Prognostic Analysis Of Patients With Coronary Artery Disease Based On Interpretable Machine Learning

Background:Coronary artery disease(CAD)is one of the largest contributors to the global burden of disease and the leading cause of death from cardiovascular disease.Due to the limited medical and health resources and the aging population,the diagnosis and treatment situation of CAD is particularly severe.Percutaneous coronary intervention(PCI)based on drugeluting stents(DES),as the mainstream treatment for CAD,has significantly improved the postoperative survival rate of patients compared with the previous bare metal stents.However,some studies have shown that even among patients with DES implantation,the incidence of in-stent restenosis(ISR)in one year is still 5.5%-12.2%,and shows a gradual upward trend.Therefore,ISR is a thorny problem that must be faced for clinicians.The death risk stratification of CAD patients after PCI is an important part of postoperative management,which helps to identify high-risk patients and make them get personalized treatment plans.However,most of the existing risk stratification tools are developed based on traditional logistic regression,which has certain deficiencies in the accuracy of prediction and treatment decisions guidance.Machine learning(ML),which is located at the confluence of computer science and statistics,has been widely used in many fields related to health care in recent years because of its unique advantages in the processing of complex clinical data.However,this excellent performance also implies extremely complex algorithm logic and variable characteristics,which makes the model less interpretable and limits its use in clinical work.In recent years,interpretable ML methods have emerged.Through Shapley’s additive explanation(SHAP),the model established by ML can be displayed visually,which is more friendly to clinical workers and patients.Objective:The aim of this study is to explore the prognostic factors of CAD patients and establish the corresponding prediction model by the method of interpretable ML,in order to provide help for individualized risk assessment and management of patients.This study mainly included the following objectives: 1.To explore the risk factors of ISR in CAD patients with DES implantation by constructing an interpretable random forest model;2.To develop different machine learning models to predict all-cause mortality and compare their predictive power and clinical applicability.The best model will be selected for risk factor analysis and risk stratification,and the results will be interpreted to confirm the feasibility of interpretable ML method.Methods:1.Results of an observational cohort study conducted in the First Affiliated Hospital of Zhengzhou University from July 2009 to August 2011 were used for secondary analysis.Patients with CAD implanted with DES were divided into ISR group and non-ISR group according to the presence or absence of ISR.Clinical characteristics,laboratory tests,lesion and surgical characteristics,and medication were included as potential predictors.Multiple imputation based on random forest was used to impute missing values.The random forest model was constructed by 5-fold cross validation method to explore the risk factors and the interpretability was analyzed by SHAP values.2.A total of 7307 coronary artery disease patients implanted with DES from the optimal Antiplatelet Therapy for Coronary heart Disease(OPT-CAD)study led by academician Han Yaling in General Hospital of Northern Theater Command were retrospectively selected and divided into training set and test set with a ratio of 7:3.Five machine learning models,including logistic regression,support vector machine,random forest,naive bayes,and extreme gradient boosting,were constructed using 10-fold cross validation in the training set.In the test set,the best model was selected by comprehensive evaluation of discrimination,calibration and clinical applicability.The risk factors of all-cause mortality were analyzed and ranked based on the optimal model,and the results were interpreted individually using SHAP values.Results:1.Of the 603 patients included,145(24.05%)were in ISR group and 458(75.95%)were in non-ISR group.46 variables,including clinical characteristics,laboratory tests,lesion and surgical characteristics,and medication,were included in the analysis.Previous PCI and restenosis were significantly different between the two groups(P < 0.05),which was consistent with previous literature reports.The area under curve(AUC)of the random forest model constructed by 5-fold cross validation was 0.64.The importance rank of variables by SHAP values showed that blood glucose,high-density lipoprotein,average stent diameter,previous PCI and systolic blood pressure were essential factors.Previous PCI,previous CABG and previous stroke promote the occurrence of ISR.Bifurcation lesions,left main coronary artery lesions,multi-vessel lesions,restenosis lesions and total occlusion lesions promote the occurrence of ISR.Femoral approach can promote the occurrence of ISR compared with radial approach.With the increase of the number of target vessels,the risk showed an increasing trend.Sirolimus stent can reduce the risk of ISR compared with paclitaxel stent or other types of DES.Bilirubin and serum creatinine are positively correlated with the risk of ISR,high-density lipoprotein is negatively correlated with the risk of ISR,and heart rate shows a potential U-shaped effect.2.A total of 7307 patients with coronary heart disease implanted with DES were included in the study.In the test dataset,the random forest model had the highest AUC of 0.758 compared with other machine learning models,and showed a highest net benefit by decision curve analysis.Gini importance,permutation importance,and SHAP feature importance consistently showed that age,left ventricular ejection fraction,heart rate,hemoglobin,creatinine,and fibrinogen were important predictors of all-cause death.The effect of different features was illustrated visually from both local(case specific)and global(model specific)aspects based on SHAP values.Kaplan-Meier curves showed that the model could effectively guide risk stratification.Conclusions:Interpretable random forest model was used to identify the risk factors of ISR and allcause death in patients with CAD after PCI.Different machine learning models were constructed and compared in the subgroup population of OPT-CAD study.Finally,the random forest model showed certain advantages in discrimination,calibration and clinical applicability.The findings of this study can help predict the risk of all-cause mortality,guide clinical decision making,optimize the allocation of health care resources,and improve the level of secondary prevention.