Interpretable machine learning analysis of immune-inflammatory biomarkers for predicting no-reflow in patients with acute myocardial infarction after percutaneous coronary intervention.

OBJECTIVE: To evaluate whether routine immune-inflammatory biomarkers predict no-reflow (NR) after percutaneous coronary intervention (PCI) in acute myocardial infarction (AMI) patients, quantify their incremental predictive value over clinical variables, and build an interpretable peri-procedural prediction model. METHODS: This prospective cohort recruited 382 AMI patients undergoing PCI. Univariable and multivariable logistic regression identified independent correlates; confounder-adjusted restricted cubic splines (RCS) assessed nonlinear associations. LASSO regression selected core predictors with variance inflation factor (VIF) assessing multicollinearity. Four machine-learning models underwent 1000-iteration bootstrap internal validation and SHAP interpretation. Incremental value was tested via ΔAUC (DeLong test), continuous NRI, and IDI. Stratified regression and interaction tests evaluated AMI-subtype effect modification. RESULTS: NR occurred in 91 patients (23.8%). Multiple neutrophil-centred indices independently correlated with NR after adjustment. Seven predictors constituted the final model (all VIF < 2.5). LightGBM yielded optimal discrimination and modest net clinical benefit (AUC 0.799, 95% CI 0.699-0.899); AMI subtype and neutrophil-related markers dominated SHAP outputs. Neutrophil count, NAR, NHR, SII, and PIV significantly improved discrimination. Seven of the 16 tested markers yielded significant continuous NRI, and ten demonstrated significant IDI gains. In stratified analyses, neutrophil count, NLR, and MHR were associated with NR in STEMI but not NSTEMI patients; interaction tests found no significant AMI-subtype effect modification. CONCLUSION: Pre-procedural neutrophil-centred inflammatory biomarkers are independent correlates of post-PCI NR and offer incremental predictive value over conventional clinical variables. The interpretable LightGBM model supports peri-procedural NR risk stratification, though external validation and recalibration are mandatory before clinical application.