Proteomic signatures and machine learning based-prediction models for cardiovascular risk in survivors of myocardial infarction

Abstract Aims: Survivors of myocardial infarction (MI) are still at risk for adverse long-term outcomes such as all-cause mortality, heart failure (HF), and ischemic stroke (IS) after acute phase treatment. This study aimed to identify specific protein markers and construct risk prediction models for the main cardiovascular events in survivors of MI. Methods: A total of 30,135 survivors of MI were included in this study, all of whom had available follow-up data from the UK Biobank (UKB). Multivariate Cox regression analysis was used to assess the clinical associations between plasma proteins and MI-related outcomes, including all-cause mortality, HF, and IS. Subsequently, prediction models with machine learning were constructed based on the plasma protein levels to further evaluate these associations. Results: We identified 570 proteins significantly associated with all-cause mortality, 172 proteins associated with incident HF, and 13 proteins associated with incident IS in survivors of MI. Among these proteins, 12 proteins were associated with three outcomes, including EDA2R, NT-proBNP and GDF15 (P < 1.71?10-5). Pathway enrichment analysis showed that these proteins were mainly involved in pathophysiological processes such as inflammatory response, fibrosis and myocardial remodeling. Machine learning models based on 117, 73 and 82 plasma protein showed good predictive performance for all-cause mortality (XGBoost: AUC = 0.79), HF (LightGBM: AUC = 0.81) and IS (Random Forest: AUC = 0.76) in survivors of MI, respectively. Finally, we systematically identified 52 mortality-associated, 14 HF-associated, and 4 IS-associated plasma protein biomarkers in survivors of MI based on the Cox regression analysis and machine learning modeling. Conclusion: Our integrated study with predictive modeling have identified the plasma protein biomarkers associated with adverse outcomes in survivors of MI, and subsequently developed predictive models to facilitate early risk stratification. Keywords: Myocardial infarction; proteomics; association study; prediction model; machine learning