Blind Source Separation (BSS) approaches for multi-channel EEG processing are popular, and in particular Independent Component Analysis (ICA) algorithms have proven their ability for artefacts removal and source extraction for this very specific class of signals. However, the blind aspect of these techniques implies well-known drawbacks. As these methods are based on estimated statistics from the data and rely on an hypothesis of signal stationarity, the length of the window is crucial and has to be chosen carefully: large enough to get reliable estimation and short enough to respect the rather non-stationary nature of the EEG signals. In addition, another issue concerns the plausibility of the resulting separated sources. Indeed, some authors suggested that ICA algorithms give more physiologically plausible results than others. In this paper, we address both issues by comparing four popular ICA algorithms (namely FastICA, Extended InfoMax, JADER and AMICA). First of all, we propose a new criterion aiming to evaluate the quality of the decorrelation step of the ICA algorithms. This criterion leads to a heuristic rule of minimal sample size that guarantees statistically robust results. Next, we show that for this minimal sample size ensuring constant decorrelation quality we obtain quasi-constant ICA performances for some but not all tested algorithms. Extensive tests have been performed on simulated data (i.i.d. sub and super Gaussian sources mixed by random mixing matrices) and plausible data (macroscopic neural population models placed inside a three layers spherical head model). The results globally confirm the proposed rule for minimal data length and show that the use of sphering as decorrelation step might significantly change the global performances for some algorithms.