AbstractWe propose an ECG denoising method based on a feed forward neuralnetwork with three hidden layers. Particulary useful for very noisy signals,this approach uses the available ECG channels to reconstruct a noisychannel. We tested the method, on all the records from Physionet MIT-BIH Arrhythmia Database, adding electrode motion artifact noise. Thisdenoising method improved the perfomance of publicly available ECGanalysis programs on noisy ECG signals. This is an oine method thatcan be used to remove noise from very corrupted Holter records. 1 Introduction The ECG is often corrupted by di erent types of noise, namely, power line inter-ference, electrode contact and motion artifacts, respiration, electrical activityof muscles in the vicinity of the electrodes and interference from other elec-tronic devices. Analysis of noisy ECGs is dicult for humans and for computerprograms. In this work we place ourselves in context of automatic and semiautomatic ECG analysis: denoising should facilitate automatic ECG analysis.General denoising signal processing methods have been applied to ECG. Lowpass linear lters are used for high frequency noise removal, namely power lineinterference and muscle activity artifacts. High pass linear lters can be appliedto cancel baseline wander. The use of neural networks to ECG denoising hasbeen, to our knowledge, limited to the removal of these two types of noise. Otherdenoising tools are median lter, wavelet transform methods, empirical modedecomposition, morphological lters, non linear bayesian ltering and templatematching. We will focus on noise introduced by electrode motion which causesmore diculties in ECG analysis[16]. Our method adapts to each particularECG channel and learns how to reproduce it from a noisy version of the di erentchannels available.In the Physionet/Cinc Challenge 2010 it was shown that we can use somephysiological signals to reconstruct another physiological signal, in particular anECG [17, 20, 21]. This approach to reconstructing the noisy ECG channel is asimpli ed version, but equally e ective, of the winning entry in that Challenge.1