Нейроинтерфейсы #1 Алексей Осадчий и Дария Клеева

Neurointerfaces (BCI): Overview

A neurointerface or brain–computer interface (BCI) is a device that reads brain activity (EEG/MEG/ECoG, invasive or non‑invasive), decodes brain states and converts them into control signals for external devices. BCIs can also deliver feedback to the brain (bidirectional BCI).

Non‑invasive paradigms

• P300 (oddball / visual‑spelling): external stimuli (matrix of symbols flashing); the target elicits a P300 event‑related potential used to select letters.
• SSVEP (steady‑state visual evoked potentials): attended visual stimuli blink at different frequencies producing corresponding peaks in occipital sensors; the frequency identifies the choice/command.
• Motor imagery (ideomotor BCI): the subject imagines movements (e.g., left/right hand, legs), producing stereotyped sensorimotor changes. Can be used for continuous or discrete control. Operation modes:
  • Synchronous (cue‑based)
  • Asynchronous (self‑paced)

---

Brain physiology and the electrophysiological substrate

• Brain structure: cortex (gray matter), white matter (axons), plus vascular and astrocyte contributions. Functional maps include the homunculus and Brodmann areas (e.g., V1).
• Neural origin of EEG/MEG: aligned pyramidal neurons produce postsynaptic currents in apical dendrites that sum to form current dipoles; these dipoles are the physical sources seen by EEG/MEG sensors.
• Volume conduction and forward physics: dipoles in a conducting head generate primary and secondary currents. The skull conductivity is much lower than brain/scalp, so only a fraction of the source reaches sensors. EEG/MEG sensitivity depends on dipole orientation. Temporal resolution is high (tens of ms); spatial resolution depends on modality and geometry.

---

Maxwell equations and the quasi‑static approximation

For EEG/MEG frequencies of interest (roughly <100–500 Hz) electromagnetic propagation can be approximated as quasi‑static:

• Time derivatives of fields are small.
• Fields can be represented via scalar potentials and simplified rotor/divergence relations.
• Sensors effectively receive simultaneous (no measurable propagation delay) contributions from sources across the brain.

---

Generative (forward) model and lead fields

• Forward model: defines the mapping from source space (many candidate dipole locations on the cortical surface) to sensor space via the lead‑field (gain) matrix G.
• Typical construction steps:
  • Acquire subject MRI and extract the cortical surface (triangulation).
  • Place oriented dipoles on vertices (source space).
  • Solve the quasi‑static electromagnetic problem using boundary‑element (BEM) or finite‑element (FEM) discretization to compute column topographies (lead fields).
• Sensor data model:
  • x(t) ≈ G · s(t) + noise
  • This linear mixture is the basis for decoding, source reconstruction (inverse problem), spatial filtering, and simulations.

---

Inverse problem and decoding approaches

• The inverse (source localization) problem is ill‑posed: many more possible sources than sensors. Solving it requires priors (smoothness, sparsity) or regularization.
• For many BCI applications researchers work directly in sensor space by learning spatial filters (weights W) so that W · x(t) isolates activity correlated with the target behavior (maximize SNR for the desired topography, suppress interferences).
• Common decoding goals:
  • Classification (discrete commands)
  • Regression / continuous control (proportional control)

---

Neural signals of interest and time–frequency concepts

• Oscillations: canonical frequency bands include delta, theta, alpha (~8–12 Hz), beta (~13–30 Hz), and gamma. Alpha rhythm was discovered by Hans Berger.
• Evoked (phase‑locked) vs induced (non‑phase‑locked) responses:
  • Evoked / ERP: reproducible, time‑locked waveform across trials; recovered by averaging epochs.
  • Induced: changes in power or envelope in frequency bands; require time–frequency analysis and power averaging.
• ERD/ERS: event‑related desynchronization (power decrease, e.g., mu rhythm during movement/imagery) and event‑related synchronization (power increase) — commonly used as BCI features.
• Instantaneous amplitude (envelope) and phase are important signal features for decoding and analysis.

---

Artifacts and preprocessing

Common artifacts

• EMG (muscle)
• EOG (eye blinks / saccades)
• ECG (cardiac)
• Sweating, bad/noisy electrodes
• Power‑line interference (50/60 Hz) and harmonics

Typical preprocessing pipeline

1. Detect and mark bad channels; interpolate if necessary (with caution).
2. For MEG: apply external interference suppression (MaxFilter / SSS) to separate internal vs external components; correct for head motion; use calibration.
3. Temporal filtering: bandpass and notch filters as appropriate.
4. Artifact detection: mark EOG/ECG events; reject grossly contaminated epochs.
5. Blind source separation (ICA) or PCA to identify and remove components corresponding to EOG/ECG/EMG; inspect component topographies, timecourses and spectra before removal.
6. Epoching and averaging for evoked analysis, or time–frequency transforms for induced/power analyses.

---

Practical data handling and tools (MNE‑Python)

• Raw data structures include channel types (magnetometers, gradiometers, EEG), sampling rate, and sensor locations.
• Useful MNE steps:
  • MaxFilter/SSS for MEG
  • Mark and interpolate bad channels
  • Bandpass / notch filtering
  • Event extraction and epoching
  • Compute evoked responses and time–frequency maps
• ICA in MNE:
  • Fit ICA on raw data or epochs
  • Inspect component topographies, timecourses, and spectra
  • Exclude artifact components and apply the decomposition back to the raw data
• Simulations:
  • Select source vertices (atlas/Brodmann/labels)
  • Create source timecourses (e.g., sinusoids aligned to events)
  • Use forward model / lead fields and a sensor noise covariance to simulate realistic sensor data for algorithm testing

---

Applications and demonstrations

• Motor imagery BCIs for cursor or object movement (discrete and continuous control)
• Competitive / multiplayer BCI demonstrations
• Rehabilitation: thought‑driven robotic gloves or exoskeletons for post‑stroke or spinal cord injury; pairing command with somatosensory feedback (proprioceptive/tactile) to accelerate relearning (closed‑loop therapy)
• Bidirectional cortical interfaces: stimulating cortex to provide tactile feedback when a prosthesis touches an object
• Speech decoding from invasive recordings (ECoG / implanted electrode arrays): decoding spoken words / motor programs from cortical activity, with evidence of decoding motor plans that can precede audio
• Research demonstrations using implanted grids in epilepsy patients (opportunities to record ECoG)

---

Conceptual and ethical remarks

• BCI states are arbitrary brain states mapped to commands via training — BCIs are not “mind reading.” Motor imagery is commonly used because it is reproducible and trainable.
• Individual variability: BCIs typically require subject‑specific training. Some generalization is possible with large datasets and improved models.
• Privacy and ethics: BCIs have clear beneficial uses (e.g., rehabilitation) but raise social and ethical questions; risks should be weighed against potential benefits.

---

Concise methodological checklist (non‑invasive BCI / EEG/MEG)

1. Acquire structural MRI → extract cortical surface → define source space (vertices).
2. Compute forward model / lead fields (BEM/FEM) → obtain G matrix.
3. Record EEG/MEG with good sensor contact; record EOG/ECG if possible.
4. Preprocessing:
   • Mark / interpolate bad channels.
   • MEG: apply SSS / MaxFilter, correct for head motion, use calibration.
   • Bandpass / notch filtering.
   • Detect artifacts (EOG/ECG) and identify artifact components (ICA/PCA).
   • Optionally reject noisy epochs.
5. Feature extraction:
   • Evoked analysis (ERP) via epoching + averaging for time‑locked stimuli.
   • Time–frequency analysis (spectrograms, wavelets) for induced activity and rhythmic power.
   • Compute envelopes / bandpower (e.g., mu/beta for motor imagery).
   • Spatial filtering (CSP, beamformers) or supervised spatial weight learning to maximize SNR.
6. Model & decode:
   • Train classifier or regressor on labeled data (synchronous or self‑paced).
   • Validate online; tune asynchronous detection (onset detection); provide neurofeedback to train the subject.
7. (Optional) Source localization / inverse modeling for interpretation or enhanced features.
8. Simulate data with forward model + noise covariance to validate algorithms.

---

Researchers, contributors and historical sources mentioned

• Алексей Осадчий (Alexey Osadchiy / Alexey Asai) — lecturer, director of the Center for Bioelectric Interfaces
• Дария Клеева (Daria Kleeva / Dasha) — graduate student; demonstrated MNE‑Python practicals and data simulation
• Никита Федосов (Nikita Fedosov) — master’s student, assisted with mathematical/code lectures
• Александр Алексеевич Фролов (Alexander Alekseevich Frolov) — collaborator (rehabilitation work)
• Павел Бобров (Pavel Bobrov) — collaborator (rehabilitation work)
• Hans Berger — historical source (discoverer of the alpha rhythm)
• Additional demo participants / lab members (briefly mentioned): Ksenia, Kolya, Dima, Denis