CSE - Minimal‑Calibration Motor‑Imagery BCI for Enhanced Neural Decoding

Motor Imagery Brain-Computer Interfaces (MI-BCIs) using EEG are systems that allow users to control external devices or communicate by imagining specific movements, such as moving a hand or foot. This mental rehearsal generates distinctive patterns in the brain’s electrical activity, which are detected through Electroencephalography (EEG). MI-BCIs decode these patterns in real-time, enabling applications like neurorehabilitation, assistive technology, and interactive systems for individuals with motor impairments. As a non-invasive method, EEG-based MI-BCIs offer a practical approach to harnessing neural signals directly for intuitive human-computer interaction.

Build and evaluate a binary MI EEG decoder (left vs. right hand) using open‑access EEG. The scientific goal is to show that simple, label‑free calibration can improve cross‑subject generalization with minimal user data and modest compute. We will use classical machine learning‑based baselines and compact CNNs to quantify trade‑offs between accuracy and calibration time. The deliverable is a reproducible pipeline and a short report or publication suitable for an undergraduate showcase or workshop poster.

Tasks and Techniques:

Data & Preprocessing

• Use open MI‑EEG datasets
• Epoching around MI cues and filtering the data
• Light artifact handling 

Classical Machine Learning and SOA Deep Learning baselines

• Training with early stopping and simple augmentations: channel dropout, small time‑shift (±100 ms)
   

Simple, Label‑Free Calibration

Feature‑space alignment : CORAL (Correlation Alignment) or any other feature alignment method, using unlabeled target‑subject features.
 

Evaluation & Analysis

• Protocols: within‑subject (session‑wise) and leave‑one‑subject‑out (LOSO).
• Calibration conditions: unlabeled adaptation, optional tuning with few labeled trials
• Metrics: balanced accuracy, mean ± SD across subjects.
• Statistics: Wilcoxon signed‑rank tests and effect sizes to compare baselines vs. calibrated models.
• Efficiency reporting: parameter counts and training time on standard hardware.

Tooling

• Python; MNE‑Python ; scikit‑learn ; PyTorch 
   

How will students participate?

Students will own clearly defined, achievable components:

1. Data & Preprocessing 

   • Download/open datasets, implement filtering/epoching, artifact checks, and configuration scripts.

    

2. Classical ML 

   • Run baselines and document hyperparameters.

    

3. Deep Learning 

   • train with early stopping and augmentations; manage training logs.

    

4. Calibration/Adaptation 

   • Add CORAL features using unlabeled target data; (optional) few‑shot tuning with 5–10 labeled trials.

    

5. Evaluation & Reporting
   • Run within‑subject and LOSO experiments; produce accuracy‑vs‑calibration curves; conduct Wilcoxon tests; create figures/tables.

   • Maintain a tidy GitHub repository (fixed seeds, configs, README, environment file) and write a concise final report/short paper.

      

Expected outcomes: a runnable codebase, clear quantitative results showing the impact of simple calibration on MI‑BCI performance, and a short paper/poster summarizing methods, findings, and limitations.