Combination approaches improve predictive performance of diagnostic rules for mass-spectrometry proteomic data

We consider a proteomic mass spectrometry case-control study for the construction of a diagnostic rule for patients' disease status allocation. We propose an approach for combining a collection of classifiers for the construction of a "combined" classification rule in order to enhance calibration and prediction ability. In a first stage this is achieved by building individual classifiers separately, each one using the entire proteomic data set. A double leave-one-out cross-validatory approach is used to estimate the class-predicted probabilities on which the combination method will be calibrated. The performance of the combination approach is examined both through a breast cancer proteomic data set and through simulation studies. Our experimental results indicate that in many circumstances gains in classification performance and predictive accuracy can be achieved.