Denoising acoustic-induced vibration artifact in intracranial EEG recordings via a phase-coupling decomposition method

VICTORIA PETERSON; MATTEO VISSANI; SHIYU LUO; QINWAN RABBANI; NATHAN E. CRONE; ALAN BUSH; MARK RICHARDSON

Bruselas

Congreso; 10th International Brain-Computer Interface Meeting 2023; 2023

BCI Society

Introduction: Intracranial electroencephalography (iEEG) recordings offer enhanced characterization in thespatial, temporal, and spectral domains of the neuronal populations supporting cognition, language andspeech. Most of the brain-computer interfaces (BCIs) for speech prosthesis are based on iEEG signaldecoding. Nevertheless, it has been shown that acoustic-induced vibration artifacts may affect up to 50%of the iEEG channels during recordings of overt-speech tasks [1]. Thus, there is a need to remove acousticinducedartifacts from iEEG signals. In this work, we present a denoising method - phase couplingdecomposition (PCD) - for artifact removal of acoustic-induced vibrations. The artifactual iEEG recordingsshow a high phase-coupling coherence in the -band (70 – 250 Hz) with respect to the produced audio [1].Thus, PCD seeks statistical components with the highest phase-coupling values with respect to the acousticsignal. Here we validate PCD as a valuable pre-processing tool for speech decoding from neural activity.Material, Methods and Results: PCD is a data-driven spatial filtering denoising method based on low-rankfactorization. Spatio-spectral decomposition (SSD) [2] is first used to enhance signal-to-noise ratio aroundthe -band and to reduce dimensionality. Phase-coupling optimization (PCO) [3] is then applied to identifysources phase-locked to the acoustic signal. Data consisted of iEEG recordings from 54 patients performinga syllable triplet repetition task [1]. Data cleaning was assessed based on the percentage of cleanelectrodes with respect to raw data (% gain). Common average reference (CAR) and independentdecomposition analysis (ICA) were applied for comparison. The effect of applying each denoising method(CAR, ICA, PCD) as a pre-processing step in a BCI deep learning model [4] for consonant decoding was alsoevaluated. Results showed that CAR can increase the number of affected electrodes by spreading theartifact presented in “common noise”. Although ICA showed the highest reduction in the number ofartifact-affected channels (% gain ICA = 21.4 > % gain PCD = 14.7), consonant decoding performance wasreduced due to strong degradation of physiological -band modulations.Discussion: While traditional denoising method can jeopardize signal quality, PCD can significantly reducethe strength and extent of the vibration artifact while preserving the underlying neural activity related tospeech.Significance: PCD is the first method specifically designed to denoise acoustic-induced vibration artifactsin brain recordings and can be safely used as a pre-processing step for iEEG-based speech decoding.