NeuroChain — Blockchain Provenance for EEG Motor Intent

Blockchain that versions EEG-BCI motor models over time — enabling tamper-proof audit, rollback to optimal neural states, and AI clinical reports for motor rehabilitation.

Comment

Inspiration

Every stroke or spinal injury patient who undergoes motor rehabilitation faces the same invisible problem: the AI model trained to decode their brain signals on week 1 is nearly useless by week 12 — because the brain changes, and that history is thrown away. There is no "version control" for neural states. We asked: what if we could Git-commit a patient's motor profile to an immutable blockchain after every session?

What it does

NeuroChain is a blockchain provenance framework for EEG-based Brain-Computer Interface (BCI) systems. It:

  • Captures EEG motor imagery signals (left hand, right hand, feet, tongue)
  • Trains a personalized EEGNet model per session
  • Hashes the model weights using SHA-256 (real cryptography, not simulated)
  • Anchors each snapshot as an immutable block in a local blockchain
  • Detects tampering in under 1ms when any data is altered
  • Rolls back to the last verified optimal neural state
  • Generates AI-powered clinical reports via Google Gemini API explaining what was compromised and recommending rehabilitation protocols

How we built it

Pipeline: EEG Signal → Butterworth Filter (8–30 Hz) → CSP Features → EEGNet Classifier → SHA-256 Hash → Blockchain Block → Gemini Report

  • Python / Google Colab for the ML pipeline and blockchain core
  • EEGNet (Lawhern et al., 2018) as the neural classifier
  • hashlib for SHA-256 — real cryptography, not simulation
  • Google Gemini API for natural language clinical audit reports
  • HTML/CSS/JS for the interactive demo interface
  • Web Crypto API for browser-side hashing in the frontend

The blockchain is a linked list where each block stores:

  • Session timestamp
  • SHA-256 hash of model weights
  • Hash of the previous block
  • Patient session metadata

Tampering any block breaks the hash chain — detectable instantly.

Challenges we ran into

  • Intra-subject variability: EEG signals from the same patient differ significantly between sessions. This is precisely why versioning matters — and also why training a robust personalized model in a hackathon timeframe required careful data augmentation.
  • No ground truth for latency: The academic literature confirms zero studies have measured combined blockchain + BCI latency. We measured our own: tamper detection runs in <1ms; block anchoring adds ~12ms per session.
  • Blockchain without a network: We implemented a single-node chain. True decentralization would require consensus (e.g., Ethereum smart contracts), which is future work.

What we learned

That the gap is real: a systematic review of 729 papers found fewer than 1% address blockchain provenance in BCI systems. The FlyWire connectome project (Nature, 2024) demonstrated that preserving synaptic weight snapshots allows reproducing complex behavior — the same principle, scaled down, is what NeuroChain does for clinical motor rehabilitation.

What's next

  • Integration with real EEG hardware (OpenBCI)
  • Ethereum smart contracts for decentralized consent management
  • Longitudinal study with stroke rehabilitation patients
  • Quantum-resistant hashing (SHA-3) for long-term provenance

Built With

Share this project:

Updates