FEATURE EXTRACTION FOR CANCER PREDICTION BY TENSOR DECOMPOSITION OF 1D PROTEIN EXPRESSION LEVELS

摘要

Tensor decomposition approach to feature extraction fromrnone-dimensional data samples is presented. Onedimensionalrndata samples are transformed into matrices ofrnappropriate dimensions that are further concatenated intorna third order tensor. This tensor is factorized according tornthe Tucker-2 model by means of the higher-orderorthogonalrniteration (HOOI) algorithm. Derived method isrnvalidated on publicly available and well known datasetsrncomprised of low-resolution mass spectra of cancerousrnand non-cancerous samples related to ovarian and prostaterncancers. The method respectively achieved, in 200rnindependent two-fold cross-validations, averagernsensitivity of 96.8% (sd 2.9%) and 99.6% (sd 1.2%) andrnaverage specificity of 95.4% (sd 3.5%) and 98.7% (sdrn2.9%). Due to the widespread significance of massrnspectrometry for monitoring protein expression levels andrncancer prediction it is conjectured that presented featurernextraction scheme can be of practical importance.