EEG-Based Motor Imagery BCI Algorithms and Technologies: A Review
Authors: Mohammad Hossein Koohi Ghamsari, Seyede Fatemeh Ghamkhari, Siavash Bayat, Ahmed Hemani
Organizations: Department of Electrical Engineering, Sharif University of Technology, Azadi St., Tehran, 14588-89694, Iran. · School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology, 100 44 Stockholm, Sweden.
Brain-computer interfaces (BCIs) have emerged as transformative technologies that enable direct communication between the brain and external devices. Among various BCI paradigms, EEG-based motor imagery (MI) has gained prominence due to its simplicity, non-invasiveness, and potential to restore motor function and facilitate rehabilitation for patients with motor impairments. This paper presents a comprehensive review of the most practical processing algorithms developed over the past decade for decoding brain sensorimotor cortex signals. Specifically, this paper discusses the integration of artificial intelligence (AI)-based algorithms, particularly machine learning and deep learning techniques, and their contributions to improving the performance and efficiency of MI-BCI systems in detail. Furthermore, the paper reviews state-of-the-art hardware platforms and emerging converging technologies, including system-on-chip (SoC) architectures, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), wearable devices, the Internet of Things (IoT), and augmented/virtual reality (AR/VR), and discusses their integration with advanced signal processing algorithms to enable next-generation MI-BCI systems. By highlighting current achievements of EEG-based MI-BCI technology and predicting future research directions that could further enhance real-time capabilities, this paper aims to provide valuable insights for researchers and practitioners, fostering innovation in high-performance EEG-based MI-BCI systems.
Figures & tables
Fig. 1: Various medical and non-medical applications of EEG-based BCI modality.
Search Strings / Search Engine
IEEE Xplore
PubMed
ScienceOPEN
Semantic Scholar
Scopus
Web of Science
Google Scholar
(EEG OR electroencephalographic OR electroencephalography) AND (BCI OR ”brain-computer interface”)
11,751
6,770
3,610
11,900
4,916
8,665
22,600
(EEG OR electroencephalographic OR electroencephalography) AND (BCI OR ”brain-computer interface”) AND (”motor imagery” OR ”motor-imagery” OR MI)
3,267
2,062
1,235
7,190
2,301
3,862
17,900
(EEG OR electroencephalographic OR electroencephalography) AND (BCI OR ”brain-computer interface”) AND (”machine learning” OR ML OR ”deep learning” OR DL OR ”neural network” OR ”artificial intelligence” OR AI)
5,230
2,961
1,547
4,040
1,101
2,449
18,100
(EEG OR electroencephalographic OR electroencephalography) AND (BCI OR ”brain-computer interface”) AND (wearable OR wireless)
806
349
287
3,710
1,318
776
17,700
(EEG OR electroencephalographic OR electroencephalography) AND (BCI OR ”brain-computer interface”) AND (”system-on-chip” OR ”system on a chip” OR SoC OR ”hardware accelerator” OR ”hardware acceleration”)
398
696
234
3,400
44
106
6,190
(EEG OR electroencephalographic OR electroencephalography) AND (BCI OR ”brain-computer interface”) AND (”Internet of Things” OR IoT)
284
29
28
301
376
99
11,200
TABLE I: Search strings and numbers of records retrieved from the selected databases for EEG-based BCI and related technologies. Searches were conducted in July 2026 and restricted to publications from January 1, 2014, to July 1, 2026.
Full Term
Abbreviation
Full Term
Abbreviation
Full Term
Abbreviation
Full Term
Abbreviation
Motor imagery-based BCI
MI-BCI
Brain-computer interface
BCI
Artificial intelligence
AI
Electroencephalography
EEG
Machine learning
ML
Artificial neural network
ANN
Deep learning
DL
Generative AI
GenAI
Classification accuracy
CA
Convolutional neural network
CNN
Feedforward neural network
FFNN
Back propagation neural network
BPNN
Probabilistic neural network
PNN
Euclidean Space
EUS
System-on-Chip
SoC
Band-pass filtering
BPF
Hybrid
HYB
Deep belief network
DBN
Transformer
TRF
Common spatial pattern
CSP
Support vector machine
SVM
Linear discriminant analysis
LDA
Neuro‑fuzzy system
NFS
Hardware accelerator
HA
TABLE II: List of frequently used abbreviations in this paper.
Fig. 2: Overview of brain activity signal acquisition modalities for BCI systems. The modalities are grouped into magnetic (MEG), electric (ECoG, LFP, and EEG), and metabolic (NIRS and fMRI) approaches, highlighting their invasive and non-invasive implementations and representative sensing hardware.
Fig. 3: Brain region labeling and the international standard EEG electrode placement systems. (a) Lateral brain view divided into the six main cortical regions referenced by electrode nomenclature: frontopolar (Fp), frontal (F), central (C), parietal (P), occipital (O), and temporal (T), with the cerebellum and spinal cord shown for anatomical context. (b) 10-20 electrode placement system (21 electrodes), with electrode labels color-coded by region and referenced to the nasion and inion. A1/A2 denote auricular (ear) reference electrodes, which are not cortical sites. (c) 10-10 electrode placement system (extended array of up to 81 electrodes), providing finer spatial resolution between the standard 10-20 sites.
Fig. 4: The MI process in the brain. The first (internal) (a) somatosensory or third (external) (b) observational perspective imagination of opening/closing the fists activates the primary sensorimotor cortex. During imagined movement (c), the specific cortical oscillation rhythms ( μ , 8–12 Hz and β , 13–30 Hz) are suppressed, known as Event-Related Desynchronization (ERD). Upon movement termination, the termination of MI elicits a distinct β -band amplitude increase, known as Event-Related Synchronization (ERS) or β -rebound.
Fig. 5: The EEG-based MI-BCI life cycle overview. (a) Brain electrical, magnetic, and metabolic signals are generated by neural activity. (b) EEG-based electrical signal acquisition method. (c) Five frequency bands of EEG signals, known as EEG rhythms. (d) MI processing pipeline: analog front end (AFE), digital signal processing (DSP), and actuator interface. (e) Conventional MI experimental paradigms. (f) Actuator to brain feedback mechanisms: Vibration, Tactile, Thermal, Electrical, Haptic, Auditory, and Visual.
Cat.
Dataset
Channels
Sampling Rate
Subjects
Classes
Task
1
BCI Competition III dataset IIIa
25
250 Hz
9
4
Both hands, one foot, and tongue MI
2
BCI Competition III dataset IVa
118
100 Hz
5
3
Both hands and right foot MI
3
BCI Competition III dataset V
32
512 Hz
3
3
Both hands and word generation MI
4
BCI Competition IV dataset 2a
22 + 3 EOG
250 Hz
9
4
Both hands, both feet, and tongue MI
5
BCI Competition IV dataset 2b
3 EEG + 3 EOG
250 Hz
9
2
Both hands MI
6
Physionet EEG Motor Movement/Imagery
64
160/128 Hz
109
4
Both fists and both feet opening/closing MI
TABLE III: The most common types of benchmark datasets used for evaluating the performance of MI algorithms.
TABLE IV: State-of-the-art EEG-based MI-BCI signal processing algorithms reported in the literature in the recent decade.
Fig. 6: The relation between various categories and sub-domains of artificial intelligence (AI).
Fig. 7: Comparison of traditional processing of EEG-based MI signals and DL-based architectures.
Fig. 8: Taxonomy of EEG pre-processing algorithms for MI-based BCI systems. EEG pre-processing techniques are broadly classified into blind source separation, transform-based, filtering, and regression-based artifact removal approaches. Typical algorithms and their corresponding signal processing mechanisms are illustrated, demonstrating how these methods enhance signal quality, reduce artifacts and noise, and facilitate reliable feature extraction and classification in MI-BCI applications.
Fig. 9: EEG signals Feature Extraction algorithms classifications based on the mathematical domain of the signal representation. The classification spans five foundational domains and a combined approach: Time-Domain Techniques, Frequency-Domain Techniques, Time-Frequency Domain Techniques, Spatial Domain Techniques, Feature Selection/Reduction Techniques, and HYB Techniques.
Fig. 10: Applying FFT conversion to EEG MI signals for extracting the delicate frequency features of MI oscillations. The FFT decomposes the raw, time-domain EEG rhythms into a frequency-domain spectrum, isolating power changes in target bands (e.g., alpha and beta rhythms) to construct a discriminative feature vector for classification.
Fig. 11: EEG wavelet transform-based feature extraction methods. Applying DWT (a) and CWT (b) to MI signals.
Fig. 12: Utilizing CSP-based methods for EEG MI signal feature extraction.
Fig. 13: Summary of prevalent AI-based MI feature classification algorithms.
Fig. 14: Schematic workflow of an EEG-based MI-BCI classification system using the SVM algorithm. The Data Acquisition & Feature Extraction shows the raw EEG signals are recorded from motor cortex channels (e.g., C3, C4) during imagined right-hand vs. left-foot movements, and transformed into a 2D feature space using the PSD. The SVM Classification shows that a linear kernel SVM, for example, identifies the optimal separating hyperplane ( H ) by maximizing the margin between the two MI classes, uniquely defined by the critical edge data points known as support vectors.
Fig. 15: Utilizing LDA-based methods for MI feature classification.
Fig. 16: Utilizing KNN-based methods for MI feature classification.
Fig. 17: Utilizing RF-based methods for MI feature classification.
CA: 93.06%; REU reduced to 2.54% area and 3.66% PC
[ 258 ]
2022
FPGA, 65 nm CMOS LP process
BPF
MC-EEGNet
MC-EEGNet
-
ARE: 87.22%, Eff: 20.77% better than a RISC-V MCU, SPU: 3.7 ×
[ 200 ]
2021
Wireless soft scalp electronics system with Samsung Gear VR; 6 flexible microneedle electrodes
BPF, BWF
PSD, WCM, CNN
Spatial CNN, Standard CNN, SVM (cubic polynomial kernel)
(8)
CA: 93.22 ± 1.33% (4-class, spatial CNN)
TABLE V: State-of-the-art EEG-based MI-BCI hardware implementations reported in the literature in recent decade.
Fig. 24: The EEGNet CNN-based processing architucture for MI signals classification.
Fig. 25: State-of-the-art EEG-based MI-BCI system. The most common and most recent algorithms reported in the literature in recent decade for each blocks of the signal processing pipeline are illustrated.
Fig. 26: Convergence of software and hardware technologies driving the commercialization of next-generation EEG-based MI-BCI applications.
Fig. 27: Different types of EEG electrodes. (a) Dry electrodes. (b) Wet electrodes. (c) Semi-dry electrodes.
Fig. 28: Intelligent integrated MI-BCI ecosystem for personalized neuro interaction.
Fig. 29: Conceptual block diagram and major subsystems of a next-generation wearable edge-processing EEG-based MI-BCI SoC.
Fig. 31: Annual number of MI-BCI publications from January 2014 to July 2026, retrieved from Scopus using the search strategy described in Section II. The connecting line represents the yearly publication counts and does not indicate statistical regression, smoothing, or extrapolation.
Fig. 32: Prevalence percentages of MI signal-processing algorithms and datasets based on the subsets of studies providing sufficient information for each analysis: (a) signal preprocessing algorithms ( N=124 studies); (b) feature-extraction algorithm categories ( N=168 studies); (c) feature-extraction algorithms ( N=286 studies); (d) feature-classification algorithms ( N=297 studies); and (e) common MI datasets ( N=301 studies).
Fig. 33: Current MI-BCI systems challenges and solutions.
Fig. 34: Trends of converging technologies for EEG-based BCI from January 2014 to July 2026 based on annual publication counts retrieved from Scopus using the search strategy described in Section II: (a) EEG-based BCI and AI. (b) EEG-based BCI and wearable technology. (c) EEG-based BCI and HA technology. (d) EEG-based BCI and IoT. (e) EEG-based BCI and EP. (f) EEG-based BCI and VR. Scopus is used as the reference (see Section II for more details). The connecting lines represent yearly publication counts and do not indicate statistical regression, smoothing, or extrapolation.
Motor Imagery (MI) is an emerging Brain-Computer Interface (BCI) paradigm where a person imagines body movements without physical action. By decoding scalp-recorded electroencephalography (EEG) signals, BCIs establish direct communication to control external devices, offering significant potential in prosthetics, rehabilitation, and human-computer interaction. However, existing solutions remain difficult to deploy. (i) Most employ independent, opaque models for each MI task, lacking a unified architectural foundation. Consequently, these models are trained in isolation, failing to learn robust representations from diverse datasets, resulting in modest performance. (ii) They primarily adopt fixed sensor deployment, whereas real-world setups vary in electrode number and placement, causing models to fail across configurations. (iii) Performance degrades sharply under low-SNR conditions typical of consumer-grade EEG. To address these challenges, we present NeuroPath, a neural architecture for robust MI decoding. NeuroPath takes inspiration from the brain's signal pathway from cortex to scalp, utilizing a deep neural architecture with specialized modules for signal filtering, spatial representation learning, and feature classification, enabling unified decoding. To handle varying electrode configurations, we introduce a spatially aware graph adapter accommodating different electrode numbers and placements. To enhance robustness under low-SNR conditions, NeuroPath incorporates multimodal auxiliary training to refine EEG representations and stabilize performance on noisy real-world data. Evaluations on three consumer-grade and three medical-grade public datasets demonstrate that NeuroPath achieves superior performance.
Jiani Cao, Kun Wang, Yang Liu +1
Department of Computer Science City University of Hong Kong Hong Kong, China · Department of Computer Science Florida State University Tallahassee, United States
Brain-computer interfaces (BCIs) enable direct communication between the brain and external devices, providing critical support for individuals with motor impairments. However, accurate motor imagery (MI) decoding from electroencephalography (EEG) remains challenging due to noise and cross-session variability. This study introduces EEG-MFTNet, a novel deep learning model based on the EEGNet architecture, enhanced with multi-scale temporal convolutions and a Transformer encoder stream. These components are designed to capture both short and long-range temporal dependencies in EEG signals. The model is evaluated on the SHU dataset using a subject-dependent cross-session setup, outperforming baseline models, including EEGNet and its recent derivatives. EEG-MFTNet achieves an average classification accuracy of 58.9% while maintaining low computational complexity and inference latency. The results highlight the model's potential for real-time BCI applications and underscore the importance of architectural innovations in improving MI decoding. This work contributes to the development of more robust and adaptive BCI systems, with implications for assistive technologies and neurorehabilitation.
Panagiotis Andrikopoulos, Siamak Mehrkanoon
Department of Information and Computing Sciences, Utrecht University, Utrecht, Netherlands
Motor imagery (MI) brain--computer interfaces (BCIs) have emerged as a promising approach for establishing flexible communication pathways between the human brain and external devices , particularly for individuals affected by stroke or neurodegenerative disorders. Reliable decoding of motor-imagery electroencephalography (MI-EEG) remains challenging because EEG recordings contain substantial noise and exhibit complex, weakly informative relationships with the underlying brain activity. Although deep learning provides an effective means of learning representations directly from EEG signals, its application to MI-EEG feature learning remains comparatively limited. This study introduces a hybrid deep-learning architecture that integrates a convolutional neural network (CNN) with a bidirectional long short-term memory (bi-LSTM) network. The CNN is used to learn high-level spatial and temporal representations directly from raw MI-EEG recordings, whereas the bi-LSTM models temporal dependencies and relationships among the extracted features. The proposed approach is evaluated using both a publicly available dataset and a privately acquired dataset obtained with an EEG acquisition system. The experimental results indicate that the CNN&bi-LSTM architecture provides robust performance for both two- and three-class motor-imagery classification and demonstrates promising subject-independent decoding capability across the evaluated methods.
Athanasios Karagounis
School of Science, Department of Digital Industry Technologies GR34400, Psachna, Greece