Brain-computer interfaces (BCI) utilizing steady-state visually evoked potentials (SSVEP) recorded by electroencephalography (EEG) have exciting potential to enable new systems for disabled individuals and novel controls for robotic and computer systems. To interact with SSVEP-based BCIs, users attend to visual stimuli modulated at predetermined frequencies. A key problem for SSVEP-based BCIs is to classify which modulation frequency the user is attending, for which there is an inherent trade-off between speed and accuracy. As SSVEP signals vary with time and stimulation frequency, a fixed-length data window does not necessarily optimize this trade-off. We propose a strategy, developed from sequential analysis, to vary the window-length used for classification. Our proposed technique adapts to the data, continuing to collect data until it is confident enough to make a classification decision. Our strategy was compared to a fixed window-length method using a simple experiment involving five frequencies presented individually to three participants. Using a canonical correlation analysis classifier to compare the proposed variable-length scheme to a standard fixed-length scheme, the variable-length approach improved the classifier information transfer rate by an average of 43%.