostic analysis system constructed in this solution uses the Apriori algorithm and graph convolutional networks to mine the direct, indirect and hidden associations between health data and diseases. It uses support vector machines (optimizing penalty parameters and kernel functions) and random forests (randomly selecting features + controlling tree parameters) to achieve accurate disease screening. Then, it combines deep Q-networks (dynamically weighted state vectors + hierarchical reward functions) to predict health trends and output optimization behavior suggestions. This system overcomes the bottlenecks of traditional diagnosis, which relies solely on a single algorithm and has limited analytical dimensions. Association analysis can identify potential disease risk factors in advance (such as the implicit association between "high-salt diet + family history of hypertension + overweight" and hypertension), avoiding the one-sidedness of traditional diagnosis that focuses solely on symptoms. Pattern recognition improves disease classification accuracy through dual-model collaboration, especially in scenarios with data noise and non-linear distributions (such as rare diseases with few samples and complex features), reducing the misclassification rate by more than 30% compared to traditional single-model approaches. Deep Q-networks adapt to different health scenarios through dynamically weighted state vectors and balance health improvement and behavioral sustainability through hierarchical reward functions, enabling long-term prediction of health trends and personalized behavioral guidance, addressing the problems of traditional prediction models that "emphasize results over process" and "lack practicality in recommendations." These three elements work together to form a complete diagnostic chain of "risk mining - disease screening - trend prediction - behavioral recommendations," ensuring that diagnostic results are both medically sound and practical, providing a comprehensive and accurate analytical foundation for subsequent priority adjustments and intervention strategy generation.

[0027] 3. The diagnostic conclusion priority adjustment mechanism of this solution is based on a fundamental assessment at the bottom layer, which is based on the severity of the disease, the timeliness of intervention, and the quantitative changes in core physiological parameters. The middle layer incorporates the professional knowledge of medical expert teams (Delphi method) and authoritative rare disease institutions to optimize weights. The top layer generates personalized adjustment factors through user needs mapping, and finally calculates priorities using a multi-dimensional formula. This mechanism overcomes the shortcomings of traditional diagnostic conclusions that are "fixedly ordered and ignore individual differences": the bottom layer is data-driven, combined with health trend data predicted by deep Q-networks, ensuring the objectivity of the assessment and avoiding the subjective bias of traditional experience-based judgments; the middle layer of expert knowledge is synchronized with data from authoritative institutions through quarterly seminars, adapting in real time to updates to treatment guidelines (such as adjustments to the weight of glycated hemoglobin in diabetes) and rare disease diagnostic standards, solving the problems of traditional weights being "difficult to update once set" and "poor adaptability to rare diseases"; the top layer of user needs adaptation uses a visual interface and mapping matrix to transform the types of diseases and treatment expectations that users are concerned about into quantifiable adjustment factors, making the priorities more aligned with users' actual needs (e.g., for users concerned about occupational diseases, the weight of occupational exposure-related indicators is automatically increased). The dynamic integration of these three factors can accurately identify high-priority health issues, avoiding situations where "urgent issues are delayed in being addressed" or "non-concern issues are over-prioritized." For example, the priority of intermediate-risk but rapidly progressing diseases can be increased by adjusting the coefficient of change in core physiological parameters, ensuring that medical resources are tilted towards key health issues. At the same time, it enhances users' understanding and trust in the diagnostic process and improves the subsequent adherence to intervention strategies. Attached Figure Description

[0028] Figure 1 A flowchart of a method for prioritizing the conclusions of a physical examination report based on the user's medical history;

[0029] Figure 2 This is a flowchart illustrating the adjustment process of a method for prioritizing the conclusions of a physical examination report based on the user's medical history. Detailed Implementation

[0030] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0031] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0032] like Figure 1 and Figure 2 As shown, this solution first collects multi-source data on individual basic characteristics, health behaviors, clinical diagnoses, and genetic medical data. After differential preprocessing, a structured dataset is generated. Then, multi-dimensional diagnostic analysis is carried out through association mining, pattern recognition, and deep Q-network (DQN) trend prediction. After outputting the diagnostic results, the priority of the diagnostic conclusions is dynamically adjusted by combining "bottom-level data assessment (disease severity, etc.), mid-level expert knowledge (Delphi method to optimize weights), and top-level user needs (mapping adjustment factors)". Ultimately, it supports the generation of multi-dimensional risk warning and tiered intervention strategies, realizing intelligent medical diagnosis and health management.

[0033] 1. Collection of basic individual characteristics data

[0034] Information such as gender, age, height, weight, and occupation is collected; this data forms the basis of disease risk assessment. The data sources and processing methods are as follows:

[0035] Stable data acquisition: Information such as age, gender, height, and weight is guaranteed to be 100% obtained through integration with medical record systems (such as electronic medical records) and real-name authentication systems. During data integration, age is recorded in full years; gender uses binary encoding, with males recorded as 1 and females as 0; height is uniformly measured in meters (m), and weight is uniformly measured in kilograms (kg), using formulas... Calculate Body Mass Index (BMI) to help assess obesity levels and the risk of related diseases.

[0036] Questionnaire-dependent data: Data such as occupation, if not collected in the questionnaire, is directly marked as "unknown". This situation has been considered during model training and has little impact on disease risk assessment. During data processing, "unknown" is treated as an independent category in subsequent analysis. Furthermore, these data have specific associations with diseases; for example, age and gender affect the incidence of cardiovascular disease and cancer, while occupation is associated with the risk of specific diseases, such as teachers being more prone to vocal cord diseases.

[0037] 2. Health behavior data collection

[0038] Dietary Habits: Dietary patterns, such as taste preferences (light, high-salt, high-sugar, high-fat), frequency of fruit and vegetable consumption, and special behaviors such as dieting and binge eating, were collected through multiple-choice questions. If dietary preference data is unavailable, the average dietary pattern of the population will be used by default when analyzing the risk of diet-related diseases. For the frequency of fruit and vegetable consumption, if data is missing, it can be obtained from statistical data of the same region and age group using a formula. Fill in, where For fill value, This represents the average consumption frequency among people of the same age group in the region. This is a random error term (with values ​​ranging from [−0.5, 0.5]), used to simulate individual differences.

[0039] Exercise-Sleep Assessment: Data on exercise type, days per week, duration, intensity, sleep onset time, number of awakenings during the night, total sleep time, and sleep problems are collected through questionnaires or wearable devices. If data on exercise duration and intensity is unavailable, estimates can be made using available data such as step count. For example, for moderate-intensity exercise, assuming an average walking speed of [missing data]. (m / min), stride length is (m / step), total number of steps is Then estimate the duration of the exercise. Exercise intensity index ( According to heart rate during exercise Calculate, when ,but ;when It was determined to be high-intensity exercise. Other situations =0. Sleep quality index ( ) through formula calculate( Total sleep time (h) (Number of times you wake up at night). If sleep data is missing, the weight of sleep-related indicators in the comprehensive assessment will be reduced to 30% of the original weight.

[0040] Smoking, alcohol, and medication history: Statistics on the quantity, frequency, and duration of smoking and alcohol consumption; recording information on currently used and previously used medications. Smoking index ( ) through formula calculate( This refers to the number of cigarettes smoked per day. (Smoking age); alcohol consumption converted to standard drinking units ( ) through formula calculate( The amount of alcohol consumed per instance ( ), (This refers to alcohol concentration (%)). If data on tobacco and alcohol consumption is unavailable, refer to the average level of people in the same region and occupation, and use the formula... Make an estimate ( This is an estimated value. This represents the average level of people in the same profession in the region. The adjustment coefficient has a range of [−0.2, 0.2]).

[0041] 3. Clinical diagnostic data collection

[0042] Symptoms and signs: Detailed data on symptoms such as pain, fever, and cough, as well as vital signs such as blood pressure and heart rate obtained during physical examinations, are collected to provide direct evidence for disease diagnosis. In the data cleaning stage, box plots are used to identify outliers. Taking systolic blood pressure data as an example, a dataset composed of the systolic blood pressure values ​​of all individuals in a group is used as an indicator. The specific calculation process is as follows:

[0043] Data sorting: sorting the dataset The systolic blood pressure values ​​are arranged in ascending order.

[0044] Calculate the lower quartiles If the number of samples in the dataset It is an odd number. For sorted positions The numerical value of the position; if Even number, For sorted positions and The average of the positional values. For example, given 11 systolic blood pressure data points, the average of the positional values ​​is calculated by sorting the data. The data is If there are 12 data points, This is the average of the 3rd and 4th data points after sorting.

[0045] Calculate the upper quartiles Similarly, if It is an odd number. For sorted positions The numerical value of the position; if Even number, For sorted positions and The average value of the location.

[0046] Calculate interquartile range : Through formula The interquartile range is obtained.

[0047] Identify outliers: Values ​​exceeding... or below Data that is missing is considered outlier. Data marked as "unknown" or "unanswered" is not included in the outlier assessment.

[0048] Imaging analysis: Integrating laboratory test results such as complete blood count and urinalysis, as well as imaging data such as X-rays and CT scans, is used to diagnose diseases and assess conditions. When processing laboratory test results, normal reference ranges are set for each indicator; data exceeding these ranges are specially marked and further verified. For unstructured imaging report text, Named Entity Recognition (NER) technology is used to extract key information, such as lesion location, size, and morphology. The specific workflow is as follows:

[0049] Text preprocessing: Denoising, word segmentation, and other operations are performed on the imaging report to remove redundant information.

[0050] Entity recognition: Using a trained NER model to identify medical entities in text, such as "lung", "nodule", "shadow", etc.

[0051] Information extraction and structuring: Identified entities and their related attributes (such as size "diameter 2cm") are extracted and transformed into structured data. If text is missing, the location and type of the missing information are recorded, and the diagnostic report indicates "Related text data is missing, which may affect some diagnostic results." Simultaneously, for imaging data, image recognition algorithms are used to extract features to assist doctors in diagnosis.

[0052] 4. Genetic medical data collection

[0053] The questionnaire gathers information on family history of genetic diseases, and genetic testing data is obtained for eligible users. It also connects with hospital systems to obtain outpatient and inpatient medical records to aid in disease diagnosis and treatment. If the questionnaire does not collect family history data, attempts are made to supplement it from the "Family History" field in medical records or genetic testing data. If completely missing, the model defaults to "no clear family history," which only affects the assessment of rare diseases or diseases with high genetic relevance. In risk assessment, the Family Genetic Risk Index (FGRI) is defined as follows: if there is a clear family history of genetic diseases, FGRI=1; if not, FGRI=0; if data is missing, FGRI is set to 0 in the initial assessment but marked as pending confirmation in subsequent analyses, and updated immediately once relevant data is obtained.

[0054] 5. Data integration and processing

[0055] After data collection, techniques such as data cleaning, natural language processing, and standardization were used to integrate multi-source heterogeneous data into a structured dataset.

[0056] Data cleaning: For records containing missing questionnaire data, prioritize retaining records with high completeness. When the missing rate of numerical data exceeds 30%, mark it as missing without imputation and process it using an adaptation algorithm; when the proportion of "unknown" or "unanswered" data in categorical data is large, analyze it separately by category.

[0057] Natural Language Processing: For unstructured medical records, imaging reports, etc., key information is extracted through techniques such as Named Entity Recognition (NER) and semantic analysis, and then transformed into structured data.

[0058] Data standardization: Numerical questionnaire data marked as "unknown" or "unanswered" will not be standardized; targeted strategies will be used in subsequent analysis. For non-missing numerical data, Z-score standardization will be used, with the formula: ( This is the original data. The mean, The standard deviation is set to 0, and the mean and standard deviation are set to 1, which makes the data easier to process by the model.

[0059] After completing data collection and integration, transforming multi-source heterogeneous data into structured datasets, the next step is intelligent diagnostic analysis. This stage utilizes a professional diagnostic engine and various algorithm models to perform in-depth analysis of the integrated data, achieving accurate diagnosis and prediction through standardized and quantifiable steps. Considering the multi-source nature and complexity of the data in this solution, as well as the specific characteristics of the medical scenario, each processing method has been adaptively modified. The specific process is as follows:

[0060] 1. Association Analysis

[0061] By leveraging association rule mining and graph analysis techniques, we can explore the potential connections between individual characteristics, health behaviors, genetic data, and diseases, and uncover the logical relationships between symptoms, signs, examination results, and disease diagnoses.

[0062] Association rule mining: Employing the Apriori algorithm, and considering the multi-dimensional nature of the data in this approach, including individual basic characteristics and health behaviors, transactional processing is performed on the data before mining, grouping data with close medical correlations into transactions. For example, dietary preferences, exercise habits, and indicators related to metabolic diseases are combined to improve mining efficiency and accuracy.

[0063] This study explores association relationships between data by setting minimum support (MinSupport) and minimum confidence (MinConfidence). It involves collecting extensive medical data covering different disease types and patient groups, including multi-source data such as individual basic characteristics, health behaviors, clinical diagnoses, and genetic medicine. Association rule mining experiments are conducted on this data, trying different combinations of MinSupport and MinConfidence values. For example, starting with MinSupport=0.01 and MinConfidence=0.5, MinSupport is gradually increased in increments of 0.01, and MinConfidence in increments of 0.05. After each combination, the number of association rules obtained, the range of disease types covered by the rules, and the coverage of the rules in the data are analyzed. Combinations that yield a certain number of rules (e.g., no less than 50) and are relatively evenly distributed across different diseases and data dimensions are retained to initially screen suitable parameter ranges. Taking the association between dietary habits and hypertension as an example, the support calculation formula is:

[0064] ;

[0065] The confidence level is calculated using the following formula:

[0066] ;

[0067] in, Indicates the preceding item (e.g., a high-salt diet). Indicates the following item (e.g., hypertension). Representative affairs, It is a set of transactions. This represents the number of elements in the set. A rule is considered statistically significant when its support is greater than or equal to MinSupport and its confidence is greater than or equal to MinConfidence.

[0068] Graph analysis technology: When constructing a knowledge graph, this solution combines data sources to transform information such as disease diagnoses and examination results from electronic medical records, and family medical history collected from questionnaires, into nodes and edges. For example, the "hypertension" node is connected to symptom nodes such as "headache" and "dizziness." The weight of the edge is calculated based on the frequency of both occurring simultaneously in clinical data, and is also adjusted by referring to the probability of related symptoms occurring in medical literature, thereby enhancing the medical reliability of the graph.

[0069] Graph Convolutional Networks (GCNs) are used to learn from knowledge graphs and update the feature representations of nodes, thereby uncovering complex relationships. The node feature update formula is:

[0070] ;

[0071] in, and They are the first Layer and first The node feature matrix of the layer, It is the adjacency matrix after adding self-loops. yes The degree matrix, It is a learnable weight matrix. It is the activation function. When training the GCN model, a pre-trained model from the medical field is used as the initialization parameter to accelerate the model convergence speed, and transfer learning is used to adapt to the data characteristics of this scheme.

[0072] 2. Pattern Recognition and Classification

[0073] Based on a large number of disease data samples, classification models such as support vector machine (SVM) and random forest are trained to automatically identify user data feature patterns and achieve preliminary disease screening and classification diagnosis.

[0074] Support Vector Machine (SVM): Addressing the issues of noise, outliers, and non-linear distributions in the data presented in this scheme, SVM minimizes the objective function when processing linearly separable data. Solve and The penalty parameter Based on the degree of dispersion of the data distribution, cross-validation is used to optimize the selection, so as to balance the model's tolerance for outliers and classification accuracy.

[0075] For nonlinear data, radial basis functions are used. Mapping data to a high-dimensional space and finding the optimal hyperplane in that space. Kernel function parameters. Similarly, cross-validation was used to ensure that the model has good generalization ability on the data of this scheme.

[0076] Random Forest: Considering the high dimensionality of the data features in this scheme, a subset of features is randomly selected from all features when constructing the decision tree (the size of the feature subset is usually [missing information]). , The total number of features is used, and based on the characteristics of medical data, features strongly correlated with disease diagnosis, such as blood glucose and blood pressure, are prioritized when selecting the optimal splitting features. Simultaneously, to avoid overfitting, parameters such as the maximum tree depth and minimum number of sample splits are set, and optimization is performed using grid search and cross-validation. The final classification result is determined by majority voting, i.e., the category that appears most frequently among the predictions of multiple decision trees is used as the final classification result.

[0077] 3. Forecasting Trend Analysis

[0078] By using a Deep Q-Network (DQN) to track user health data over a long period, the system predicts disease risk and progression trends, and outputs behavioral suggestions to optimize health status. The specific implementation process is as follows:

[0079] Data and Definition of State / Action / Reward

[0080] State design: From the structured data integrated with the solution, select core data that comprehensively reflects the health status to construct a state vector. Assume there are a total of With feature dimensions, the state vector can be represented as: The data covers basic physiological indicators (such as body temperature and basal metabolic rate), health behavior data (intensity of daily activities and dietary balance score), individual characteristics (age and body type), health risk factors (history of allergies and occupational exposure), and recent health events (such as whether there has been an infection or during surgical recovery). All data has been standardized, and numerical characteristics are... The standardized formula is:

[0081] ;

[0082] in, and Features The mean and standard deviation of the values. Feature selection algorithms (such as chi-square test, mutual information method) are used to screen key features, assuming that the selected features will retain... These features form the final state vector. To further enhance state representation capabilities, a dynamic weighting mechanism is introduced to automatically adjust the weights of each feature in the state vector based on different health scenarios. Let the weight vector be... The dynamically weighted state vector is The weight Based on scene characteristics Through function The calculation yielded the result.

[0083] Action design: The action space is constructed around interveneable health management behaviors, following general health promotion principles. Exercise behaviors are categorized into low-intensity, moderate-intensity, and high-intensity activities; dietary behaviors are classified as nutritionally unbalanced, balanced, and targeted nutritional supplementation; lifestyle habits include irregular sleep patterns, regular sleep patterns, and strengthening healthy habits (such as increasing meditation and social activities); medical interventions include routine monitoring, preventative checkups, and early intervention measures. These various action dimensions are combined to form a rich and practical action set. Assume that the motor behavior has Types of eating behaviors Types, lifestyle habits There are various types of medical interventions. If there are a certain number of such actions, then the size of the action set is... To improve the interpretability and applicability of the movements, a mapping relationship between the movements and clinical guidelines should be established to ensure that each movement has a clear medical basis and health management goals.

[0084] Reward Design: The reward function adopts a hierarchical design, including rewards for improved health. Risk reduction reward Sustainable behavioral rewards ,Right now The health improvement reward is calculated based on changes in the comprehensive health score, assuming the comprehensive health score function is... ,but ,in and These are the state vectors for the current time step and the next time step, respectively. The risk reduction reward is determined based on the change in risk level in the disease risk prediction model, assuming the risk level is 1. ,but (Positive when risk decreases, negative otherwise). Sustainable behavior rewards provide feedback by assessing the fit between the action and the user's historical behavior patterns, as well as the difficulty of maintaining the behavior. Let the fit function be... The difficulty function is ,but ,in and This is the adjustment coefficient. To enhance the guiding effect of reward signals, a credit allocation mechanism is introduced to rationally allocate long-term health benefits to short-term action decisions. Assume the long-term benefit is... , assigned to The reward for each moment is ,but ,in The attenuation coefficient is... This represents the total number of time steps.

[0085] DQN model training

[0086] Data preprocessing: The training and validation sets are cleaned, and density-based outlier detection algorithms are used to remove noisy data. For missing data, multiple imputation and model-based prediction imputation methods are flexibly used depending on the data type and correlation. To improve training efficiency, the data is normalized and discretized, mapping numerical data to appropriate intervals and encoding categorical data. Data augmentation techniques (such as shifting and scaling time-series data) are used to expand the dataset. To address the data imbalance problem, a combination of oversampling (SMOTE algorithm) and undersampling is used to adjust the proportion of data from different health states or risk levels, making the model's learning of various health problems more balanced.

[0087] Model structure optimization: A multi-head attention mechanism is introduced into the basic network architecture, enabling the model to capture the correlations between different health features from multiple perspectives. Assume the input state vector is... After passing through the multi-head attention mechanism, the output is The calculation process is as follows:

[0088] ;

[0089] in, , , , and The weight matrix is ​​a learnable matrix. The number of heads is a factor. Dedicated feature processing modules are designed for different types of health data (such as time-series data and static data) to perform feature extraction and fusion respectively. Simultaneously, a batch normalization layer is added to the network for the input... The batch normalization formula is:

[0090] ;

[0091] in, and Inputs The mean and variance, To prevent constants with zero denominators, hyperparameters such as the number of network layers, neurons, and activation functions are tuned using hyperparameter optimization algorithms (such as Bayesian optimization and genetic algorithms). To further improve the model's generalization ability, knowledge distillation is employed to transfer knowledge learned from complex, large networks to lightweight networks, facilitating deployment on resource-constrained medical devices.

[0092] The training process is refined: the experience replay pool employs a priority experience replay combined with a hierarchical storage strategy. It not only sorts samples based on TD error but also stratifies them according to health scenarios and data types, ensuring that samples from various health issues can be effectively learned. The TD error calculation formula is:

[0093] ;

[0094] in, As a reward, As a discount factor, For the target network output, This is the current network output. During training, the learning rate is dynamically adjusted using an adaptive learning rate algorithm (such as AdamW). The learning rate update formula is:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] in, For gradient, and For momentum parameters, For learning rate, The learning rate is constant. Combined with a periodic learning rate adjustment strategy, the model learns quickly in the early stages of training and is finely adjusted later. The exploration rate uses an entropy-based dynamic adjustment method, dynamically adjusting the exploration level according to the uncertainty of the model's output. Furthermore, a curriculum learning strategy is introduced, allowing the model to learn from simple health scenarios and data patterns first, gradually transitioning to complex scenarios, reducing learning difficulty and improving training stability and efficiency.

[0101] Prediction and Output: The DQN model processes health data using a sliding window approach. Each input is the current state vector, and the output is the Q-value of each action, selecting the optimal action. Based on this action, multiple future health development paths are simulated using methods such as Monte Carlo tree search to predict the probability distribution of health status changes over different periods (short-term, medium-term, and long-term). To improve the reliability of the prediction results, an ensemble learning method is used to fuse the prediction results of multiple DQN models with different initialization parameters or structures. Let there be a total... The model's prediction results are fused using the following formula:

[0102] ;

[0103] in, For the first The prediction results of each model The results are the fused predictions. The predictions and recommended actions are presented in the form of visualized health trend charts, risk assessment radar charts, and detailed strategy reports. The reports include health benefit analysis of the actions, assessment of implementation difficulty, and potential risk warnings. This information enters the comprehensive diagnostic decision-making module to help doctors develop personalized health management plans. Simultaneously, the system triggers different levels of health alerts based on the predicted risk level, reminding users and medical personnel to take timely measures. To enhance user interactivity, an interactive interface is provided, allowing users to input their own health management preferences (such as exercise time limits, dietary restrictions, etc.), and the system dynamically adjusts the recommended actions accordingly.

[0104] 4. Comprehensive diagnostic decision-making

[0105] Based on the above analysis results, and with reference to medical knowledge bases, clinical guidelines, and expert experience, a comprehensive diagnostic decision-making model is constructed to derive diagnostic conclusions such as disease name, stage, and severity, and to provide evidence and suggestions.

[0106] Diagnostic Conclusion Generation: The system takes as input information the potential disease risk obtained from association analysis, preliminary diagnostic results from pattern recognition classification, and health trend predictions and behavioral suggestions output by the deep Q-network. Based on disease diagnostic criteria and clinical guidelines in the medical knowledge base, a rule engine performs matching. The rule engine integrates the latest medical research findings and clinical practice experience, setting differentiated diagnostic rule weights for different disease types to comprehensively assess the disease status and determine its severity.

[0107] Diagnostic Basis and Suggested Output: The system automatically generates a detailed diagnostic report, listing various data indicators supporting the diagnostic conclusions, algorithm analysis results, and corresponding medical evidence. Simultaneously, based on the disease diagnosis results and the patient's individual condition, it extracts relevant treatment plans and health management suggestions from the medical knowledge base, and deeply integrates behavioral suggestions from a deep Q-network for personalized adjustments, providing comprehensive and practical references for patients and doctors.

[0108] Dynamic and flexible conclusion priority adjustment mechanism

[0109] This mechanism is built on a hierarchical weight model, deeply integrating actual clinical needs with intelligent diagnostic analysis results to achieve dynamic optimization of diagnostic conclusion priorities, ensuring that decisions are more aligned with individual patient conditions and medical scenario needs.

[0110] Underlying general feature library: Data-driven basic assessment

[0111] A foundational assessment system is constructed based on health trend prediction data output by deep Q-networks, risk factors mined through association analysis, and pattern recognition classification results. Using the severity and critical illness grading standards in the "Detailed Rules for the Implementation of the Regulations on the Administration of Medical Institutions" as a reference, and combining the trends in health indicators predicted by DQN (such as the continuous rate of decline in the comprehensive health score), the severity of the disease is assessed. Disease progression is quantified by calculating the magnitude and speed of change in key health indicators (such as core physiological parameters in the state vector). Complication risks are determined based on the multi-system disease associations identified through association analysis. The timeliness of intervention measures is classified according to the intervention time windows specified in clinical guidelines (such as the golden rescue time for acute illnesses and the regular monitoring cycle for chronic diseases).

[0112] In the specific calculation process, the disease severity score Using a weighted summation method:

[0113] ;

[0114] in, The weights for each assessment indicator (such as the rate of decline in the overall health score, the number of complications, etc.) This represents the normalized value of the corresponding indicator. This is to assess the total number of indicators. For example, if the normalized value of the rate of decline in the patient's overall health score is 0.8, with a weight of 0.4; and the normalized value of the number of complications is 0.6, with a weight of 0.3, then... =0.4×0.8+0.3×0.6=0.5, and the severity level is determined based on the preset threshold. When A value ≥0.7 is considered high-risk; a value ≤0.4 is considered low-risk. 0.7 indicates a medium risk level; 0.4 indicates a low risk level.

[0115] The timeliness level of intervention is determined by the urgency of the time window. Division:

[0116] ;

[0117] in, The intervention cutoff time specified in clinical guidelines. This is the current time. At 0.3, it is considered an emergency (requiring immediate intervention); 0.3 ≤ 0.7 indicates a secondary emergency (requiring immediate action). ≥0.7 is considered normal (can be arranged according to plan).

[0118] In addition, the rate of disease progression is determined by calculating the change in core physiological parameters per unit time. To measure:

[0119] ;

[0120] in, express Key physiological parameter values ​​at any given time express Key physiological parameter values ​​at any given time For time intervals. If Exceeding the threshold for parameter changes in the same healthy population (This threshold was derived from statistical analysis of a large amount of health data from normal populations. Different core physiological parameters correspond to different thresholds, such as body temperature.) for / day, blood pressure for If the disease progresses rapidly (e.g., every week), it is marked as a rapid progression state, and the priority of the diagnostic conclusion is increased accordingly.

[0121] Mid-level disease specificity: Deep integration of professional knowledge

[0122] A team of authoritative medical experts from various fields has been assembled to manually set and regularly update the weighting parameters of disease-related characteristics for common diseases such as diabetes and cardiovascular diseases, based on the latest treatment guidelines (such as the "Guidelines for the Prevention and Treatment of Type 2 Diabetes in China" and the "Guidelines for the Prevention of Cardiovascular Diseases in China"). For rare diseases, a data and knowledge sharing mechanism has been established with authoritative rare disease diagnosis and treatment institutions at home and abroad (such as the National Rare Disease Medical Center) to optimize the weighting calculation model for rare diseases using their professional databases and research results.

[0123] In terms of specific operational procedures, the expert team holds quarterly seminars to adjust the weights of common disease features based on the latest clinical research findings and actual treatment cases. Taking diabetes as an example, if new research indicates that short-term changes in glycated hemoglobin are more critical for disease assessment, the weight of this indicator in the diagnostic conclusion is increased. The adjustment process uses the Delphi method, reaching expert consensus through multiple rounds of anonymous voting and feedback. For rare diseases, the system automatically synchronizes the latest research data and diagnostic criteria released by authoritative institutions, triggering updates to the weight calculation model. The model is constructed using a Bayesian network, dynamically updating the conditional probabilities of each feature based on new evidence. For example, when a new association is found between a rare disease and a specific gene mutation, the weights of gene detection indicators are adjusted accordingly, making the diagnostic conclusions more consistent with disease characteristics and cutting-edge medical developments.

[0124] Top-level user needs adaptation: precise response to personalized preferences

[0125] At the user interaction level, in addition to implicitly mining needs through users' historical health data and behavioral patterns, a visual front-end selection interface is designed. Users can directly select their preferred disease type (such as hereditary diseases or occupational diseases), treatment expectations (such as prioritizing symptom control or pursuing a radical cure), and recovery cycle preferences (short-term rapid improvement or long-term stable recovery), among other personalized needs. The system maps user selections to weight adjustment factors and integrates them with the underlying data evaluation results and the mid-level disease-specific weights for calculation.

[0126] Weight adjustment factor acquisition process

[0127] Disease type weight adjustment factor The system has a built-in dimension. Disease type-indicator correlation matrix ,in Set the preset number of disease types (e.g., covering 50 common disease types). This represents the total number of health indicators included in the assessment (e.g., including 100 basic physiological, examination, and test indicators). Matrix elements. Indicates the first The first disease type The basic weights of each health indicator are pre-set by medical experts based on clinical experience and disease treatment guidelines. When a user selects the type of disease they are interested in on the front-end interface (let's say it's the [specific disease]), these weights are then used to define the underlying health indicators. After identifying the disease type, the system extracts a matrix. The row vector Then, using the normalization formula: Calculations yielded The elements of the vector, where This indicates that for this type of disease, the first Adjustment factors for each health indicator. For example, if a user selects occupational disease, the system extracts the corresponding row vector, increases the weights of lung function indicators and heavy metal detection indicators related to occupational exposure, and decreases the weights of other irrelevant indicators. After normalization, the final result is formed. The vector is used for subsequent calculations.

[0128] Treatment expectation weight adjustment factor Based on clinical guidelines and expert experience, the system has developed separate treatment plans for different treatment expectations, such as "prioritizing symptom control" and "pursuing a radical cure." Set (assumption) Weight allocation templates, each template is a dimension. The vectors represent the weight allocations of various treatment-related indicators under different treatment expectations. When a user selects a specific treatment expectation (let's say it's the first one), the vectors represent the weight allocations of various treatment-related indicators under different treatment expectations. After obtaining the expected treatment, the system obtains the corresponding weight allocation template vector. The same normalization process is also performed: ,get The elements of the vector, where This indicates that under the expected treatment, the first Adjustment factors for treatment-related indicators. For example, when a user chooses to pursue a radical cure, the system obtains the corresponding template, increases the weight of indicators related to etiological diagnosis such as gene testing and pathological analysis, and decreases the weight of indicators such as the frequency of pain medication use that only relieves symptoms. After normalization, the result is... .

[0129] Rehabilitation cycle preference weight adjustment factor The system employs two sets of weight adjustment rules to address two rehabilitation cycle preferences: "short-term rapid improvement" and "long-term stable recovery." For the "short-term rapid improvement" preference, the rules tend to increase the weight of indicators related to fast-acting treatments such as medication and physical therapy. For the "long-term stable recovery" preference, the rules emphasize increasing the weight of indicators related to measures requiring long-term adherence, such as rehabilitation training and lifestyle interventions. Taking "short-term rapid improvement" as an example, the system provides a single weight adjustment vector. After a user selects this preference, the system normalizes the vector: ,get The elements of the vector, where This indicates a preference for "rapid short-term improvement". Adjustment factors for each rehabilitation-related indicator. If the user selects "long-term stable recovery," then a different set of weighted adjustment vectors will be used. Perform a normalization operation to generate the corresponding .

[0130] Calculation of the priority of the final diagnosis conclusion

[0131] set up The adjustment factor is ( The settings are dynamically configured by the system based on the type and severity of the disease. For example, for high-risk and rapidly progressing diseases, The value ranges from 0.5 to 1; for low-risk diseases, If the value ranges from 0.1 to 0.3, then the priority score for the final diagnostic conclusion is... The calculation formula is:

[0132] ;

[0133] In this formula, when The larger the size, the faster the disease progresses. The larger the value, the more significantly it will improve. For example, if a user chooses to focus on hereditary diseases and seeks a radical cure, with a preference for long-term stable recovery, and their core physiological parameters change beyond a threshold, the system will comprehensively calculate based on the aforementioned formula. This calculation considers not only the weighting adjustments of disease type, treatment expectations, and recovery period preferences, but also amplifies the impact of rapid disease progression, resulting in an accurate diagnostic priority and providing strong support for subsequent treatment decisions. Simultaneously, the system offers a needs explanation function. When a user clicks on an adjustment factor, they can view a detailed explanation of the impact of that selection on the diagnostic conclusion, including changes in relevant health indicators and the direction of adjustments to recommended measures, enhancing the user's understanding and trust in the diagnostic process. Furthermore, based on users' historical selection data, the system will utilize collaborative filtering algorithms to provide personalized needs recommendations for new users, helping them express their needs more efficiently.

[0134] Multi-dimensional risk warning and tiered intervention strategy generation mechanism

[0135] Based on a dynamic and flexible conclusion priority adjustment mechanism, a multi-dimensional risk early warning and tiered intervention strategy generation mechanism is constructed to further strengthen the precise control and effective intervention of individual health risks.

[0136] (a) Data Input

[0137] This system integrates basic individual data (age, gender, body type, etc.), health behavior data (daily exercise, diet, sleep patterns, etc.), clinical diagnostic data (symptoms, signs, examination and test results, disease diagnosis conclusions, etc.), and genetic medical data (family medical history, gene testing results, etc.). Through data cleaning, standardization, and other preprocessing operations, it unifies data formats and measurement standards to ensure data accuracy and consistency, providing solid data support for subsequent risk assessment. Simultaneously, a data update mechanism is established to access the latest health data in real time or periodically, ensuring the timeliness of the assessment basis.

[0138] (II) Processing Procedure

[0139] Data fusion scoring

[0140] Risk baseline score: A risk baseline assessment model is constructed based on individual baseline data and genetic medical data, referencing medical standards and population statistics. For example, for age factors, corresponding risk weights are assigned based on disease incidence statistics for different age groups; for family history, if there is a family history of a specific genetic disease, the baseline risk value for that disease is increased. The risk baseline score is calculated using a weighted summation method. :

[0141] ;

[0142] in, The weights of each basic and genetic factor, These are the quantified values ​​for the corresponding factors. To assess the total number of factors.

[0143] Current Score: Based on health behavior data and clinical diagnostic data, combined with real-time health status indicators, the current health risk is assessed. For health behavior data, such as insufficient exercise or an unhealthy diet, points are deducted according to preset rules. For clinical diagnostic data, corresponding risk scores are assigned based on disease severity and symptom presentation. The current score is obtained by summing the scores of all indicators. :

[0144] ;

[0145] in, The weights of each health behavior and clinical indicator, The score is for the corresponding indicator. This represents the total number of indicators.

[0146] Trend coefficient: Calculates the trend coefficient of health indicators by using long-term tracking and prediction results of historical health data through a deep Q-network. The specific acquisition process is as follows:

[0147] Data filtering and preprocessing: Time series data of target health indicators (such as blood pressure, blood sugar, etc.) are filtered from historical health data. Missing values ​​are imputed and outliers are handled. For missing values, linear interpolation or mean-based imputation based on similar population data is used; for outliers, the 3σ principle is used to identify and correct them. Then, the data is normalized to map the data to the [0,1] interval, using the formula: ,in This is the original data. This is the normalized data.

[0148] Deep Q-Network Training and Prediction: Preprocessed time-series data is divided into training and test sets, and a supervised learning approach is used to train the deep Q-network. During training, time steps are used to predict the network. state vector As input, the network outputs the predicted time step. state vector The objective function for network training is: ,in For the sample size, and The first and second halves of the true state vector and the predicted state vector are respectively the first and second halves of the true state vector and the predicted state vector. Each element is used. The network parameters are optimized using the stochastic gradient descent algorithm. The model is considered trained when the loss function converges. The trained model is then used to predict health indicators for a future period, resulting in a sequence of predicted values. .

[0149] Trend coefficient calculation: The predicted value sequence is fitted using linear regression to obtain the fitted straight line equation. ,in For time steps, For predicted values, The slope Intercept. Trend coefficient. The value is the slope. ,like This indicates that health indicators are showing a deteriorating trend, and The higher the value, the faster the deterioration; if This indicates that health indicators are showing an improving trend. The higher the value, the faster the improvement; if This indicates that health indicators are stabilizing.

[0150] Comprehensive Risk Score: The comprehensive risk score is obtained by weighting and integrating the baseline risk score, current score, and trend coefficient. :

[0151] ;

[0152] in, , , The weighting coefficients are determined through extensive training with historical data and expert experience to balance the impact of each scoring item on the overall risk.

[0153] Risk correlation mining

[0154] Starting with core diseases, and based on a constructed knowledge graph, graph analysis algorithms (such as breadth-first search and depth-first search) are used to uncover direct, indirect, and hidden risk associations between diseases. Direct associations refer to explicit connections between diagnosed diseases and other related diseases; indirect associations are established through intermediate diseases or risk factors; and hidden risk associations are discovered by mining potential gene-disease, environment-disease, and other relationships. During the mining process, each association edge is assigned a weight based on its frequency, strength, and the support of medical evidence, constructing a risk hierarchy. Simultaneously, the latest medical research findings and clinical case data are regularly synchronized to update the knowledge graph, ensuring the accuracy and timeliness of risk associations.

[0155] Layered strategy generation

[0156] Based on a comprehensive risk score, users are categorized into three risk levels: high, medium, and low. A risk threshold is set: when... High risk; Medium risk; Low risk. Appropriate intervention strategies are matched to users with different risk levels:

[0157] High-risk users: Match detailed chronic disease management pathways, including regular specialist check-ups, personalized drug treatment plans, and multidisciplinary joint diagnosis and treatment plans; at the same time, develop strict health management plans, covering measures such as high-intensity exercise intervention, precise dietary control, and psychological intervention.

[0158] For users at medium risk: Customized health management plans, such as moderate exercise programs, dietary adjustment suggestions, and guidance on improving lifestyle habits, and regular monitoring plans to closely track changes in health indicators.

[0159] For low-risk users: Develop a basic monitoring plan, including routine health checkups and regular testing of key indicators, while providing health education and prevention advice to help maintain good health.

[0160] Output

[0161] A personalized risk assessment report is generated, including a comprehensive risk score, disease-related risk levels, and tiered intervention strategies. The report uses a combination of visual charts (such as a risk radar chart showing the components of the comprehensive risk score and a knowledge graph visualizing the disease-related risk levels) and detailed text descriptions to intuitively present the user's health risk status and intervention recommendations. Simultaneously, the report is synchronized to user terminals and medical personnel systems to guide users in self-health management and assist medical personnel in making treatment decisions. Furthermore, a report feedback mechanism is established to dynamically adjust intervention strategies based on the user's actual implementation and changes in health indicators, achieving closed-loop optimization of health management.

[0162] For example, Mr. Wang, a 35-year-old male teacher, participated in this health management program because of "recurrent dizziness for 3 months". Mr. Wang has a family history of hypertension (his father has had hypertension for 10 years), exercises for an average of 20 minutes a day, has a salty diet, and his blood pressure has fluctuated between 145-155 / 90-95 mmHg in the past 3 months. He has no other underlying diseases.

[0163] Data collection and integration

[0164] Individual basic characteristic data collection

[0165] Stable data acquisition: Wang's gender (male, code 1), age (35 years old), height (1.75m), and weight (80kg) were obtained through the electronic medical record system. His BMI was calculated. This falls within the overweight range.

[0166] Questionnaire data: Occupation (Teacher), no missing values, directly included in the analysis, combined with occupational characteristics to focus on vocal cord health risks.

[0167] Health behavior data collection

[0168] Dietary habits: The questionnaire showed a preference for salty food (high-sodium diet), and that the individual consumed fruits and vegetables three times a week. There was no dieting or overeating. No missing data was found, so no data needs to be filled.

[0169] Exercise-based sleep: The sleep pattern is based on 5 days of exercise per week, 20 minutes each time, mainly brisk walking; the sleep time is 23:30, with one awakening during the night, and a total sleep time of 6.5 hours.

[0170] Exercise intensity estimation: walking speed stride Daily average steps Steps, exercise duration Heart rate during exercise times / minute times / minute times / minute Therefore =0 (low-intensity exercise).

[0171] Sleep quality index: Sleep quality is good.

[0172] History of smoking, alcohol, and medication: No history of smoking. Drinking alcohol twice a week, 500ml of beer (4% alcohol content) each time, calculate... Standard drinking unit; no history of special medication use.

[0173] Clinical diagnostic data collection

[0174] Symptoms and signs: Dizziness (occurring 2-3 times per week, lasting 10-15 minutes); physical examination revealed a systolic blood pressure of 150 mmHg and a diastolic blood pressure of 92 mmHg. Box plot method was used to identify abnormal blood pressure values: after sorting the systolic blood pressure data of a certain group... , , , , The value is not abnormal; however, if it is higher than the normal range (120 / 80 mmHg), it is marked as abnormal blood pressure.

[0175] Imaging findings: Blood routine and liver and kidney function tests were normal; carotid ultrasound showed mild intimal thickening without plaque. The imaging report extracted the key information "mild carotid intimal thickening" using NER technology, and incorporated the data after structuring.

[0176] Genetic medical data collection

[0177] The questionnaire collected family history of genetic diseases (father has hypertension, FGRI=1), but there was no genetic testing data. The model defaulted to "no other clear history of genetic diseases".

[0178] Data integration and processing

[0179] Data cleaning: All data has a 95% integrity rate, with no records having more than 30% missing values, so there is no need to mark missing data.

[0180] Natural Language Processing: Transform unstructured text such as "mild carotid intima-media thickening" and "dizziness" into structured tags.

[0181] Data standardization: Z-score standardization was performed on blood pressure (systolic pressure 150 mmHg, mean 120 mmHg, standard deviation 10 mmHg): BMI (mean 24, standard deviation 2): .

[0182] Intelligent diagnostic analysis

[0183] Association analysis

[0184] Association rule mining: The Apriori algorithm was used, with MinSupport=0.05 and MinConfidence=0.6. During transactional processing, "high-salt diet, overweight, family history of hypertension, and abnormal blood pressure" were grouped into transactions.

[0185] Support calculation: .

[0186] Confidence calculation .

[0187] Graph analysis technology: A knowledge graph is constructed, where the node "hypertension" is connected to "dizziness," "carotid intimal thickening," "high-salt diet," and "family history," with edge weights of 0.8 (80% clinical frequency), 0.6 (60% clinical frequency), 0.7 (70% clinical frequency), and 0.9 (90% clinical frequency), respectively. Node features are updated using GCN. Ultimately, a complex link was uncovered between "family history of hypertension + high-salt diet + overweight → hypertension".

[0188] Pattern recognition and classification

[0189] Support Vector Machine (SVM): An SVM model is trained using "BMI, blood pressure, family history, diet, and exercise" as features. Penalty parameter. (Cross-validation optimization), kernel function The model output that Wang's hypertension risk classification was "prehypertension", with an accuracy of 92%.

[0190] Random Forest: Total Number of Features Feature subset size The maximum tree depth is 8, and the minimum number of sample splits is 5. 100 decision trees voted, and 85 were classified as "prehypertension," resulting in the final classification as "prehypertension."

[0191] Predictive Trend Analysis (Deep Q-Network, DQN)

[0192] Data and Definition of State / Action / Reward

[0193] State: Construct an 8-dimensional state vector After dynamic weighting (Blood pressure, family history, and high-salt diet have higher weighting).

[0194] Action: Choose "Moderate-intensity exercise (A2: 150 minutes of brisk walking per week) + low-salt diet (B2: daily salt intake)". 5g) + regular sleep schedule (C2: fall asleep before 11 pm) + regular monitoring (D1: measure blood pressure 3 times a week), i.e. Action = A2B2C2D1.

[0195] Reward: (Current health score 1.2, predicted score 0.8 after performing the action); (Current risk level: 2; predicted risk level: 1) (Compatibility 0.9, Difficulty 0.5) , ),total .

[0196] DQN model training: Experience replay pool size 10000, learning rate 0.001, exploration rate decayed from 0.9 to 0.1. After 1000 training rounds, the loss function converged. ).

[0197] Prediction and Output: The model predicts that if the A2B2C2D1 action is performed within the next month, there is a 70% probability that the systolic blood pressure will drop to 135-145 mmHg and the diastolic blood pressure will drop to 85-90 mmHg; the risk level will drop from "prehypertension" to "high normal".

[0198] Comprehensive diagnostic decision

[0199] Diagnostic conclusion: Association analysis showed a strong association between "high-salt diet + family history + overweight" and hypertension. The pattern recognition classification was "prehypertension". The DQN predicted a decrease in risk after the action was performed. Combined with the "Chinese Guidelines for the Prevention and Treatment of Hypertension", the comprehensive diagnosis was "prehypertension", with no complications and low severity.

[0200] Diagnostic basis and recommendations: The diagnostic report lists evidence such as "abnormal blood pressure, family history, and high-salt diet" and recommends "moderate-intensity exercise, low-salt diet, and regular monitoring". It also incorporates DQN exercise recommendations to customize a personalized plan.

[0201] Dynamic and flexible conclusion priority adjustment mechanism

[0202] Low-level general feature library

[0203] Disease severity score: (Blood pressure abnormality weight 0.4, normalized value 0.8) + 0.3 × 0.6 (family history weight 0.3, normalized value 0.6) + 0.2 × 0.7 (overweight weight 0.2, normalized value 0.7) + 0.1 × 0.5 (carotid artery thickening weight 0.1, normalized value 0.5) = 0.32 + 0.18 + 0.14 + 0.05 = 0.69, which is classified as intermediate risk (0.4 ≤ 0.69). 0.7).

[0204] Intervention timeliness: Intervention time window for prehypertension Months , , routine intervention.

[0205] Disease progression rate: Change in systolic blood pressure / week (1 month ago 148) ), =10 / week, Progress was not rapid.

[0206] Mid-layer disease specificity

[0207] The expert team, focusing on hypertension, increased the weighting of "blood pressure change trend and family history" (from 0.3 to 0.4). (Wang) The recalculation is 0.4×0.8+0.4×0.6+0.2×0.7+0.1×0.5=0.32+0.24+0.14+0.05=0.75, and it is adjusted to a high-risk level (≥0.7).

[0208] Top-level user needs adaptation

[0209] Obtaining weight adjustment factor

[0210] Users are interested in "cardiovascular diseases". The association matrix's cardiovascular disease row vectors are extracted, with increased weights for blood pressure and family history, and then normalized. , .

[0211] Users select "Prioritize Prevention Progress" to obtain the corresponding template; the weight of the "Lifestyle Intervention" indicator is increased, and after normalization... , .

[0212] User preference for "long-term stable recovery" led to an increased weighting of rehabilitation training indicators after normalization. .

[0213] Final priority calculation: (High-risk diseases, Take 0.3). It has a medium priority and requires priority implementation of lifestyle interventions.

[0214] Multi-dimensional risk warning and tiered intervention strategy generation mechanism

[0215] Data input

[0216] Integrate all of Wang's standardized data and access weekly blood pressure monitoring data in real time (148 / 90 mmHg in week 1, 145 / 88 mmHg in week 2).

[0217] Processing

[0218] Data fusion scoring

[0219] Risk baseline score: 0.3×0.5 (age 35, weight 0.3, quantified value 0.5) + 0.5×1 (family history, weight 0.5, quantified value 1) + 0.2×0.8 (overweight, weight 0.2, quantified value 0.8) = 0.15 + 0.5 + 0.16 = 0.81.

[0220] Current score: =0.4×0.8 (abnormal blood pressure, weight 0.4, score 0.8)+0.3×0.5 (insufficient exercise, weight 0.3, score 0.5)+0.2×0.7 (high-salt diet, weight 0.2, score 0.7)+0.1×0.3 (carotid artery thickening, weight 0.1, score 0.3)=0.32+0.15+0.14+0.03=0.64.

[0221] Trend coefficient: DQN predicts a downward trend in blood pressure, and the slope of the fitted straight line is... =-0.05, =-0.05 (improving trend).

[0222] Overall risk score: , , , Medium risk (0.3≤0.57) 0.7).

[0223] Risk association mining: Taking "prehypertension" as the core symptom, we used knowledge graph to mine the indirect association of "prehypertension → carotid artery thickening → coronary heart disease" (edge ​​weight 0.4) and the direct association of "high salt diet → high blood pressure" (edge ​​weight 0.8).

[0224] Strategies generated: For medium-risk users, customized intervention strategies include: 150 minutes of exercise per week (brisk walking + swimming), daily salt intake <5g, monthly carotid ultrasound monitoring, and follow-up visits every 2 weeks.

[0225] Output

[0226] Generate a personalized risk assessment report, including:

[0227] The overall risk score is 0.57 (medium risk), and the risk radar chart shows that "blood pressure and family history" are the main risk items.

[0228] Disease-related risk levels: "Prehypertension → Carotid artery thickening → Coronary heart disease" (low-risk association).

[0229] Stratified intervention strategy: exercise, diet, and monitoring plan details. The report was synchronized to Mr. Wang's mobile app and the community doctor system. One month later, at the follow-up visit, his blood pressure was 140 / 85 mmHg, and his comprehensive risk score had decreased to 0.45. The intervention strategy was adjusted to "maintain current exercise and reduce dietary salt..." 4g.

[0230] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0231] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0232] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A method for prioritizing medical examination report conclusions based on user medical history, characterized in that, The method includes: Collect multi-source data and preprocess the multi-source data to obtain a structured multi-source dataset; Based on structured multi-source datasets, the correlation between health data and diseases is explored through association analysis, preliminary screening is carried out using pattern recognition classification, and a health trend prediction model is constructed by combining deep Q-networks to obtain multi-dimensional diagnostic analysis results. By combining the results of multi-dimensional diagnostic analysis, a hierarchical weighting model is used, along with a priority calculation formula, to determine the priority of diagnostic conclusions. Input structured multi-source datasets and multi-dimensional diagnostic analysis results, calculate a comprehensive risk score through data fusion scoring, and classify risk levels according to the comprehensive risk score to match corresponding phased intervention strategies; By inputting multi-dimensional diagnostic analysis results, intervention strategy execution feedback data, and user health preferences into a comprehensive diagnostic decision model, and combining the priority of diagnostic conclusions, personalized control coefficients adapted to different health scenarios are calculated to obtain differentiated health-related strategies. The specific process for obtaining the multidimensional diagnostic analysis results is as follows: This process involves selecting disease-related data dimensions from structured multi-source datasets and combining them into transaction units. The Apriori algorithm is used to mine association rules that meet thresholds, and a knowledge graph is constructed. Graph convolutional networks are then used to mine direct, indirect, and hidden associations, clarifying the relationship between health data and disease. Based on the structured multi-source datasets and the association results, support vector machines (SVMs) and random forest classification models are trained. The SVMs handle data noise and non-linear distributions by optimizing penalty parameters and kernel functions, while the random forest classification models avoid overfitting by randomly selecting features and setting tree parameters. Together, they output preliminary disease screening results. Core health indicators are selected from the structured data to construct state vectors. A health management-related action space and hierarchical reward function are defined, and a deep Q-network model is trained. The training process is optimized through experience replay and dynamic parameter tuning, outputting disease risk, progression trends, and suggestions for optimized health behaviors. Finally, the potential disease risks obtained from association mining, the preliminary diagnostic results from pattern recognition classification, and the health trends and behavioral suggestions output by the deep Q-network are summarized to form a multi-dimensional diagnostic analysis result. The specific process for obtaining the priority of the diagnostic conclusions is as follows: Based on multi-dimensional diagnostic analysis results, the severity of the disease is assessed with reference to the disease severity and criticality grading standards. Intervention timeliness levels are classified in conjunction with clinical guidelines. The rate of disease progression is quantified by changes in core physiological parameters, forming basic assessment data. A team of medical experts manually sets and regularly updates the weights of common disease characteristics based on the latest treatment guidelines. Simultaneously, they collaborate with rare disease treatment institutions to optimize the rare disease weight calculation model using professional data, thus correcting the basic assessment weights. The front-end interface collects user preferences regarding disease types, treatment expectations, and recovery periods. These user choices are mapped to three weight adjustment factors: disease type, treatment expectation, and recovery period. After normalization, personalized adjustment vectors are generated. The underlying basic assessment results, mid-level optimized weights, and top-level adjustment factors are integrated, substituted into the priority calculation formula, and combined with the disease progression rate adjustment coefficient to obtain a diagnostic conclusion priority score, determining the priority ranking.

2. The method for prioritizing medical examination report conclusions based on user medical history according to claim 1, characterized in that, The specific process for obtaining the structured multi-source dataset is as follows: Individual basic characteristic data, health behavior data, clinical diagnostic data, and genetic and medical data were collected separately. Records with high data completeness were prioritized for retention. Numerical data with a missing rate exceeding 30% were marked as missing. Categorical data with a large proportion were classified separately, and outliers were identified and processed. For unstructured data, key information was extracted using natural language processing technology and transformed into structured data. Non-missing numerical data were processed using standardization methods to make the data meet the input requirements of subsequent analysis models. Numerical questionnaire data was not standardized, and subsequent analysis adopted targeted strategies to generate structured multi-source datasets.

3. The method for prioritizing medical examination report conclusions based on user medical history according to claim 2, characterized in that, The specific process of mining the correlation between the health data and diseases is as follows: From structured multi-source datasets, data dimensions closely related to diseases are selected, and data with medical relevance are grouped into transaction units to form transaction sets. Using the Apriori algorithm, with minimum support and minimum confidence parameters set, the transaction sets are mined to identify association rules whose support and confidence meet threshold requirements, thus preliminarily identifying potential associations between health data and diseases. Disease diagnoses and examination results from electronic medical records, as well as family medical histories collected from questionnaires, are transformed into nodes and edges of a knowledge graph. The weights of the edges are adjusted based on the frequency of associations in clinical data and medical literature. The knowledge graph is trained using a graph convolutional network to update node feature representations, mining direct, indirect, and hidden associations between health data and diseases. The results of association rule mining are verified and supplemented to clarify the association between health data and diseases.

4. The method for prioritizing medical examination report conclusions based on user medical history according to claim 3, characterized in that, The specific process for determining the staged intervention strategy is as follows: Collect structured multi-source datasets and multi-dimensional diagnostic analysis results; based on the input data, calculate the risk baseline score and current score by weighted summation, and combine the health indicator change trend coefficients output by the deep Q network to obtain a comprehensive risk score through weighted fusion; Set risk thresholds and classify users corresponding to the comprehensive risk scores into three risk levels: high, medium, and low. Starting with core symptoms, and based on knowledge graphs, graph analysis algorithms are used to mine direct, indirect, and hidden risk associations between symptoms and construct a risk hierarchy. For high-risk users, a chronic disease management pathway and a strict health management plan are matched; for medium-risk users, a personalized health management plan and a regular monitoring plan are customized; and for low-risk users, a basic monitoring plan and health prevention recommendations are developed, forming a phased intervention strategy. Based on the results of the intervention strategy implementation and changes in user health indicators, feedback is used to adjust the comprehensive risk score calculation parameters and the risk correlation mining model, and the phased intervention strategy is updated.

5. The method for prioritizing medical examination report conclusions based on user medical history according to claim 4, characterized in that, The specific process for obtaining the differentiated health association strategy is as follows: Collect high-precision diagnostic conclusions, system optimization data, and user tags to obtain integrated data; use the integrated data as input to build a collaborative strategy model, and embed health scenario classification logic into the model; The model determines core health goals based on the priority of diagnostic conclusions for different health scenarios, adjusts the adaptability parameters of the strategy with reference to system optimization data, matches personalized needs based on user tags, and calculates personalized control coefficients that are suitable for each scenario. Based on the personalized control coefficient, the basic intervention strategy is adjusted, intervention measures are strengthened for high-priority health issues, the feasibility of the strategy is optimized in combination with user preferences, and differentiated health management strategies adapted to different scenarios are generated. The generated differentiated strategies are applied to actual health management, and user feedback and health indicator changes are collected to feed back into the collaborative strategy model and update the personalized control coefficient calculation logic.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement the method for prioritizing the conclusions of a physical examination report based on a user's medical history as described in any one of claims 1-5.

Images