Learning from the Past: Postprocessing of Classification Scores to Find a More Accurate and Earlier Movement Prediction

Abstract

Brain–computer interfaces performing movement prediction are useful in a variety of application fields from telemanipulation to rehabilitation. However, current systems still struggle with a level of unreliability requiring improvement, so that the full potential of these systems can be used in the future. Here, we suggest to improve the performance and robustness of classification outcomes by postprocessing the raw score values with the history of previous classifications. For this several postprocessing methods that operate on the classification outcomes are investigated. In particular, the data was classified after preprocessing using a support vector machine (SVM). The output of the SVM, i.e. the raw score values, were postprocessed using previously obtained scores to account for trends in the classification result. The respective methods differ in the way the transformation is performed. The idea is to use trends, like the rise of the score values approaching an upcoming movement, to yield a better prediction in terms of detection accuracy and/or an earlier time point. We present results from different subjects where upcoming voluntary movements of the right arm were predicted using movement related cortical potentials from the EEG. The results illustrate that better and earlier predictions are indeed possible with the suggested methods. However, the best postprocessing method was rather subject-specific. Finally, we use straightforward ensemble approaches to exemplify how the methods can be directly used in an application and how this can influence the overall movement prediction performance. Depending on the requirements of the application at hand, postprocessing the classification scores as suggested here can be used to find the best compromise between prediction accuracy and time point.