Non-Invasive Brain–Computer Interfaces

• Non-invasive BCIs are systems that enable direct communication between the human brain and external devices without surgical implants, using techniques like EEG, fNIRS, and MEG.
• They employ advanced signal processing and machine learning methods to decode neural activity, improving real-time control and enhancing system robustness.
• Applications span assistive devices, smart home technologies, and clinical tools, emphasizing user safety, comfort, ethical data handling, and practicality.

Non-invasive brain-computer interfaces (BCIs) are systems that enable communication and control between the human brain and external devices without requiring surgical implantation of electrodes or sensors. Predominantly based on techniques such as electroencephalography (EEG), functional near-infrared spectroscopy (fNIRS), and magnetoencephalography (MEG), these interfaces have become central to a diverse range of scientific, clinical, and consumer applications. Their development is driven by advancements in neural signal acquisition, signal processing, and machine learning, as well as by considerations of user comfort, safety, privacy, and usability.

1. Principles and Modalities of Non-Invasive BCIs

Non-invasive BCIs acquire neural signals from sensors placed on the scalp or external to the head, eliminating the need for surgery. This approach contrasts with invasive techniques (e.g., electrocorticography or implanted electrodes), which offer higher signal fidelity but are associated with surgical risks and potential long-term tissue reactions (Ghosh, 2023). The main modalities in use include:

• Electroencephalography (EEG): The most prevalent modality, employing metal electrodes to capture the brain’s electrical activity with high temporal resolution but low spatial resolution. EEG is portable, cost-effective, and widely adopted for real-time control and monitoring (Ghosh, 2023).
• Functional Near Infrared Spectroscopy (fNIRS): Utilizes near-infrared light to measure local hemodynamic responses associated with neural activity. fNIRS achieves moderate spatial and temporal resolution and is suitable for wearable, movement-robust systems (Ghosh, 2023).
• Magnetoencephalography (MEG): Measures the magnetic fields associated with neuronal currents. Non-invasive and offering high temporal and spatial resolution, MEG is primarily used in research due to equipment complexity (Saha et al., 2019, Jia et al., 15 Jun 2025).
• Hybrid Systems: Combine modalities (such as EEG+fNIRS) to leverage complementary strengths, improving decoding performance and system robustness (Ghosh, 2023, Ortega et al., 2021).

Signal Characteristics. Neural signals are characterized by specific frequency bands (e.g., delta, theta, alpha, beta, gamma) and event-related potentials (e.g., P300, SSVEPs) that are exploited by different BCI paradigms (Ghosh, 2023, 1211.2737).

2. Signal Processing, Feature Extraction, and Classification

The BCI information processing pipeline is typically organized into acquisition, pre-processing, feature extraction, and classification:

• Acquisition: Signals are recorded by multi-channel devices (e.g., Emotiv EPOC+ with 14 electrodes (Faruk et al., 2022), scientific-grade EEG with up to 64 channels (Cho et al., 2020)).
• Pre-processing: Signal quality is improved by amplification (EEG amplitudes ≈ 100 µV, amplified by ≈10^4 (1211.2737)), band-pass, and notch filtering to isolate relevant frequency bands and mitigate power line interference and artifacts (e.g., EMG, EOG). Advanced methods such as adaptive filtering (e.g., Recursive Least Squares), wavelet transforms, and independent component analysis (ICA) are used for robust de-noising (1211.2737, Basit et al., 23 May 2025).
• Feature Extraction: Extracted features may include time-frequency decomposition (wavelet transforms), spatial filtering (e.g., common spatial pattern—CSP, with optimization: $$\operatorname{argmax}_w \frac{w^\top C_1 w}{w^\top (C_1+C_2) w}$$ (Saha et al., 2019)), and component-based summaries (e.g., logarithmic variances from CSP (Cho et al., 2020)).
• Classification: Machine learning models—linear discriminant analysis, support vector machines, deep neural networks (including CNNs such as EEGNet (Cho et al., 2020)), and ensemble methods—map features to control commands or cognitive/affective states (Faruk et al., 2022, Basit et al., 23 May 2025).

Multimodal and Adaptive Approaches. Novel frameworks integrate multiple signals (e.g., EEG+fNIRS in CNNATT (Ortega et al., 2021)), employ deep learning with self-attention, and adapt online to nonstationarities using reinforcement learning agents and error-related potential signals (Fidêncio et al., 25 Feb 2025).

3. Paradigms and Application Domains

Non-invasive BCIs support both discrete and continuous paradigms, each matched to specific use cases:

• Discrete BCIs: Classical paradigms such as P300 spelling matrices, SSVEP-based selection, or RSVP (rapid serial visual presentation) systems for communication (Chen et al., 2023, Sunger et al., 20 Dec 2024). Advanced models, such as MarkovType (a POMDP-based recursive classifier), balance speed and accuracy in symbol selection (Sunger et al., 20 Dec 2024).
• Continuous BCIs: Systems for cursor control and force tracking use spatial encoding and projection methods to estimate continuous variables (e.g., movement velocity vectors in painting/gaming (Huang et al., 2023), continuous hand force in CNNATT (Ortega et al., 2021)).
• Hybrid Control: Systems like BRAVE combine EEG-based intent decoding with integrated voice commands for mode switching in prosthetic arm control (Basit et al., 23 May 2025).

Key Application Areas:

Application Area	Example Modalities	Key Technical Features
Assistive Devices	EEG, fNIRS, MEG	Robust intent decoding, real-time control, adaptive interfaces (Ghosh, 2023, Saha et al., 2019, Jia et al., 15 Jun 2025)
Smart Home/IoT	EEG	Thought-driven appliance control, focus on usability and safety for older adults (Kopeć et al., 2021)
Communication	EEG, MEG	Speller systems (P300, RSVP), speech decoding aided by text/speech features (1211.2737, Sunger et al., 20 Dec 2024, Jia et al., 15 Jun 2025)
Mobility and Robotics	EEG	Wheelchair control (Emotiv-based with digital twin simulation (Ghasemi et al., 27 Apr 2024)); scalable robot swarm control via posterior matching (Canal et al., 2022)
Neurohaptics/VR	EEG	Texture sensation decoding (CSP, EEGNet), bi-directional haptic feedback (Cho et al., 2020)
Gaming and Entertainment	EEG	Game-state adaptation, continuous BCI control in real-time tasks (Huang et al., 2023)

4. Performance, Usability, and User-Centered Considerations

Performance of non-invasive BCI systems is characterized by classification accuracy, information transfer rate (ITR), latency, and robustness to signal and user variability (Huang et al., 2023, Sunger et al., 20 Dec 2024, Basit et al., 23 May 2025).

• Decoding Accuracy: Ensemble learning and deep neural architectures can achieve high accuracies (e.g., 96% with LSTM-CNN-RF ensembles in BRAVE (Basit et al., 23 May 2025); 65%+ for deep learning tactile classification (Cho et al., 2020); >85% symbol recognition in POMDP-based RSVP typing (Sunger et al., 20 Dec 2024)).
• Adaptation and Calibration: Adaptive and transfer learning strategies address non-stationarity, reducing user calibration time (Saha et al., 2019, Fidêncio et al., 25 Feb 2025). Domain alignment networks, such as SSVEP-DAN, transform source user data to new users' templates, reducing calibration (Chen et al., 2023).
• User-Centered Design: Preferences trend toward non-invasive or minimally invasive solutions, particularly in patient groups such as multiple sclerosis (MS) (Russo et al., 7 Apr 2024), older adults (Kopeć et al., 2021), or persons with physical disabilities (Ghasemi et al., 27 Apr 2024). Trade-offs between performance (e.g., speed of communication) and invasiveness/risk are commonly accepted in favor of safety and comfort (Russo et al., 7 Apr 2024). User feedback also highlights concerns over usability, the naturalness of control paradigms, and the burden of device wearability.
• Limitations with Target Users: Severely motor-impaired patients may exhibit degraded BCI performance due to "negative plasticity"—maladaptive neural changes that diminish the attentional and cognitive processes exploited by BCI systems (Séguin et al., 2023).

5. Privacy, Security, and Ethical Considerations

Non-invasive BCIs inherently collect sensitive information beyond the primary control or communication task, including personal identity, gender, and cognitive states (Meng et al., 29 Nov 2024). This raises important privacy and trust issues:

• Privacy-Preserving BCIs: Perturbation-based algorithms can mask multiple types of private information in EEG data (user identity, gender, BCI-experience) while minimally impacting primary task performance. Perturbed signals $x' = x + \sum_{m} \delta_{m,p_m}$ can render privacy attributes unlearnable by adversarial classifiers while preserving utility (Meng et al., 29 Nov 2024).
• Security and Safety: As BCIs are deployed in mobility or smart environments, concerns over data interception, command misclassification, and accidental actuation are becoming more significant. Encryption, robust protocol design, and feedback mechanisms to verify user intent are recommended to address these risks (Ghasemi et al., 27 Apr 2024).
• User Consent and Data Governance: The participatory design with explicit consent and ongoing evaluation standards is emphasized, especially for clinical, home, or consumer BCI applications (Kopeć et al., 2021, Russo et al., 7 Apr 2024).
• Standardization and Legal Issues: International collaborative efforts promote benchmarks for signal quality, user data protection, and interoperability. Examples include EU’s BNCI Horizon 2020 and the Human Brain Project (Saha et al., 2019).

6. Recent Advances and Research Directions

Major recent advancements and ongoing research avenues include:

• Modality and Feature Expansion: The combination of EEG/fNIRS, high-density dry electrode cap development, and the use of MEG for non-invasive speech decoding (including tonal languages like Chinese) (Jia et al., 15 Jun 2025).
• Data-Driven Adaptation and Personalization: Adaptive algorithms utilizing reinforcement learning (with error-related potentials as feedback), as well as domain alignment and transfer learning to minimize calibration requirements (Fidêncio et al., 25 Feb 2025, Chen et al., 2023).
• Continuous Control and Multimodal Applications: Transition from discrete to smooth, continuous BCI control in both 2D navigation (painting, gaming) and force decoding for natural prosthetic control (Huang et al., 2023, Ortega et al., 2021, Basit et al., 23 May 2025).
• Hybrid and Multi-Modal Systems: Integration of voice commands and text or speech-assistive modalities to aid BCI decoding and enrich human-machine interaction (Basit et al., 23 May 2025, Jia et al., 15 Jun 2025).
• Real-Time, Embedded Implementation: Systems optimized for low-power embedded hardware, achieving high performance on resource-constrained platforms and supporting broader clinical or consumer deployment (Basit et al., 23 May 2025).
• Neurohaptics and Sensory Feedback: Decoding tactile and touch imagery to enhance control of advanced touch displays, haptic devices, and AR/VR interfaces (Cho et al., 2020).

Ongoing Challenges. Despite significant progress, remaining obstacles relate to inter-user variability, nonstationary brain dynamics, BCI illiteracy, data privacy, and the ethical implications of decoding or interpreting subconscious neural information (Saha et al., 2019, Séguin et al., 2023, Meng et al., 29 Nov 2024). Future research will likely focus on improved neural interface design, scalable multimodal signal processing, privacy safeguards, and ethically aligned frameworks for widespread adoption.

---

Non-invasive BCIs represent a rapidly evolving intersection of neuroscience, engineering, and data science. Their continued advancement, guided by robust technical development, ethical considerations, and user-centered design, is poised to substantially augment human–machine interaction across assistive, rehabilitative, and mainstream domains.