Obesity predisposes to cardiovascular diseases (CVD); however not all obese subjects will develop CVD. In search of predictive biomarkers to stratify the obese, we considered lipids. Baseline plasma samples from the Cohort On Diabetes and Atherosclerosis Maastricht (CODAM) were profiled for lipids. 88 individuals were divided into three groups: obese with (n=39) and without (n=39) CVD after 7-year follow-up and lean controls (n=10). Lipidomics analysis identified 868 lipids according to their exact mass-to-charge ratios and retention time. Not all lipids were detectable in all samples, resulting in 32% of missing values (MVs).  Using a novel machine learning approach, we produced 160.000 datasets on 6 imputation value thresholds (ranging from 20 to 70% imputation of MVs) and assessed these with logistic regression. Imputation of MVs did not introduce bias in our dataset, as the AUCs generated remained similar (20% MVs included, AUC:0.53. 60% MVs included, AUC:0.55). To avoid overfitting, we performed feature selection on all imputed datasets. Univariate feature selection reduced the feature set to an average of 50 features, dramatically improving the prediction model (average AUC:0.89). Subsequent multivariate feature selection further reduced the dataset to an average of 20 lipids mostly composed of triglycerides, sphingolipids and ceramides* (average AUC:0.96). Importantly, our model significantly outperforms classical CVD risk prediction in the obese population (using systolic blood pressure, total cholesterol and fasting plasma glucose levels, AUC:0.53). Taken together, we have identified a lipid signature with unprecedented high predictive power for cardiovascular events in obese adults.

04 - BARCODING IN OBESE: WHO’S AT INCREASED RISK FOR THE DEVELOPMENT OF CARDIOVASCULAR DISEASE?