Application of spatiotemporal Gaussian Process model for fMRI data analysis

Application of spatiotemporal Gaussian Process model for fMRI data analysis

Background: Statistical analysis is base on preparing brain maps. Their accuracy and reliability are essential. Adjusting models for considering spatiotemporal correlation that is embedded in fMRI data can increase accuracy, but it introduces a high computational cost. Material and methods: We applied a spatiotemporal Gaussian process model (STGP) for task-based fMRI data. This model modified common group-level-GLM for spatiotemporal correlation in a reasonably fast way that can solve the underestimation of parameter estimation variation and leads to a more accurate result with a less false positive rate. A simulation study was conducted to assess the accuracy of these models. Results: Proposed model and group-level-GLM were fitted to a memory tfMRI data. The main activated area was the frontal brain lobe, as mentioned in previous studies. Z-score was computed for all voxels, and functional and activation maps for both models were calculated. The STGP model increased the absolute maximum Z-score by about 18 and 13 units compared to the group-level GLM. In the simulation study, the STGP model resulted in more accurate results (higher accuracy; lower: FPR) compared to the GLM. Conclusion: The STGP model was applied to denoise group-level GLM for valid inference. This model resulted in a higher Z-score and more accurate results for experimental and simulated data.