Respiratory gating and motion correction can increase resolution in PET chest imaging, but require a respiratory signal. Data-Driven (DD) methods aim to produce a respiratory signal from PET data, avoiding the use of external devices. Principal Component Analysis (PCA) is an easy to implement DD algorithm whose signals, however, are determined up to an arbitrary factor. The direction of the motion represented by its signal has to be determined. In this work we present the extension to TOF data of a previously presented sign-determination method. Furthermore, we propose the application of a selection process in sinogram space, to automatically select the areas of the data mostly affected by respiratory motion. The performance of the updated signdetermination method is evaluated on patient data, and the effect of TOF information and masking process is investigated also in terms of quality of the PCA respiratory signal.