Functional Magnetic Resonance Imaging ( fMRI) is an exciting technology which allows neuroscientists to gather data on activity within the brain. This activity corresponds to neural processes such as emotion or motor activities. By applying machine learning techniques to fMRI data, the patterns corresponding to these processes can be recognised and classified.
The ability to classify neural processes opens up a wealth of opportunities to neuroscientists. Early fMRI experiments focus on identifying regions of the brain involved in processes such as pain or emotion. Having identified these regions, it is possible to see how they react to stimuli differently in participants with different conditions, for example depression, autism or attachment disorder.
Most of this work is exploratory in nature, with analysis being carried out offline, that is, once the fMRI data collection is complete. More recent advances in classification speed and accuracy, and in fMRI technology, have allowed for real time experiments. During real time fMRI experiments the classifier is trained and then used during the course of the experiment. In what is termed a neurofeedback loop, the stimuli presented to the participant can be updated or altered dependent upon the classification result of the output data. Real time fMRI has been used in many proof of concept type experiments, such as navigating mazes or balancing a pendulum by using different thought processes.
In order to better facilitate real time fMRI classification, it is proposed here that an online classifier will be advantageous. During the course of a real time fMRI experiment, training data is often very limited, therefore the ability of a classifier to learn from new data during the course of the experiment will be of benefit. In order to maintain speed and accuracy, we propose a random subspace ensemble of linear classifiers.
Further to this, it is noted that in many cases, during the online phase, true class labels may not be known. The use of an online 'naive labelling' classifier within an ensemble framework is proposed as an alternative to a fixed pre-trained classifier. This is extended by the introduction of a 'guided update' strategy for the ensemble, whereby the classifiers within the ensemble are updated using the ensemble decision, rather than the individual decisions. Comparison of this strategy with a fixed classifier ensemble and an ensemble of classifiers with individual 'naive' updates is provided. Variations of the guided update strategy are also proposed, whereby classifiers within the ensemble are only updated when specific criteria are fulfilled. These criteria are based upon agreement with the
ensemble decision, and confidence in the ensemble decision.
The proposed methods are shown to provide more accurate results than
using a fixed classifier, and are tested across a variety of emotion based fMRI data sets.