Thinking out Loud

Inspiration

ALS is a progressive neurodegenerative disease that leads to severe muscle atrophy and paralysis. In advanced stages, patients may lose speech, limb movement, and even reliable eye control.

Most assistive communication systems rely on residual muscle activation. We wanted to explore whether intentional neural signals alone, without any motor movement, could enable structured communication.

Our goal was not just to decode brain signals, but to build a deployment-ready, hardware-agnostic software layer that could realistically integrate with existing EEG systems.

What it does

Thinking out Loud is a muscle-independent brain-computer interface prototype that translates imagined directional intent (“up”, “down”, “left”, “right”) into structured communication.

The system includes:

A clinical dashboard for patient setup and session tracking Real-time EEG acquisition with stimulus-triggered markers An action-aligned neural decoding pipeline A hierarchical communication interface powered by neural selection and an LLM backend

Rather than free-form thought-to-speech, it enables structured question selection through purely neural interaction.

How I built it

We trained our decoding model on an open-source 128-channel imagined speech EEG dataset containing directional thought labels.

Our pipeline consists of:

Strict preprocessing (referencing, notch + bandpass filtering, action-aligned epoching) A filter-bank Riemannian feature extraction approach capturing spatial interaction patterns across frequency bands Regularized logistic regression classification

To ensure realistic performance, we used group-aware cross-validation to prevent session-level data leakage.

We then simulated an 8-channel wearable EEG configuration (matching our Unicorn headset) to validate portability.

On a 4-class imagined direction task (chance = 25%), the model achieved ~45% accuracy under strict validation.

Challenges I ran into

EEG signals are highly subject-specific and noisy. Limited calibration data makes robust decoding difficult. Preventing data leakage while maintaining enough data for validation required careful grouping. Designing a structured communication interface that balances reliability and expressiveness was non-trivial. Integrating real-time EEG acquisition, stimulus triggering, and backend processing within 30 hours required tight coordination.

Accomplishments that I'm proud of

Building a full end-to-end system (not just a model) within 30 hours. Implementing group-aware validation to avoid inflated performance estimates. Maintaining decoding performance when reducing from 128 to 8 channels. Designing a hardware-agnostic architecture rather than a device-specific demo. Delivering a working clinical-style dashboard and live communication prototype.

What I learned

Deployment architecture matters as much as decoding accuracy. EEG decoding benefits significantly from spatial interaction modeling rather than raw waveform analysis. Preventing data leakage is critical for honest performance reporting. Structured communication design can reduce cognitive load compared to letter-by-letter approaches. Building interdisciplinary systems requires strong coordination between signal processing, UX, and backend logic.