2025-26
EEG flow-state detection - funded summer research
Python, MNE/DSP, EEG-BIDS, scikit-learn, Optuna, MLflow, LSL/XDF
A 10-week funded research project (“In the Zone: Decoding Flow from Wearable EEG”), continuing into my current honours dissertation. The value of this project is the research process, not a leaderboard number: real-world data gathered and processed at scale, a pipeline built from scratch, and a labelling protocol that was my own idea.
- Designed and built a complete pipeline from raw multi-modal lab recordings (16-channel EEG plus accelerometer) to a standards-compliant EEG-BIDS dataset and an ML-ready feature set: 306,682 windows, 119 features per window, zero missing or invalid values, reproducible via a config hash.
- Ran real experimental sessions (5 participants, 58 usable sessions) and helped design the protocol; devised a semi-supervised labelling strategy (confident labels, then ensemble pseudo-labelling and retraining with temporal smoothing).
- Automated marker cleaning that validated 58/58 against manual review; root-caused a six-order-of-magnitude accelerometer unit bug; used leave-session-out grouped cross-validation to avoid scaler leakage.