Dimensionality reduction of fMRI time series data using locally linear embedding.

OBJECTIVE: Data-driven methods for fMRI analysis are useful, for example, when an a priori model of signal variations is unavailable. However, activation sources are typically assumed to be linearly mixed, although non-linear properties of fMRI data, including resting-state data, have been observed. In this work, the non-linear locally linear embedding (LLE) algorithm is introduced for dimensionality reduction of fMRI time series data. MATERIALS AND METHODS: LLE performance was optimised and tested using simulated and volunteer data for task-evoked responses. LLE was compared with principal component analysis (PCA) as a preprocessing step to independent component analysis (ICA). Using an example data set with known non-linear properties, LLE-ICA was compared with PCA-ICA and non-linear PCA-ICA. A resting-state data set was analysed to compare LLE-ICA and PCA-ICA with respect to identifying resting-state networks. RESULTS: LLE consistently found task-related components as well as known resting-state networks, and the algorithm compared well to PCA. The non-linear example data set demonstrated that LLE, unlike PCA, can separate non-linearly modulated sources in a low-dimensional subspace. Given the same target dimensionality, LLE also performed better than non-linear PCA. CONCLUSION: LLE is promising for fMRI data analysis and has potential advantages compared with PCA in terms of its ability to find non-linear relationships.