Abstract 17570: Combining Genomic Profiling and Clinical Data Through Machine Learning Modeling for the Prediction of Coronary Artery Disease Severity: Insights From the Genetic SYNTAX Score (GESS) Trial

Abstract

Background: Genomic profiling emerges as a powerful partner in the prediction of CAD development. Hypothesis: Machine learning (ML) algorithms integrating genomic profiling and clinical data could improve patient outcomes by contributing to earlier non-invasive diagnosis of obstructive CAD. Aims: The Genetic Syntax Score (GESS) prospective cohort study aimed to create a ML-algorithm able to predict the severity of CAD based on the analysis of 228 single nucleotide polymorphisms (SNPs) and relevant clinical and demographic data. Ultimate goal of this analysis was to compare the predictive capacity of a SNP-including model with that of a clinical model without SNPs. Methods: Patients with suspected CAD (n=919, mean age: 64±12 years) underwent invasive coronary angiography (from 2019 to 2021), and, thereby, their SYNTAX score was determined. Peripheral blood samples were drawn for genomic profiling (next generation sequencing analysis of a custom-made panel consisting of 228 SNPs) under standardized methods. After the ML-based selection of the most important subset of features (clinical/SNPs), two competing meta-learners were fitted, called Model A (including clinical/demographic predictors) and Model B (Model A plus SNPs predictors). The difference in their predictive capacities for the precise assessment of CAD risk (whether a patient presents with SYNTAX score >0 or =0) and CAD severity (whether a patient presents with higher SYNTAX score) was investigated via Delong’s test and Wilcoxon signed-rank test, respectively. Results: Our ML-algorithm identified a total of 8 SNPs (rs11984041, rs41291556, rs2023938, rs2107595, rs216172, rs1800562, rs3732379, and rs964184) as significant predictors of CAD risk, and 7 SNPs (rs4845625, rs6689306, rs2046934, rs1332844, rs6801273, rs663129, and rs870142) as significant predictors of CAD severity. Model B demonstrated superior performance compared to Model A (Z=2.451, p=0.014 for CAD risk assessment, and V=16731, p=0.002 for CAD severity assessment). Conclusion(s): The combined use of clinical/demographic parameters and genomic profiling might be deployed to improve the prediction of CAD risk and severity under the guidance of ML-techniques. Registration: NCT03150680