Semi-supervised ensemble update strategies for on-line classification of fMRI data

• We apply classifier ensembles with online updates to streaming fMRI data.
• We compare five update strategies using naive labels, predicted by the classifiers.
• Ensembles which update using the ensemble decision are shown to be most accurate.
• Updating only when the ensemble is confident in the decision improves performance.

Real time classification of fMRI data allows for neurofeedback experiments, whereby stimuli are updated in accordance with the response of the brain. In order to better facilitate real time fMRI classification, we propose a random subspace ensemble of online linear classifiers. In the absence of true class labels, classifiers are updated using the ‘naive’ label – the label predicted by the classifier. We propose three new ensemble update strategies, using the ensemble decision for updates. Our methods are tested on two emotion based fMRI data sets. We show that the best results are produced by an ensemble which updates using the ensemble decision, constrained by ensemble confidence.