Temporal pattern based classification of independent components in resting state fMRI

The analysis of the resting state fMRI is hampered by the confounding presence of the artefacts Independent component analysis (ICA) presents a data-driven approach, ideally, separating noise and independent components (IC) of interest. The automatic identification of meaningful ICs in the resting state fMRI is done using three classification algorithms: multi-layer perceptron (MLP), support vector machines (SVM), and random forest (RF) based only on temporal IC patterns. The algorithms' performance was evaluated using manually labeled resting state fMRI data of 13 subjects. The achieved accuracy on group level is 91%, 85, 77% and 89,83% for MLP, SVM and RF, respectively. MLP performed the best on the reduced feature set, providing the best recall of 89% for the meaningful class and the best individual accuracy of 96%.