An Efficient P300-based Brain-Computer Interface with Minimal Calibration Time Fabien LOTTE, Cuntai GUAN Brain-Computer Interfaces laboratory Institute for Infocomm Research (I2R) Singapore
Introduction Brain-Computer Interfaces (BCI) [wolpaw02] Communication devices based on brain activity only Very promising tools for severely disabled people Mostly based on ElectroEncephaloGraphy (EEG)
P300-based BCI The P300 Positive potential occurring ~ 300 ms after a rare and relevant stimulus Very useful for assistive BCI applications Spellers [Farwell88, Nijboer08] Wheelchair control [Rebsamen07, Iturrate09] P300 Average EEG signal at electrode Cz
limitation and objective Current BCI require long calibration times Many examples of EEG signals needed to calibrate the BCI (supervised learning) Re-calibration may be necessary on a regular-basis Inconvenient and uncomfortable for the user A problem for disabled users with limited attention span [birbaumer06] Objective of this paper: Reducing the calibration time of P300-based BCI
State-of-the-art Few attempts to reduce calibration time in P300-based BCI All based on online adaptation [Li08, Liu09] Standard initial training with few examples Online adaptation based on semi-supervised learning Limitation Poor initial performances We propose a simple but efficient method to design P300-based BCI from few EEG examples
Our P300-based BCI design Preprocessing Segment from 150 ms to 500 ms after the stimulus Segment around the P300, if any Low-pass filtering below 25 Hz The P300 is a slow wave Downsampling to 50 Hz A first dimensionality reduction Classical preprocessing [Krusienski06, Thulasidas06] Such processed EEG are usually directly used as features For P300 DETECTION
Feature Extraction with Canonical Correlation Analysis Canonical Correlation Analysis (CCA) [Hardoon04] Find the directions wx and wy maximizing the correlation between the variables X and Y Solved by eigenvalue decomposition Feature extraction Use CCA to find the directions weeg which maximize the correlation between the EEG and the class labels The features are the EEG projected on weeg The features are linear combinations of the initial EEG We obtain new features, most likely most discriminative We can obtain less features, but using only a small number of directions => better for small sample size learning
Regularized CCA Solving CCA requires the estimation of the data covariance matrix C Problem Few training examples => poor estimation Solution Regularization Automatic process with [Ledoit & Wolf 04]
Classification Linear Discriminant Analysis (LDA) Most used and efficient classifier for P300-based BCI [Krusienski06] Also based on covariance matrix estimation Regularized LDA (RLDA) Automatic regularization using [Ledoit & Wolf 04]
Evaluation Single trial analysis of P300 data [Thulasidas06] 10 healthy subjects 8 EEG channels 41 training & testing characters 1 character includes 20 P300 EEG epochs 100 non-P300 EEG epochs Evaluation for various training set sizes ROC analysis: Area Under the ROC curve Comparison of CCA with PCA 50 features extracted from 136 initial EEG samples
Results Results averaged over the 10 subjects
Results
Results
Results
Results
Conclusion New design for P300-based BCI Regularized CCA for feature extraction Regularized LDA for classification Simple to implement and computationally efficient Requires much less training data for high performance => reduced calibration time Future work Online evaluation with disabled users Joint CCA-LDA optimization
Thank you for your attention! Any question? Acknowledgements Dr. David R. HARDOON Dr. Brahim HAMADICHAREF Fabien LOTTE http://sites.google.com/site/fabienlotte/ fprlotte@i2r.a-star.edu.sg
All curves