EEG Finger Movement Detection
Lightweight CNN with attention for real-time BCI — 85% accuracy, F1 0.75
During my internship at CSIR-Central Electronics Engineering Research Institute (CEERI), Rajasthan, I worked on brain-computer interface (BCI) research using multi-channel EEG data.
Problem
Real-time finger movement classification from EEG signals — a core BCI problem with applications in motor rehabilitation and assistive technology. Challenges: high noise, inter-session variability, and strict latency requirements for real-time deployment.
Work
Exploratory analysis:
- Analyzed multi-channel EEG time-series for temporal correlations across electrode channels during different motor imagery tasks
- Built an interactive visualization dashboard (Plotly/Dash) for exploring channel-level activation patterns and temporal dynamics
CNN classifier with attention:
- Lightweight architecture designed for low-latency inference (real-time constraint)
- Temporal convolutions to capture short-range EEG dynamics
- Channel attention mechanism to weight informative electrode channels
- Input: raw EEG segments (multi-channel time windows)
Results:
- 85% accuracy on held-out test sessions
- F1-score: 0.75 across 5 finger movement classes
- Inference latency compatible with real-time deployment (<10ms per segment)
Stack: Python · PyTorch · NumPy · Plotly/Dash · MNE (EEG processing)
Period: May – July 2023 · CSIR-CEERI, Rajasthan, India