Monitoring of neural function in newborns may enable the early prediction of the long-term neurodevelopmental outcome of infants. Artefacts, like eye movements or muscular artefacts are common during long-term recording of neural activity and may lead to erroneous results. Fourier analysis, wavelet analysis and principal component analysis (PCA) are some of major approaches used to extract the artefacts from the time series. The limitation of stationarity, time resolution (wavelet) or the lack of sufficient number of sources (PCA) are some of the motivations for the use of new analysis techniques for the identification and classification of such artefacts prior to further signal processing. In this paper we have developed and studied the effectivness of an Empirical Mode Decomposition (EMD) -based method for online processing and characterisation of EEG data contaminated with different types of artefacts like ocular and muscle artefacts or power line noise. The novel approach allows the detection of different types of artefacts based on the characterization of the intrinsic oscillatory modes, adaptively extracted from the EEG signal. The performance of the model, as concerns the computational time, allows the real time processing and classification up to 4 channel EEG data acquired at 256Hz (the time needed for the decomposition of a 2048 samples data segment is in average 0.1s on a Pentium4 at 3GHz PC).