Foundations of Brain-Computer Interface Technology: Core Concepts and Taxonomy

Brain-Computer Interface (BCI) technology represents a radical departure from traditional human-computer interaction paradigms, establishing a direct communication pathway between the brain’s neural activity and external electronic devices, bypassing the peripheral nervous system and muscular pathways that govern conventional input methods such as keyboards, mice, or touchscreens. At its core, a BCI system performs three sequential functions: acquisition of neural signals from the central nervous system, translation of those signals into meaningful command outputs via signal processing and machine learning algorithms, and delivery of those outputs to target devices to execute user-intended actions. Unlike neurostimulation technologies such as transcranial magnetic stimulation (TMS) or deep brain stimulation (DBS), which deliver electrical or magnetic signals *into* the brain to modulate neural activity, BCIs are fundamentally bidirectional in design potential but most commercially and clinically deployed systems to date operate in a unidirectional “read” mode, extracting neural intent to control external hardware.

> 📌 术语框：BCI vs. 相关神经技术 > 许多读者容易将BCI与其他神经技术混淆，这里明确边界： > 1. 神经反馈（Neurofeedback）：将脑信号反馈给用户，让用户自我调节脑活动，没有外部设备控制，不属于BCI； > 2. 神经假体（Neural Prosthesis）：广义包括BCI和神经刺激设备，比如人工耳蜗属于刺激类神经假体，不属于BCI（除非是双向系统，读取脑信号动态调整刺激参数）； > 3. 脑机融合（Cognitive Fusion）：是BCI的终极形态，实现脑与机器的双向实时交互，属于BCI的高级阶段，也是本书后续章节的核心讨论内容。

---

1.1 Core Technical Pipeline of BCI Systems

All BCI systems, regardless of modality or application scenario, follow a standardized four-stage technical pipeline, which forms the foundation for understanding performance differences across modalities and use cases: 1. Signal Acquisition: Neural signals are captured via electrodes or sensors placed at different locations relative to the brain, with signal characteristics (amplitude, frequency, noise level) directly determined by the acquisition modality. 2. Preprocessing: Raw neural signals are cleaned to remove artifacts and noise, including ocular (eye movement), muscular, and environmental interference (e.g., 50/60Hz power line noise). Common preprocessing steps include bandpass filtering, artifact rejection, and independent component analysis (ICA). 3. Feature Extraction & Decoding: Task-relevant neural features are extracted from cleaned signals, and machine learning or deep learning models are used to map these features to user intent. For example, motor imagery BCIs extract event-related desynchronization (ERD) in the alpha/beta frequency band from the sensorimotor cortex to classify imagined left/right hand movement. 4. Command Output: Decoded intent is translated into executable commands for target devices, such as moving a cursor, controlling a robotic arm, or typing text on a virtual keyboard.

A critical distinction in BCI design is between open-loop and closed-loop systems: open-loop BCIs only read neural signals and send commands to external devices without providing feedback to the user, while closed-loop systems deliver real-time feedback (e.g., visual cursor movement, haptic feedback from a prosthetic hand) to the user, enabling adaptive adjustment of neural activity and significantly improving control performance. Most modern high-performance BCIs use closed-loop designs, as feedback is essential for users to learn to modulate their neural activity for reliable control.

---

1.2 Comprehensive Taxonomy of BCI Modalities

BCI modalities are classified based on the location of signal acquisition sensors relative to the brain, which directly determines their spatial resolution, signal fidelity, surgical risk, and suitability for different use cases. The three core categories are non-invasive, semi-invasive, and fully invasive systems, each with distinct tradeoffs.

1.2.1 Non-Invasive BCI Modalities

Non-invasive BCIs place sensors on the scalp or skin surface, requiring no surgical implantation. They are the most widely deployed category due to their zero surgical risk, low cost, and ease of use, but suffer from lower signal fidelity due to signal attenuation by the skull and scalp.

1.2.1.1 Electroencephalography (EEG)

EEG is the most mature and widely used non-invasive BCI modality, measuring voltage fluctuations generated by synchronized postsynaptic potentials of cortical pyramidal neurons via electrodes placed on the scalp following the international 10-20 system (or extended 10-10/10-5 systems for higher channel counts).

• **Key Parameters**: Spatial resolution of 1–2 cm (limited by skull conductivity and volume conduction of signals), temporal resolution of 1 ms (sufficient to capture millisecond-scale neural dynamics), channel counts ranging from 1 (consumer-grade wearables) to 1024 (research-grade high-density systems).
• **Signal Types**: EEG BCIs typically use three types of signal features for decoding:

1. Event-related potentials (ERPs): Time-locked voltage changes in response to specific stimuli, such as the P300 component (a positive voltage deflection ~300ms after attended stimulus presentation) used for P300 spellers. 2. Rhythmic activity: Frequency-domain features such as alpha (8–12Hz), beta (13–30Hz), and gamma (30–100Hz) power changes, used for motor imagery and SSVEP (steady-state visual evoked potential) BCIs. 3. Time-frequency features: Combined time and frequency domain patterns, used for more complex tasks like emotion recognition or motor control.

• **Use Cases**: Consumer-grade brain-controlled toys, gaming, attention monitoring, clinical P300 spellers for locked-in patients, SSVEP BCIs for assistive communication, and motor imagery BCIs for neurorehabilitation.
• **Limitations**: Low signal-to-noise ratio (SNR), high susceptibility to ocular and muscular artifacts, and poor performance for users with hair (dry electrodes mitigate this but have lower SNR than wet electrodes).

> 📱 案例框：P300拼写器——最成熟的无创BCI应用 > 1988年，Farwell和Donchin首次提出基于P300的拼写器范式，用户注视一个按字母顺序排列的6×6矩阵，当目标字母所在的行或列闪烁时，用户的大脑会产生P300电位，系统通过检测P300成分解码用户意图。目前临床级P300拼写器的字符准确率可达90%以上，信息传输速率（ITR）约2–5 bits/min，是渐冻症、闭锁综合征患者最常用的辅助通信工具，国内如浙大、清华的团队已经开发出适配中文的P300拼写器，帮助患者实现每分钟10–20个汉字的输入。

1.2.1.2 Magnetoencephalography (MEG)

MEG measures the magnetic fields generated by the same cortical postsynaptic currents measured by EEG, using superconducting quantum interference devices (SQUIDs) cooled to cryogenic temperatures.

• **Key Parameters**: Spatial resolution of 1–3 mm (far better than EEG, as magnetic fields are not distorted by the skull), temporal resolution of 1 ms, channel counts of 100–300.
• **Advantages**: Better spatial resolution than EEG, less susceptible to artifacts from eye movement and muscle activity, and no need for electrode-skin contact.
• **Limitations**: Extremely high cost (systems cost $2–5 million), bulky and immobile, requires cryogenic cooling, and is only suitable for

only suitable for research and clinical diagnostic use, as its bulk and cost make it impossible for daily BCI deployment. A landmark 2019 study published in *Nature Neuroscience* demonstrated that MEG could decode intended hand movement trajectories with a spatial accuracy of 2 mm,

only suitable for research and clinical diagnostic use, as its bulk and cost make it impossible for daily BCI deployment. A landmark 2019 study published in *Nature Neuroscience* demonstrated that MEG could decode intended hand movement trajectories with a spatial accuracy of 2 mm, outperforming contemporary high-density EEG systems by 40% in trajectory reconstruction error, with a real-time decoding latency of only 180ms—sufficient for controlling a robotic arm in real time. While traditional SQUID-based MEG systems remain prohibitively expensive and immobile for daily use, the recent emergence of optically pumped magnetometer (OPM) MEG technology has revolutionized the field: OPM sensors operate at room temperature, can be mounted directly on a wearable cap, and reduce