Objective For multi-class motor imagery tasks in brain computer interface (BCI), this paper presents a novel recognition method of electroencephalography (EEG) by combining RLS-ICA-SampEn ［RLS (recursive least-squares), ICA (independent component analysis), SampEn (sample entropy)］, multi-class CSP (common spatial patterns) and ISVM (incremental support vector machine). Methods In the RLS-ICA-SampEn, Firstly, the ICA is used to decompose the contaminated EEG signals into independent components (IC). Then, the sample entropy is used to automatically identify the noise signal in the IC. Next, the RLS adaptive filters are applied to the identified noise in IC to remove noise further. Finally, the processed ICs are then projected back to reconstruct the noise-free EEG signals. The RLS-ICA-SampEn is used to preprocess EEG signals to get pure EEG signals, in which some noise signals can be removed. The multi-class CSP combines the CSP and the multi-band filtering technology, in which the CSP uses the ‘one versus one’ strategy. The multi-class CSP is used to extract features for pure EEG signals. The obtained features are input to the ISVM for classification. The ‘one versus rest’ strategy is applied to classify three-class EEG signals. In order to verify the effectiveness of the proposed novel method, it is compared with other two methods including multi CSP+ISVM(method 1), RLS-ICA + multi CSP + ISVM(method 2). ResultsThe result shows that the recognition accuracy obtained by the proposed method is higher about 8% than other two methods. Conclusions Compared with method 1 and 2, the proposed method is better suited for the recognition of multi-class motor imagery tasks in BCI.