Stacked Feature Selection in Liver Disease Using IMR-MS Analysis

The combination of different feature selection approaches has shown to produce feasible feature subsets with high predictive value. We here introduce a modified version of Stacked Feature Ranking (SFR), using a two level learning architecture with a suggestion and a decision layer aggregating different feature selectors to a consensus feature ranking. Ion molecule reaction mass spectrometry (IMR-MS) was applied to breath gas samples of a total of 57 patients suffering from alcoholic fatty liver disease (AFLD) and nonalcoholic fatty liver disease (NAFLD), and 35 healthy controls with the objective of identifying breath gas marker candidates at disease versus non-disease state. We compared SFR with four common feature selection methods and one ensemble-based approach, indicating a significantly higher discriminatory ability of up to 10% for the selected subsets using ROC analysis. SFR is a powerful tool for the identification of highly discriminating biomarkers in complex biological mixtures.