Machine learning-driven lipidomics for non-invasive assessment of coronary lesion grading

Background Phospholipids and sphingolipids, such as Phosphatidylcholine, Phosphatidylethanolamine,and Ceramide, are closely associated with cardiovascular risk and have been established in the diagnosis and prognosis of coronary artery disease (CAD). However, there is limit studies that focused on the correlation between plasma lipidomics and the severity of coronary artery lesions. Purpose This study aims to develop non-invasive models for classifying the severity of coronary artery lesions using serum lipidomic biomarkers and clinical variables. Methods We prospectively enrolled 257 patients with acute coronary syndrome (ACS) from a single center between August 2023 and January 2024. Based on coronary angiography and quantitative flow ratio (QFR), we calculated the functional SYNTAX score (FSS) and categorized patients into low-risk (≤ 22 points, n = 232) and intermediate-high-risk (> 22 points, n = 25) groups. Serum lipidomic and clinical variables were selected using univariate logistic regression (p < 0.1), volcano plot analysis (fold change >1.2, p < 0.1), and partial least squares discriminant analysis (PLS-DA, VIP > 2). Predictive models were developed using Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), optimized through oversampling, recursive feature elimination (RFE), and hyperparameter tuning. Results A total of 312 distinct lipid species spanning 22 major lipid classes were identified by Lipidomic profiling. Through integrative analysis with 57 clinical variables, we identified 8 crucial predictive factors, including hemoglobin, immature granulocyte ratio, and 6 signature lipids (LPC 20:0, LPC 20:2, PC 36:5, ePC 38:2, PG 38:2, and PG 38:3) (Figure 1). The SVM model incorporating these 8 biomarkers demonstrated superior predictive capability, achieving an area under the curve (AUC) of 0.97 with 95% accuracy, 100% sensitivity, 63% specificity, and an F1-score of 74%, significantly outperforming RF (AUC 0.64) and XGBoost (AUC 0.78) (Table 1). Conclusion We proposed a model that combined lipids and clinical variables, which accurately identified high-risk coronary artery lesions. It is important that the validation was limited to ACS patients. To improve the clinical applicability, future multicenter studies should aim to validate this model in cohorts encompassing the full spectrum of CAD.Figure1.SVM + RFE for Feature Selection Table1.3 models in the test set