Mechanical activity from chest compressions and ventilations during cardiopulmonary resuscitation (CPR) introduces artefact components into the electrocardiogram (ECG). CPR must therefore be discontinued for reliable shock advice analysis in automated external defibrillators. Reducing or eliminating this detrimental "hands-off" time by removing the CPR artefacts, should significantly improve the defibrillation success rate. The feasibility of this was tested by removing the CPR artefacts using a multichannel adaptive filter, the multichannel recursive adaptive matching pursuit (MC-RAMP) algorithm. Human ECG and reference channel data from episodes with both shockable and non-shockable underlying heart rhythms were recorded from 105 patients with out-of-hospital cardiac arrest. The performance of a shock advice algorithm was evaluated before and after artefact removal using the MC-RAMP algorithm. From a test set consisting of 92 shockable and 174 non-shockable episodes a sensitivity of 96.7% and specificity of 79.9% was achieved, an increase of approximately 15 and 13%, respectively, compared to no filtering. Good sensitivity was achieved, enabling ECG analysis during CPR that would reduce the hands-off time on patients with shockable rhythms. However, CPR artefact removal on non-shockable rhythms proved a more difficult problem. We need a better understanding of the physiological mixing of artefacts and the underlying heart rhythm and suggest clinical trials to investigate the nature of CPR artefacts further.