Search

The high dimensionality and noisy spectra of Mass Spectrometry (MS) data are two of the main challenges to achieving high accuracy recognition. The objective of this work is to produce an accurate prediction of class content by employing compressive sensing (CS). Not only can CS significantly reduce MS data dimensionality, but it will also allow for full reconstruction of original data. We are proposing a weighted mixing of L1- and L2-norms via a regularization term as a classifier within compressive sensing framework. Using performance measures such as OSR, PPV, NPV, Sen and Spec, we show that the L2-algorithm with regularization terms outperforms the L1-algorithm and Q5 under all applicable assumptions. We also aimed to use Block Sparse Bayesian Learning (BSBL) to reconstruct the MS data fingerprint which has also shown better performance results that those of L1-norm. These techniques were successfully applied to MS data to determine patient risk of prostate cancer by tracking Prostate-specific antigen (PSA) protein, and this analysis resulted in better performance when compared to currently used algorithms such as L1 minimization. This proposed work will be particularly useful in MS data reduction for assessing disease risk in patients and in future personalized medicine applications.