Project Logs

Collapse

• Open-Loop Alpha Monitoring with Light Stimulations

  Neuradock001 • 5 days ago • 0 comments

  What We Built

  This week we put together an open-loop alpha monitoring workflow using the NeuraDock EEG Workstation (7-channel dry-electrode) paired with a wearable light stimulation device.

  
  The idea is simple: run a predefined light stimulation script while recording EEG in real time, then inspect how the brain responds. No closed-loop feedback yet — just clean observation, solid data, and a dual-pipeline setup that gives us both live process monitoring and deep offline analysis.

  System Architecture

  Three units, one stream, two pipelines:
  

  Unit	Role
  Sensing	NeuraDock — 7-channel dry EEG, posterior visual-area focus
  Modulation	Wearable light stim — runs preset scripts
  Control	Software layer — sends commands, records event markers, splits EEG into two analysis paths

  The key design choice: One EEG stream feeds two pipelines simultaneously.

  • Online pipeline → real-time alpha trend dashboard (process monitoring, not diagnosis)
  • Offline pipeline → post-session power, coherence, and asymmetry analysis

  This split lets us validate the experiment while it runs and still dig deep after the session ends.

  Online Pipeline: Real-Time Alpha Trend

  During the session, we visualize an alpha-related proxy metric live. Think of it as a flight dashboard — it won't tell you the full story, but it immediately tells you if the session is behaving as expected.
  We treat this as a process-monitoring signal, not a clinical endpoint. If the trend flatlines when it shouldn't, we know something is off with the hardware or the subject's state before we waste a full session.

  Offline Pipeline: The Real Work

  After the session, we analyze three conditions side by side:

  1. Eyes open (baseline)
  2. Eyes closed (natural alpha rebound)
  3. Eyes closed + light stimulation (target condition)

  Metrics we ran:

  • Alpha power — standard bandpower inspection across conditions
  • Alpha-band coherence — phase consistency between channel pairs; higher coherence = more synchronized oscillation
  • Hemispheric alpha asymmetry — log(right α) - log(left α) as an exploratory feature

  In our single-session demo, the stimulation condition showed stronger alpha-band synchronization in selected posterior channel pairs. We’re not calling this a therapeutic result — it’s an exploratory workflow for linking stimulation protocols to measurable EEG dynamics.

  offline alpha power analysis: eye open

  offline alpha power analysis: eye close

  Why Open-Loop First?

  Everyone wants closed-loop neurofeedback. But before you close the loop, you need to answer a more basic question:

  What EEG changes actually show up when you apply a predefined stimulation script?

  Open-loop monitoring is the honest first step. It lets us:

  • Observe stimulation-associated EEG changes without confounding feedback effects
  • Verify that the experiment protocol runs as designed
  • Record clean, marker-aligned raw data for future model training
  • Build the data foundation that closed-loop systems actually need

  What We Learned

  • Dry EEG + event markers work. The 7-channel posterior setup captures usable alpha dynamics without gel or prep.
  • Dual-pipeline is the right split. Real-time dashboard prevents bad sessions; offline analysis extracts the science.
  • Coherence is sensitive to stimulation. In our demo, light stimulation during eyes-closed showed elevated alpha-band coherence between selected channels — worth expanding with more subjects and controlled protocols.
  • Asymmetry is exploratory. Interesting, but interpretation-heavy. We keep it in the offline toolbox, not the live dashboard.

  Limitations (We're Honest About This)

  This is an exploratory demo, not a clinical validation. Single session, no statistical power, no therapeutic claims. What we proved is the workflow — not the therapy.
  To move from "interesting demo" to "validated protocol," we need: more participants, repeated sessions, sham controls, and proper statistical testing.

  What's Next

  We're stacking more workflows on this foundation:

  • SSVEP and motor imagery pipelines ...

  Read more »

• Motor Imagey Paradigm Demo

  Neuradock001 • 7 days ago • 0 comments

  We’ve been obsessed with a question: Can we decode motor imagery without asking the user to "just imagine" moving their hand? Traditional MI-BCI suffers from one brutal flaw — it relies on the user’s willingness and ability to produce clear motor imagery, which many patients (especially post-stroke) struggle with.

  
  So we built a hack. Instead of going directly after the motor cortex, we’re piggybacking on the dorsal visual stream — the brain’s "Where/How" pathway that processes visual motion and spatial planning. By tracking how the occipital and temporo-parietal areas respond to a falling ball visual stimulus, we can infer motor intent indirectly. Here’s how we wired it up.

  The Neuroscience Hack

  The dorsal stream runs V1 → V5/MT → posterior parietal cortex (SPL/IPL) → premotor cortex. When a patient watches a ball fall and receives a "grasp" cue, this whole highway lights up:
  
  

  Stage	Region	What It Does	Our Electrodes
  Visual input	V1 (Occipital)	Processes motion stimulus	O1, O2, Oz
  Motion/spatial planning	SPL/IPL (Parietal)	Integrates visual space + motor intent	via coherence analysis
  Attention relay	TPJ (Temporo-parietal junction)	Coordinates visual attention & motor prep	T5, T6

  Our 7-channel layout (O1, O2, Oz, T5, T6, plus references) spans the occipital → temporo-parietal segment of the dorsal stream. We’re not reading motor cortex directly — we’re sniffing the traffic on the visual-motor integration highway.
  

  The Experimental Paradigm

  We designed a 3-phase trial that forces the brain into the right state without relying on the user’s "imagination discipline":
  表格
  

  Time	Phase	Visual Stimulus	User State	EEG Purpose
  0.0–2.0 s	Baseline	Ball falls, no flashing	Relax, passive watch	REST reference
  2.0–4.0 s	Motor Imagery	Ball flashes as it nears the hand	Actively imagine grasping	ERS + coherence features
  4.0 s+	Feedback	Robotic hand moves or animation plays	Relax / observe	Backup data

  Why this works:

  • 0–2s: The moving (but non-flashing) ball pre-activates the visual pathway. We use this as a within-trial baseline — no more fighting individual alpha differences across sessions.
  • 2–4s: The flashing ball at "grasp distance" triggers instinctive motor anticipation. The user doesn’t have to "try" to imagine — the brain prepares to grasp automatically. This is the hack: we hijack the visual affordance of a falling object to force motor planning.
  • Net task duration: 1.8s (minus 200ms VEP onset transient). At 250 Hz, that gives us 450 clean samples per trial.

  The Algorithm

  We use a two-stage pipeline to extract intent from this visual-motor traffic:

  Stage 1: SSVEP-Driven ERS

  We measure Event-Related Synchronization (ERS) during the MI phase relative to baseline. The flashing ball drives SSVEP in the occipital channels, and we track how alpha-band power changes as the dorsal stream engages.

  Stage 2: Cross-Channel Coherence (Alpha & Beta)

  We compute spectral coherence between electrode pairs to track information flow along the dorsal stream:
  表格
  

  Electrode Pair	Direction	Band	Neural Meaning
  O1 – T5	Left dorsal stream	Alpha (8–13 Hz)	Left visual → temporo-parietal activation
  O2 – T6	Right dorsal stream	Alpha (8–13 Hz)	Right visual → temporo-parietal activation
  Oz – T5	Midline → left temporal	Alpha (8–13 Hz)	Bilateral convergence
  Oz – T6	Midline → right temporal	Alpha (8–13 Hz)	Bilateral convergence

  For single-trial power spectra, we use Welch’s method — more stable than raw FFT for our short 1.8-second windows.
  
  

  What’s Next

  This paradigm is currently running on our 7-channel NeuraDock hardware. The next log will cover the real-time implementation: how we pipeline the 250 Hz stream through a lightweight online classifier, and whether we can hit <500ms latency from flash onset to intent detection.

  
  We’re also exploring whether adding a beta-band (13–30 Hz) coherence layer can separate "true motor planning" from pure visual attention — but that’s for the next hack.

  
  Questions? Want to replicate this? Our EEG workstation examples are on GitHub released very soon:https://github.com/Neuradock...

  Read more »

• SSVEP Paradigm Demo with Data Processing

  Neuradock001 • 06/04/2026 at 14:46 • 0 comments

  SSVEP data results
  

  SSVEP experiment  and data process video:

• Replicating a NeurIPS Visual Decoding Paper with 7-Channel Dry Electrodes

  Neuradock001 • 05/27/2026 at 01:58 • 0 comments

  # Replicating a NeurIPS Visual Decoding Task with 7-Channel Dry Electrodes

  *Can wearable EEG hardware match lab-grade equipment for brain-state decoding? We ran the experiment.*

  **Tags:** EEG · BCI · Visual Decoding · NeurIPS 2024 · Dry Electrodes · ADS1299

  ---

  ## Overview

  EEG-based visual decoding has long been the domain of well-funded neuroscience labs. The bottleneck isn't the algorithms — it's the hardware. A clinical-grade wet-electrode amplifier costs upward  $28,000 USD, requires conductive gel, and is far from portable.

  We built NeuraDock as a wearable, dry-electrode EEG platform and asked a straightforward question: **can it actually hold up against lab-grade equipment on a state-of-the-art decoding task?**

  > **Short answer:** Yes. Using 7-channel dry electrodes in an everyday environment, we replicated the NeurIPS 2024 visual decoding benchmark and matched the paper's reported Top-1 accuracy at equivalent channel counts — at roughly 1/7th the hardware cost.

  ---

  ## The Benchmark: NeurIPS 2024 Visual Decoding

  We replicated *Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion* (NeurIPS 2024). The paper's core task: decode and reconstruct natural images purely from EEG signals recorded while subjects viewed photos.

  The framework has three main technical components:

  - **Multimodal latent alignment (CLIP):** EEG features are aligned to CLIP image embeddings via contrastive learning — bridging brain signals and image semantics.
  - **ATM encoder (Adaptive Temporal-spatial Modeling):** A custom EEG encoder that extracts spatiotemporal features at state-of-the-art efficiency.
  - **Two-stage image generation:** Separate extraction of high-level semantic features ("a cat") and low-level visual features ("edges, color") — followed by a lightweight prior diffusion model — achieving reliable reconstructions from under 500ms of EEG data.

  The original paper tested multiple channel configurations. Critically, it reported that **7-channel EEG achieves 16–25% Top-1 accuracy** in image classification — giving us a meaningful comparison point.

  The live demo is as below:

  ---

  ## Hardware Comparison

  | Dimension | Original paper | NeuraDock (this work) |
  |---|---|---|
  | Channel count | 64-ch subset → 7 for comparison | 7 channels |
  | Electrode type | Wet (conductive gel) | Dry (comb-tooth) |
  | Hardware | BrainVision actiCHamp | NeuraDock ADS1299-based |
  | Portability | Lab-only | Wearable, everyday use |
  | Setup time | 20–40 min (gel application) | <5 min |

  **NeuraDock device specs:**

  | Parameter | Spec |
  |---|---|
  | Channels | 7 (+ REF + GND) |
  | Electrode positions | T5, T6, PO3, PO4, O1, Oz, O2 (occipital + temporal) |
  | Sample rate | 250 Hz |
  | Resolution | 24-bit |
  | Connectivity | Bluetooth / USB |
  | Mounting | Wraparound headband, Velcro |

  ---

  ## Key Results

  | Metric | Original paper (7-ch) | NeuraDock (7-ch) |
  |---|---|---|
  | Top-1 accuracy | ~20% | 20.5% |
  | Top-5 accuracy | ~55% (full config) | ~45% |
  | 2-way retrieval | Comparable | Comparable |
  | 4-way retrieval | Comparable | Comparable |
  | Data trials | Single-trial | Multi-trial average |

  On 2-way and 4-way retrieval tasks, NeuraDock's 7-electrode configuration performed on par with the paper's 64-electrode results. Individual image reconstructions capture the primary semantic content and basic contours of the original stimuli — consistent with the paper's 7-channel results.

  ---

  ## Engineering Decisions

  ### 1. Signal Filtering and Timing Calibration

  Because we use a software-based marker system (USB + software timestamps) rather than a hardware trigger line, we needed to characterize and compensate for system latency. We swept across delay offsets and three filter bands:

  - **1–20 Hz:** Significant accuracy drop — gamma-band (>30 Hz) content is essential for visual decoding.
  - **1–45 Hz:** Best trade-off — preserves visual features, rejects power-line and high-frequency muscle artifacts.
  - **1–100...

  Read more »

View all 4 project logs