Biol. Cybern. 91, 63–75 (2004) DOI 10.1007/s00422-004-0500-8 © Springer-Verlag 2004 An adaptive neuro-fuzzy method (ANFIS) for estimating single-trial movement-related potentials D. D. Ben Dayan Rubin 1,2 , G. Baselli 1 , G. F. Inbar 2 , S. Cerutti 1 1 Department of Biomedical Engineering, Politecnico di Milano, 20133 Milano, Italy 2 Department Electrical Engineering, Technion, Israel Institute of Technology, 32000 Haifa, Israel Received: 22 July 2002 / Accepted: 22 June 2004 / Published online: 21 August 2004 Abstract. This study aims to recover transient, trial- varying evoked potentials (EPs), in particular the move- ment-related potentials (MRPs), embedded within the background cerebral activity at very low signal-to-noise ratios (SNRs). A new adaptive neuro-fuzzy technique will attempt to estimate movement-related potentials within multi-channel EEG recordings, enabling this method to completely adapt to each input sweep without system training procedures. We assume that one of the sensors is corrupted by noise deriving from other sensors via an un- known function that will be estimated. We will approach this problem by: (1) spatially decorrelating the sensors in the preprocessing phase, (2) choosing the most infor- mative of the filtered channels that will permit the best MRP estimation (input-selection phase) and (3) train- ing the neuro-fuzzy model to fit the noise over the cho- sen sensor and therefore estimating the buried MRP. We tested this framework with simulations to validate the ana- lytical results before applying them to the real biologi- cal data. Whenever it is applied to biological data, this method improves the SNR by more than 12dB, even to very low SNRs. The processing method proposed here is likely to complement other estimation techniques and can be useful to process, enhance and analyse single-trial MRPs. Keywords: Movement-related-potential – Single-trial estimation – Nonlinear estimation methods – Evoked potentials – Neuro-fuzzy models 1 Introduction When approaching the problem of analysing evoked potentials (EPs) or event-related potentials (ERPs), two major issues arise. The first issue stems from the extremely low signal-to-noise ratio (SNR) with overlapping spectra Correspondence to: D. B. D. Rubin (e-mail: dbd@biomed.polimi.it) of the evoked response embedded within the background EEG brain activity, ranging from 0 to −20 dB, depending on the type of evoked signals. The second one concerns possible vectorial signal summation, resulting in compo- nent overlap, which may cause partial or total occlusion of the desired features. Usually these field potentials are aver- aged to increase the SNR and other phase-locked EEG activity. The averaging methods do not take into account that in single epochs response activity may vary widely in both time course and scalp distribution (Popivanov and Krekule 1983), depending on the external experimental conditions as well as on the subject’s performance and state of mind (Schwent and Hillyard 1975). Single-trial- analysis methods can avoid problems due to time and/or phase shifts and can potentially reveal richer information about event-related brain dynamics. On the other hand, these methods suffer from pervasive artefacts associated with blinks, eye movements, muscle noise and SNR aris- ing from the fact that non-phase-locked background EEG activities often are larger than phase-locked response com- ponents. The focus of this study deals with recovering specifically movement-related potentials (MRPs); never- theless the proposed methodology can be applied gener- ally to ERP estimations. The movement potentials relate to the planning and execution of voluntary movements (Boschert and Deecke 1986) and have usually been studied in the context of simple movements, commonly of single limbs (Deecke et al. 1976). Many of the studies address the focus on MRPs because of their importance in both clin- ical and research purposes. Many algorithms have been proposed to detect trial-to-trial variability (Bartnik et al. 1982; Birch et al. 1993; Thakor 1993). Most of these algo- rithms work well for cognitive evoked potentials with a non-negative SNR but fail whenever applied to MRPs that have a highly negative SNR. The few algorithms that succeed in these cases of highly negative SNR have to rely heavily on the average MRP (Lange et al. 1997; Cerutti et al. 1988). Other works investigate the use of physiological signals, usually from multi-electrode EEG, for communication and operation of devices for both healthy subjects and patients with severe motor impair- ments in many international groups (Birbaumer et al.