Feature Extraction for Cancer Prediction by Tensor Decomposition of 1D Protein Expression Levels

Ivica Kopriva, Ante Jukić, and Andrzej Cichocki


Cancer prediction, mass spectrometry, feature extraction, tensor decomposition, pattern recognition


Tensor decomposition approach to feature extraction from one-dimensional data samples is presented. One-dimensional data samples are transformed into matrices of appropriate dimensions that are further concatenated into a third order tensor that is factorized according to the Tucker-2 model by means of the higher-order-orthogonal iteration (HOOI) algorithm. Derived method is validated on publicly available and well known datasets comprised of low-resolution mass spectra of cancerous and non-cancerous samples related to ovarian and prostate cancers. The method respectively achieved, in 200 independent two-fold cross-validations, average sensitivity of 96.8% (sd 2.9%) and 99.6% (sd 1.2%) and average specificity of 95.4% (sd 3.5%) and 98.7% (sd 2.9%). Due to the widespread significance of mass spectrometry for monitoring protein expression levels and cancer prediction it is conjectured that presented feature extraction scheme can be of practical importance.

Important Links:

Go Back