Abstract 4144730: Combining novel lipid biomarkers with deep learning algorithms to develop an initial non-invasive screening approach for ruling out obstructive coronary artery disease

Background: A personalized, non-invasive assessment approach for evaluating the risk of obstructive coronary artery disease (CAD) is crucial for patients with an intermediate or low clinical likelihood of CAD before undergoing invasive coronary angiography (ICA). This method allows clinicians to effectively rule out the presence of obstructive CAD without the need for ICA or to determine if a referral for ICA is warranted. Emerging lipidomics biomarkers may be valuable in this process. However, technological challenges in detecting structurally similar lipids and the requirement for advanced computational tools have so far impeded the clinical application of lipidomics research. Hypothesis: Our study aims to develop an innovative non-invasive diagnostic test utilizing novel lipidomics biomarkers, potentially revolutionizing current risk classification schemes for CAD. Methods: In this post-hoc analysis of the CorLipid trial (NCT04580173), we employed extreme gradient boosting (XGBoost) machine learning to assess the predictive power of a lipidomics panel for obstructive CAD risk. Liquid chromatography-mass spectrometry analyzed lipid profiles from 146 individuals undergoing ICA. SYNTAX Score (SS) was used to define obstructive CAD as SS>0 versus non-obstructive CAD (SS=0). Results: Of the 146 participants (25% female, mean age: 61 ±11 years old), 55% had obstructive CAD (SS>0). Lipidome changes [phosphatidylinositols, (lyso-)phosphatidylethanolamine, (lyso-)phosphatidylcholine, triglycerides, diglycerides, and sphingomyelins] were investigated to identify lipids potentially associated with the phenotype and complexity of CAD. Using this information, 290 quantified serum lipid species were utilized to develop an XGBoost algorithm with 17 serum biomarkers ( consisting of sphingolipids, glycerophospholipids, triacylglycerols, galectin-3, glucose, low-density lipoprotein, and lactate dehydrogenase) with very good discriminative ability [ROC AUC: 0.875 (95%CI: 0.867-0.883)], excellent sensitivity (100%) but moderate specificity (62.1%) for the prediction of obstructive CAD. Conclusions: These findings indicate that a deep-learning-based non-invasive diagnostic test, using lipidomics serum biomarkers, could reliably rule-out obstructive CAD without necessitating ICA. To enhance generalizability, these results should be validated in larger and similar cohorts. Further research, particularly leveraging machine learning, is promising for refining risk stratification.