ERP Detection Based on Smoothness Priors

Abstract

Detection of event-related potentials (ERPs) in electroencephalography (EEG) is of great interest in the study of brain responses to various stimuli. This is challenging due to the low signal-to-noise ratio of these deflections. To address this problem, a new scheme to detect the ERPs based on smoothness priors is proposed. Methods: The problem is considered as a binary hypothesis test and solved using a smooth version of the generalized likelihood ratio test (SGLRT). First, we estimate the parameters of probability density functions from the training data under Gaussian assumption. Then, these parameters are treated as known values and the unknown ERPs are estimated under the smoothness constraint. The performance of the proposed SGLRT is assessed for ERP detection in poststimuli EEG recordings of two oddball settings. We compared our method with several powerful methods regarding ERP detection. Results: The presented method outperforms the competing algorithms and improves the classification accuracy. Conclusion:The proposed SGLRT could be employed as a powerful means for different ERP detection schemes. Significance: ERP-based systems (e.g. brain-machine interfaces) mainly suffer from the lack of classification accuracy, hence, the proposed method is an important step toward real-life applicability of these systems.

Improving the performance of P300-based BCIs by mitigating the effects of stimuli-related evoked potentials through regularized spatial filtering

“…Another simple strategy to enhance ERP is to pass the data through a bandpass filter (BPF). ERPs have slower fluctuations compared to wide-band background EEG (0-60 Hz) [10]. Therefore, applying a BPF to raw data can improve the SNR.…”

“…They demonstrated that the whitening of EEG data can improve P300 detection in a subsequent phase. In [10], the authors formulated a P300 detection method based on temporal smoothness priors, enabling the incorporation of spatial distribution of P300 signals through a structural matrix as well. In [19], the authors proposed a modified spatiotemporal filter that extracted the data windows along both space and time dimensions to maximize the P300 SNR.…”