Prediction of post-PCI angina risk using machine learning-based magnetocardiography model

Background Recurrent angina after percutaneous coronary intervention (PCI) impairs quality of life and poses a clinical challenge. Magnetocardiography (MCG), as a noninvasive tool, its role in predicting symptomatic outcomes post-PCI remains undefined. Objective This study sought to develop and validate a machine learning-based MCG model, both alone and combined with clinical biomarkers, to predict angina after PCI. Methods 110 patients with coronary artery disease undergoing successful PCI were included. MCG was performed pre-PCI and within 72 h post-PCI. MCG was performed to derive quantitative scores for all patients both prior to and following percutaneous coronary intervention (PCI). Angina status was assessed within 3 months using the Seattle Angina Questionnaire-Angina Stability (SAQ-AS) and -Angina Frequency (SAQ-AF) domains. Based on SAQ-AS scores, patients were stratified into an AS-negative group ( n = 105, mean age 62 ± 8 years, 59% male) and an AS-positive group ( n = 5, mean age 65 ± 15 years, 40% male). Similarly, SAQ-AF scores categorized patients into an AF-negative group ( n = 101, mean age 63 ± 8 years, 60.4% male) and an AF-positive group ( n = 9, mean age 60 ± 12 years, 33.3% male). Serum biomarkers, including low-density lipoprotein cholesterol (LDL-C), cardiac troponin I (cTnI), and N-terminal pro-B-type natriuretic peptide (NT-proBNP), were measured. An MCG score was measured for each patient both pre- and post-PCI. An MCG model and a combined model integrating MCG with biomarkers were developed using multivariable logistic regression. Performance was evaluated by the area under the receiver operating characteristic curve (AUC), sensitivity, and F1-score. Prediction models were visualized using nomograms. Results MCG score decreased from 0.783 pre-PCI to 0.616 post-PCI. The combined model demonstrated superior discriminatory ability for SAQ-AS. For SAQ-AF, the combined model achieved higher sensitivity (0.663 vs. 0.605), F1-score (0.797 vs. 0.754), and AUC (0.813 vs. 0.801) compared to MCG alone. The nomogram provided broader risk stratification (SAQ-AS: 0.7–0.9; SAQ-AF: 0.5–0.9). Calibration was satisfactory (Hosmer–Lemeshow p > 0.05). Conclusion A machine learning-based MCG model effectively predicts post-PCI angina risk. Integrating MCG with clinical biomarkers enhances risk stratification, offering a noninvasive strategy for identifying high-risk patients.