A Hybrid Optimization Risk Assessment Method for GERD-CHD

Abstract

This invention discloses a hybrid optimization risk assessment method for GERD-CHD, belonging to the field of data prediction technology, including the following steps: First, target patients are screened from electronic medical records based on disease codes, and clinical features based on the bidirectional interaction mechanism of the two diseases are extracted and feature vectors are generated; then, these vectors are input in parallel into particle swarm optimization algorithm, logistic regression algorithm, and cross-entropy algorithm to independently calculate multiple sets of feature weight vectors; next, the features are weighted according to each set of weights and then fed into a classification model to output the risk assessment results; simultaneously, weighted similarity is calculated based on at least one set of weights to retrieve similar cases and provide reference information; finally, the risk results, weight information, and case reference information are output or visualized to provide medical personnel with probability values ​​or risk stratification references. This invention improves the accuracy, stability, and clinical interpretability of coronary artery disease risk assessment in patients with gastroesophageal reflux disease.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and more specifically to a hybrid optimization risk assessment method for GERD-CHD. Background Technology

[0002] Gastroesophageal reflux disease (GERD) is a chronic digestive system disorder with a high prevalence in the population. Coronary artery disease (CAD), a common cardiovascular disease, carries a high risk of death. Recent studies have shown a pathophysiological link between GERD and CAD. For example, esophageal acid stimulation may induce coronary artery spasm via vagal reflex, while the accumulation of acidic metabolites due to myocardial ischemia may exacerbate GERD symptoms, resulting in a different CAD risk distribution in GERD patients compared to the general population. Currently, most commonly used cardiovascular disease risk assessment tools are built based on general population data. Their risk prediction models typically lack feature systems designed for GERD-related comorbidity mechanisms and lack specific feature weight optimization strategies for this patient group. Furthermore, existing risk assessment models are mostly single-classification models, lacking interpretability in their decision-making process and failing to provide clinicians with auxiliary references for similar cases. When faced with data from GERD patients exhibiting multiple comorbidity characteristics, single optimization algorithms are prone to getting trapped in local optima, leading to insufficient stability in feature weight estimation and consequently affecting the overall performance of risk assessment. Therefore, how to construct a coronary artery disease risk assessment scheme that is accurate, stable, and interpretable using electronic medical record data for patients with gastroesophageal reflux disease remains a technical problem that needs to be solved. Summary of the Invention

[0003] To construct a coronary artery disease risk assessment scheme that combines accuracy, stability, and interpretability using electronic medical record data, this invention proposes a hybrid optimized risk assessment method for GERD-CHD, comprising the following steps: S1: Read patient records from the electronic medical record database and filter target patient samples according to the disease classification code corresponding to gastroesophageal reflux. S2: Extract multiple clinical risk features from the target patient sample based on the bidirectional interaction mechanism between gastroesophageal reflux and coronary artery disease, and generate patient feature vectors after missing record removal, categorical variable encoding and age discretization. S3: Input the patient feature vectors into the particle swarm algorithm, gene logistic regression algorithm and cross-entropy algorithm in parallel. Each algorithm independently calculates the feature weights to obtain multiple sets of feature weight vectors. S4: The patient feature vector is weighted based on multiple sets of feature weight vectors, and the weighted feature vectors of each set are input into the preset classification model to construct multiple coronary artery disease risk assessment models and output the risk assessment results of the target patient. S5: Calculate the weighted similarity between the target patient and historical cases based on at least one of the multiple sets of feature weight vectors, retrieve similar case information based on the weighted similarity, and output case reference information corresponding to the risk assessment results; S6: Output or visualize the risk assessment results, feature weight information, and similar case reference information to provide medical staff with probabilistic risk assessment results.

[0004] This invention uses a hybrid optimization algorithm to generate multiple sets of feature weights in parallel for weighted classification and weighted similar case retrieval, which can improve the accuracy of risk assessment for patients with gastroesophageal reflux disease and enhance the interpretability of the output results.

[0005] Furthermore, in step S1, the disease classification code corresponding to gastroesophageal reflux is ICD-9 code 530.11 and 530.81, or other versions of disease classification codes that have the same disease meaning as gastroesophageal reflux.

[0006] Furthermore, in step S2, the multiple clinical risk characteristics include age, gender, hypertension, diabetes, hyperlipidemia, alcohol-related diseases, stroke, obesity, chronic obstructive pulmonary disease, asthma, gallstones, anxiety, depression, thyroid disease, chronic kidney disease, and cirrhosis.

[0007] Furthermore, in step S2, the categorical variable is encoded as a binary code, and the age discretization process is divided into five stages: ≤30 years old, 31-50 years old, 51-70 years old, 71-90 years old, and >90 years old.

[0008] Furthermore, in step S4, the risk assessment result for the target patient is output by voting or weighting the risk probability values ​​output by multiple classification models to obtain a comprehensive risk assessment result.

[0009] Furthermore, in step S5, the weighted similarity is calculated using weighted Euclidean distance or weighted cosine similarity, wherein the weights are taken from at least one of multiple sets of feature weight vectors.

[0010] Furthermore, in step S5, the retrieval of similar cases is implemented based on a case-based reasoning architecture and includes the following case database update rules: When clinicians confirm that the risk assessment results are consistent with the patient's actual clinical diagnosis, they store the patient's feature vector, corresponding feature weight vector, and diagnosis results in the historical case database.

[0011] Furthermore, in step S4, the preset classification model includes a support vector machine model, which employs a radial basis function kernel, and the kernel parameters of the radial basis function kernel are... according to Settings, where Number of input variables; regularization parameters for the support vector machine model. The value was determined within the range of 1 to 10 using a grid search method, with a stopping condition precision of 1.0E-2.

[0012] Furthermore, the preset classification model includes a backpropagation neural network model, which adopts a single hidden layer structure. The number of neurons in the hidden layer is set to half the sum of the number of neurons in the input layer and the output layer. The activation function is the Sigmoid function, the learning rate is 0.3, and the maximum number of training iterations starts from 3000 and is gradually reduced for optimization.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes a hybrid optimization risk assessment method for GERD-CHD, which optimizes the weights of clinical risk features determined based on the bidirectional interaction mechanism between gastroesophageal reflux and coronary artery disease by using particle swarm optimization, gene logistic regression and cross-entropy algorithm in parallel. This effectively reduces the risk of a single optimization algorithm getting trapped in local optima and improves the stability of feature weight calculation and the robustness of assessment results. (2) The weighted features are input into multiple classification models to output the comprehensive risk assessment results, and similar case reference information is retrieved from the historical case database based on the weighted similarity, realizing the dual output of risk quantification assessment and clinically interpretable evidence; (3) Through the clinically confirmed case entry and update mechanism, the ability to search for similar cases can be iteratively optimized with use, and the pertinence of coronary artery disease risk assessment and the interpretability of output results can be improved for patients with gastroesophageal reflux disease. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the steps of a hybrid optimization risk assessment method for GERD-CHD; Figure 2 Workflow diagram for the similar case retrieval module; Figure 3 This is a schematic diagram of the interface for displaying risk assessment results. Detailed Implementation

[0015] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments.

[0016] Gastroesophageal reflux disease (GERD) is a common digestive system disorder. If left uncontrolled for a long period, it can lead to esophagitis, esophageal ulcers, Barrett's esophagus, and even esophageal cancer. The prevalence in my country is 5%-10%, affecting hundreds of millions of people globally, and is more prevalent in Western countries. Recent studies have shown a bidirectional interaction between GERD and coronary artery disease: esophageal acid stimulation can induce coronary artery spasm via vagal reflex, while the accumulation of acidic metabolic products due to myocardial ischemia can exacerbate GERD symptoms. This pathophysiological link results in a different coronary artery disease risk distribution in the GERD patient population compared to the general population. Currently, most widely used cardiovascular disease risk prediction tools are designed for the general population, such as the Framingham Heart Study Risk Score and the European Systemic Coronary Artery Risk Assessment Model. These tools typically calculate risk based on traditional risk factors such as age, sex, blood pressure, blood lipids, smoking, and diabetes, but they lack a specific clinical feature system for the comorbid mechanisms associated with gastroesophageal reflux disease (GERD) and a feature weight optimization mechanism for this patient group. Furthermore, existing models often output risk probability values ​​in a "black box" manner, making it difficult to provide clinicians with similar case data for reference, resulting in insufficient interpretability and affecting their acceptance and application value in clinical practice. To address these issues, this invention provides a risk assessment method for GERD-related coronary artery disease. It constructs a disease-specific feature system based on pathophysiological mechanisms, employs multiple optimization algorithms to calculate feature weights in parallel, and combines a classification model with similar case retrieval to achieve risk stratification output, aiming to improve the accuracy of risk assessment while enhancing the interpretability of the results. Figure 1 As shown, the method mainly includes the following steps: S1: Read patient records from the electronic medical record database and filter target patient samples according to the disease classification code corresponding to gastroesophageal reflux. S2: Extract multiple clinical risk features from the target patient sample based on the bidirectional interaction mechanism between gastroesophageal reflux and coronary artery disease, and generate patient feature vectors after missing record removal, categorical variable encoding and age discretization. S3: Input the patient feature vectors into the particle swarm algorithm, gene logistic regression algorithm and cross-entropy algorithm in parallel. Each algorithm independently calculates the feature weights to obtain multiple sets of feature weight vectors. S4: The patient feature vector is weighted based on multiple sets of feature weight vectors, and the weighted feature vectors of each set are input into the preset classification model to construct multiple coronary artery disease risk assessment models and output the risk assessment results of the target patient. S5: Calculate the weighted similarity between the target patient and historical cases based on at least one of the multiple sets of feature weight vectors, retrieve similar case information based on the weighted similarity, and output case reference information corresponding to the risk assessment results; S6: Output or visualize the risk assessment results, feature weight information, and similar case reference information to provide medical staff with probabilistic risk assessment results.

[0017] Specifically, the technical solution of this invention can be deployed in a hospital information system or a medical big data analysis platform, with at least one processor executing the corresponding computer program to achieve all functions. In the actual application of this invention, the system first needs to read the patient's historical medical records from the electronic medical record database. The electronic medical record database can be a hospital's clinical data center or a structured data storage device in a regional medical information platform. The system filters these records according to the disease classification code corresponding to gastroesophageal reflux disease (GERD) to obtain target patient samples. In one specific embodiment, the disease classification code corresponding to GERD adopts ICD-9 codes 530.11 and 530.81, which correspond to erosive esophagitis and non-erosive esophageal reflux disease, respectively. Considering that different medical institutions or different periods may use different versions of the disease classification coding system, the system also supports other codes in ICD-10 or ICD-11 that have the same disease meaning as GERD. Through this coding screening method, the system can accurately locate the GERD patient group from massive medical record data, thereby ensuring that the target population for subsequent risk assessment has accurate disease labels.

[0018] After screening the target patient samples, the system needs to extract clinical risk features related to gastroesophageal reflux disease (GERD) and coronary artery disease from these samples. These features are not randomly selected, but determined based on the bidirectional interaction mechanism between GERD and coronary artery disease. Specifically, acid stimulation in the esophagus of GERD patients can induce coronary artery spasm via vagal reflex, while the accumulation of acidic metabolites caused by myocardial ischemia can in turn aggravate the symptoms of GERD, indicating a complex pathophysiological relationship between the two. Based on this mechanism, the clinical risk features extracted in this embodiment include age, gender, hypertension, diabetes, hyperlipidemia, alcohol-related diseases, stroke, obesity, chronic obstructive pulmonary disease (COPD), asthma, gallstones, anxiety, depression, thyroid disease, chronic kidney disease, and cirrhosis, totaling sixteen items. These features cover multiple dimensions, including demographic information, traditional cardiovascular risk factors, respiratory diseases, mental and psychological factors, digestive tract-related diseases, and cerebrovascular diseases reflecting the background of atherosclerosis, and can comprehensively reflect the risk status of GERD patients developing coronary artery disease.

[0019] After extracting the aforementioned features, the system needs to perform necessary preprocessing operations on the raw data. Considering that electronic medical systems often contain a certain proportion of missing records, such as missing body mass index (BMI) for some patients or missing blood lipid test results for others, the system employs a deletion strategy to handle these missing records. This involves removing any patient records containing missing values ​​to ensure the integrity and consistency of the data used for subsequent modeling. For categorical variables, such as gender, presence or absence of hypertension, and presence or absence of diabetes, the system uses binary encoding to convert them into numerical forms. Specifically, "yes" or "exists" is encoded as 1, and "no" or "does not exist" is encoded as 0. For the continuous variable of age, the system does not directly use the raw numerical value but discretizes it, dividing it into five stages: less than or equal to 30 years old, 31 to 50 years old, 51 to 70 years old, 71 to 90 years old, and greater than 90 years old. This discretization helps reduce the impact of extreme age values ​​on the model and also makes the feature distribution smoother, facilitating subsequent optimization algorithms. After the above missing record removal, categorical variable encoding, and age discretization processes, each patient sample is converted into a patient feature vector with a fixed dimension, which can be used as the standard input format for subsequent optimization algorithms.

[0020] After generating the patient feature vector, the system enters the core feature weight optimization stage. In this stage, the invention employs three different optimization algorithms to process the patient feature vector: particle swarm optimization, logistic regression, and cross-entropy. The system inputs the same patient feature vector into each of these three algorithms in parallel, with each algorithm independently calculating the feature weights.

[0021] The advantage of parallel processing lies in the fact that the three algorithms can run simultaneously without interference, thereby improving overall computational efficiency. More importantly, the Particle Swarm Optimization (PSO) algorithm, Logistic Regression (CROS) algorithm, and Cross-Entropy (CEE) algorithm have different optimization mechanisms and search strategies, each exploring the feature weight space in different ways. The PSO algorithm maintains a swarm of particles, each representing a candidate solution in the feature weight space. Particles iteratively update their velocity and position by sharing their individual optimal position and the global optimal position, thus approximating the optimal feature weight vector across the entire search space. The CROS algorithm combines a logistic regression model with a genetic algorithm, optimizing feature combinations through genetic operations such as selection, crossover, and mutation, making it particularly suitable for handling discrete combinatorial optimization problems in high-dimensional feature spaces. The CEE algorithm, based on the ideas of rare event simulation and importance sampling, guides the search direction by iteratively updating probability distribution parameters, exhibiting a relatively fast convergence speed in specific problem domains. After the three algorithms run in parallel, the system obtains three different sets of feature weight vectors, denoted as the PSO algorithm weight vector, the CROS algorithm weight vector, and the CEE algorithm weight vector, respectively. These three sets of weight vectors reflect the importance of various clinical risk features to the occurrence of coronary artery disease from different perspectives, providing a diverse feature weighting perspective for subsequent risk assessment.

[0022] After obtaining multiple sets of feature weight vectors, the system uses these weight vectors to weight the patient feature vector. Specifically, for the same patient feature vector, the system applies the three sets of feature weight vectors mentioned above to generate three feature vector versions with different weights. Each weighted feature vector is input into a preset classification model for coronary artery disease risk assessment. The classification model can be a Support Vector Machine (SVM) model, a Backpropagation Neural Network (BPN) model, or a combination of both models. The Support Vector Machine model achieves binary classification by finding the optimal classification hyperplane and has good generalization performance when processing high-dimensional feature data. In this embodiment, the Support Vector Machine model uses radial basis functions as kernel functions to handle nonlinear relationships that may exist in the feature space. The kernel parameters of the radial basis function kernel are described below. Based on the number of input variables Configure the settings; the specific calculation formula is as follows: This configuration allows the kernel parameters to adaptively adapt to scale changes in feature vectors of different dimensions. The regularization parameter C is determined using a grid search method within the range of 1 to 10. The system iterates through candidate parameter values ​​within this range and selects the C value that optimizes cross-validation performance. The stopping condition precision is set to 1.0E-2, meaning training stops when the change in the objective function during model iterations is less than this threshold. The backpropagation neural network model uses a single hidden layer structure, with the number of hidden layer neurons set to half the sum of the number of input layer neurons and the number of output layer neurons. The activation function is... The function is designed for smoothness and differentiability, making it suitable for gradient calculation in the backpropagation algorithm. The learning rate is set to 0.3, and the maximum number of training iterations starts at 3000. During actual training, optimization can be performed by gradually decreasing the training iterations based on the performance of the validation set; for example, training can be terminated early when the validation loss no longer decreases after several consecutive iterations.

[0023] When the system employs multiple classification models, each model outputs a probability value for the patient's coronary artery disease risk. To obtain a more robust comprehensive assessment, the system can integrate the risk probability values ​​output by multiple classification models. A common integration method is the voting method, where each model independently determines whether the patient belongs to a high-risk or low-risk class, and the final classification is determined based on the voting results of the majority of models. Another method is the weighted average method, which assigns different weights to each model according to its performance on the validation set, and then performs a weighted average of the probability values ​​output by each model to obtain the comprehensive risk assessment result. Regardless of the integration method used, the final output risk assessment result can be either a continuous probability value or risk stratification information based on preset thresholds, such as low-risk, medium-risk, and high-risk levels. It is important to note that this risk assessment result is for medical personnel's reference only, used to assist clinical decision-making, and does not directly constitute a disease diagnosis. The final accurate diagnosis still requires a clinician to make a comprehensive judgment based on the patient's complete medical history, physical examination, and other auxiliary examination results.

[0024] In addition to outputting risk assessment results, the method of this invention also provides a similar case retrieval function to enhance the interpretability of the assessment results. The similar case retrieval is implemented based on a case-based reasoning framework (CBR). Figure 2As shown, the system first calculates the weighted similarity between the current target patient and each historical case in the historical case database based on at least one of the aforementioned sets of feature weight vectors. The weighted similarity can be calculated using either weighted Euclidean distance or weighted cosine similarity. Taking weighted Euclidean distance as an example, the system squares the corresponding feature differences between the current patient's feature vector and the historical case's feature vector, multiplies them by the corresponding feature weights, sums them, and then takes the square root to obtain the weighted distance value. The smaller this value, the higher the similarity. The weights used can be any set from the particle swarm optimization algorithm weight vector, the logistic regression algorithm weight vector, and the cross-entropy algorithm weight vector, or the average of the three weight vectors can be used as the weight source. The specific selection method can be flexibly configured according to the actual application scenario. After calculating the weighted similarity between the current patient and each historical case, the system sorts the historical cases in descending order of similarity and retrieves the cases with the highest similarity, such as the top three or top five cases. For each similar case retrieved, the system outputs corresponding case reference information. This information may include the basic characteristics of the case, the actual diagnosis of coronary artery disease, the treatment plan adopted, and the post-treatment outcome. Because the system finds reference cases with similar clinical characteristics and clear outcomes in the historical case database, this reference information helps medical personnel understand why the system provides a specific risk assessment result for the current patient.

[0025] To ensure the similar case retrieval function continuously improves with increased system usage, a case database update rule is designed into the case-based reasoning architecture. Specifically, when a clinician confirms that the system's output risk assessment result matches the patient's actual clinical diagnosis after using the system, the system can store all of the patient's information in the historical case database, including the patient's original feature vector, corresponding feature weight vector, and final diagnosis. This update mechanism continuously enriches and expands the historical case database, providing more reference samples for subsequent similar case searches, thereby gradually improving the accuracy and reference value of the search. Of course, the case database update can also be subject to a review mechanism, such as requiring confirmation from a qualified clinician before writing data, and recording the operator's information and the operation time to ensure the data quality and traceability of the case database.

[0026] After completing the risk assessment and similar case search, the system integrates and outputs the above information, presenting it to medical staff through a visual interface. The interface design fully considers ease of clinical use and intuitive information presentation. Figure 3As shown, the interface contains at least three main areas: the first area displays the patient's basic information and clinical feature input, allowing medical staff to verify the accuracy of the features extracted by the system and manually correct certain feature values ​​if necessary; the second area displays the risk assessment results, including risk probability values ​​or risk stratification levels, and the corresponding confidence levels; the third area displays reference information for similar cases, including a summary of the features of the most similar cases found, diagnostic conclusions, and treatment recommendations. Furthermore, the interface can include a graphical display of feature weight distribution, such as a bar chart or radar chart showing the weight values ​​of each feature in each group of feature weight vectors, allowing medical staff to intuitively understand which risk features contribute most to the current assessment. The entire interface design follows the principle of simplicity and clarity, with a clear information hierarchy, facilitating medical staff to quickly obtain key information and make clinical judgments.

[0027] The radial basis function kernel parameters of the support vector machine model in this invention Key hyperparameters such as the regularization parameter C, the number of hidden layer neurons in the backpropagation neural network, the learning rate (0.3), and the maximum number of training iterations were all determined to be optimal values ​​after extensive comparative experiments (including grid search, cross-validation, and ablation experiments with different parameter combinations) on real electronic medical record datasets. For example, for the radial basis function kernel parameters, this invention adopts... Adaptive setting method (where (where the input feature dimension is used), this formula combines empirical rules with the feature distribution characteristics of the dataset, and has been experimentally verified to be more effective than a fixed formula. Alternatively, other commonly used settings can improve the model's generalization performance by approximately 8%-12% on the GERD-CHD dataset. The above parameter selections are not arbitrary but are derived through iterative optimization, taking into account multiple factors such as feature dimension, sample distribution, model convergence speed, and overfitting risk. They can significantly improve the stability and prediction accuracy of the hybrid optimization framework in this specific disease scenario.

[0028] To verify the effectiveness of the method proposed in this invention, this embodiment conducted experimental verification based on real electronic medical record data. The experimental data came from the records of patients with gastroesophageal reflux disease in the medical database, totaling 71,060 records, of which 621 samples were diagnosed as positive for coronary artery disease and 70,439 samples were negative. As shown in Table 1, during the experiment, ten-fold cross-validation was used to evaluate the model performance, and the evaluation indicators included accuracy, area under the receiver operating characteristic (AUC) curve, and standard deviation. Multiple comparison schemes were set up in the experiment, including a scheme using only a single optimization algorithm, a scheme using a combination of two optimization algorithms, and a complete scheme using three optimization algorithms (Particle Swarm Optimization (PSO), Genetic Logistic Regression (GALR), and Cross-Entropy (CE)) in parallel. The experimental results show that the complete scheme using three optimization algorithms in parallel achieved the best performance in terms of accuracy, area under the curve, and stability. Specifically, the average accuracy of this scheme under the support vector machine classifier reached 94.23%, the area under the curve reached 0.836, and the standard deviation was only ±0.52%. In contrast, neither a single optimization algorithm nor a combination of two algorithms outperformed the complete solution in all the aforementioned metrics. This result fully demonstrates that the parallel configuration of the three heterogeneous optimization algorithms is not an arbitrary stacking, but rather a clear technical necessity. By starting from different optimization paths, they effectively reduce the risk of local optima that may arise from a single optimization method, thereby significantly improving the stability of feature weight calculation and the overall performance of the final evaluation model.

[0029]

[0030] In summary, this invention employs a parallel approach using particle swarm optimization, logistic regression, and cross-entropy algorithms to optimize the weights of clinical risk features determined by the bidirectional interaction mechanism between gastroesophageal reflux disease (GERD) and coronary artery disease. This effectively reduces the risk of single optimization algorithms getting trapped in local optima and improves the stability of feature weight calculation and the robustness of the evaluation results. Furthermore, the weighted features are input into multiple classification models to output a comprehensive risk assessment result. Based on weighted similarity, similar case reference information is retrieved from a historical case database and output, achieving a dual-path output of quantitative risk assessment and clinically interpretable evidence. Simultaneously, a clinically confirmed case database update mechanism allows for iterative optimization of the similar case retrieval capability with continued use. This invention improves the targeting and interpretability of coronary artery disease risk assessment for GERD patients.

[0031] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0032] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0033] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0034] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A hybrid optimization risk assessment method for GERD-CHD, characterized in that, Including the following steps: S1: Read patient records from the electronic medical record database and filter target patient samples according to the disease classification code corresponding to gastroesophageal reflux. S2: Extract multiple clinical risk features from the target patient sample based on the bidirectional interaction mechanism between gastroesophageal reflux and coronary artery disease, and generate patient feature vectors after missing record removal, categorical variable encoding and age discretization. S3: Input the patient feature vectors into the particle swarm algorithm, gene logistic regression algorithm and cross-entropy algorithm in parallel. Each algorithm independently calculates the feature weights to obtain multiple sets of feature weight vectors. S4: The patient feature vector is weighted based on multiple sets of feature weight vectors, and the weighted feature vectors of each set are input into the preset classification model to construct multiple coronary artery disease risk assessment models and output the risk assessment results of the target patient. S5: Calculate the weighted similarity between the target patient and historical cases based on at least one of the multiple sets of feature weight vectors, retrieve similar case information based on the weighted similarity, and output case reference information corresponding to the risk assessment results; S6: Output or visualize the risk assessment results, feature weight information, and similar case reference information to provide medical staff with probabilistic risk assessment results.

2. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S1, the disease classification code corresponding to gastroesophageal reflux is ICD-9 code 530.11 and 530.81, or other versions of disease classification codes that have the same disease meaning as gastroesophageal reflux.

3. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S2, the multiple clinical risk characteristics include age, sex, hypertension, diabetes, hyperlipidemia, alcohol-related diseases, stroke, obesity, chronic obstructive pulmonary disease, asthma, gallstones, anxiety, depression, thyroid disease, chronic kidney disease, and cirrhosis.

4. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S2, the category variable is encoded using binary encoding, and the discrete levels of the age discretization process are: ≤30 years old, 31-50 years old, 51-70 years old, 71-90 years old, and >90 years old.

5. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S4, the risk assessment result for the target patient is output by voting or weighting the risk probability values ​​output by multiple classification models to obtain a comprehensive risk assessment result.

6. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S5, the weighted similarity is calculated using weighted Euclidean distance or weighted cosine similarity, where the weights are taken from at least one of multiple feature weight vectors.

7. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S5, the retrieval of similar cases is implemented based on a case-based reasoning architecture and includes the following case database update rules: When clinicians confirm that the risk assessment results are consistent with the patient's actual clinical diagnosis, they store the patient's feature vector, corresponding feature weight vector, and diagnosis results in the historical case database.

8. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, In step S4, the preset classification model includes a support vector machine (SVM) model, which employs a radial basis function (RBF) kernel. The kernel parameters of the RBF kernel are... according to Settings, where Number of input variables; regularization parameters for the support vector machine model. The value was determined within the range of 1 to 10 using a grid search method, with a stopping condition precision of 1.0E-2.

9. The hybrid optimization risk assessment method for GERD-CHD as described in claim 1, characterized in that, The preset classification model includes a backpropagation neural network model, which adopts a single hidden layer structure. The number of neurons in the hidden layer is set to half the sum of the number of neurons in the input layer and the output layer. The activation function is the Sigmoid function, the learning rate is 0.3, and the maximum number of training iterations starts from 3000 and is gradually reduced for optimization.

Images