Coronary artery calcification prediction and analysis system based on receiver operating characteristic curve

Abstract

The present invention discloses a coronary artery calcification prediction and analysis system based on the receiver operating characteristic curve, which relates to the field of coronary artery calcification detection technology, including: an integrated processing unit for collecting clinical data, laboratory data, and imaging data, and generating standardized data through data preprocessing; an evaluation and prediction unit for predicting the risk probability of coronary artery calcification under different physiological feature combinations and supporting risk stratification; an analysis and display unit for generating transparent explanations and assisting decision-making through interactive visualization tools; a decision optimization unit for generating intervention recommendations based on the prediction results and optimizing blood magnesium management strategies. The present invention significantly improves the prediction accuracy and clinical practicality through an improved random forest model that integrates dynamic integrated selection and weighted probability. Through algorithm innovation and visual closed-loop design, it achieves full-chain optimization from data to decision-making, providing a precise, transparent, and dynamic intelligent tool for the prevention and treatment of CAC in CKD patients.

Description

Technical Field

[0001] The present invention relates to the field of coronary artery calcification detection technology, and in particular to a coronary artery calcification prediction and analysis system based on a receiver operating characteristic curve. Background Art

[0002] Chronic kidney disease (CKD) is a global public health issue, imposing significant economic costs on healthcare systems. With 14 out of every 100 adults worldwide suffering from CKD, the prevention and treatment of CKD remains a daunting task.

[0003] End-stage renal disease (ESRD) often presents with mineral metabolism disorders, such as hyperphosphatemia, elevated levels of intravenous parathyroid hormone (iPTH), elevated fibroblast growth factor-23 (FGF-23), and hypocalcemia. This often leads to valvular and vascular calcification (VC), accelerating the development of coronary artery disease, congestive heart failure, arrhythmias, and sudden cardiac death in CKD patients. The mortality rate from cardiovascular disease (CVD) in CKD patients is 10-30 times higher than that in the general population, accounting for 40% of all-cause mortality in CKD patients. VC, a pathophysiological characteristic of CVD in CKD patients, deserves attention and is a common pathophysiological feature of conditions such as diabetes, hypertension, and aging. Calcified blood vessel walls become more rigid and less elastic, leading to ventricular hypertrophy, myocardial ischemia, and heart failure, thrombosis, and plaque rupture. They are important markers of acute endpoint events of cardiovascular and cerebrovascular diseases and the occurrence of diabetic peripheral vascular disease, and are also the main independent risk factors for the high incidence of CVD and death in CKD patients.

[0004] Currently, several coronary artery calcification prediction technologies are available, primarily combining imaging studies (such as CT scans and X-rays) with biomarkers (such as serum phosphorus and iPTH) to assess CAC risk. Some existing technologies apply deep learning algorithms to automated calcified plaque detection based on imaging data and combine them with patient clinical data for risk assessment. Furthermore, traditional statistical methods such as regression analysis and logistic regression are also used to construct risk prediction models. These technologies achieve relatively accurate calcification risk prediction through data fusion and machine learning models.

[0005] Although existing technologies have made progress in some aspects, there are still some defects or shortcomings. First, the acquisition and analysis process of imaging data is relatively complicated, requiring expensive equipment and patients to be exposed to radiation, which may lead to poor popularization. Secondly, some existing prediction models mainly rely on a single indicator or biomarker, ignoring the integration of multi-dimensional data and individual differences, and the accuracy and generalization ability of the model are limited. Furthermore, existing technologies still have room for improvement in dealing with data imbalance and improving model stability. Some models may have overfitting phenomena, resulting in increased errors in prediction results. Finally, the clinical application translation of existing technologies is poor, especially the adaptability and popularization for different populations still need to be further optimized. Summary of the Invention

[0006] Based on this, it is necessary to provide a coronary artery calcification prediction and analysis system based on the receiver operating characteristic curve to address the above technical issues.

[0007] The present invention provides a coronary artery calcification prediction and analysis system based on a receiver operating characteristic curve, comprising:

[0008] An integrated processing unit is used to collect clinical data, laboratory data, and imaging data, and generate standardized data after data preprocessing;

[0009] An evaluation and prediction unit is used to predict the risk probability of coronary artery calcification under different combinations of physiological characteristics and support risk stratification;

[0010] An analysis and presentation unit, used to generate transparent explanations and assist decision-making through interactive visualization tools;

[0011] A decision optimization unit is used to generate intervention recommendations based on the prediction results and optimize the blood magnesium management strategy;

[0012] Among them, the evaluation and prediction unit includes:

[0013] Feature engineering module, used to extract static features, dynamic features and image-derived features corresponding to various types of data in standardized data and merge them into physiological features;

[0014] The model building module is used to build a risk probability prediction model for coronary artery calcification based on a random forest fused with dynamic integrated weighted probability, including:

[0015] Generate several decision trees by resampling and build the original random forest model;

[0016] Based on the dynamic ensemble selection of k-nearest neighbor-hierarchical clustering algorithm, the optimal performance decision tree is selected to build a risk probability prediction model;

[0017] The weighted probability fusion rule is used to obtain the prediction results of risk probability.

[0018] Furthermore, the integrated processing unit, the evaluation and prediction unit, the analysis and display unit, and the decision optimization unit are sequentially connected;

[0019] Clinical data included age, gender, chronic kidney disease stage, medical history, and medication data; laboratory data included blood magnesium, blood phosphorus, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and intact parathyroid hormone; and imaging data included original coronary CT images.

[0020] Furthermore, the evaluation and prediction unit also includes: a feature engineering module, a model building module, a dynamic update module and a calculation layer module;

[0021] The dynamic update module is used to set the conditions for updating new data. When the conditions are met, the model incremental training is triggered, the model parameters are updated, and a preset proportion of historical data is retained as a replay buffer.

[0022] The calculation and stratification module is used to input real-time standardized data into the risk probability prediction model, output the risk probability of coronary artery calcification, and divide the corresponding risk levels according to the preset threshold;

[0023] And the feature engineering module, model building module, dynamic update module and calculation layer module are connected in sequence.

[0024] Furthermore, based on the dynamic ensemble selection of the k-nearest neighbor-hierarchical clustering algorithm, the optimal decision tree is selected to construct a risk probability prediction model including:

[0025] Obtain test samples according to the data type of physiological characteristics, calculate the Euclidean distance between the test samples and the training set, and select the first k nearest neighbors in ascending order to form a neighborhood sample;

[0026] Calculate the inconsistency measure between any two decision trees in the neighborhood samples, construct the difference matrix, obtain the hierarchical clustering results of the decision trees, and form a cluster set;

[0027] Calculate the classification accuracy of each decision tree in the random forest on the cluster set, select the decision tree with the highest accuracy from each cluster as the best performing decision tree, and output the dynamically selected optimal subtree set;

[0028] The optimal subtree set was integrated into the original random forest model to generate a risk probability prediction model for coronary artery calcification.

[0029] Furthermore, the weighted probability fusion rule is adopted to obtain the risk probability prediction results including:

[0030] The performance of each decision tree in the optimal subtree set is evaluated through the test set to obtain the prediction results of each decision tree, and the prediction results matrix of each decision tree is recorded;

[0031] Based on Bayesian theory, the probability of the decision tree prediction results is calculated, and the total probability of the true category is calculated through weighted probability random forest to obtain the final coronary artery calcification risk probability prediction result of the test sample.

[0032] Furthermore, the calculation formula for the probability of the decision tree prediction result is:

[0033] ;

[0034] Where, Represents a test sample x test Select m The probability of the true category of a decision tree; F i represents the true category; x test represents the test sample; m represents a decision tree, and m =1, 2, 3, ..., M ; N Indicates the total number of categories in the test set; Represents a test sample x test Select m The conditional probability of a decision tree;

[0035] The total probability of the true category is calculated as:

[0036] ;

[0037] Where, Represents the true category F i The total probability of W Represents weighted weight.

[0038] Furthermore, the analysis and display unit includes: a curve analysis module, a feature analysis module, an integrated selection module and a graphic display module;

[0039] The curve analysis module is used to evaluate the prediction effect of the risk probability prediction model by drawing the receiver operating characteristic curve and select the optimal threshold to set the risk stratification;

[0040] The feature analysis module is used to analyze the impact of various physiological characteristics on risk probability prediction and display the contribution of physiological characteristics in combination with the receiver operating characteristic curve;

[0041] The ensemble selection module is used to introduce dynamic ensemble selection into the risk probability prediction model, visualize the optimization process of the test samples, and verify the prediction results;

[0042] Graphics display module, used to create heat maps, contour maps and coronary artery 3D models, annotating calcification locations and severity;

[0043] The curve analysis module, the feature analysis module, the integrated selection module and the graphic display module blocks are connected in sequence.

[0044] Furthermore, by drawing the receiver operating characteristic curve, the prediction effect of the risk probability prediction model was evaluated, and the optimal threshold was selected to set the risk stratification, including:

[0045] Obtain the prediction results of the risk probability prediction model for the test sample, set different thresholds to convert the probability values ​​of the prediction results into prediction labels;

[0046] Calculate the true positive rate and false positive rate for each threshold, plot the false positive rates and true positive rates calculated under multiple thresholds into a curve to obtain the receiver operating characteristic curve, and calculate the area under the curve of the receiver operating characteristic curve;

[0047] The Youden index of the receiver operating characteristic curve was calculated, and the optimal threshold of risk probability was selected based on the area under the curve and the Youden index. The patients were then divided into multiple risk levels based on the optimal threshold.

[0048] Furthermore, we analyze the impact of various physiological characteristics on risk probability prediction, and combine the receiver operating characteristic curve to show the contribution of physiological characteristics, including:

[0049] The contribution of each physiological feature was calculated based on the SHAP value and a bar graph was generated by arranging them in descending order;

[0050] For a single test sample, extract the decision paths of each decision tree in the risk probability prediction model, count the occurrence frequencies of feature splitting nodes, and generate a Sankey diagram to display the decision logic from input features to risk probability.

[0051] After removing individual physiological features one by one, the risk probability prediction model was retrained, and the decrease in the area under the curve was observed to quantify the criticality of other physiological features.

[0052] Furthermore, dynamic ensemble selection is introduced into the risk probability prediction model to visualize the optimization process of the test samples and verify the prediction results including:

[0053] Select any test sample and mark the k nearest neighbor samples in the two-dimensional t-SNE dimensionality reduction graph to distinguish the calcification group from the non-calcification group and reflect the local data distribution characteristics;

[0054] The dynamic merging process animation of the cohesive layer clustering is used to demonstrate the clustering process of neighborhood samples, annotate the calcification ratio of samples within the cluster, and list the best-performing decision trees in each cluster in tabular form.

[0055] The beneficial effects of the present invention are:

[0056] 1. The improved random forest model, which combines dynamic ensemble selection with weighted probability, significantly improves prediction accuracy and clinical practicality. The KNN-HC algorithm is introduced to dynamically select the optimal decision tree, and the model is adjusted based on the local data distribution of different patients to adapt to the heterogeneity of CKD patients. Dynamic threshold recommendations based on ROC curves, SHAP value-driven feature contribution analysis, and risk heat maps are used to transform complex models into actionable intervention recommendations for physicians (such as magnesium supplementation target values). The integration of biochemical indicators, imaging calcification scores, and dynamic time series features overcomes the limitations of traditional models that rely on a single data source. Through algorithmic innovation and a visual closed-loop design, full-chain optimization from data to decision-making is achieved, providing a precise, transparent, and dynamic intelligent tool for the prevention and treatment of CAC in CKD patients.

[0057] 2. This invention uses the KNN-HC algorithm to hierarchically cluster the neighborhood of each test sample and dynamically select the optimal decision tree within each cluster, enabling the model to adaptively optimize within the local data distribution. For example, for high-risk patients with hypomagnesemia (<0.75 mmol / L) and CKD stages 4-5, subtrees with an accuracy rate >90% within clusters of similar patients are prioritized over trees with average global performance, thereby reducing the risk of missed diagnoses. Furthermore, inter-cluster diversity is used to control overfitting and ensure generalizability.

[0058] 3. Quantify feature contributions through SHAP values ​​and recommend optimal thresholds in combination with ROC curves, directly guiding magnesium supplementation targets. In addition, the risk heat map displays the synergistic effect of blood magnesium and LDL, helping doctors develop joint intervention plans. Translate AI outputs into clinical actions, effectively improving model adoption rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0060] Figure 1 4 is a system principle block diagram of a coronary artery calcification prediction and analysis system based on a receiver operating characteristic curve according to an embodiment of the present invention.

[0061] Figure numbers: 1. Integration and processing unit; 2. Evaluation and prediction unit; 3. Analysis and display unit; 4. Decision optimization unit. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0063] See also Figure 1 , provides a coronary artery calcification prediction and analysis system based on the receiver operating characteristic curve, including: an integrated processing unit 1, an evaluation and prediction unit 2, an analysis and display unit 3 and a decision optimization unit 4; and the integrated processing unit, the evaluation and prediction unit, the analysis and display unit and the decision optimization unit are connected in sequence.

[0064] The integrated processing unit 1 is used to collect clinical data, laboratory data and imaging data, and generate standardized data after data preprocessing.

[0065] In the present invention, clinical data includes age, gender, chronic kidney disease stage, medical history, and medication data. Laboratory data includes serum magnesium, serum phosphorus, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and intact parathyroid hormone. Imaging data includes coronary CT raw images.

[0066] Among them, age: the patient's actual age or the age corresponding to the date of birth. Gender: male or female, used to distinguish the differences in disease progression and medication between patients of different genders. Chronic kidney disease (CKD) staging: CKD is divided into stages 1 to 5 based on indicators such as the patient's glomerular filtration rate (GFR) or urine protein quantification, which is used to assess the degree of renal function damage. Medical history data: including past history information such as diabetes, hypertension, cardiovascular disease, and other disease histories that may affect kidney and cardiovascular health. Medication data: records information such as the type, dosage, and course of treatment of drugs that the patient is currently using or has used, such as antihypertensive drugs, hypoglycemic drugs, lipid-lowering drugs, etc.

[0067] Blood Magnesium: Serum magnesium concentration (Serum Magnesium). Abnormal magnesium metabolism is closely related to the risk of vascular calcification in CKD patients. Blood Phosphorus: Serum phosphorus concentration (Serum Phosphorus). Hyperphosphatemia is common in the middle and late stages of CKD and is closely related to vascular calcification. Low-density lipoprotein cholesterol (LDL): Also known as "bad" cholesterol, elevated levels can easily lead to atherosclerosis. High-density lipoprotein cholesterol (HDL): Also known as "good" cholesterol, high levels generally have a cardiovascular protective effect. Intact parathyroid hormone (iPTH): Secreted by the parathyroid glands, it regulates calcium and phosphorus balance in the body. Abnormally elevated iPTH is often seen in CKD patients, leading to bone and vascular lesions.

[0068] It is important to note that age, CKD stage, pulse pressure, diabetes, low serum magnesium, low cholesterol (LDL), and left atrial diameter (LA) are independent risk factors for CAC in CKD patients; however, high serum magnesium and high HDL cholesterol are protective factors for CAC. Serum magnesium levels are negatively correlated with varying degrees of coronary artery calcification (CAC). Serum magnesium levels have a predictive value for severe coronary artery calcification, severe thoracic aortic calcification, and heart valve calcification. Increasing serum magnesium levels is an important approach for preventing and treating VC and a potential option for improving the prognosis of dialysis patients.

[0069] Data preprocessing includes data cleaning, format conversion, standardization, and feature selection and fusion.

[0070] Data cleaning requires processing missing values ​​and outliers, removing or correcting unreasonable data, and ensuring data quality. Format conversion converts data from different systems or file formats into a unified data format. Standardization normalizes or standardizes numerical features to eliminate the impact of different dimensions and enable the model to better adapt to the data. For feature selection and fusion, based on research needs, the most meaningful clinical, laboratory, and imaging-derived features for predicting coronary artery calcification risk are screened and data fused. Ultimately, the preprocessed clinical data, laboratory data, and imaging data (or their extracted features) are merged and stored in a unified data structure to form a standardized dataset that can be used for subsequent modeling and analysis.

[0071] Evaluation and prediction unit 2 is used to predict the risk probability of coronary artery calcification under different combinations of physiological characteristics and support risk stratification.

[0072] In the description of the present invention, the evaluation and prediction unit 2 includes: a feature engineering module (not shown in the figure), a model building module (not shown in the figure), a dynamic update module (not shown in the figure) and a calculation layer module (not shown in the figure); and the feature engineering module, the model building module, the dynamic update module and the calculation layer module are kept connected in sequence.

[0073] The feature engineering module is used to extract the corresponding static features, dynamic features and image-derived features of various types of data in the standardized data and merge them into physiological features.

[0074] Specifically, static features refer to features that remain largely unchanged throughout the observation period, or are updated only rarely. Examples include the patient's gender, age at initial diagnosis, chronic kidney disease stage, and medical history. Static features typically reflect the patient's baseline condition and past health status, providing baseline information for subsequent risk prediction models.

[0075] Dynamic features represent characteristics that change over time and require recording multiple time points or test values. Examples include blood magnesium, blood phosphorus, LDL, HDL, and iPTH test values ​​at different time periods, as well as patients' medication records and changes in their condition. Dynamic features can capture trends and fluctuations in patients' physiological indicators, providing a basis for identifying disease progression or intervention effectiveness.

[0076] Imaging-derived features (IDFs) are features extracted from medical imaging data, such as coronary CT, through algorithms or manual interpretation. Examples include the Agatston score, vessel wall thickness, calcified plaque volume and distribution, and 3D-reconstructed vascular morphology. IDFs can intuitively quantify the extent and spatial distribution of lesions and, when combined with other indicators, provide a more accurate assessment of cardiovascular risk.

[0077] Static, dynamic, and image-derived features are matched and integrated within the same patient dimension to form a unified data structure. These features are then normalized, encoded, or time-series processed as necessary to remove redundancy and noise, retaining the key information that best reflects the patient's condition and risk. Ultimately, a complete and usable set of physiological features is obtained, providing high-quality input for subsequent risk prediction models or other analytical modules, thereby improving model accuracy and stability.

[0078] In order to meet the needs of subsequent risk probability prediction model construction and training, the physiological feature set needs to be divided into a test set, a training set, and a validation set, which are used for model construction operations at different progress levels.

[0079] A model building module is used to construct a risk probability prediction model for coronary artery calcification based on a random forest fused with dynamic integrated weighted probability.

[0080] In the description of the present invention, the risk probability prediction model of coronary artery calcification based on the random forest fusion of dynamic integrated weighted probability includes:

[0081] Step S21: Generate several decision trees by resampling to construct the original random forest model.

[0082] Specifically, the present invention utilizes the Bootstrap resampling idea and random sub-feature strategy to train multiple decision trees, thereby forming a random forest model with good generalization ability.

[0083] First, prepare the original training dataset. Assume this dataset contains several samples, each with a certain number of features and a corresponding target label. The basic idea of ​​a random forest is to generate several different sub-datasets by sampling the training data multiple times with replacement, and then train multiple decision trees on each of these sub-datasets.

[0084] Next, resampling begins. During each resampling, a new subset of the training data is created by extracting, with replacement, a number of samples equal to the original dataset. Because of the use of replacement, some samples may appear multiple times in the same subset, while others may not be extracted. By repeating this process, multiple different subsets of the dataset can be obtained.

[0085] Secondly, a separate decision tree is trained for each subset of the dataset. During training, in addition to using the resampled subsets, the decision tree also randomly selects features for evaluation at each node split. This aims to further increase the diversity between decision trees during training, preventing them from becoming too similar and thus reducing the model's over-reliance on certain features or noise.

[0086] Ultimately, after the above steps, several trained decision trees are obtained. This collection of trees is called the original random forest model. The essence of a random forest is to integrate these relatively independent decision trees. Each tree makes its own prediction for a new sample, and the predictions of all trees are ultimately combined. For classification tasks, voting is generally used to determine the final category; for regression tasks, the final prediction value is usually obtained by averaging.

[0087] Step S22: Based on the dynamic ensemble selection of the k-nearest neighbor-hierarchical clustering algorithm (KNN-HC algorithm), the decision tree with the best performance is selected to construct a risk probability prediction model.

[0088] In the description of the present invention, based on the dynamic ensemble selection of the k-nearest neighbor-hierarchical clustering algorithm, the optimal performance decision tree is selected and the risk probability prediction model is constructed, which includes:

[0089] Step S221: Acquire a test sample according to the data type of the physiological feature, calculate the Euclidean distance between the test sample and the training set, and select the first k nearest neighbors in ascending order to form a neighborhood sample.

[0090] Specifically, the physiological characteristics of the test sample, such as blood magnesium, blood phosphorus, age, and CKD stage, are read. For each sample in the training set, the same physiological characteristics are extracted and the distance between them and the test sample is calculated. This distance is the Euclidean distance, which is considered a measure of similarity.

[0091] The calculated distances are sorted from smallest to largest, and the first k training samples closest to the test sample are selected to form a set of "neighborhood samples." Neighborhood samples can better represent local information with similar characteristics to the test sample, providing targeted data support for subsequent decision tree difference measurement.

[0092] Step S222: Calculate the inconsistency measure between any two decision trees in the neighborhood samples, construct a difference matrix, obtain the hierarchical clustering result of the decision trees, and form a cluster set.

[0093] Specifically, the inconsistency metric measures the difference in predictions between two decision trees on the same set of samples. When two trees produce consistent outputs on most samples, the inconsistency is low; when two trees frequently produce different predictions, the inconsistency is high. By summarizing the inconsistency between each pair of trees, a dissimilarity matrix is ​​generated. Hierarchical clustering algorithms are then used to group trees based on the degree of dissimilarity.

[0094] For each neighborhood sample, record the predictions of each decision tree for those samples. For any two decision trees, count whether their predictions on the neighborhood samples are consistent. If the two trees predict the same thing on most samples, the inconsistency is low; if they often predict opposite things, the inconsistency is high. The inconsistency results between all trees are entered into a dissimilarity matrix, which reflects the degree of difference between the trees. A hierarchical clustering algorithm is used to perform cluster analysis on this dissimilarity matrix, grouping trees with small differences into the same cluster and trees with large differences into different clusters, thus forming a cluster set.

[0095] Step S223: Calculate the classification accuracy of each decision tree in the random forest on the cluster set, select the decision tree with the highest accuracy from each cluster as the optimal decision tree, and output the dynamically selected optimal subtree set.

[0096] Specifically, for each cluster, the classification accuracy or other performance metrics of each tree in the cluster are calculated on neighboring samples. The tree with the highest accuracy is selected from that cluster and becomes the representative tree for that cluster, known as the optimal decision tree. The optimal decision trees from all clusters are combined to form a set of optimal subtrees. This set better matches the test sample and maintains the diversity of the random forest.

[0097] Step S224: Integrate the optimal subtree set into the original random forest model to generate a risk probability prediction model for coronary artery calcification.

[0098] Specifically, the original random forest structure is retained, but the optimal subtree set is prioritized for weighted voting or probabilistic fusion during prediction. When predicting a new sample, if a similar distribution of neighboring samples is found to the test sample, the dynamic ensemble selection mechanism is triggered again to continuously update or replace the optimal subtree set. The final prediction result can be the voting result of the optimal subtree set, weighted fusion based on the original forest, to generate the risk probability of coronary artery calcification.

[0099] Step S23: Use the weighted probability fusion rule to obtain the prediction result of the risk probability.

[0100] In the description of the present invention, the weighted probability fusion rule is adopted to obtain the prediction result of risk probability including:

[0101] Step S231: Evaluate the performance of each decision tree in the optimal subtree set using a test set to obtain a prediction result for each decision tree, and record the results to form a prediction result matrix for each decision tree.

[0102] Specifically, the test set is fed into each decision tree in the optimal subtree set, and each decision tree is asked to make predictions. For each test sample, a prediction output from a different decision tree is obtained. The prediction results for all test samples from each tree are recorded to form a decision tree prediction matrix. The rows of this matrix can represent test samples, and the columns can represent the prediction outputs of each decision tree, or vice versa. While recording, metrics such as prediction accuracy, sensitivity, and specificity can be calculated for each decision tree, helping to assign more appropriate weights during subsequent weighting.

[0103] Step S232: Based on the Bayesian theory, the probability of the decision tree prediction result is calculated, and the total probability of the true category is calculated by weighted probability random forest to obtain the final coronary artery calcification risk probability prediction result of the test sample.

[0104] Specifically, the weight of each decision tree in the ensemble is set based on its accuracy, sensitivity, or other indicators on the test set. Generally, the better the performance of the tree, the higher the weight.

[0105] For each test sample, the predictions from all decision trees in the optimal subtree set are collected. For classification problems (for example, determining whether there is coronary artery calcification), the prediction from each tree can be considered the probability that the sample belongs to a certain category.

[0106] The outputs of each decision tree are weighted and synthesized using Bayesian theory or other probability weighting methods. For example, the conditional probability of each tree for a certain category can be calculated, multiplied by its weight, and finally summed to obtain the total probability that the sample belongs to that category.

[0107] This total probability is considered the final risk probability prediction value for coronary artery calcification. If further classification is required, a threshold can be set to classify samples above the threshold as high-risk calcification patients, and otherwise as low-risk. This ultimately completes the weighted probability fusion prediction of the test samples, resulting in a more accurate and robust coronary artery calcification risk assessment result.

[0108] In the description of the present invention, the calculation formula for calculating the probability of the decision tree prediction result is:

[0109] ;

[0110] Where, Represents a test sample x test Select m The probability of the true category of a decision tree; F i represents the true category; x test represents the test sample; m represents a decision tree, and m =1, 2, 3, ..., M ; N Indicates the total number of categories in the test set; Represents a test sample x test Select m The conditional probability of a decision tree;

[0111] The total probability of the true category is calculated as:

[0112] ;

[0113] Where, Represents the true category F i The total probability of W Represents weighted weight.

[0114] The dynamic update module is used to set the conditions for updating new data. When the conditions are met, it triggers incremental model training, updates model parameters, and retains a preset proportion of historical data as a replay buffer.

[0115] Specifically, after the model is deployed, the system continuously monitors the quantity and quality of new data (such as new clinical indicators, laboratory results, and imaging information). When the amount of new data reaches a certain threshold, or the model's predictive performance (such as accuracy or AUC) on the new data drops beyond a pre-set tolerance, the update condition is met.

[0116] Once an update condition is triggered, the system executes an incremental training process instead of retraining the entire model from scratch. This incremental training allows the model to quickly learn the distribution and feature changes of the new data, significantly reducing training time and computing power consumption.

[0117] During incremental training, a certain percentage of historical data must be retained to prevent the model from focusing solely on the latest data and forgetting previously learned knowledge. This historical data, called the "replay buffer," can be used alongside new data during incremental training to train the model, allowing the model to adapt to the new distribution while retaining its memory of past data. After incremental training is complete, the replay buffer is maintained and updated to ensure that the proportion of historical data within it meets pre-defined standards and to eliminate the oldest or unrepresentative data.

[0118] The calculation stratification module is used to input real-time standardized data into the risk probability prediction model, output the risk probability of coronary artery calcification, and divide the corresponding risk levels according to preset thresholds.

[0119] Specifically, in the system, all raw data (clinical, laboratory, imaging, etc.) is preprocessed and standardized to ensure consistent model input dimensions and formats. Once new patient data or updated data for existing patients enters the system, it is passed to this module.

[0120] Risk probability prediction models (such as random forests and deep learning) receive standardized data and output a coronary artery calcification risk score between 0 and 1. A higher score indicates a higher risk. For clinical convenience, the system sets several thresholds (such as 0.3 and 0.7) to categorize the output risk score into different risk levels (e.g., low risk, medium risk, and high risk). This stratification helps physicians quickly determine the next steps for intervention, such as whether further testing, adjustments to treatment plans, or increased monitoring frequency are necessary.

[0121] Analysis and display unit 3 is used to generate transparent explanations and assist decision making through interactive visualization tools.

[0122] In the description of the present invention, the analysis and display unit 3 includes: a curve analysis module (not shown in the figure), a feature analysis module (not shown in the figure), an integrated selection module (not shown in the figure) and a graphic display module (not shown in the figure); and the curve analysis module, the feature analysis module, the integrated selection module and the graphic display module blocks are connected in sequence.

[0123] The curve analysis module is used to evaluate the predictive effect of the risk probability prediction model by drawing the receiver operating characteristic curve and select the optimal threshold to set the risk stratification.

[0124] In the description of the present invention, by drawing the receiver operating characteristic curve, evaluating the prediction effect of the risk probability prediction model, and selecting the optimal threshold to set the risk stratification include:

[0125] Step S301: Obtain the prediction results of the risk probability prediction model for the test sample, set different thresholds to convert the probability values ​​of the prediction results into prediction labels.

[0126] Specifically, suppose there is a test set containing 100 test samples. The model outputs a coronary artery calcification risk probability for each sample, such as 0.05, 0.27, 0.64, 0.90, etc. Multiple thresholds can be set, ranging from zero to one in steps (for example, 0.1, 0.2, 0.3, ..., 0.9), or finer intervals can be selected based on actual needs.

[0127] For each threshold, samples with a predicted probability above the threshold are classified as positive, and those with a predicted probability above the threshold are classified as negative. For example, for a threshold of 0.3, if a sample has a predicted probability of 0.27, it is labeled negative; if it has a predicted probability of 0.64, it is labeled positive. These results are recorded at different thresholds and used to calculate the true positive rate and false positive rate.

[0128] Step S302: Calculate the true positive rate and false positive rate for each threshold, plot the false positive rates and true positive rates calculated under multiple thresholds into a curve to obtain a receiver operating characteristic curve, and calculate the area under the curve of the receiver operating characteristic curve.

[0129] Specifically, the receiver operating characteristic curve demonstrates a model's discriminatory ability by comparing its true positive rate and false positive rate at different thresholds. The true positive rate refers to the percentage of samples that are correctly predicted as positive out of actual positive samples; the false positive rate refers to the percentage of samples that are incorrectly predicted as positive out of actual negative samples. Using the false positive rate as the horizontal axis and the true positive rate as the vertical axis, a curve is formed. The larger the area under the curve, the stronger the model's discriminatory ability.

[0130] For each threshold, the predicted label output by the model is compared with the actual label to calculate the true positive rate and false positive rate. For example, when the threshold is 0.3, the proportion of positive samples correctly identified (true positive rate) and the proportion of negative samples mistakenly classified as positive (false positive rate) at that threshold are counted.

[0131] The false positive rate is plotted on the X-axis and the true positive rate on the Y-axis. The curves are connected to form curves at multiple thresholds. The curves typically start at the lower left corner (0, 0) and end near the upper right corner (1, 1). Different models have different curve shapes.

[0132] By numerically integrating the curve or using specialized tools, the area under the curve can be calculated. If the area is close to one, the model has high predictive discrimination ability; if the area is close to 0.5, the model performance is similar to random guessing.

[0133] Step S303: Calculate the Youden index of the receiver operating characteristic curve, select the optimal threshold of risk probability according to the area under the curve and the Youden index, and divide the risk levels into multiple levels according to the optimal threshold.

[0134] Specifically, the Youden Index is an indicator that comprehensively considers the true positive rate and true negative rate. By finding the threshold that maximizes this index, we can determine an optimal decision point while maintaining high sensitivity and specificity. This threshold can then be used to stratify samples into multiple levels of risk.

[0135] The Youden Index is the sum of the true positive rate and the true negative rate minus one at a certain threshold. The true negative rate is the proportion of samples that are actually negative that are correctly predicted to be negative. The Youden Index can be calculated at each threshold, and the threshold corresponding to its maximum value can be found. For example, the Youden Index at a threshold of 0.3 might be 0.5, at a threshold of 0.4 it might be 0.6, at a threshold of 0.5 it might be 0.55, and so on. Ultimately, the Youden Index is found to be highest at a threshold of 0.4.

[0136] A large area under the curve indicates good overall model performance. Selecting the threshold that maximizes the Youden Index based on this area can achieve a good balance. For example, if the area under the curve is 0.85, and a threshold of 0.4 is found to have the highest Youden Index, then 0.4 is set as the optimal threshold. Using this optimal threshold as a benchmark, you can further set multiple threshold levels for stratification.

[0137] For example, samples with a risk probability below 0.3 are classified as low risk, those between 0.3 and 0.7 as medium risk, and those above 0.7 as high risk. This multi-level classification can help clinicians take differentiated intervention measures based on different risk levels.

[0138] The feature analysis module is used to analyze the impact of various physiological characteristics on risk probability prediction and display the contribution of physiological characteristics in combination with the receiver operating characteristic curve.

[0139] In the description of the present invention, the impact of various physiological characteristics on risk probability prediction is analyzed, and the contribution of physiological characteristics is displayed in combination with the receiver operating characteristic curve, including:

[0140] Step S311: Calculate the contribution of each physiological feature based on the SHAP value, and generate a bar chart by arranging them in descending order.

[0141] Specifically, the SHAP value (Shapley Additive Explanations) is a game theory-based method that calculates a contribution for each feature, indicating its role in a given prediction. By considering all possible feature combinations, the SHAP value quantifies the positive or negative contribution of a single feature to the model's output. By calculating each feature's SHAP value, we can determine its contribution to the model's predictions. Features with high contribution values ​​have a greater impact on model predictions, while features with low contribution values ​​have a smaller impact.

[0142] The test set is input into the trained risk probability prediction model, and the SHAP algorithm is used to calculate the contribution of each feature in each prediction sample. For example, if the SHAP value of feature A is +0.1 and the SHAP value of feature B is -0.05, this means that feature A has a positive contribution to the final prediction result, while feature B has a negative contribution.

[0143] Summarize the SHAP values ​​of all features and sort them by contribution from highest to lowest. Use a bar chart to display the contribution of each feature, such as blood magnesium, blood phosphorus, and age. The length of each feature's bar indicates the degree of influence on the prediction result, with longer bars indicating a greater contribution to the model output.

[0144] Step S312: For a single test sample, extract the decision path of each decision tree in the risk probability prediction model, count the occurrence frequency of feature splitting nodes, and generate a Sankey diagram to display the decision logic from input features to risk probability.

[0145] Specifically, each decision tree splits along a path within the tree based on the feature values ​​of the input sample until it reaches a leaf node. Therefore, the decision path of each tree can be extracted to understand how it makes classifications based on the features. A Sankey diagram is a visualization tool used to show flows and relationships, intuitively illustrating how different input features, through the decision paths of each tree, ultimately influence the prediction result.

[0146] For a single test sample, we input it into each decision tree in the random forest, extracting the splitting node information (e.g., features like blood magnesium level, blood phosphorus concentration, and age) for each tree layer by layer. Each decision path shows the splitting process from the root node to the leaf node, revealing the role each feature plays in the final decision.

[0147] The splitting nodes in all decision trees are summarized, and the frequency of each feature appearing in different trees is counted. For example, blood magnesium levels appear more frequently as splitting nodes in some trees, while other features participate less frequently. The counted splitting nodes are combined with the prediction results to generate a Sankey diagram. This diagram visually illustrates how features flow through different decision tree paths and influence the final risk probability. The width of the Sankey diagram represents the flow rate, or frequency, showing how different features contribute to the final risk prediction.

[0148] Step S313: After removing individual physiological features in turn, retrain the risk probability prediction model, observe the decrease in the area under the curve, and quantify the criticality of other physiological features.

[0149] Specifically, we remove individual physiological features one by one and retrain the model to observe changes in model performance. If model performance significantly decreases after removing a feature, it indicates that the feature is critical to the model's predictions. The change in area under the curve (AUC) quantifies the contribution of each feature to the model's predictive ability by calculating the change in the model's AUC value after each feature removal.

[0150] During each training run, remove one feature (e.g., blood magnesium, blood phosphorus, or age) from the original model and retrain the risk prediction model. Calculate the performance of the new model on the test set after removing the feature, paying particular attention to the change in the Area Under the Curve (AUC). Compare the AUC values ​​before and after removing the feature. If the AUC value decreases significantly, it indicates that the feature makes a significant contribution to the model's predictions. Conversely, if the change in the AUC value is small, the feature has little impact on the prediction results.

[0151] By comparing the changes in AUC across all features, we can quantify the criticality of each physiological feature. If removing a feature results in the largest decrease in AUC, that feature is most important to the model's predictions. For example, in a set of test samples, removing the blood magnesium feature decreased the AUC from 0.85 to 0.75, while removing the age feature only decreased the AUC from 0.85 to 0.82. This indicates that blood magnesium contributes more to the model than age, making it a more critical feature.

[0152] The ensemble selection module is used to introduce dynamic ensemble selection into the risk probability prediction model, visualize the optimization process of the test samples, and verify the prediction results.

[0153] In the description of the present invention, dynamic integrated selection is introduced into the risk probability prediction model, the optimization process of the test samples is visualized, and the verification of the prediction results includes:

[0154] Step S321: Select any test sample, mark k nearest neighbor samples in the two-dimensional t-SNE dimensionality reduction graph, distinguish the calcification group from the non-calcification group, and reflect the local data distribution characteristics.

[0155] Specifically, t-Distributed Stochastic Neighbor Embedding (t-SNE) is a nonlinear dimensionality reduction technique that maps high-dimensional data to a low-dimensional space while preserving the relative distances and local structure between data points. It clusters similar samples together in the low-dimensional space, thereby reflecting the local characteristics of the data.

[0156] A test sample is randomly selected from the dataset. Assume that this sample's features include data such as blood magnesium, blood phosphorus, age, and CKD stage, representing a patient's health status. Use the k-nearest neighbor algorithm to calculate the similarity between the test sample and all other samples in the training set (using methods such as Euclidean distance). The k closest samples are selected, which are close to the test sample in feature space. The high-dimensional feature data (e.g., blood magnesium, blood phosphorus, age, etc.) of the test sample and its k neighboring samples are then fed into the t-SNE dimensionality reduction algorithm to obtain a two-dimensional representation.

[0157] This allows us to visually observe the distribution of samples on a two-dimensional plane, further determining the position of the test sample in the data space. Based on the model's predicted risk, the test sample and its k neighboring samples are color-coded as either "calcified" (high risk) or "non-calcified" (low risk). This clearly demonstrates the similarity between the test sample and its neighboring samples, as well as whether they belong to the same group, reflecting the distribution characteristics of the local data.

[0158] Step S322: Use the dynamic merging process animation of the cohesive layer clustering to display the neighborhood sample clustering process, mark the calcification ratio of the samples in the cluster, and list the best performance decision tree in each cluster in a table.

[0159] Specifically, agglomerative hierarchical clustering is a bottom-up clustering method that starts with each data point and gradually merges the most similar samples into clusters. During each merge, the clustering algorithm selects the cluster with the smallest distance based on similarity calculations (such as Euclidean distance and Manhattan distance). This ultimately generates a cluster tree (dendrogram). Visualization techniques dynamically display the progress of each cluster merging process, helping us intuitively understand how samples are gradually clustered.

[0160] The calcification ratio of samples within the cluster. In each cluster, the ratio of samples in the "calcification group" (high risk) to the "non-calcification group" (low risk) is counted to help understand the overall risk characteristics of the cluster.

[0161] The test sample selected in step S321 and its k nearest neighbor samples are clustered using an agglomerative hierarchical clustering algorithm. Each time clustering occurs, the algorithm merges clusters based on the similarity between samples (such as Euclidean distance) and dynamically displays the merging process.

[0162] Use dynamic visualization tools (such as Plotly or Matplotlib in Python) to animate the clustering process, reflecting the changes at each merging step. As each cluster merge occurs, you can see that different samples gradually come together, eventually forming several clusters.

[0163] For each cluster, the proportion of samples in the "calcification group" (high risk) to the "non-calcification group" (low risk) is calculated. This allows us to understand the risk characteristics of each cluster. For example, if 90% of the samples in a cluster are high-risk calcification patients, the cluster can be considered high-risk.

[0164] Graphics display module, used to create heat maps, contour maps and three-dimensional coronary artery models, and annotate the location and severity of calcifications.

[0165] Specifically, a heatmap is a graphical representation of data values ​​through color intensity. It's often used to display two-dimensional matrix data, where the color depth reflects the magnitude of the values. Heatmaps are effective in demonstrating patterns and relationships in data, and are suitable for displaying correlations between features, the distribution of calcification areas, and other statistical analysis results.

[0166] A contour map is a graphical representation of three-dimensional data projected onto a two-dimensional plane. Contour lines are used to show the range of variation of a variable, often used to represent areas of variation such as height or density.

[0167] The 3D coronary artery model is constructed based on 3D imaging data (such as CT scan data) and provides a spatial understanding of the coronary artery structure. This display helps doctors observe calcification areas and their severity from multiple angles, allowing them to make more accurate assessments.

[0168] Using coronary CT image data, a three-dimensional structural model of the coronary arteries is obtained through image segmentation and three-dimensional reconstruction. The degree of calcification in each vascular region can be annotated using the calcification integral. The coronary arteries are reconstructed from CT images into a three-dimensional model using professional three-dimensional modeling software (such as 3DSlicer, ITK-SNAP, etc.). Combined with the calcification integral, the severity of calcification can be annotated for each vascular segment. The severity of calcification is displayed on the model through color marking or three-dimensional visualization, with heavily calcified areas represented by dark colors and mildly calcified areas by light colors. The generated three-dimensional coronary model is interactively operated by the physician, who can rotate, zoom in, or zoom out the three-dimensional model to more clearly see the calcification status of each vascular segment.

[0169] The decision optimization unit 4 is used to generate intervention recommendations based on the prediction results and optimize the blood magnesium management strategy.

[0170] In the description of the present invention, the decision optimization unit 4 includes: a blood magnesium management optimization module (not shown in the figure), a treatment response tracking module (not shown in the figure) and an intervention effect prediction and optimization module (not shown in the figure).

[0171] The blood magnesium management optimization module is used to optimize the blood magnesium management strategy based on the patient's risk prediction results and health status, including adjusting the target range and monitoring frequency of blood magnesium levels and selecting appropriate intervention methods (such as medication, dietary adjustments, etc.) to achieve the purpose of preventing and controlling coronary artery calcification.

[0172] Specifically, a personalized target range for blood magnesium levels is set for each patient based on risk assessment results. Low-risk patients may maintain a normal range (e.g., 0.75-1.05 mmol / L), while high-risk patients may require a slightly higher or more precise target (e.g., 1.0-1.2 mmol / L), adjusted based on the patient's health status and medical history.

[0173] For high-risk patients, a magnesium supplementation or adjustment program is developed, such as the use of magnesium supplements or dietary adjustments (increasing magnesium-containing foods) to increase blood magnesium levels. If the patient has hypermagnesemia (such as dialysis patients), the blood magnesium level may need to be lowered to reduce the risk of calcification. The monitoring frequency is personalized based on the patient's risk stratification and assessment of treatment response. For example, for high-risk patients, blood magnesium levels may need to be monitored monthly or weekly, while low-risk patients can be checked quarterly.

[0174] The treatment response tracking module is used to monitor the patient's treatment response after implementing the blood magnesium management strategy and adjust the intervention plan. If the blood magnesium level does not reach the expected target or the risk profile changes, the intervention strategy is updated based on real-time data.

[0175] Specifically, the effectiveness of the magnesium management strategy should be assessed by regularly monitoring blood magnesium levels and other relevant physiological indicators (such as serum phosphorus and iPTH) to determine whether the ideal blood magnesium range has been achieved. Treatment plans should be adjusted promptly based on changes in the patient's blood magnesium level and other risk indicators (such as the degree of vascular calcification and coronary imaging results). If blood magnesium levels do not reach the target range, adjustments to medication dosage, dietary regimen, or increased intervention intensity may be necessary.

[0176] Considering the different responses of individual patients to treatment, such as age, renal function, comorbidities and other factors, these factors may lead to different responses of patients to blood magnesium intervention, so personalized adjustments to the intervention strategy are needed.

[0177] The intervention effect prediction and optimization module is used to predict the impact of long-term intervention on coronary calcification risk based on the results of optimized blood magnesium management. It then optimizes the intervention plan using a dynamic model. This predictive model further improves the accuracy of the intervention plan, ensuring that blood magnesium management produces the best results in preventing coronary calcification and improving patients' cardiovascular health.

[0178] Specifically, the study simulated different magnesium management strategies and risk probability changes to predict the impact of long-term intervention on patients' coronary artery calcification risk. Based on historical data and clinical trial results, the study evaluated the effectiveness of different intervention strategies in controlling the calcification process and predicted risk trends at different magnesium levels.

[0179] Combined with the patient's long-term monitoring data, the blood magnesium management strategy can be dynamically adjusted. If the patient's coronary artery calcification risk does not decrease after a period of intervention, the intervention plan may need to be reassessed and strengthened. For example, new medications may be introduced, magnesium intake may be increased, or lifestyle interventions may be strengthened.

[0180] Based on the predicted results and clinical experience, personalized treatment recommendations are generated and recommended to physicians for adjusting blood magnesium management. For patients who do not respond well to treatment, other non-drug treatments (such as electrotherapy and lifestyle interventions) may need to be considered.

[0181] In summary, with the help of the above-mentioned technical solutions of the present invention, the prediction accuracy and clinical practicality are significantly improved through the improved random forest model with dynamic integrated selection and weighted probability fusion. The KNN-HC algorithm is introduced to dynamically select the optimal decision tree, and the model is adjusted according to the local data distribution of different patients to adapt to the heterogeneity of CKD patients. Through the dynamic threshold recommendation of the ROC curve, SHAP value-driven feature contribution analysis and risk heat map, the complex model is converted into actionable intervention recommendations for doctors (such as magnesium supplementation target value). By integrating biochemical indicators, image calcification scores and dynamic time series features, the limitation of traditional models relying on a single data source is broken through. Through algorithm innovation and visual closed-loop design, the whole chain optimization from data to decision-making is achieved, providing a precise, transparent and dynamic intelligent tool for the prevention and treatment of CAC in CKD patients.

[0182] This invention uses the KNN-HC algorithm to hierarchically cluster the neighborhood of each test sample and dynamically select the optimal decision tree within each cluster, enabling the model to adaptively optimize within the local data distribution. For example, for high-risk patients with low blood magnesium (<0.75 mmol / L) and CKD stages 4-5, subtrees with an accuracy rate >90% within similar patient clusters are prioritized over trees with average global performance, thereby reducing the risk of missed diagnoses. Inter-cluster diversity is also used to control overfitting and ensure generalizability. SHAP values ​​are used to quantify feature contributions, and optimal thresholds are recommended in conjunction with ROC curves, directly guiding magnesium supplementation targets. Furthermore, risk heat maps demonstrate the synergistic effects of blood magnesium and LDL cholesterol, helping physicians develop combined intervention plans. Translating AI output into clinical action effectively increases model adoption.

[0183] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

Claims

1. A coronary artery calcification prediction and analysis system based on a receiver operating characteristic curve, characterized by: include: An integrated processing unit is used to collect clinical data, laboratory data, and imaging data, and generate standardized data after data preprocessing; An evaluation and prediction unit is used to predict the risk probability of coronary artery calcification under different combinations of physiological characteristics and support risk stratification; An analysis and presentation unit, used to generate transparent explanations and assist decision-making through interactive visualization tools; A decision optimization unit is used to generate intervention recommendations based on the prediction results and optimize the blood magnesium management strategy; Wherein, the evaluation and prediction unit includes: Feature engineering module, used to extract static features, dynamic features and image-derived features corresponding to various types of data in standardized data and merge them into physiological features; The model building module is used to build a risk probability prediction model for coronary artery calcification based on a random forest fused with dynamic integrated weighted probability, including: Generate several decision trees by resampling and build the original random forest model; Based on the dynamic ensemble selection of k-nearest neighbor-hierarchical clustering algorithm, the optimal performance decision tree is selected to build a risk probability prediction model; Use weighted probability fusion rules to obtain the prediction results of risk probability; The dynamic ensemble selection based on the k-nearest neighbor-hierarchical clustering algorithm, selecting the optimal performance decision tree, and constructing the risk probability prediction model include: Obtain test samples according to the data type of physiological characteristics, calculate the Euclidean distance between the test samples and the training set, and select the first k nearest neighbors in ascending order to form a neighborhood sample; Calculate the inconsistency measure between any two decision trees in the neighborhood samples, construct the difference matrix, obtain the hierarchical clustering results of the decision trees, and form a cluster set; Calculate the classification accuracy of each decision tree in the random forest on the cluster set, select the decision tree with the highest accuracy from each cluster as the best performing decision tree, and output the dynamically selected optimal subtree set; The optimal subtree set was integrated into the original random forest model to generate a risk probability prediction model for coronary artery calcification; The weighted probability fusion rule is used to obtain the prediction result of risk probability, including: The performance of each decision tree in the optimal subtree set is evaluated through the test set to obtain the prediction results of each decision tree, and the prediction results matrix of each decision tree is recorded; Based on Bayesian theory, the probability of the decision tree prediction results is calculated, and the total probability of the true category is calculated through weighted probability random forest to obtain the final coronary artery calcification risk probability prediction result of the test sample; The calculation formula for the probability of the decision tree prediction result is: ； Where, Represents a test sample x test Select m The probability of the true category of a decision tree; F i represents the true category; x test represents the test sample; m represents a decision tree, and m =1, 2, 3, ..., M ; N Indicates the total number of categories in the test set; Represents a test sample x test Select m The conditional probability of a decision tree; The total probability of the true category is calculated as follows: ； Where, Represents the true category F i The total probability of W Represents weighted weight.

2. The coronary artery calcification prediction and analysis system based on the receiver operating characteristic curve according to claim 1, characterized in that: The integration processing unit, the evaluation and prediction unit, the analysis and display unit, and the decision optimization unit are connected in sequence; The clinical data include age, gender, chronic kidney disease stage, medical history data and medication data; the laboratory data include blood magnesium, blood phosphorus, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol and intact parathyroid hormone; the imaging data include original coronary CT images.

3. The coronary artery calcification prediction and analysis system based on receiver operating characteristic curve according to claim 1, characterized in that: The evaluation and prediction unit also includes: a feature engineering module, a model building module, a dynamic update module and a calculation layer module; The dynamic update module is used to set the conditions for updating new data. When the conditions are met, the model incremental training is triggered, the model parameters are updated, and a preset proportion of historical data is retained as a replay buffer. The calculation and stratification module is used to input real-time standardized data into the risk probability prediction model, output the risk probability of coronary artery calcification, and divide the corresponding risk levels according to the preset threshold; And the feature engineering module, the model building module, the dynamic update module and the calculation layer module are connected in sequence.

4. The coronary artery calcification prediction and analysis system based on receiver operating characteristic curve according to claim 1, characterized in that: The analysis and display unit includes: a curve analysis module, a feature analysis module, an integrated selection module and a graphic display module; The curve analysis module is used to evaluate the prediction effect of the risk probability prediction model by drawing the receiver operating characteristic curve and select the optimal threshold to set the risk stratification; The feature analysis module is used to analyze the impact of various physiological characteristics on risk probability prediction and display the contribution of physiological characteristics in combination with the receiver operating characteristic curve; The ensemble selection module is used to introduce dynamic ensemble selection into the risk probability prediction model, visualize the optimization process of the test samples, and verify the prediction results; Graphics display module, used to create heat maps, contour maps and coronary artery 3D models, annotating calcification locations and severity; The curve analysis module, the feature analysis module, the integrated selection module and the graphic display module blocks are connected in sequence.

5. The coronary artery calcification prediction and analysis system based on receiver operating characteristic curve according to claim 4, characterized in that: The method of drawing a receiver operating characteristic curve to evaluate the prediction effect of the risk probability prediction model and selecting the optimal threshold to set the risk stratification includes: Obtain the prediction results of the risk probability prediction model for the test sample, set different thresholds to convert the probability values ​​of the prediction results into prediction labels; Calculate the true positive rate and false positive rate for each threshold, plot the false positive rates and true positive rates calculated under multiple thresholds into a curve to obtain the receiver operating characteristic curve, and calculate the area under the curve of the receiver operating characteristic curve; The Youden index of the receiver operating characteristic curve was calculated, and the optimal threshold of risk probability was selected based on the area under the curve and the Youden index. The patients were then divided into multiple risk levels based on the optimal threshold.

6. The coronary artery calcification prediction and analysis system based on receiver operating characteristic curve according to claim 5, characterized in that: The analysis of the impact of each physiological characteristic on risk probability prediction and the display of the contribution of the physiological characteristics in combination with the receiver operating characteristic curve include: The contribution of each physiological feature was calculated based on the SHAP value and a bar graph was generated by arranging them in descending order; For a single test sample, extract the decision paths of each decision tree in the risk probability prediction model, count the occurrence frequencies of feature splitting nodes, and generate a Sankey diagram to display the decision logic from input features to risk probability. After removing individual physiological features one by one, the risk probability prediction model was retrained, and the decrease in the area under the curve was observed to quantify the criticality of other physiological features.

7. The coronary artery calcification prediction and analysis system based on receiver operating characteristic curve according to claim 6, characterized in that: The introduction of dynamic integrated selection into the risk probability prediction model, the visualization of the optimization process of the test samples, and the verification of the prediction results include: Select any test sample and mark the k nearest neighbor samples in the two-dimensional t-SNE dimensionality reduction graph to distinguish the calcification group from the non-calcification group and reflect the local data distribution characteristics; The dynamic merging process animation of the cohesive layer clustering is used to demonstrate the clustering process of neighborhood samples, annotate the calcification ratio of samples within the cluster, and list the best-performing decision trees in each cluster in tabular form.

Images