Subject-Independent Brain Computer Interface through Boosting Shijian Lu, Cuntai Guan, and Haihong Zhang Institute for Infocomm Research 1, Fusionopolis Way, #21-01 Connexis Singapore 138632 slu,ctguan,hhzhang@i2r.a-star.edu.sg Abstract This paper presents a subject-independent EEG (Electroencephalogram) classification technique and its application to a P300-based word speller. Due to EEG variations across subjects, a user calibration procedure is usually required to build a subject-specific classifi- cation model (SSCM). We remove the user calibration through the boosting of a committee of weak classifiers learned from EEG of a pool of subjects. In particular, we ensemble the weak classifiers based on their confi- dence that is evaluated according to the classification consistency. Experiments over ten subjects show that the proposed technique greatly outperforms the super- vised classification models, hence making P300-based BCIs more convenient for practical uses. 1 Introduction The emerging technology of brain-computer inter- face (BCI) has attracted increasing interest from multi- disciplinary domains [1]. The technology directly trans- lates brain signals into communication messages while bypassing normal neuromuscular pathways. Thus it po- tentially provides severely paralyzed people with com- munication, control or rehabilitation tools to help com- pensate or restore their lost capabilities. P300 is an endogenous, positive polarity component of the event-related brain potential (ERP) and it has been widely used for the purpose of brain computer interface (BCI). Farwell and Donchin [2] first demon- strate the use of P300 in a so-called oddball paradigm. In the paradigm, the computer displays a matrix of cells and flashes each row and column shown in Fig. 1 alter- nately in a random order. Subjects needs to focus on a cell for a short while, meanwhile a P300 ERP will be elicited in the subject’s EEG (Electroencephalogram) when the row or the column specifying the focused cell flashes. The elicited P300 can then be identified by sig- Figure 1. Interface of P300-based speller. nal processing and machine learning algorithms [3, 4]. Many studies [7, 8] have shown variations of P300 across subjects. In particular, P300 amplitude and la- tency vary among both normal and clinical populations shown in Fig. 2. As a result, P300 models learned from one subject would not apply well to another subject. To deal with such EEG variations, most P300-based BCIs usually perform a user calibration to build a subject- specific classification model (SSCM). But the user cali- bration makes BCIs inconvenient for practical uses. This paper presents a subject-independent EEG clas- sification technique that does not require the user cal- ibration. The proposed technique is based on the ob- servation that P300 of different subjects usually share common waveform characteristics as defined, namely, a positive peak after around 300 ms of the external stim- uli. It directly classifies EEG a new subject by boost- ing multiple weak EEG classifiers that are learned from EEG of a pool of existing subjects. 2 Proposed Techniques This section presents our proposed EEG classifica- tion technique including the EEG preprocessing, the EEG classification by using linear discriminant, and the boosting classification, respectively. 978-1-4244-2175-6/08/$25.00 ©2008 IEEE