In this paper a new version, more immune to noise, of the nonlinear projective filtering is presented. The method employs an algorithm of robust principal component analysis to signal subspaces construction and as a result it achieves high performance in real electromyographic noise environment. Two aspects of the method's action are investigated: its ability to suppress noise and its influence on the precision of the QT interval measurement. Then the method influence on evaluation of the beat-to-beat variability of the repolarization duration is presented. A comparison to the previous versions of the nonlinear projective filtering and to the classical linear one is carried out.