A method for constructing a post-PCI patient all-cause mortality risk prediction model

Abstract

The application provides a method for constructing a PCI postoperative patient all-cause death risk prediction model, which comprises the following steps: obtaining a training data set containing clinical electronic medical records, blood cell count and blood biochemical data of patients undergoing selective PCI, and at least 5 years of postoperative follow-up and all-cause death label; performing Z-score standardization on continuous variables and one-hot encoding on category variables for pretreatment; screening 43 core features through LASSO regression combined with ten-fold cross-validation; constructing a model based on a LightGBM algorithm, taking a binary logarithmic loss function as a measurement standard and optimizing hyperparameters through grid search; finally outputting a prediction model which can receive core feature data of a new patient and output the postoperative all-cause death probability of the new patient; and the application can realize accurate prediction of long-term death risk of PCI postoperative patients.

Description

Technical Field

[0001] This application relates to the field of medical risk prediction technology, and in particular to a method, device, electronic device and storage medium for constructing a prediction model of all-cause mortality risk in patients after PCI. Background Technology

[0002] Coronary artery disease (CAD) is one of the leading causes of death from cardiovascular diseases worldwide. Percutaneous coronary intervention (PCI), as a core clinical treatment for CAD, can effectively improve myocardial blood supply and alleviate symptoms. However, the long-term prognosis of patients after PCI varies significantly among individuals, and some patients may die from all causes due to postoperative complications, progression of underlying diseases, or other reasons. Therefore, accurately identifying patients at high risk of all-cause mortality after PCI is crucial for optimizing clinical follow-up strategies, adjusting intervention plans in a timely manner, and reducing mortality.

[0003] Existing methods for predicting post-PCI risk have several significant drawbacks: First, they rely on single or limited clinical indicators (such as age and blood pressure) without comprehensively considering multiple factors such as the patient's hematological characteristics, biochemical status, and lifestyle habits, resulting in a one-sided prediction. Second, they employ traditional models such as logistic regression and simple decision trees, which struggle to capture complex nonlinear relationships in high-dimensional data, leading to limited prediction accuracy. Third, some methods have not undergone rigorous validation on independent datasets, exhibiting poor generalization ability and making them difficult to apply directly to clinical practice. Furthermore, the lack of standardized feature selection processes in current technologies makes them prone to overfitting due to redundant features, further reducing prediction reliability.

[0004] Therefore, there is an urgent need for a method to construct a prediction model for all-cause mortality risk after PCI that can integrate multi-dimensional data, use scientific feature screening and high-performance model construction, and undergo rigorous validation, in order to fill the gap in existing technology and meet the clinical need for accurate prediction. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method, apparatus, electronic device and storage medium for constructing a prediction model of all-cause mortality risk for patients after PCI, which can achieve accurate prediction of long-term mortality risk for patients after PCI.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for constructing a prediction model for all-cause mortality risk in patients after PCI, comprising the following steps: Obtain a training dataset containing clinical electronic medical record data, blood cell count data, and blood biochemistry data of patients admitted for elective PCI. All patients have follow-up records for at least 5 years after surgery and are labeled with whether all-cause mortality occurred. The training dataset is preprocessed, including Z-score standardization for continuous variables and one-hot encoding for categorical variables; The LASSO regression algorithm was used to select features from the preprocessed dataset. The optimal regularization penalty coefficient was determined by ten-fold cross-validation, and the core features with non-zero regression coefficients were selected as a combination of modeling markers. Based on the core features obtained through screening, a prediction model is constructed using the LightGBM algorithm, with the binary log loss function as the loss metric, and the model hyperparameters are optimized using the grid search method. The completed prediction model is output. The model can receive new core feature data of PCI patients and output the probability value of all-cause mortality after the operation.

[0007] Secondly, embodiments of this application also provide an apparatus for constructing a prediction model for all-cause mortality risk in patients after PCI, the apparatus comprising: The acquisition module is used to acquire the training dataset, which includes clinical electronic medical record data, blood cell count data and blood biochemistry data of patients admitted for elective PCI. All patients have follow-up records for at least 5 years after the operation and are labeled with whether all-cause mortality occurred. The preprocessing module is used to preprocess the training dataset, including Z-score standardization for continuous variables and one-hot encoding for categorical variables. The filtering module is used to filter features of the preprocessed dataset using the LASSO regression algorithm. It determines the optimal regularization penalty coefficient through ten-fold cross-validation and selects the core features with non-zero regression coefficients as a combination of modeling markers. The optimization module is used to build a prediction model based on the core features obtained from the screening, using the LightGBM algorithm, with the binary log loss function as the loss metric, and to optimize the model hyperparameters through a grid search method. The output module is used to output the completed prediction model. The model can receive new core feature data of PCI patients and output the probability value of all-cause mortality after the operation of the patient.

[0008] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the method for constructing a prediction model for all-cause mortality risk of PCI patients as described in any of the first aspects.

[0009] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the method for constructing a prediction model for all-cause mortality risk of PCI patients as described in any one of the first aspects.

[0010] The embodiments of this application have the following beneficial effects: By limiting the training data to include three core data categories—clinical electronic medical records, blood cell counts, and blood biochemistry—and requiring at least 5 years of follow-up and mortality labels, the system ensures multi-dimensional coverage and a strong correlation between admission characteristics and long-term prognosis from the source, avoiding the predictive bias caused by a single data dimension. Targeted Z-score standardization is used to eliminate magnitude differences in continuous variables, and one-hot encoding is used to convert categorical variables, effectively addressing the heterogeneity of the original medical data and laying a standardized data foundation for subsequent modeling. Combining LASSO regression and 10-fold cross-validation to screen core features not only eliminates redundant features to simplify the model structure but also retains key predictive information to reduce the risk of overfitting. The LightGBM algorithm, adapted to high-dimensional medical data, is selected and paired with a binary logarithmic loss function, significantly improving the model's predictive accuracy and training efficiency. The final output is the patient's postoperative all-cause mortality probability value rather than a simple classification result, providing a direct quantitative reference for clinical decision-making. This achieves precise empowerment across the entire chain from data preparation to clinical application, comprehensively outperforming existing fragmented and one-dimensional PCI postoperative risk prediction methods. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating steps S101-S105 provided in the embodiments of this application; Figure 2 The LASSO regression regularization penalty coefficient λ and the characteristic coefficients, as well as the cross-validation error diagram, are provided in the embodiments of this application. Figure 3 These are the ROC curves and AUC values ​​of the LightGBM prediction model provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the device for constructing a prediction model of all-cause mortality risk in patients after PCI provided in the embodiments of this application; Figure 5 This is a schematic diagram of the composition structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0014] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0015] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0016] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0017] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application and is not intended to limit this application.

[0019] See Figure 1, Figure 1 This is a flowchart illustrating steps S101-S105 of the method for constructing a prediction model for all-cause mortality risk in patients after PCI provided in this application embodiment. Figure 1 Steps S101-S105 are explained below.

[0020] In step S101, a training dataset is obtained. The training dataset includes clinical electronic medical record data, blood cell count data, and blood biochemistry data of patients admitted for elective PCI. All patients have follow-up records for at least 5 years after the operation and are labeled with whether all-cause mortality occurred.

[0021] Here, the first step is to obtain qualified and high-quality training data, which is a prerequisite for the accuracy of model predictions. The three types of data (clinical electronic medical records, blood cell counts, and blood biochemistry tests) directly reflect the patient's basic condition upon admission and are strongly correlated with long-term postoperative prognosis. Only long-term follow-up data can accurately establish the association between "admission characteristics" and "long-term mortality risk," avoiding prediction bias caused by short-term follow-up. Therefore, "at least 5 years of follow-up records (at least 5 years)" and "all-cause mortality labels" are selected. The "all-cause mortality labels" clarify the model's prediction target (binary classification: occurrence / absence of all-cause mortality), providing clear label guidance for subsequent modeling.

[0022] In step S102, the training dataset is preprocessed, including Z-score standardization for continuous variables and one-hot encoding for categorical variables.

[0023] The purpose of data preprocessing is to eliminate the "heterogeneity" and "unavailability" of the original data. In the original data, continuous variables (such as age, systolic blood pressure, blood glucose, etc.) vary greatly in magnitude (for example, age is measured in "years," while blood glucose is measured in "mmol / L," with a numerical range that differs by tens of times). If directly input into the model, the large-magnitude variables will dominate model training, causing small but important variables to be ignored. Z-score standardization transforms all continuous variables into standard normal distribution data with a mean of 0 and a standard deviation of 1 through a formula, fundamentally eliminating the magnitude difference and allowing the model to treat each feature fairly. Categorical variables (such as "whether or not one has diabetes" or "whether or not one drinks alcohol") are non-numerical information that the model cannot directly identify. One-hot encoding transforms these binary variables into two independent binary features (e.g., "if one has diabetes = 1, the corresponding feature value is 1; if one does not have diabetes, the feature value is 0; and vice versa), completely avoiding the model's misclassification of categorical variables as ordinal variables (e.g., misinterpreting "yes = 1, no = 0" as a quantitative relationship where "yes" is more risky than "no"), ensuring the accurate transmission of categorical information.

[0024] In step S103, the LASSO regression algorithm is used to screen features in the preprocessed dataset. The optimal regularization penalty coefficient is determined by ten-fold cross-validation, and the core features with non-zero regression coefficients are selected as the modeling marker combination.

[0025] Here, the LASSO regression algorithm is used for feature selection to address the problem of "high-dimensional data redundancy." The original three types of data contain dozens of features, some of which may be correlated (e.g., serum creatinine and creatinine clearance both reflect kidney function). Using all of these features for modeling would increase model complexity, the risk of overfitting, and reduce model interpretability. LASSO regression introduces an L1 regularization term, which can compress the regression coefficients of redundant features to 0, thus achieving feature "selection." The optimal regularization penalty coefficient λ is determined through 10-fold cross-validation because the value of λ directly affects the selection results. If λ is too small, redundant features cannot be effectively removed; if λ is too large, key features may be removed. 10-fold cross-validation divides the dataset into 10 parts, using 9 parts for training and 1 part for validation in turn, ultimately determining the λ value that optimizes the model's generalization ability. This ensures that the selected core features are neither redundant nor do they completely retain the key prediction information, forming the optimal combination of modeling markers.

[0026] In step S104, based on the core features obtained through screening, a prediction model is constructed using the LightGBM algorithm, with the binary logarithmic loss function as the loss metric, and the model hyperparameters are optimized using a grid search method.

[0027] Here, the LightGBM algorithm is chosen to build the model. Compared to traditional algorithms such as logistic regression and random forest, LightGBM has the core advantages of "efficiently handling high-dimensional data, fast training speed, and strong generalization ability," making it particularly suitable for high-dimensional data scenarios with core features in this solution. The binary logarithmic loss function is used as the loss metric because this problem belongs to the "binary classification problem" (death / not dead), and the binary logarithmic loss function can accurately measure the deviation between the model's predicted probability and the true label, making it the standard loss function for binary classification tasks. Optimizing hyperparameters through the grid search method is a key means to improve model performance. Hyperparameters are the model's "configuration parameters" (not automatically learned during model training), and their values ​​directly determine model performance. Grid search iterates through preset hyperparameter combinations to select the combination that optimizes model performance, ensuring that the model achieves the best prediction results.

[0028] In step S105, the completed prediction model is output. The model can receive new core feature data of PCI patients and output the probability value of all-cause mortality after the operation of the patient.

[0029] Here, the model input is the core feature data of new PCI patients, which is consistent with the core features of the training data to ensure the adaptability of the input data; the output is the probability value of all-cause mortality after surgery (rather than a simple "death / survival" classification), which can provide clinicians with a more accurate risk quantification reference, and facilitate the development of personalized follow-up and intervention plans based on the probability value (such as patients with a probability higher than 80% need to increase the frequency of follow-up).

[0030] In some embodiments, the clinical electronic medical record data includes: patient age, whether taking statins, cancer status, diabetes status, systolic blood pressure, diastolic blood pressure, waist-to-hip ratio, body fat percentage, Townsend index, drinking habits, and smoking habits; The blood cell count data include: eosinophil percentage, absolute lymphocyte count, lymphocyte percentage, mean corpuscular hemoglobin level, mean corpuscular volume, mean reticulocyte volume, absolute monocyte count, monocyte percentage, absolute neutrophil count, percentage of nucleated red blood cells, and red blood cell distribution width. The blood biochemical test data include: creatinine clearance, alanine aminotransferase, albumin, alkaline phosphatase, aspartate aminotransferase, C-reactive protein, calcium, serum creatinine, cystatin C, gamma-glutamyl transferase, glucose, high-density lipoprotein cholesterol, direct measurement of low-density lipoprotein cholesterol, phosphate, sex hormone-binding globulin, testosterone, total bilirubin, total protein, triglycerides, uric acid, and urea.

[0031] In this application embodiment, the specific indicators selected for the clinical electronic medical record data are all supported by clear clinical logic: "Patient age" is a prognostic risk factor (elderly patients have declining organ function and higher postoperative risks); "whether or not statin drugs are taken" is related to blood lipid control (statin drugs can lower blood lipids and reduce the risk of cardiovascular event recurrence); "cancer status" and "diabetes status" are important underlying diseases. Cancer directly affects the patient's survival time, and diabetes damages blood vessels and increases the risk of postoperative complications; "systolic blood pressure and diastolic blood pressure" reflect basic cardiovascular function. Hypertension is an important cause of coronary heart disease and also affects postoperative recovery; "waist-to-hip ratio and body fat percentage" reflect the degree of obesity (central obesity is a risk factor for cardiovascular disease); "Townsend index" is a core indicator for measuring socioeconomic status (socioeconomic status affects access to medical resources and lifestyle habits, thus affecting prognosis); "drinking habits and smoking habits" are clear cardiovascular disease risk factors that can aggravate vascular damage and reduce postoperative survival rate.

[0032] Blood cell count data focuses on the "state of blood cells," and the selection of indicators is directly related to the patient's immune function, hematopoietic function, and inflammation level. Inflammatory response and immune function are key influencing factors for post-PCI recovery and long-term survival: "Eosinophil percentage" is related to allergic reactions and inflammation; "Absolute lymphocyte count and lymphocyte percentage" reflect the body's immune function (patients with low immune function have a higher risk of postoperative infection and a worse prognosis); "Mean red blood cell hemoglobin level and mean corpuscular volume" reflect hematopoietic function (anemic patients have poor postoperative tolerance and a higher risk); "Mean reticulocyte volume" reflects erythropoietic activity; "Absolute monocyte count and monocyte percentage" participate in inflammatory response and immune regulation; "Absolute neutrophil count" is an important marker of inflammation (excessive postoperative inflammatory response can damage tissues); "Percentage of nucleated red blood cells and red blood cell distribution width" reflect abnormal erythropoiesis and are related to organ function damage.

[0033] Blood biochemistry data comprehensively covers core dimensions such as "organ function, inflammatory state, metabolic level, and nutritional status," providing multi-dimensional support for risk prediction: "creatinine clearance rate, serum creatinine, and cystatin C" reflect renal function (renal insufficiency is an important risk factor for poor prognosis after PCI); "alanine aminotransferase, alkaline phosphatase, aspartate aminotransferase, and gamma-glutamyl transferase" reflect liver function (the liver is the core metabolic organ, and abnormal liver function can affect drug metabolism and bodily recovery); "C-reactive protein" is a key marker of systemic inflammation (inflammation levels are positively correlated with the risk of recurrence of postoperative cardiovascular events); "calcium and phosphate" The following indicators are used: "Reflecting electrolyte balance (electrolyte imbalances can affect heart function); "Glucose, high-density lipoprotein cholesterol, directly measured low-density lipoprotein cholesterol, and triglycerides" reflect metabolic status (hyperlipidemia and hyperglycemia are core contributing factors to the progression of coronary heart disease); "Albumin and total protein" reflect nutritional status (malnourished patients have slower wound healing and a higher risk of infection after surgery); "Sex hormone-binding globulin and testosterone" (especially for male patients) are related to cardiovascular function (abnormal testosterone levels can affect vascular elasticity); "Total bilirubin" participates in metabolism, and its abnormal levels are related to liver function and oxidative stress; "Uric acid and urea" are supplementary indicators of kidney function and metabolic status. All indicators were selected based on extensive clinical research findings, ensuring that each indicator has a clear association with the risk of all-cause mortality after PCI.

[0034] In some embodiments, the Z-score standardization process is specifically implemented by making the processed data distribution have a mean of 0 and a standard deviation of 1. The specific method of one-hot encoding is to convert the binary categorical variable into two independent binary features.

[0035] For Z-score standardization, the processed data distribution has a mean of 0 and a standard deviation of 1. The specific mathematical formula is: z = (x - μ) / σ, where x is the original continuous variable value, μ is the mean of the variable in the entire training dataset, and σ is the standard deviation of the variable in the entire training dataset. Taking a specific indicator as an example, if the mean age of patients in the training set is 65 years and the standard deviation is 10 years, and a patient's age is 75 years, then the standardized value is (75 - 65) / 10 = 1; if a patient's age is 55 years, the standardized value is (55 - 65) / 10 = -1. The significance of this process is that it transforms indicators of different magnitudes and units (such as age: 20-80 years, systolic blood pressure: 90-180 mmHg, blood glucose: 3.9-11.1 mmol / L) into the same numerical range (most data are concentrated between -3 and 3), completely eliminating the impact of magnitude differences. Without standardization, the model may overemphasize the impact of systolic blood pressure because the numerical range of systolic blood pressure (90-180) is much larger than the numerical range of age (20-80), thus ignoring age as a key risk factor. After standardization, the weights of all features are more balanced, and the model is better able to learn the true risk associations.

[0036] One-hot encoding transforms binary categorical variables into two independent binary (0 or 1) features, specifically for "yes / no" binary variables (such as "whether or not one has diabetes," "whether or not one drinks alcohol," or "whether or not one takes statins"). Taking "whether or not one has diabetes" as an example, the original data might use "1" to represent "yes" and "0" to represent "no." If directly input into the model, the model might incorrectly interpret this as an "ordinal variable," meaning it considers the risk of "1" to be higher than the risk of "0" (although this logic is coincidental, it doesn't apply to all binary variables), and it fails to reflect the "mutual exclusion" between categories. After one-hot encoding, two new features are generated: "Feature 1: Has diabetes" (yes = 1, no = 0) and "Feature 2: Does not have diabetes" (yes = 1, no = 0). These two features are independent and mutually exclusive, allowing the model to clearly identify category information and avoid misjudging ordinal categories. This approach is a standard method for handling binary categorical variables in machine learning, and it is particularly suitable for scenarios with multiple binary clinical indicators in this application's embodiments, ensuring the accurate transmission of category information.

[0037] In some embodiments, the objective function of the LASSO regression algorithm is:

[0038] Where N is the sample size and K is the number of coefficients. This is the predicted value for this sample. This is the true value of the sample. This is the regularization penalty coefficient, used to adjust the coefficient. The severity of the punishment.

[0039] Here, in this embodiment of the application, 74 preprocessed indicators are used as independent variables, and whether the patient experienced all-cause mortality within 5 years after surgery (coded as 1 and 0) is used as the dependent variable. The Least Absolute Contraction and Selection Operator (LASSO) regression algorithm is applied for feature selection. LASSO regression, by introducing an L1 regularization term into its loss function, can compress the coefficients of some unimportant or redundant features to exactly 0.

[0040] The optimal penalty coefficient is determined using a 10-fold cross-validation method. The value of . At the optimal penalty coefficient In this step, features with regression coefficients non-zero are considered the core biomarkers most relevant to the predicted all-cause mortality. From the initial 74 indicators, this step selects a more concise and efficient subset of features for subsequent modeling. Ultimately, 43 features are obtained as shown in the above embodiment.

[0041] In some embodiments, the hyperparameters optimized by grid search include the learning rate, the maximum number of leaf nodes per tree, the maximum depth of the tree, the minimum number of samples per leaf node, the data sampling ratio, and the feature sampling ratio. The optimized hyperparameter combination is as follows: learning rate 0.05, maximum number of leaf nodes per tree 15, maximum depth of the tree set to -1 (depth is not limited), minimum number of samples per leaf node 20, data sampling ratio 0.9, and feature sampling ratio 0.7.

[0042] In this application embodiment, the types of hyperparameters that need to be optimized are all core parameters in the LightGBM algorithm that have the greatest impact on model performance. The function of each parameter is as follows: Learning rate `learning rate`: Controls the weight contribution of each decision tree, directly affecting the model's convergence speed (a larger learning rate leads to faster convergence but is prone to oscillations and non-convergence; a smaller learning rate leads to slower convergence but is more stable); Maximum number of leaf nodes per tree (`num_leaves`): Controls the complexity of a single decision tree (more leaf nodes mean a more complex tree with stronger fitting ability, but also a higher risk of overfitting); Maximum tree depth (`max_depth`): Further limits the tree's complexity to prevent excessively deep branches from causing overfitting (in some scenarios, this is set to -1, meaning no depth limit; it needs to be combined with other parameters to control complexity); Minimum number of samples per leaf node (`min_child_samples`): Controls the purity of leaf nodes (fewer samples mean more refined leaf nodes, making overfitting easier; more samples mean more stable nodes and stronger generalization ability); Data sampling ratio (`subsample`): The proportion of samples randomly sampled when training each tree, reducing the risk of overfitting by introducing randomness; Feature sampling ratio (`colsample_bytree`): The proportion of features randomly sampled when training each tree, also used to introduce randomness and prevent a single feature from dominating the model.

[0043] This application embodiment determines the optimal hyperparameter combination using a grid search method. The core logic of grid search is: pre-setting a reasonable value range for each hyperparameter (e.g., learning rate preset to 0.01, 0.05, 0.1), then traversing all parameter combinations, training a model for each combination, evaluating model performance (e.g., AUC value) using a validation set, and finally selecting the combination with the best performance as the final hyperparameters. The optimal combination and selection logic are defined here as follows: Learning rate = 0.05: This balances convergence speed and prediction accuracy. A learning rate of 0.01 is stable but has extremely low training efficiency (requiring a large number of decision trees), while a learning rate of 0.1 is fast but prone to oscillations and non-convergence. 0.05 ensures both training efficiency and stable convergence to the optimal solution. Maximum leaf nodes per tree = 15: LightGBM typically doesn't recommend too many leaf nodes (to avoid overfitting). A value of 15 ensures the fitting ability of a single tree without making the model overly complex due to too many nodes. Maximum tree depth = -1 (no depth limit): This complements the settings of "maximum leaf nodes per tree = 15" and "single..." The minimum number of samples per leaf node = 20” controls the complexity and avoids underfitting caused by limiting the depth alone; the minimum number of samples per leaf node = 20: ensures that each leaf node has enough samples to support it, avoids node prediction bias caused by insufficient sample size, and improves model stability; data sampling ratio = 0.9: retains 90% of the samples for training, which not only ensures the representativeness of the data, but also introduces randomness through 10% sampling missing, reducing the risk of overfitting; feature sampling ratio = 0.7: retains 70% of the features for training each tree, avoids some strong features (such as age) from playing a dominant role in all trees, and improves the generalization ability of the model.

[0044] The optimal hyperparameter combination provided in this application embodiment gives the embodiment strong repeatability (those skilled in the art can directly use this combination to train the model), and further ensures the predictive performance of the model - this combination is the optimal configuration verified by a large number of experiments, which can enable the LightGBM model to achieve the best balance between "fitting effect" and "generalization ability".

[0045] In some embodiments, the model output is a classification probability, which is a probability that satisfies the logistic function mapping, as described below:

[0046] in, is the predicted probability of all-cause mortality after PCI, x is the input feature vector that includes all feature variables used for prediction, and z is the sum of all valid decision tree output values.

[0047] Here, the model does not directly output probability values. Instead, it first generates intermediate results through the LightGBM algorithm, and then maps them to a final output that conforms to the probability range of (0,1) through a logistic function. As an ensemble decision tree algorithm, LightGBM's core output is the sum of the output values ​​of multiple decision trees (i.e., z in the formula). The range of this summation value may be any real number (-∞, +∞), and it cannot be directly used as a probability. The core function of the logistic function is to "compress" the z value from any real number range into the (0,1) interval, making it a value that conforms to the definition of probability (the probability value must be between 0 and 1, corresponding to the extreme cases of "absolutely will not happen" and "absolutely will happen," respectively).

[0048] P(y=1│x) represents "the probability of a patient experiencing all-cause mortality (y=1) post-surgery under the condition that the input feature vector is x". This clarifies the physical meaning of the output probability and is directly related to the core prediction objective of this scheme, allowing clinicians to clearly understand the meaning of the output result. P∈(0,1) clarifies the range of output probability values, conforms to the mathematical definition of probability, and avoids invalid values ​​greater than 1 or less than 0. x, as the input feature vector, corresponds to the 43 core features obtained through screening, covering three key data categories: clinical electronic medical records, blood cell counts, and blood biochemistry, ensuring the completeness and relevance of the input information. z is the cumulative value of the output values ​​of all effective decision trees. Here, "effective decision trees" refers to all decision trees trained by the LightGBM algorithm (each tree learns for the core features and outputs the risk score of the corresponding sample). The cumulative value z integrates the learning results of all decision trees, which can more comprehensively and accurately reflect the patient's mortality risk level.

[0049] From a clinical application perspective, this logistic function mapping mechanism ensures the "interpretability" and "practicality" of the output probability. For example, if a patient's z-value is 0, after calculation by the logistic function, P=1 / (1+e^0)=0.5, indicating that the patient's probability of death and survival are each 50%; if the z-value is 2 (positive, representing higher risk), then P=1 / (1+e^-2)≈0.88, indicating that the patient's risk of death is as high as 88%, requiring close attention from clinicians and the development of intensive intervention plans; if the z-value is -2 (negative, representing lower risk), then P=1 / (1+e^2)≈0.12, indicating that the patient's risk of death is only 12%, with a relatively good prognosis. This quantitative probability output, compared to a simple "high risk / low risk" classification, provides a more accurate basis for clinical decision-making, further enhancing the clinical practical value of the model.

[0050] In some embodiments, the training process of the LightGBM algorithm is achieved by minimizing the objective function, which is:

[0051] in, Let be the objective function. Let n be the true label of the i-th sample, and n be the total number of samples. The regularization term represents the positive class probability predicted by the model. Describe the model complexity of the t-th decision tree. :

[0052] in, Let be the complexity coefficient of the leaf node. is the L2 regularization coefficient for the weights, and T is the number of leaf nodes in the current tree. This represents the output weight value of leaf node j.

[0053] The core of the LightGBM algorithm is an optimized version of Gradient Boosting Decision Tree (GBDT). Its training process essentially involves iteratively generating multiple decision trees to gradually reduce the model's loss function value. In this embodiment, the objective function is the "loss metric" for each iteration of training a new decision tree, consisting of two parts: a "binary cross-entropy loss term" and a "model complexity regularization term," suitable for the binary classification task of this scheme. The first part is the binary cross-entropy loss term, whose core function is to measure the "degree of deviation" between the model's predicted probabilities and the true labels. Among them, is the objective function value (which needs to be minimized), and n is the total number of samples (the number of patients in the training set). Let n be the true label of the i-th sample, and n be the total number of samples. This represents the positive class probability predicted by the model. Minimizing this value essentially aims to make the model's predicted probability as close as possible to the true label, thereby improving the model's fit.

[0054] The second part is the model complexity regularization term for the t-th decision tree, whose core function is to "control model complexity and avoid overfitting". LightGBM improves the fitting effect by iteratively generating multiple decision trees, but too many trees or an overly complex single tree can cause the model to "overlearn" noise (rather than the true patterns) in the training data, resulting in decreased generalization ability (i.e., poor prediction performance for new samples). The regularization term Ω balances "fitting effect" and "generalization ability" by penalizing the complexity of the tree.

[0055] In summary, the training process of the LightGBM model in this application embodiment is as follows: in each iteration, based on the current model state, the negative gradient of the objective function is calculated, a new decision tree is generated (to fit the negative gradient and reduce loss), and then the new tree is integrated into the existing model through weight updates. Finally, through multiple iterations, the objective function value is minimized. At this point, the model has both good fitting effect (small prediction bias) and strong generalization ability (accurate prediction of new samples).

[0056] Please see Figure 2 , Figure 2 This application provides an embodiment of the relationship between the LASSO regression regularization penalty coefficient λ and the characteristic coefficients, as well as a cross-validation error graph. Figure 2 As shown, when the logλ value is small (i.e., λ itself is large), the regularization penalty is strong, and the coefficients of all features are compressed to 0. At this time, the model is the simplest but the fitting effect is poor (high cross-validation error). Excessive removal of features leads to the loss of key information. As the logλ value increases (i.e., λ itself decreases), the penalty weakens, and the coefficients of some features begin to "thaw" from 0 and gradually increase. The cross-validation error decreases accordingly, and the effective features are gradually incorporated, improving the model's fitting effect. The optimal regularization penalty coefficient is determined by ten-fold cross-validation, which is the logλ value corresponding to the minimum cross-validation error. When logλ exceeds the optimal value and continues to increase (λ further decreases), the penalty is too weak, more feature coefficients are retained, and the coefficients of some redundant features also begin to increase. The cross-validation error no longer decreases or even increases slightly, indicating that the model is overfitting.

[0057] The above embodiments can intuitively verify the feature selection logic of LASSO regression, that is, the process of feature coefficients gradually unfreezing from 0 as λ decreases; and can provide a visual basis for the selection of the optimal λ value, avoiding subjective judgment; the 43 features with non-zero regression coefficients selected by the curve can provide an accurate input feature set for the subsequent construction of the LightGBM model.

[0058] Please see Figure 3 , Figure 3 These are the ROC curves and AUC value graphs of the LightGBM prediction model provided in the embodiments of this application, as shown below. Figure 3 As shown, the main ROC curve (Recipient Operating Characteristic curve) is as follows: the horizontal axis is the "False Positive Rate (FPR)," which is the percentage of patients who "did not actually die but were predicted to die by the model"; the vertical axis is the "True Positive Rate (TPR)," which is the percentage of patients who "did actually die and were predicted to die by the model"; the diagonal reference line is the "Random Guessing Line," and its corresponding AUC (Area Under the Curve) = 0.5, which represents that the model has no discriminative ability and is used to intuitively compare and evaluate the model performance in this case.

[0059] In conjunction with the embodiments of this application, the ROC curve of the model provided in the embodiments of this application is clearly located above the diagonal and tends to move towards the upper left corner. This indicates that under any false positive rate, the true positive rate of the model is higher than the random guessing level, and it has a good ability to "distinguish between dead patients and living patients".

[0060] In summary, the embodiments of this application have the following beneficial effects: (1) Comprehensive data dimensions: It comprehensively incorporates three core data categories: clinical electronic medical records, blood cell counts, and blood biochemistry, covering multiple dimensions of patient information, including basic information, underlying diseases, physiological indicators, immune status, organ function, and metabolic level. This avoids the one-sidedness of prediction caused by a single data dimension and provides sufficient data support for accurate prediction.

[0061] (2) Scientific feature selection: The core features are selected by using the LASSO regression algorithm combined with ten-fold cross-validation, which effectively eliminates redundant features, simplifies the model structure, retains key prediction information, reduces the risk of model overfitting, and improves model interpretability.

[0062] (3) Excellent model performance: The LightGBM algorithm was selected to build the model, and the hyperparameters were optimized by grid search. This algorithm has the advantages of efficient processing of high-dimensional data and strong generalization ability. It has been verified by independent datasets and the prediction accuracy is significantly better than that of traditional models.

[0063] (4) Practical output results: The intermediate output of the model is mapped to the probability value in the range of 0 to 1 through the logic function, which intuitively reflects the patient's mortality risk level, making it easy for clinicians to quickly formulate personalized follow-up and intervention plans, realize the precise allocation of medical resources, and has extremely high clinical application value.

[0064] (5) High repeatability: The key details such as data acquisition standards, preprocessing methods, algorithm parameters, and hyperparameter combinations are clearly defined. Those skilled in the art can accurately reproduce the model construction process according to the contents of the manual, which is convenient for technology promotion.

[0065] Based on the same inventive concept, this application also provides a device for constructing a PCI post-operative all-cause mortality risk prediction model, which corresponds to the construction method of the PCI post-operative all-cause mortality risk prediction model in the first embodiment. Since the principle of the device in this application is similar to the above-mentioned PCI post-operative all-cause mortality risk prediction model construction method, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0066] like Figure 4 As shown, Figure 4This is a schematic diagram of the structure of the device 400 for constructing a prediction model of all-cause mortality risk for patients after PCI, as provided in this application embodiment. The device 400 for constructing a prediction model of all-cause mortality risk for patients after PCI includes: The acquisition module 401 is used to acquire a training dataset, which includes clinical electronic medical record data, blood cell count data and blood biochemistry data of patients admitted to elective PCI. All patients have follow-up records for at least 5 years after the operation and are labeled with whether all-cause mortality occurred. The preprocessing module 402 is used to preprocess the training dataset, including Z-score standardization for continuous variables and one-hot encoding for categorical variables. The filtering module 403 is used to filter features of the preprocessed dataset using the LASSO regression algorithm, determine the optimal regularization penalty coefficient through ten-fold cross-validation, and filter out the core features with non-zero regression coefficients as a combination of modeling markers. Optimization module 404 is used to build a prediction model based on the core features obtained from screening, using the LightGBM algorithm, with the binary log loss function as the loss metric, and to optimize the model hyperparameters through a grid search method. The output module 405 is used to output the completed prediction model, which can receive new core feature data of PCI patients and output the probability value of all-cause mortality of the patient after surgery.

[0067] Those skilled in the art should understand that Figure 4 The functions of each unit in the PCI post-operative patient all-cause mortality risk prediction model construction device 400 shown can be understood by referring to the relevant description of the PCI post-operative patient all-cause mortality risk prediction model construction method described above. Figure 4 The functions of each unit in the PCI post-operative patient all-cause mortality risk prediction device 400 shown can be realized by a program running on a processor or by specific logic circuits.

[0068] In one possible implementation, the clinical electronic medical record data includes: patient age, whether taking statins, cancer status, diabetes status, systolic blood pressure, diastolic blood pressure, waist-to-hip ratio, body fat percentage, Townsend index, drinking habits, and smoking habits. The blood cell count data include: eosinophil percentage, absolute lymphocyte count, lymphocyte percentage, mean corpuscular hemoglobin level, mean corpuscular volume, mean reticulocyte volume, absolute monocyte count, monocyte percentage, absolute neutrophil count, percentage of nucleated red blood cells, and red blood cell distribution width. The blood biochemical test data include: creatinine clearance, alanine aminotransferase, albumin, alkaline phosphatase, aspartate aminotransferase, C-reactive protein, calcium, serum creatinine, cystatin C, gamma-glutamyl transferase, glucose, high-density lipoprotein cholesterol, direct measurement of low-density lipoprotein cholesterol, phosphate, sex hormone-binding globulin, testosterone, total bilirubin, total protein, triglycerides, uric acid, and urea.

[0069] In one possible implementation, the Z-score standardization process is specifically performed by making the processed data distribution have a mean of 0 and a standard deviation of 1. The specific method of one-hot encoding is to convert the binary categorical variable into two independent binary features.

[0070] In one possible implementation, the objective function of the LASSO regression algorithm is:

[0071] Where N is the sample size and K is the number of coefficients. This is the predicted value for this sample. This is the true value of the sample. This is the regularization penalty coefficient, used to adjust the coefficient. The severity of the punishment.

[0072] In one possible implementation, the hyperparameters optimized by grid search include the learning rate, the maximum number of leaf nodes per tree, the maximum depth of the tree, the minimum number of samples per leaf node, the data sampling ratio, and the feature sampling ratio. The optimized combination of hyperparameters is: learning rate 0.05, maximum number of leaf nodes per tree 15, maximum tree depth set to -1 (depth not limited), minimum number of samples per leaf node 20, data sampling ratio 0.9, and feature sampling ratio 0.7.

[0073] In one possible implementation, the model outputs a classification probability, which is a probability that satisfies a logistic function mapping, as described below:

[0074] in, is the predicted probability of all-cause mortality after PCI, x is the input feature vector that includes all feature variables used for prediction, and z is the sum of all valid decision tree output values.

[0075] In one possible implementation, the training process of the LightGBM algorithm is achieved by minimizing the objective function, which is:

[0076] in, Let be the objective function. Let n be the true label of the i-th sample, and n be the total number of samples. The regularization term represents the positive class probability predicted by the model. Describe the model complexity of the t-th decision tree. :

[0077] in, Let be the complexity coefficient of the leaf node. is the L2 regularization coefficient for the weights, and T is the number of leaf nodes in the current tree. This represents the output weight value of leaf node j.

[0078] The device for constructing the above-mentioned prediction model for all-cause mortality risk in patients after PCI has the following beneficial effects: (1) Comprehensive data dimensions: It comprehensively incorporates three core data categories: clinical electronic medical records, blood cell counts, and blood biochemistry, covering multiple dimensions of patient information, including basic information, underlying diseases, physiological indicators, immune status, organ function, and metabolic level. This avoids the one-sidedness of prediction caused by a single data dimension and provides sufficient data support for accurate prediction.

[0079] (2) Scientific feature selection: The core features are selected by using the LASSO regression algorithm combined with ten-fold cross-validation, which effectively eliminates redundant features, simplifies the model structure, retains key prediction information, reduces the risk of model overfitting, and improves model interpretability.

[0080] (3) Excellent model performance: The LightGBM algorithm was selected to build the model, and the hyperparameters were optimized by grid search. This algorithm has the advantages of efficient processing of high-dimensional data and strong generalization ability. It has been verified by independent datasets and the prediction accuracy is significantly better than that of traditional models.

[0081] (4) Practical output results: The intermediate output of the model is mapped to the probability value in the range of 0 to 1 through the logic function, which intuitively reflects the patient's mortality risk level, making it easy for clinicians to quickly formulate personalized follow-up and intervention plans, realize the precise allocation of medical resources, and has extremely high clinical application value.

[0082] (5) High repeatability: The key details such as data acquisition standards, preprocessing methods, algorithm parameters, and hyperparameter combinations are clearly defined. Those skilled in the art can accurately reproduce the model construction process according to the contents of the manual, which is convenient for technology promotion.

[0083] like Figure 5 As shown, Figure 5 This is a schematic diagram of the composition structure of the electronic device 500 provided in the embodiments of this application. The electronic device 500 includes: The device includes a processor 501, a storage medium 502, and a bus 503. The storage medium 502 stores machine-readable instructions executable by the processor 501. When the electronic device 500 is running, the processor 501 communicates with the storage medium 502 via the bus 503. The processor 501 executes the machine-readable instructions to perform the steps of the method for constructing a prediction model for all-cause mortality risk of patients after PCI as described in the embodiments of this application.

[0084] In practical applications, the various components in the electronic device 500 are coupled together via a bus 503. It is understood that the bus 503 is used to achieve communication between these components. In addition to a data bus, the bus 503 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 5 The general designated all buses as Bus 503.

[0085] The above-mentioned electronic devices have the following beneficial effects: (1) Comprehensive data dimensions: It comprehensively incorporates three core data categories: clinical electronic medical records, blood cell counts, and blood biochemistry, covering multiple dimensions of patient information, including basic information, underlying diseases, physiological indicators, immune status, organ function, and metabolic level. This avoids the one-sidedness of prediction caused by a single data dimension and provides sufficient data support for accurate prediction.

[0086] (2) Scientific feature selection: The core features are selected by using the LASSO regression algorithm combined with ten-fold cross-validation, which effectively eliminates redundant features, simplifies the model structure, retains key prediction information, reduces the risk of model overfitting, and improves model interpretability.

[0087] (3) Excellent model performance: The LightGBM algorithm was selected to build the model, and the hyperparameters were optimized by grid search. This algorithm has the advantages of efficient processing of high-dimensional data and strong generalization ability. It has been verified by independent datasets and the prediction accuracy is significantly better than that of traditional models.

[0088] (4) Practical output results: The intermediate output of the model is mapped to the probability value in the range of 0 to 1 through the logic function, which intuitively reflects the patient's mortality risk level, making it easy for clinicians to quickly formulate personalized follow-up and intervention plans, realize the precise allocation of medical resources, and has extremely high clinical application value.

[0089] (5) High repeatability: The key details such as data acquisition standards, preprocessing methods, algorithm parameters, and hyperparameter combinations are clearly defined. Those skilled in the art can accurately reproduce the model construction process according to the contents of the manual, which is convenient for technology promotion.

[0090] This application also provides a computer-readable storage medium storing executable instructions. When the executable instructions are executed by at least one processor 501, the method for constructing a prediction model for all-cause mortality risk in patients after PCI as described in this application embodiment is implemented.

[0091] In some embodiments, the storage medium may be a magnetic random access memory (FRAM), a read-only memory (ROM), or a programmable read-only memory (PROM). Erasable Programmable Read-Only Memory (EPROM) Electrically Erasable Programmable Read-Only Memory (EEPROM) Read-only memory, flash memory, magnetic surface storage, optical disc, or CD-ROM ROM, Compact Disc Read It can be a memory such as a memory only; or it can be a device that includes one or any combination of the above-mentioned memories.

[0092] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0093] As an example, executable instructions may, but do not necessarily, correspond to files in the file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0094] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0095] The aforementioned computer-readable storage media have the following beneficial effects: (1) Comprehensive data dimensions: It comprehensively incorporates three core data categories: clinical electronic medical records, blood cell counts, and blood biochemistry, covering multiple dimensions of patient information, including basic information, underlying diseases, physiological indicators, immune status, organ function, and metabolic level. This avoids the one-sidedness of prediction caused by a single data dimension and provides sufficient data support for accurate prediction.

[0096] (2) Scientific feature selection: The core features are selected by using the LASSO regression algorithm combined with ten-fold cross-validation, which effectively eliminates redundant features, simplifies the model structure, retains key prediction information, reduces the risk of model overfitting, and improves model interpretability.

[0097] (3) Excellent model performance: The LightGBM algorithm was selected to build the model, and the hyperparameters were optimized by grid search. This algorithm has the advantages of efficient processing of high-dimensional data and strong generalization ability. It has been verified by independent datasets and the prediction accuracy is significantly better than that of traditional models.

[0098] (4) Practical output results: The intermediate output of the model is mapped to the probability value in the range of 0 to 1 through the logic function, which intuitively reflects the patient's mortality risk level, making it easy for clinicians to quickly formulate personalized follow-up and intervention plans, realize the precise allocation of medical resources, and has extremely high clinical application value.

[0099] (5) High repeatability: The key details such as data acquisition standards, preprocessing methods, algorithm parameters, and hyperparameter combinations are clearly defined. Those skilled in the art can accurately reproduce the model construction process according to the contents of the manual, which is convenient for technology promotion.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed methods and electronic devices can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0101] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0102] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0103] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0104] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing a prediction model for all-cause mortality risk in patients after PCI, characterized in that, Includes the following steps: Obtain a training dataset containing clinical electronic medical record data, blood cell count data, and blood biochemistry data of patients admitted for elective PCI. All patients have follow-up records for at least 5 years after surgery and are labeled with whether all-cause mortality occurred. The training dataset is preprocessed, including Z-score standardization for continuous variables and one-hot encoding for categorical variables; The LASSO regression algorithm was used to select features from the preprocessed dataset. The optimal regularization penalty coefficient was determined by ten-fold cross-validation, and the core features with non-zero regression coefficients were selected as a combination of modeling markers. Based on the core features obtained through screening, a prediction model is constructed using the LightGBM algorithm, with the binary log loss function as the loss metric, and the model hyperparameters are optimized using the grid search method. The completed prediction model is output. The model can receive new core feature data of PCI patients and output the probability value of all-cause mortality after the operation.

2. The method according to claim 1, characterized in that, The clinical electronic medical record data includes: patient age, whether they are taking statins, cancer status, diabetes status, systolic blood pressure, diastolic blood pressure, waist-to-hip ratio, body fat percentage, Townsend index, drinking habits, and smoking habits; The blood cell count data include: eosinophil percentage, absolute lymphocyte count, lymphocyte percentage, mean corpuscular hemoglobin level, mean corpuscular volume, mean reticulocyte volume, absolute monocyte count, monocyte percentage, absolute neutrophil count, percentage of nucleated red blood cells, and red blood cell distribution width. The blood biochemical test data include: creatinine clearance, alanine aminotransferase, albumin, alkaline phosphatase, aspartate aminotransferase, C-reactive protein, calcium, serum creatinine, cystatin C, gamma-glutamyl transferase, glucose, high-density lipoprotein cholesterol, direct measurement of low-density lipoprotein cholesterol, phosphate, sex hormone-binding globulin, testosterone, total bilirubin, total protein, triglycerides, uric acid, and urea.

3. The method according to claim 1, characterized in that, The specific method of Z-score standardization is to make the processed data distribution have a mean of 0 and a standard deviation of 1. The specific method of one-hot encoding is to convert the binary categorical variable into two independent binary features.

4. The method according to claim 1, characterized in that, The objective function of the LASSO regression algorithm is: Where N is the sample size and K is the number of coefficients. This is the predicted value for this sample. This is the true value of the sample. This is the regularization penalty coefficient, used to adjust the coefficient. The severity of the punishment.

5. The method according to claim 1, characterized in that, The hyperparameters optimized through grid search include the learning rate, the maximum number of leaf nodes per tree, the maximum depth of the tree, the minimum number of samples per leaf node, the data sampling ratio, and the feature sampling ratio. The optimized hyperparameter combination is as follows: learning rate 0.05, maximum number of leaf nodes per tree 15, maximum tree depth set to -1 (no depth limit), minimum number of samples per leaf node 20, data sampling ratio 0.9, and feature sampling ratio 0.

7.

6. The method according to claim 1, characterized in that, The model outputs classification probabilities, which are probabilities that satisfy the logistic function mapping, as described below: in, is the predicted probability of all-cause mortality after PCI, x is the input feature vector that includes all feature variables used for prediction, and z is the sum of all valid decision tree output values.

7. The method according to claim 1, characterized in that, The training process of the LightGBM algorithm is achieved by minimizing the objective function, which is: in, Let be the objective function. Let n be the true label of the i-th sample, and n be the total number of samples. The regularization term represents the positive class probability predicted by the model. Describe the model complexity of the t-th decision tree. : in, Let be the complexity coefficient of the leaf node. is the L2 regularization coefficient for the weights, and T is the number of leaf nodes in the current tree. This represents the output weight value for leaf node j.

8. A device for constructing a predictive model for all-cause mortality risk in patients after PCI, characterized in that, The device includes: The acquisition module is used to acquire the training dataset, which includes clinical electronic medical record data, blood cell count data and blood biochemistry data of patients admitted for elective PCI. All patients have follow-up records for at least 5 years after the operation and are labeled with whether all-cause mortality occurred. The preprocessing module is used to preprocess the training dataset, including Z-score standardization for continuous variables and one-hot encoding for categorical variables. The filtering module is used to filter features of the preprocessed dataset using the LASSO regression algorithm. It determines the optimal regularization penalty coefficient through ten-fold cross-validation and selects the core features with non-zero regression coefficients as a combination of modeling markers. The optimization module is used to build a prediction model based on the core features obtained from the screening, using the LightGBM algorithm, with the binary log loss function as the loss metric, and to optimize the model hyperparameters through a grid search method. The output module is used to output the completed prediction model. The model can receive new core feature data of PCI patients and output the probability value of all-cause mortality after the operation of the patient.

9. An electronic device, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is in operation, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the method for constructing a prediction model for all-cause mortality risk of patients after PCI as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the method for constructing a prediction model for all-cause mortality risk in patients after PCI as described in any one of claims 1 to 7.

Images