Alkhoury L, Scanavini G, Louviot S, Radanovic A, Shah S A, Hill N J
Department of Radiology, Weill Cornell Medicine, New York, NY 10065, United States of America.
National Center for Adaptive Neurotechnologies, Stratton VA Medical Center, Albany, NY, United States of America.
J Neural Eng. 2025 Jun 26;22(3). doi: 10.1088/1741-2552/ade566.
. We present a novel and lightweight method that removes ocular artifacts from electroencephalography (EEG) recordings while demanding minimal training data.. A robust, cross-validated thresholding procedure automatically detects the times at which eye blinks occur, then a linear scalp projection is estimated by regressing a simplified time-locked reference signal against the multi-channel EEG.. Performance was compared against four commonly-used and readily available blink removal methods: signal subspace projection and forward regression (Reg) from the MNE toolbox, EEGLab's independent component analysis (ICA) combined with ICLabel for automated component identification, and Artifact Subspace Reconstruction (ASR) Python implementation compatible with MNE. On semi-synthetic blink-contaminated EEG data, our method exhibited better reconstruction of the ground truth than the two MNE native methods, and comparable (or better in some scenarios) performance to ASR algorithm and ICA+IClabel. We also examined a real EEG dataset from 16 human participants, where the ground truth was unknown. Our method affected contaminated channels in blink intervals more than the two MNE native methods and ASR, while having a smaller impact on non-blink intervals, uncontaminated channels, and higher-frequency amplitudes, than the two MNE methods; its performance was again similar to ICA+ICLabel. On a second real dataset from 42 human participants, we showed that ARMBR removed the unwanted blink artifacts while successfully preserving the desired event-related-potential signals.. The proposed algorithm exhibited comparable, and in some scenarios better performance relative to readily-available implementations of other widely-used methods. Another feature of our method is its potential as method for online applications. Therefore, it stands to make valuable contributions towards the automation of neural-engineering technologies and their translation from laboratory to clinical and other real-world usage.
我们提出了一种新颖且轻量级的方法，该方法能从脑电图（EEG）记录中去除眼部伪影，同时所需的训练数据极少。一种稳健的、经过交叉验证的阈值化程序会自动检测眨眼发生的时间，然后通过将简化的锁时参考信号与多通道EEG进行回归来估计线性头皮投影。将该方法的性能与四种常用且易于获得的眨眼去除方法进行了比较：MNE工具箱中的信号子空间投影和正向回归（Reg）、EEGLab的独立成分分析（ICA）结合ICLabel进行自动成分识别，以及与MNE兼容的Artifact Subspace Reconstruction（ASR）Python实现。在半合成的受眨眼污染的EEG数据上，我们的方法比两种MNE原生方法能更好地重建真实情况，并且在性能上与ASR算法和ICA + ICLabel相当（或在某些情况下更好）。我们还研究了来自16名人类参与者的真实EEG数据集，其中真实情况未知。我们的方法对眨眼间隔中受污染通道的影响比两种MNE原生方法和ASR更大，而对非眨眼间隔、未受污染通道和高频幅度的影响比两种MNE方法更小；其性能再次与ICA + ICLabel相似。在来自42名人类参与者的第二个真实数据集上，我们表明ARMBR去除了不需要的眨眼伪影，同时成功保留了所需的事件相关电位信号。所提出的算法与其他广泛使用方法的现有实现相比，表现出相当的性能，并且在某些情况下更好。我们方法的另一个特点是其作为在线应用方法的潜力。因此，它有望为神经工程技术的自动化以及从实验室到临床和其他实际应用的转化做出有价值的贡献。
