NeuroLink
A brain–computer interface that turns imagined movements into commands for a robotic car. EEG is recorded with an 8-channel dry-electrode headset, classified by a lightweight EEGNet model, and sent over WiFi to an ESP32-controlled vehicle. Built as a team of four at Umm Al-Qura University.
My role: I was responsible for the complete EEGNet model, from building it to improving it, and for the windowing of the EEG data.
Case study
How it works
Decisions
I evaluated with Leave-One-Subject-Out cross-validation, so every score comes from a person the base model had never seen. It gives lower numbers than a random split, but it shows how the system really generalizes.
Each user records a short calibration session. Freezing the first EEGNet block and retraining the rest raised mean accuracy from 25.3% to 73.8%.
While testing, we found data leakage: windowing happened before the trial-level split, and evaluation data was used for validation during fine-tuning. We fixed both and enforced strict data separation.
The recording tool rejects poor-quality trials automatically, using band power in the mu and beta ranges and a check on the key motor-cortex channels.
Results and limits
On the validation set, recall was 97% for stop, 89% for right, 88% for left, 87% for backward, and 78% for forward. Foot imagery is the hardest to separate with only 8 electrodes.
The dataset is small: 5 participants with 3 sessions each. In the semi-real-time test on unseen recorded files, a new user reached 92% (23/25) and a previously seen user 64% (16/25), so results vary a lot between people. The test replays recorded EEG rather than a live stream.