Cardiopulmonary Resuscitation (CPR) introduces strong interference on the surface ECG, corrupting any diagnosis of the cardiac rhythm. In this work a filtering method is proposed to remove the CPR artefact from Ventricular Fibrillation (VF) signals. It is a notch filter adapted to the fundamental frequency of the CPR artefact. It is designed using the mean time interval between compressions characteristic of each CPR artefact. The method has been tested with a set of 17 artefact signals and 191 VF signals, recorded in real out-of-hospital assistances. For the different corruption levels tested (i.e. SNR of -10, -6, -3, 0, 3 and 3 dB), the mean SNR values restored are: -3.16, 0.39, 2.77, 4.79 and 6.38 dB. The positive effect of the filtering on the rhythm classification algorithm of an AED has been evaluated. The sensitivities values after filtering are: 52.26%, 79.3%, 90.48%, 94.91% and 96.64% respectively