Editorial: Explainable and advanced intelligent processing in the brain-machine interaction

COPYRIGHT © 2023 Xie, Zhang, Yu, Duan, Daly and He. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Editorial: Explainable and advanced intelligent processing in the brain-machine interaction

Editorial on the Research Topic Explainable and advanced intelligent processing in the brain-machine interaction Brain-machine interfaces (BMIs), also known as brain-computer interfaces (BCIs), allow their users to control external devices directly using brain signals without relying on the peripheral nervous system and muscles (Wolpaw et al., 2002).As a new bioengineering technology, BCI has a great potential in motor function enhancement to help patients disabled by diseases, such as stroke.In particular, BCI systems with interactive brain stimulation, such as deep brain stimulation (DBS), provide an active way to reveal relationships underlying the interplay between body control and brain activities, as well as to better understand pathological mechanisms underlying diseases such as Parkinson's disease (PD) and essential tremor (ET).While developing such advanced BCI systems, one of the major challenges is to apply explainable and advanced intelligent processes to decode the information embedded in brain signals such as EEG and functional Near-Infrared Spectroscopy (fNIRS) recorded non-invasively from scalp, or local field potential (LFP) recorded invasively from cortical (e.g., motor cortex) or subcortical (e.g., thalamus) brain structures (Heldman et al., 2006;Opri et al., 2020;He et al., 2021).
The goal of this Research Topic for the Brain-Computer Interface section of Frontiers in Human Neuroscience is to collect the current developments of explainable and advanced intelligent methods in BCIs based on EEG, fNIRS, or LFPs.With this aim, we collected five original research articles focusing on different aspects in developing an explainable BCI, including artifact rejection, feature extraction, classification using explainable algorithms, and hyper-parameter tuning.
Artifact removal is a common topic in the BCI community.In particular, it becomes more challenging to remove artifacts such as electrooculogram (EOG) when the number of recorded EEG signals is limited.To deal-with the over-complete issue while applying independent component analysis (ICA) to remove EOG artifacts from single-channel EEG recordings, Hu et al. proposed a method called DWT-CEEMDAN-ICA, in which a complete empirical mode method proposed by Torres et al. (2011) that can adapt to noise (CEEMDAN) is used to decompose the discrete wavelet transformation output of the raw EEG signals, ICA is then applied on the decomposed intrinsic mode functions (IMFs) to identify and remove EOG artifacts.This approach might be of interest to those dealing with EOG artifacts in EEG recordings, especially for those focusing on single-channel EEG-based BCIs, further evaluation of the effectiveness of this approach on bigger dataset is still needed though.
To improve the generalizability of motor imagery (MI)-based BCIs, Wang et al. proposed an EEG joint feature classification algorithm based on instance transfer and ensemble learning.In their method, spatial and frequency domain features are extracted through common spatial pattern (CSP, Ramoser et al., 2000) and power spectral density (PSD) analysis, then the MI classification is achieved using an ensemble learning algorithm based on kernel mean matching (KMM, Huang et al., 2007) and adaptive enhancement of transfer learning (TrAdaBoost, Dai et al., 2007).Experimental results using BCI Competition IV Dataset 2a and 2b showed that their method achieved higher decoding accuracies compared with some state-of-the-art methods, demonstrated its effectiveness.However, as the authors acknowledged, this method may not be applicable to other EEG data such as the P300 eventrelated potential.
Recently, researchers interested in BCI studies have paid great attention to deep leaning-based approaches, in particular, convolutional neural networks (CNN)-based methods.However, there tends to be a trade-off between the accuracy and the interpretability of the trained models, and in many BCI applications, the former is more important.Focusing on this tradeoff, Shibu et al. proposed an explainable artificial intelligence (xAI) system that attempts to decompose the CNN model's output onto the input variables (i.e., channels) of fNIRS signals recorded during motor execution or motor imagery.Specially, a method called DeepShap is applied to compute the shapley values, which are further used to explain the model's output (Lundberg and Lee, 2017;Alsuradi et al., 2020).In contrast, Rodriguez et al. took another path and systematically compared the accuracy of decoding movement states and the interpretability between endto-end CNN-based methods and feature-based methods, i.e., a support vector machine (SVM), using both simulated and real LFP data recorded from ET patients.The synthetic data consisted of 7 different oscillatory patterns including power changes in the beta and gamma bands, beta waveform sharpness, nonlinear phase, theta-gamma phase amplitude coupling (PAC), crosschannel phase shifts, and beta burst length.The LFP data were recorded from the bilateral ventra intermediate (VIM) nucleus of the thalamus of ET patients while they perform self-paced upper limb movement tasks (He et al., 2021).The experimental results suggested that end-to-end deep leaning-based methods can yield This study provides a promising solution to deal with the challenges surrounding hyper-parameter tuning while developing BCI systems.
We believe explainable and advanced intelligent processing is an important perspective while developing BCI systems, which can help to improve the performance and interpretability of the developed systems, and will further facilitate the understanding of disease mechanisms (Figure 1).In collating a few good examples of this, we hope this special topic can act as a resource for those interested in this topic, trigger further discussion, and eventually push forward development in this area.

FIGURE
FIGUREImproving brain-computer interface (BCI) performance and facilitating the understanding of disease mechanisms using explainable intelligent algorithms.