Model-Based Fitting of EEG Signals With Uncertainty Quantification for Robust Alpha Wave Identification

Electroencephalography (EEG) is a noninvasive technique widely used to assess the brain’s electrical activity, suitable for a variety of applications in clinical and nonclinical environments. However, the reliable identification of specific EEG features still remains a challenge, especially if the signals are recorded from wearable devices, notably less accurate than their hospital counterparts. The uncertainty analysis can help to address this challenge, by providing a robust and rigorous tool to assess the validity of the information extracted from the signals. This is particularly important in automated processes, such as brain–machine interfaces (BMIs), to avoid misclassifications and misinterpretation. This article proposes the use of a model-based fitting of prefiltered EEG signals, combined with uncertainty quantification, to extract the alpha amplitude oscillation with an optimal trade-off between accuracy and time resolution and, importantly, to allow the identification of those parts of the signal that do not follow the expected alpha dynamics, e.g., because affected by artifacts. This is achieved through a metrology-sound analysis of compatibility between measurement and model, taking their respective uncertainties into consideration. The proposed method has been successfully tested on real EEG signals and shown to have significant advantages in terms of time resolution and interpretability, compared with more traditional techniques, such as the independent component analysis (ICA) by temporal decorrelation, especially when applied to single-channel signals.