Eye State Prediction from EEG Data Using Boosted Rotational Forests

With the advent of brain-computer interfacing (BCI) technology, the need for fast and accurate classification and prediction of brain activity from electrode recording data is becoming more critical. The classifier systems developed in the current study improve upon the previous state-of-the-art developed by Rösler & Suendermann's (2013) for detecting whether an individual's eyes are open or closed based on EEG recordings. It was hypothesized that a system could be constructed with comparatively accurate classifications to Rösler & Suendermann's K*-based system, but with enough speed to be utilized within a BCI by constructing ensembles from eager learners (e.g. decision trees). To meet these requirements, three ensemble learners were developed: a rotational forest that implements random forests as its base classifiers (RRF), a rotational forest that implements J48 trees as its base classifiers and is boosted by adaptive boosting (ada(RJ48F)), and an ensemble of the RRF model with Rösler & Suendermann's K* model (RRF+K*). The RRF and ada(RJ48F) models were shown to have a classification speed faster than the K* model by a factor of seventeen and thirteen respectively, though the RRF model was significantly less accurate and the ada(RJ48F) model had an accuracy comparable to K*. The results of this study indicate that instance-based learners like the K* algorithm are likely to be too slow to be used within a BCI device, while the ada(RJ48F) model performed accurate classifications well within the time constraints of real-time classification and control.