Metabolomics in the prediction of prodromal stages of carotid artery disease using a hybrid ML algorithm

Carotid artery disease (CAD) may be responsible for a stroke with fatal consequences for the patients. Early and non-invasive diagnosis and prediction of significantly high carotid intima media thickness (IMT) can reduce the death rates caused by cardiovascular disease. Machine learning can be applied for the development of robust models for this purpose when adequate data are available. In this work, we utilized metabolomics data from 2,147 patients in the Young Finns Study clinical trial to predict the high intima media thickness as a prodromal stage of the atherosclerotic carotid disease. An explainable AI based pipeline was developed which includes a novel employment of the Gradient Boosted Trees (GBT). More specifically, a hybrid loss function was used to adjust the effect of the dropout rates in the 'dart' booster in the loss function topology. The results of our analysis demonstrate that the novel implementation of the GBT improves the results in terms of the sensitivity which is the most important requirement to our analysis (accuracy 0.80, sensitivity 0.86, AUC 0.85). Moreover, it is shown that metabolomics can be used to increase sensitivity in predicting the increased IMT.