| Using a classification guided feature selection (wrapper method) in conjunction with a new performance metric, I present a solution to multi-class subject invariant Brain Computer Interface (BCIs) using electroencephalography (EEG) and near infrared spectroscopy (NIRS) signals, a complex problem known to be prone to trivial classification. In a data-driven multi-class BCI, evaluation of the one versus rest (OVR) classifier is a major challenge using error rate. The hence derived multiclass OVRs using wrapper methods with error rate as the classifier feedback can show degeneracy in terms of imbalance in sensitivity and specificity, leading to trivial classification. This imbalance can be removed by the usage of a scalar quality factor as the performance metric. The error rate is replaced by a simple scalar quality factor that adjusts the simple correct rate with the ratio of sensitivity and specificity. A 4-class subject invariant EEG-based BCI using signals from 10 untrained subjects is presented here to prove the efficacy of the quality metric. Left hand, right hand, left leg, and right leg movements are classified using Naive Bayesian, Gaussian SVM, Polynomial SVM, and k-Nearest neighbor classifiers. Extending the same method to optical signals, here I present an NIRS-based BCI using signals from two subjects to classify left hand and right hand movements. The same quality-metric based wrapper methods could identify the salient time samples of oxy-hemoglobin (HbO) and deoxy-hemoglobin (Hb) channels from NIRS signals to achieve 100% classification rate, sensitivity, and specificity. |
