Core Terminology and Metrics

This section gives precise definitions of the 20+ most central terms and metrics in BCI. Subsequent chapters use them without explanation.

1. System-Role Terminology

Decoder vs Encoder

Decoder: Infers the user's intent or state from neural activity. Mathematically, a function \(f: \text{neural signal} \to \text{intent}\). This is "reading the brain."
Encoder: Transforms external information (visual images, tactile events) into stimulation patterns written into the nervous system. Mathematically, \(g: \text{info} \to \text{stim pattern}\). This is "writing to the brain."

Note: in the ML context, "encoder" and "decoder" usually refer to the two halves of an autoencoder. The BCI context strictly follows the definitions above — easy to confuse, so identify by context.

Forward vs Inverse Model

Forward model: Predicts "given stimulus, what the neural activity will be," i.e., \(P(\text{neural} | \text{stim})\). Used mainly for scientific understanding.
Inverse model: Infers "given neural activity, what the stimulus/intent is," i.e., \(P(\text{intent} | \text{neural})\). BCI decoders are essentially inverse models.

2. Signal Performance Metrics

Bit Rate / ITR

Information Transfer Rate (ITR) is the standard metric for BCI communication capacity. Wolpaw's formula:

\[B = \log_2 N + P \log_2 P + (1 - P) \log_2 \frac{1 - P}{N - 1}\]

where \(N\) is the number of target classes and \(P\) is classification accuracy. ITR is measured in bits/trial, multiplied by trials/min to yield bits/min.

System	Typical ITR
P300 speller (EEG)	20–60 bits/min
SSVEP speller (EEG)	30–120 bits/min
BrainGate cursor (invasive)	50–90 bits/min
Willett 2023 speech BCI	~400 bits/min equivalent

Key insight: invasive BCI's advantage in ITR over non-invasive reaches 5–10×.

Latency

End-to-end latency is the time from neural event to external-device response. Typical breakdown:

Stage	Typical latency
Signal acquisition + ADC	1–10 ms
Feature extraction + decoding	5–50 ms
Wireless transmission	1–20 ms
Device response	1–10 ms
Total	~50–200 ms

For closed-loop control, users perceive discontinuity beyond ~300 ms total latency. Speech BCI and fine robotic-arm control are especially latency-sensitive.

Signal yield: How many usable units (spikes) each electrode sees; reported as "units/electrode" in invasive BCI. Typical Utah array: 1–3 units/electrode.
SNR: Ratio of signal to noise power. Typical ECoG high-band SNR 20–40 dB; typical EEG SNR 5–15 dB.
Effective channel count: Not every implanted channel yields usable signal. "Effective channel yield" is a pragmatic metric.

3. Learning / Training Terminology

Calibration

Users collect a small amount of data before each use to train or fine-tune the decoder. Calibration time is the key bottleneck for BCI practicality.

Traditional calibration: 10–30 min per session
ReFIT / CLDA: Online calibration, gradually improving
Foundation-model era: With cross-subject transfer, new-user calibration drops to 1–5 min

Co-adaptation

The process in which user and decoder learn each other simultaneously. The user learns to produce neural patterns recognizable by the decoder; the decoder simultaneously adapts to the user's neural drift. It is a bidirectional meta-learning process.

Closed-loop

The system contains a complete "neural → decode → action → sensory feedback → neural" loop. Open-loop BCI only has the forward direction (neural → action); closed-loop BCI additionally feeds sensory or visual signals back.

Visual closed-loop: The user's eyes observe the robotic-arm motion (most common)
ICMS closed-loop: Direct microstimulation into S1 somatosensory cortex (more natural)
Full closed-loop: Bidirectional BCI, where read and write happen on the same cortical patch

4. Intent-Level Terminology

These terms are often conflated; distinctions:

Term	Meaning	Example
Movement intent	Wanting to perform some action	"reach out"
Kinematic decoding	Decoding continuous variables of velocity/position/force	\(v_t = [0.2, 0.5, -0.1]\) m/s
Goal-level intent	High-level goal, abstracted from specific action	"drink water"
Semantic intent	Semantic-level intent	"I am thirsty"

Traditional BCI did kinematic decoding; modern BCI increasingly moves toward goal-level / semantic — the core theme of Chapter 06 Intention-to-Action.

5. Data and Evaluation Terminology

Classification metrics

Accuracy: Most common, but can mislead on class-imbalanced data
WER (Word Error Rate): Standard metric for speech BCI
CER (Character Error Rate): Common in handwriting BCI
PER (Phoneme Error Rate): Intermediate metric in speech decoding

Regression metrics

R² (coefficient of determination): Standard metric for kinematic decoding, measuring explained variance
Correlation: Correlation between predicted and true trajectory

Benchmark datasets

FALCON: NeurIPS 2024 cross-session transfer benchmark
Neural Latents Benchmark (NLB): Latent-dynamics modeling benchmark
BCI Competition IV: Classic EEG motor-imagery dataset

6. Engineering Constraint Terminology

Power

The power consumption of implanted BCIs directly determines tissue heating. FDA requires brain-tissue temperature rise under 1°C, roughly corresponding to < 10 μW per channel. Neuralink N1 consumes ~6 mW total (1024 channels), requiring a wireless charging + intermittent use strategy.

Size

Utah array: 4×4 mm², 96 channels
Neuralink N1: 23 mm × 8 mm, 1024 channels
Neuropixels: 10 mm-long probe, 960 recording sites
Precision Layer 7: thin film 1/5 the thickness of a human hair, 1024 microelectrodes

Biocompatibility

Foreign Body Response (FBR): Glial cells encapsulate the electrode after implantation, significantly degrading signal quality after 6–12 months. This is the main challenge for long-term stability of invasive BCI.

7. Ethics and Legal Terminology

Neurorights: A concept first introduced into law by Chile, including mental privacy, cognitive liberty, psychological continuity, and mental integrity.
Cognitive Liberty: The right to be free from involuntary neural interference. Core to Minnesota 2024 legislation.
Cognitive Biometrics: The ability to infer an individual's identity, preferences, or emotions from neural activity — more private than DNA.

8. Logical Chain

BCI performance triad: accuracy × latency × ITR. The three trade off against each other.
The BCI training bottleneck is calibration time; foundation models reduce it from 30 min to under 5 min.
Intent hierarchy is central to distinguishing traditional BCI from modern BCI: kinematic → goal-level → semantic.
Closed-loop is not icing but essential: without sensory feedback, BCI is only a remote control.
Engineering constraints (power, size, biocompatibility) ultimately determine the form factor of invasive BCI.
Neurorights terminology (mental privacy, cognitive liberty) is BCI-specific — DNA legal frameworks do not apply.

References

Wolpaw et al. (2002). Brain-computer interfaces for communication and control. Clinical Neurophysiology.
Yuan & He (2014). Brain-computer interfaces using sensorimotor rhythms. IEEE TBME. — ITR formula details
Zander & Kothe (2011). Towards passive brain-computer interfaces. J. Neural Engineering.
Gilja et al. (2012). A high-performance neural prosthesis enabled by control algorithm design. Nature Neuroscience.
Shenoy & Carmena (2014). Combining decoder design and neural adaptation in BCIs. Neuron. — Classic co-adaptation review