PaperlayerView exact revision on arXiv

Source-linked AI summary

EEG-based Brain-Computer Interfaces (BCIs): A Survey of Recent Studies on Signal Sensing Technologies and Computational Intelligence Approaches and their Applications

Xiaotong Gu, Zehong Cao, Alireza Jolfaei, Peng Xu, Dongrui Wu, Tzyy-Ping Jung, Chin-Teng Lin

arXiv:2001.11337v1eess.SPcs.AIcs.HC

TL;DR

EEG-based BCIs face variability across subjects, sessions, and headsets, complicating consistent classification and calibration. This paper surveys recent sensing, signal-enhancement, and computational-intelligence advances, finding improved system performance alongside persistent real-world usability challenges.

  • Problem

    EEG distributions vary across subjects, sessions, and time, while headset-independent operation without recalibration remains challenging.

  • Method

    The paper surveys recent dry and wearable sensing, signal enhancement, transfer learning, deep learning, interpretable fuzzy models, and healthcare applications for EEG-based BCIs.

  • Results

    The reviewed advances have lifted EEG-based BCI performance across sensing technologies, signal enhancement, transfer learning, deep learning, and interpretable fuzzy models.

  • Takeaways & Limitations

    The survey identifies computational-intelligence approaches and improved sensing technologies as central components of ongoing EEG-based BCI development and healthcare applications.

  • Takeaways & Limitations

    Real-world usability challenges remain, including prediction or classification capability and stability in complex BCI scenarios.

Abstract

from arXiv · show

Brain-Computer Interface (BCI) is a powerful communication tool between users and systems, which enhances the capability of the human brain in communicating and interacting with the environment directly. Advances in neuroscience and computer science in the past decades have led to exciting developments in BCI, thereby making BCI a top interdisciplinary research area in computational neuroscience and intelligence. Recent technological advances such as wearable sensing devices, real-time data streaming, machine learning, and deep learning approaches have increased interest in electroencephalographic (EEG) based BCI for translational and healthcare applications. Many people benefit from EEG-based BCIs, which facilitate continuous monitoring of fluctuations in cognitive states under monotonous tasks in the workplace or at home. In this study, we survey the recent literature of EEG signal sensing technologies and computational intelligence approaches in BCI applications, compensated for the gaps in the systematic summary of the past five years (2015-2019). In specific, we first review the current status of BCI and its significant obstacles. Then, we present advanced signal sensing and enhancement technologies to collect and clean EEG signals, respectively. Furthermore, we demonstrate state-of-art computational intelligence techniques, including interpretable fuzzy models, transfer learning, deep learning, and combinations, to monitor, maintain, or track human cognitive states and operating performance in prevalent applications. Finally, we deliver a couple of innovative BCI-inspired healthcare applications and discuss some future research directions in EEG-based BCIs.

1 INTRODUCTION

BCI provides a direct communication pathway between brain activity and external systems, with EEG supporting non-invasive, portable measurement for interaction and cognitive-state monitoring. This survey addresses gaps in recent reviews by integrating sensing technologies, signal enhancement, computational intelligence, applications, and healthcare developments.

  • What is BCI: BCI translates brain activity into commands or implicit information for controlling external devices and enriching human-computer interaction.
  • Application areas: BCI applications span human-computer interaction, entertainment, health, assistive systems, and monitoring cognitive states during demanding activities.
  • Brain imaging techniques: EEG is a widely used non-invasive technique that directly measures cortical electrical activity with high temporal resolution and portable headset implementations.EEG headsets can capture multiple frequency bands associated with different behavioural states.
  • Research gap: Recent reviews lacked a comprehensive synthesis covering EEG sensing, signal enhancement, interpretable fuzzy models, deep learning, specific applications, and healthcare systems.The survey includes recently released studies from 2019.
  • Contributions: The survey reviews advances in sensors, online signal processing, machine learning and interpretable fuzzy models, deep learning combinations, and healthcare applications.

2 ADVANCES IN SENSING TECHNOLOGIES

Recent sensing advances aim to make EEG acquisition more wearable, wireless, portable, and usable in everyday or medical settings. Dry and noncontact sensors reduce setup burdens while retaining competitive task performance, and available headsets differ in channels, sampling, mobility, and connectivity.

  • Sensing technologies: Advanced sensing technologies have enabled smaller, smarter, wearable, and wireless EEG devices using wet, dry, and noncontact sensors.The survey compares devices by channels, sampling rate, portability, and company-specific features.
  • Wet sensor technology: Wet electrodes provide a clean conductive path through gels, but gel application can be uncomfortable, inconvenient, time-consuming, and laborious for everyday use.Removing the gel can compromise EEG signal quality because electrode-skin impedance cannot be measured.
  • Dry sensor technology: 74.23% averaged SSVEP-detection accuracy across all subjects was comparable between proposed dry sensors and commercially available Cognionics sensors.
  • Dry sensor technology: 100% accuracy was achieved as the maximum information transfer rate in an experiment using data collected from a noncontact electrode positioned on top of the hair.The result supports the reported promise of dry and noncontact electrodes for mobile BCI applications.
  • Augmented BCIs: Augmented BCIs use non-intrusive, quick-setup EEG solutions with real-time biosignal processing for comfortable, stable, robust, and long-term monitoring.
  • Device comparison: The reviewed EEG headsets vary in wearability, wireless or tethered transmission, and channel count, which affect how users can move while monitored.

3 SIGNAL ENHANCEMENT AND ONLINE PRO-

EEG signal enhancement combines blind source separation, component-based artifact correction, and online processing to clean recordings for BCI analysis. The section emphasizes ICA-based removal, ASR limitations, and hybrid real-time methods.

  • Artifact removal: Blind Source Separation estimates original sources and mixing parameters to remove artifacts such as eye blinks and movement.Prevalent algorithms include PCA, CCA, and ICA.
  • Artifact removal: ICA decomposes observed EEG into independent components and reconstructs cleaner signals by removing artifact-containing components.The survey identifies ICA as the predominant approach for EEG artifact removal.
  • Artifact removal: ASR removes large-amplitude or transient artifacts but is not applicable to single-channel recordings and has limited effectiveness for regularly occurring eye artifacts without suitable cutoffs.ICA-based removal is proposed as a complement to ASR.
  • Online processing: Online ASR, recursive ICA, and an IC classifier can jointly remove transients and classify artifact components during near-real-time EEG processing.The altered EyeCatch measure remained imperfect for eliminating eye-movement artifacts.
  • Artifact removal: Muscle artifacts arise from nearby contractions and are commonly addressed with regression, CCA, EMD, BSS, or EMD-BSS methods.Talking, sniffing, and swallowing can generate these artifacts, whose amplitude and waveform vary with muscle activity.
  • Implementation: EEGLAB and related toolboxes provide interactive or automatic preprocessing functions, including ICA, ASR, ARfit, and ADJUST-based artifact correction.These tools support cleaning continuous data and removing artifact independent components.

4.1 An overview of machine learning

Machine learning methods for EEG-based BCIs include supervised and unsupervised paradigms, with models trained on extracted EEG features after preprocessing. Common model families range from linear classifiers to neural, Bayesian, nearest-neighbor, and combined classifiers.

  • Learning paradigms: Supervised learning trains models using training and test subsets, whereas unsupervised learning operates without classified or labelled training data.Unsupervised methods describe hidden structures through grouping or clustering.
  • Model families: EEG-based BCIs commonly use linear classifiers, neural networks, non-linear Bayesian classifiers, nearest-neighbour classifiers, and classifier combinations.Examples of linear classifiers include LDA, Regularized LDA, and Support Vector Machines.
  • Processing pipeline: Machine-learning pipelines preprocess EEG signals and extract features such as frequency-band power and connectivity between channels before pattern recognition.Figure 4 presents this processing sequence as a compact EEG-based BCI pipeline.

4.2 Transfer learning

Transfer learning addresses EEG distribution differences across subjects, sessions, tasks, and headsets by reusing knowledge from related source settings. The survey describes transfer categories, domain transformations, calibration reduction, and challenges caused by task and domain mismatch.

  • Why transfer learning: Transfer learning addresses violations of the shared-distribution assumption caused by human variability in EEG data.It reuses knowledge from a related task to improve a learned classifier for another task, session, or subject.
  • Transfer settings: Transfer learning is categorized into inductive, transductive, and unsupervised settings according to source and target tasks, domains, and label availability.Inductive transfer may use multitask or self-taught learning, while transductive transfer includes sample selection bias and domain adaptation.
  • Transfer settings: BCI transfer learning can move information across tasks, subjects, or sessions by finding a transformation space that benefits prediction in a new dataset.Effectiveness depends strongly on how related the source and target circumstances are.
  • Task transfer: Task-to-task transfer can alter EEG feature distributions and classifier performance because mental and operational tasks may differ while remaining dependent.Reported accuracy based on the baseline descended when operational tasks and subtasks were generalised.
  • Subject transfer: Cross-subject transfer can reduce data-collection time and the number of SSVEP training templates, but conventional approaches require pilot data because of inter-subject variability.Subject-to-subject transfer among the same tasks is reported as the most frequently investigated form.
  • Subject transfer: Feature weighted episodic training eliminates the need for labelled or unlabelled calibration data from a new subject in subject-to-subject EEG drowsiness estimation.FWET combines feature weighting with episodic training for domain generalization and is presented as useful for plug-and-play BCIs.
  • Headset transfer: Headset-independent BCI remains challenging, although historical data from the same user can reduce calibration effort when changing EEG headsets.The goal is to allow headset replacement or upgrading without recalibration.

4.3 Interpretable Fuzzy Models

Interpretable fuzzy models are presented as alternatives to opaque machine-learning systems for understanding EEG-based BCI decisions. The survey covers fuzzy sets, fuzzy inference rules, fuzzy integrals, and hybrid fuzzy-neural architectures across several applications.

  • Motivation: Fuzzy models are introduced to improve understanding of BCI systems that otherwise behave like unexplained black boxes.Interpretability may support understanding and improvement of automatically learned EEG-based BCIs.
  • Fuzzy sets: Fuzzy sets represent gradual category boundaries through membership functions rather than strict true-or-false values.Their flexible boundary conditions support applications in BCI.
  • Fuzzy inference: Fuzzy Inference Systems extract interpretable If-Then rules linking input feature values to output categories for EEG-pattern classification.Fuzzy integrals support data fusion when interactions among multiple information sources must be considered.
  • Hybrid models: Fuzzy neural networks combine neural-network learning with fuzzy inference by fuzzifying inputs or weights and tuning membership functions through connection weights.SONFIN uses a dynamic self-adaptive architecture to identify fuzzy models while retaining neural learning capability.
  • BCI applications: Fuzzy models have been applied to EEG classification, regression, reaction-time estimation, calibration reduction, brain-state-drift detection, and motor-imagery BCI.Reported approaches combine fuzzy sets, Riemannian features, domain adaptation, fuzzy rules, or fuzzy integrals.
  • BCI applications: Fuzzy-neural methods have been used to identify sleep stages and predict EEG-based driving fatigue.Examples include fuzzy C-means, Takagi-Sugeno models, and recurrent self-evolving fuzzy neural networks.

5 DEEP LEARNING ALGORITHMS WITH BCI AP-

Deep learning jointly learns features and classifiers from data, with CNNs applied broadly to EEG-based BCI tasks because static-data methods are poorly suited to rapidly changing brain signals.

  • Deep learning jointly learns features and classifiers through cascaded trainable feature extractors and nonlinearities.Representative architectures include CNN, GAN, RNN, and DNN.
  • CNNs use convolutional, pooling, and fully connected layers to reduce EEG inputs while capturing distinctive spatial dependencies.CNN applications include automatic feature extraction and diagnosis from epileptic intracortical data.
  • CNN-based EEG BCIs have been applied to fatigue detection, stress recognition, sleep-stage classification, motor imagery, and emotion recognition.These applications address sustained attention, health monitoring, movement imagery, and affective-state classification.
  • Stress-related EEG: 86.62% maximum accuracy was obtained by a CNN framework for EEG-based stress recognition using cortisol-labeled tasks from 10 construction workers.The study used cortisol levels to label task stress levels.
  • Motor-imagery EEG: CNN applications also include practical motor-imagery classification evaluated in both online and offline BCI settings.

5.2 Generative Adversarial Networks (GAN)

GANs address limited EEG training data by generating synthetic or interpolated samples, but their broader BCI use remains constrained because time-sequence generation requires further evaluation.

  • GANs use a generator and discriminator trained together to model input distributions, create fake samples, and distinguish them from true data.Generated samples can support downstream functions such as classification.
  • EEG data augmentation: GAN-based EEG augmentation addresses insufficient training data and has been reported to outperform other generative models for classification improvement.
  • EEG data augmentation: GAN-based EEG super-resolution generates high-spatial-resolution data from low-resolution samples by interpolating missing channels.This approach targets limitations of low-density EEG devices.
  • EEG data augmentation: Generated EEG data have resembled real signals in spatial, spectral, and temporal characteristics in one CNN-based GAN study.
  • GANs remain comparatively less studied in BCIs because the feasibility of generating time-sequence data has not been fully evaluated.

5.3 Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM)

RNNs incorporate prior information for EEG time series, while LSTMs add gated memory to learn long-term dependencies that conventional RNNs cannot retain.

  • RNNs pass information through recurrent loops, allowing current outputs to use recent inputs and hidden states.
  • LSTMs extend RNNs with four gates that control data flow and enable learning of long-term dependencies.The gates are input, output, forget, and input-modulation gates, using sigmoid and tanh operations.
  • 93.0% average classification accuracy was achieved using RNN with LSTM architectures for EEG time-series signals.
  • 83.2% average accuracy was reported for classifying three English vowels with a regulated RNN reservoir, outperforming a deep neural network method.
  • RNN-based methods have been applied to identification, hand-motion recognition, sleep staging, emotion analysis, and hybrid CNN-RNN classification.LSTM has also been combined with CNNs for feature extraction, temporal-rule learning, and non-stationary EEG prediction.

5.4 Deep Transfer Learning (DTL)

Deep transfer learning reuses knowledge across EEG domains to reduce retraining and calibration demands, with CNN, GAN, and RNN-based approaches supporting cross-subject and cross-session BCI applications.

  • Deep transfer learning is categorized as instance-based, mapping-based, network-based, or adversarial-based transfer.These categories adjust source-instance weights, map domain similarities, reuse pretrained network parts, or use adversarial features.
  • Network-based transfer learning reuses pretrained source-domain layers, while adversarial transfer uses GAN technology to find features suitable for two domains.
  • MI EEG classification: MI EEG classification is a major application area for deep transfer learning, including subject-to-subject and session-to-session knowledge transfer.
  • MI EEG classification: A pretrained VGG-16 and target CNN framework transferred, froze, and fine-tuned parameters, exceeding standard CNN and SVM in efficiency and accuracy.
  • Deep transfer learning is intended to avoid time-consuming retraining and improve accuracy over solitary CNN and transfer learning approaches.
  • GAN-based transfer learning has been proposed to restrain domain divergence and improve domain adaptation through domain transformation.

5.5 Adversarial Attacks to Deep Learning Models in BCI

Deep learning models used in EEG-based BCIs are vulnerable to adversarial perturbations that can degrade performance and, in safety-critical applications, cause malfunction or misdiagnosis. Studies have demonstrated effective attacks across white-box, black-box, and gray-box scenarios, while subsequent work explores attack-efficient and regression-targeted methods.

  • Small adversarial perturbations can mislead deep learning models and cause dramatic performance degradation.These perturbations may be difficult for humans or computer programs to detect.
  • Adversarial attacks in EEG-based BCIs could malfunction wheelchair or exoskeleton control and cause harm, while clinical attacks could produce serious misdiagnosis.
  • White-box, black-box, and gray-box attacks effectively targeted EEGNet, DeepCNN, and ShallowCNN.The attack scenarios differ according to the attacker’s access to model architecture, parameters, responses, or training data.
  • Query-synthesis active learning can reduce training EEG trials for black-box attacks, while tiny perturbations can alter estimated driver drowsiness or user reaction time in regression tasks.

6 BCI-BASED HEALTHCARE SYSTEMS

EEG-based BCIs and computational intelligence methods have been applied across neurological and mental-health screening, diagnosis, prediction, and rehabilitation. Reported studies cover epilepsy, Parkinson’s disease, Alzheimer’s disease, schizophrenia, depression, migraine, pain, and motor impairment rehabilitation.

  • EEG-based BCI research increasingly classifies and predicts cognitive states while supporting mental-health and productivity monitoring.EEG and MEG contain information related to brain health and disease conditions.
  • Neurological disorders: A gated recurrent unit RNN achieved approximately 98% accuracy in epileptic seizure detection, while an LSTM methodology outperformed traditional machine learning and CNN in seizure prediction.
  • Neurological disorders: CNN-based EEG analysis demonstrated potential clinical use for Parkinson’s disease detection, while deep learning classified mild cognitive impairment and healthy controls with an averaged 80% accuracy.
  • Neurological disorders: Multiple EEG biomarkers improved Alzheimer’s disease classification and supported disease identification accuracy and clinical trials.
  • Neurological disorders: Sensor- and source-level EEG features were used to classify schizophrenia patients and healthy controls, with the resulting tool described as promising for diagnosis.
  • Rehabilitation and other healthcare applications: Non-invasive EEG-based BCI supports volitional brain-signal transmission for hand movement and has been investigated for stroke-related motor impairment rehabilitation.
  • Rehabilitation and other healthcare applications: Neural-network EEG classification has been proposed for migraine detection, and BCI signal processing has been used in phantom-limb-pain control training.
  • Mental-health applications: CNN architecture with transfer learning indicated that EEG spectral information is critical for depression recognition.

7 DISCUSSION AND CONCLUSION

The review synthesizes more than 150 studies from 2015–2019 on EEG sensing, signal enhancement, and computational intelligence for BCI applications. It identifies advances in sensors and learning methods alongside persistent real-world usability challenges and future priorities including adaptive, hybrid, and healthcare-oriented systems.

  • The review analyzed over 150 studies published between 2015 and 2019 on EEG sensing technologies and computational intelligence approaches.
  • Dry sensors, wearable devices, signal-enhancement tools, transfer learning, deep learning, and interpretable fuzzy models have improved EEG-based BCI performance.
  • Real-world usability remains constrained by prediction or classification capability and stability in complex BCI scenarios.
  • Future sensor research emphasizes signal quality, improved materials, and user experience, including wearable designs that maintain scalp contact on hair-covered sites.
  • Adaptive EEG-based BCI training may benefit from transfer learning, deep transfer learning, and reinforcement learning, while hybrid BCIs are proposed to improve artefact removal.
  • The survey presents computational intelligence approaches as methods for learning reliable brain-cortex features and understanding human knowledge from EEG signals.
Loading 2001.11337v1…