Platelet-to-Lymphocyte Ratio-Driven Machine Learning Model for Predicting In-Hospital Major Adverse Cardiovascular Events After Percutaneous Coronary Intervention in Acute Myocardial Infarction: A Retrospective Cohort Study

<title>Abstract</title> <p> <bold>Background</bold> Accurate risk assessment of major adverse cardiovascular events (MACE) following percutaneous coronary intervention (PCI) in patients with acute myocardial infarction (AMI) is central to optimizing clinical decision-making. Although inflammatory mechanisms play a crucial role in ischemia-reperfusion injury, previous studies have largely been limited to single inflammatory markers, lacking systematic integration and comparative validation of multidimensional inflammatory indices. Moreover, which inflammatory marker possesses the optimal driving efficacy within a machine learning framework remains unclear. <bold>Objective</bold> This study aimed to systematically integrate five commonly used multidimensional inflammatory indices (PLR, MLR, MHR, SII, and SIRI), construct risk prediction models for in-hospital MACE after PCI in AMI patients based on traditional logistic regression and nine machine learning algorithms, establish PLR as a core driving factor, and explore nonlinear relationships between key clinical indicators and MACE. <bold>Methods</bold> A total of 1,163 consecutive AMI patients who underwent PCI at the Second Affiliated Hospital of Shenyang Medical College from 2020 to 2025 were retrospectively enrolled. Patients were randomly divided into a training set (n = 814) and a test set (n = 349) at a 7:3 ratio. Five multidimensional inflammatory indices were calculated: platelet-to-lymphocyte ratio (PLR), monocyte-to-lymphocyte ratio (MLR), monocyte-to-high-density lipoprotein ratio (MHR), systemic immune-inflammation index (SII), and systemic inflammation response index (SIRI). LASSO regression was used for variable selection, and multivariable logistic regression was performed to identify independent risk factors and construct a nomogram. Restricted cubic spline (RCS) analysis was employed to explore nonlinear relationships. The Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training set to balance class distribution. Based on the identified independent predictors, nine machine learning models (XGBoost, Random Forest, LightGBM, MLP, SVM, AdaBoost, Decision Tree, Gaussian Naive Bayes, and Logistic Regression) were constructed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). The SHAP method was used for model interpretability. Robustness was verified through hard endpoint sensitivity analysis and subgroup analyses. <bold>Results</bold> The training and test sets were well-balanced in baseline characteristics. The incidence of MACE was 17.2% (140/814). Univariate analysis showed that PLR, SIRI, SII, and MLR were significantly elevated in the MACE group (P &lt; 0.05), while MHR showed no significant difference between groups. LASSO regression (λ_1se = 0.02) selected 11 candidate variables, with PLR being the only retained multidimensional inflammatory index. Multivariable logistic regression revealed that age (OR = 1.076, 95% CI: 1.054–1.101, P &lt; 0.001), Gensini score (OR = 1.014, 95% CI: 1.009–1.020, P &lt; 0.001), PLR (OR = 1.010, 95% CI: 1.006–1.014, P &lt; 0.001), and AST (OR = 1.004, 95% CI: 1.002–1.005, P &lt; 0.001) were independent risk factors, whereas EF (OR = 0.965, 95% CI: 0.946–0.983, P &lt; 0.001) and statin use (OR = 0.165, 95% CI: 0.057–0.473, P = 0.001) were independent protective factors. A significant nonlinear relationship was identified between EF and MACE risk (P for non-linearity &lt; 0.001), with risk rising sharply when EF fell below approximately 50%. The training set AUC was 0.842, and the Hosmer-Lemeshow test yielded P = 0.615. After SMOTE resampling (n = 1,348), comparison of nine machine learning models on the independent test set showed that XGBoost performed best (AUC = 0.937, 95% CI: 0.914–0.959, F1 = 0.859, Brier Score = 0.100), followed by MLP (AUC = 0.906) and Random Forest (AUC = 0.903). DeLong test demonstrated that XGBoost significantly outperformed traditional logistic regression (Z = 5.451, P &lt; 0.001). In the independent test set, XGBoost achieved optimal performance (AUC = 0.937, 95% CI: 0.914–0.959, F1 = 0.859, Brier Score = 0.100), significantly superior to traditional logistic regression (AUC = 0.846, DeLong test: Z = 5.451, P &lt; 0.001), with ΔAUC = 0.091. SHAP analysis revealed that age (39.2%), Gensini score (21.9%), and AST (19.8%) were the most important predictive features, with PLR contributing approximately 10.3%. In hard endpoint sensitivity analysis, XGBoost maintained optimal performance (AUC = 0.965). All five subgroup analyses yielded interaction P-values &gt; 0.05. <bold>Conclusion</bold> Among the five multidimensional inflammatory indices evaluated, PLR was the only one that retained independent predictive value after LASSO regularization and multivariable adjustment, highlighting its potential as a core inflammatory driver within the prediction framework. EF exhibits a nonlinear threshold effect with MACE risk at approximately 50%. The XGBoost model significantly outperforms traditional logistic regression in predictive performance, and when combined with SHAP analysis, enables individualized risk visualization. This PLR-driven machine learning model can serve as a precise and robust quantitative decision-making tool for early risk stratification and individualized intervention in AMI patients after PCI. </p>