Improved Manual Annotation of EEG Signals through Convolutional Neural Network Guidance

Abstract

Visual Abstract

Robin’s Viewer: Using Deep-Learning Predictions to Assist EEG Annotation

“…This is done to allow for custom preprocessing of the model’s input, independent of the preprocessing applied to the viewed data. In the case of the integrated model by Diachenko et al (2022), for example, this is necessary because this model specifically requires time-frequency plots generated from segments of the time-series EEG data as input, as mentioned in Section 2.2. Note that, because the model is sensitive to bad channels, they should be marked before running the model and evaluating the predictions.…”

“…A convolutional neural network was trained on expert-annotated resting-state EEG data from typically developing children and children with neurodevelopmental disorders, with the purpose of recognizing artifacts (Diachenko et al, 2022). This model was integrated into RV as a proof-of-concept decision-support system for EEG artifact annotation.…”

“…Machine-learning techniques trained to detect artifacts in EEG data can address the aforementioned issues, by providing an objective standard for artifact marking, and speeding up the marking process. The most common machine-leaning techniques used for artifact detection in EEG data, are based on support vector machines (SVMs) (Barua and Begum, 2014; Sai et al, 2017; Shao et al 2008), k-nearest neighbor classifiers (k-NN) (Barua and Begum, 2014; Roy, 2019), independent component analysis (ICA) (Barua and Begum, 2014; Radüntz et al, 2015; Sai et al, 2017), and, as of recently, various deep-learning models (e.g., autoencoders (Roy et al, 2019; Yang et al, 2016; 2018), convolutional neural networks (CNNs) (Diachenko et al, 2022; Jurczak et al, 2022; Roy et al, 2019; Sun et al, 2020), and recurrent neural networks (RNNs) (Liu et al, 2022; Roy et al, 2019)). Deep-learning solutions, in particular, have increased in popularity for artifact handling (Craig et al, 2019; Roy et al, 2019) due to the minimal preprocessing they require and because of their ability to learn very complex functions between input data and the desired output classification (LeCun et al, 2015).…”

“…In RV, the user can load EEG recordings, apply common preprocessing methods (e.g., bandpass filtering, re-referencing, and down-sampling), plot the data, use the GUI to navigate through the signal, and mark bad channels and artifacts. To provide an example of how the viewer can be used with integrated deep-learning predictions, we have additionally included with RV a proof-of-concept deep-learning model trained to recognize artifacts in resting-state, multichannel EEG recordings (Diachenko et al, 2022). RV can be downloaded for free at https://github.com/RobinWeiler/RV.git, where we also provide code documentation, demos, and example data.…”

Robin’s Viewer: Using deep-learning predictions to assist EEG annotation

“…Machine-learning techniques trained to detect artifacts in EEG data can address the aforementioned issues, by providing an objective standard for artifact marking, and speeding up the marking process. The most common machine-leaning techniques used for artifact detection in EEG data, are based on support vector machines (SVMs) (Shao et al, 2008;Barua and Begum, 2014;Sai et al, 2017), k-nearest neighbor classifiers (k-NN) (Barua and Begum, 2014;Roy, 2019), independent component analysis (ICA) (Barua and Begum, 2014;Radüntz et al, 2015;Sai et al, 2017), and, as of recently, various deeplearning models [e.g., autoencoders (Yang et al, 2016(Yang et al, , 2018Roy et al, 2019), convolutional neural networks (CNNs) (Roy et al, 2019;Sun et al, 2020;Diachenko et al, 2022;Jurczak et al, 2022), and recurrent neural networks (RNNs) (Roy et al, 2019;Liu et al, 2022)]. Deep-learning solutions, in particular, have increased in popularity for artifact handling (Roy et al, 2019) due to the minimal preprocessing they require and because of their ability to learn very complex functions between input data and the desired output classification (LeCun et al, 2015).…”