Bidirectional BCI and Channel Separation
A bidirectional BCI (biBCI) does reading + writing simultaneously — both decoding intent from the brain and stimulating the brain to write signals back. The biggest engineering challenge is stimulation artifact: stimulation currents produce enormous artifacts on recording electrodes that mask neural signals. Channel separation (channel isolation) is the core technique.
1. Why Bidirectional
Limits of unidirectional BCI
- Read-only: we know what the user wants to do, but the user has no feedback perception of the machine's output
- Write-only: we can make the user feel things, but we don't know what the user wants
The power of bidirectional
Read + write = closed-loop BCI: - Motor BCI + ICMS: feel the prosthesis you control - Visual prosthesis + V1 decoding: wear a camera + see content + system learns user reactions - Memory prosthesis: read hippocampus + write hippocampus → memory enhancement - Emotion loop: read prefrontal cortex + deep stimulation → depression treatment
2. The Ganzer 2020 Cell Milestone
Ganzer et al. (2020) was the first clinical bidirectional BCI — restoring sensation and movement of the patient's own hand in a spinal-cord-injury patient:
Subject
- C5/C6 SCI (spinal-cord injury)
- Arm paralysis, partial loss of sensation
System
M1 Utah Array (read)
↓
Intent decoding → control forearm stimulation (FES) → own hand moves
↓
Pressure sensor on the hand
↓
S1 ICMS (write)
↓
Sensory feedback
Innovations
- Simultaneous read and write — M1 reads, S1 writes
- Native hand: not a robotic arm but the patient's own paralyzed hand, activated via FES
- Patient reports "feeling my own hand holding the cup"
Results
- Grasp efficiency improved by ~50%
- Tactile-event detection accuracy ~90%
- High subjective "embodiment" ratings
3. The Stimulation Artifact Problem
Scale of the problem
- Typical neural signal: 100 μV
- Typical ICMS current: 20–100 μA
- Resulting artifact: on the order of 100 mV — 1000× larger than the signal
Neighboring electrodes and even other channels on the same chip saturate and cannot record.
Why it's hard
ICMS must be triggered simultaneously with the user's action (contact = sensation) — which is exactly when M1 decoding most needs real-time signals. Stimulation time = decoding time produces a critical conflict.
4. Channel-Separation Techniques
1. Blanking
Disable acquisition during the stimulation instant: - Stimulation pulse 200 μs, 1 ms blank before and after - M1 signal briefly lost for ~2.5 ms - Decoding algorithm must be robust to dropout
Pros: simple; cons: signal loss.
2. Template subtraction
- Measure the artifact template in the absence of neural activity
- Subtract the template during real acquisition
- Requires the artifact to be repeatable
3. Differential amplification + hardware isolation
- Stimulation channels vs. acquisition channels physically separated
- Common-mode rejection cancels artifacts
- Supported by Blackrock Cerebus + Ripple systems
4. Time-division multiplexing
- Stimulation / acquisition alternate
- ~1 kHz switching
- Requires fast switching circuits
5. Adaptive filtering
- Learn artifact structure online using ICA / blind source separation
- Retain neural signal after removal
5. Closed-Loop System Architecture
Typical structure of modern bidirectional BCI:
Brain M1 → neural amplification → decoder → controls assistive device
↓
sensors (force / position / environment)
↓
sensory encoder
↓
ICMS stimulator → S1 brain
Timing budget: the entire loop < 100 ms (physiological latency), otherwise the "sense of causation" is lost.
6. Hochberg Lab biBrainGate
Brown University + Massachusetts General's BrainGate biBRain project:
- Multiple arrays (M1 + S1 + PPC)
- Software: xPC real-time system
- 2024 first demonstration of a complete bidirectional task
- Goal: long-term human trials 2025–2027
7. Neuralink's Bidirectional Direction
The N1 architecture natively supports bidirectional operation (each channel can read or stimulate): - 1024 configurable channels - Attempting S1 + M1 dual arrays - Limited public information
8. Engineering Details
Sampling rate
- Acquisition: 30 kHz / channel
- Stimulation: 10 kHz control
- Synchronization between the two requires a high-precision hardware clock
Low-latency decoding
- Transformer decoders may be too slow — RNN / CNN are more real-time
- NDT3: ~50 ms latency
- EEGNet: ~10 ms
- Choice depends on the task
Multi-array coordination
- Each array has its own amplifier
- A central processor (FPGA / embedded) does real-time fusion
- Operating system: commonly ROS2 or a custom real-time kernel
9. Closed-Loop Calibration (CLDA × Sensation)
Bidirectional BCI requires bidirectional calibration:
- M1 decoder: CLDA (ReFIT and Online Calibration)
- S1 encoder: "which electrode corresponds to which sensation" calibration
- Dual learning — both user and system
This is a complex co-optimization problem — but typically stabilizes after a few weeks of training.
10. Applications
Paralysis
M1 decoding + S1 writing → robotic arm / FES.
Bilateral amputation
M1 decoding + prosthesis + S1 writing → nearly complete hand function.
Memory disorders
Hippocampal read + write → Alzheimer's and similar diseases. See Memory Prosthesis.
Psychiatric disease
Prefrontal read + DBS write → closed-loop depression treatment (work by Mayberg and others).
11. Ethics
Read vs. write rights
The ethics of reading and writing are asymmetric: - Read: privacy concerns - Write: autonomy concerns (who controls "my brain"?)
"Out-of-control" moments
When the system writes automatically → the user may feel "taken over" — emergency-stop mechanisms are mandatory.
Regulation
The FDA regulates bidirectional BCI more strictly than unidirectional — stimulation risk stacks on decoding dependency.
12. Logical Chain
- Bidirectional BCI = read + write closed loop — more natural and more powerful than unidirectional.
- Ganzer 2020 was the first human bidirectional system: M1 → own hand → S1.
- Stimulation artifact is the core engineering challenge (1000× the signal).
- Channel-separation techniques: blanking, template subtraction, hardware isolation, time division, adaptive filtering.
- Closed-loop latency < 100 ms is a physiological requirement.
- Bidirectional calibration requires co-learning of decoder + encoder.
- Ethics: the asymmetry of read vs. write and the takeover risk of automatic writing.
References
- Ganzer et al. (2020). Restoring the sense of touch using a sensorimotor demultiplexing neural interface. Cell. https://www.cell.com/cell/fulltext/S0092-8674(20)30347-2
- O'Doherty et al. (2011). Active tactile exploration using a brain-machine-brain interface. Nature. — Early bidirectional monkey experiment
- Bouton et al. (2016). Restoring cortical control of functional movement in a human with quadriplegia. Nature.
- Flesher et al. (2021). A brain-computer interface that evokes tactile sensations improves robotic arm control. Science.
- Rao (2019). Towards neural co-processors for the brain: combining decoding and encoding in brain-computer interfaces. Current Opinion in Neurobiology.