Resting-State fMRI Data Classification of Exercise-Induced Brain Changes in Healthy Subjects Using Probabilistic Independent Component Analysis (PICA) Saman Sarraf 1 , Raghda Hasswa 2 , Carol Dematteo 3 , John Connolly 4 , and Michael D. Noseworthy 1,2 1 Electrical and Computer Engineering, McMaster University, Hamilton, Ontario, Canada, 2 School of Biomedical Engineering, McMaster University, Hamilton, Ontario, Canada, 3 School of Rehabilitation Science, McMaster University, Hamilton, Ontario, Canada, 4 Department of Linguistics & Languages, McMaster University, Hamilton, Ontario, Canada Introduction • The level of co-activation between brain regions can be measured by resting-state functional magnetic resonance imaging (rs-fMRI) [1]. These studies are all done while subjects quietly lay awake with eyes open inside the MRI. • We are interested in exercise induced brain effects and recent fMRI work has shown brain activation during a cycling ergometer paradigm [2]. • The purpose of our study was to assess whether there are any significant resting state brain changes immediately following exercise. We hypothesized motor area activation could remain even after the exercise. • We used, as a new method, Probabilistic Independent Component Analysis (PICA), to classify post-exercise rs-BOLD signal. Discussion and Conclusion • The resultant 3D-ANOVA showed there was significant differences post- exercise in the insula, claustrum, Brodmann areas 13 and 21/22 (Fig.7). • It is known the insula is involved in cardiac regulation and has a parasympathetic nervous system role. The claustrum is a communication and relay centre which is possibly more important when more activity and exertion increases (as in exercise). • The inclusion of Brodmann area 13 involves elements of cognitive processing generally which we hypothesize is involved in attention to the task. It is unclear what involvement of Brodmann areas 21/22 indicates as this typically involves language (i.e. includes Wernicke's area). • The results from the PICA approach demonstrate that numerous non-motor areas in the brain become important post-exercise. • The PICA method could clearly distinguish the different activated areas in the brain and showed a high level of consistency. References [1] van den Heuvel MP, Hulshoff Pol HE (2010) Eur. Neuropsychopharm. 20:519–534. [2] Mehta JP, et al. (2009) J. Neurosci. Meth. 179:230–239. [3] Woolrich MW, et al. (2001) NeuroImage 14:1370-1386. [4] RW Cox. (1996) Comp. Biomed. Res. 29:162-173. [5] Smith SM, et al. (2004) NeuroImage 23:208- 219. Results The average number of components extracted by PICA was 54 using the multi- session temporal concatenation method. The first component of the pre and post exercise data are shown in Figs. 3 and 4, respectively, including their threshold IC maps (p<0.05) and raw z-transformed maps (1-99 percentile). FSL–MELODIC orders the components by decreasing amounts of uniquely explained variance. Therefore, several first components with high explained variance values were selected for both groups of study. The Talairach coordinate system was used to define the brain regions. Figure 1. Multi-Session Temporal Concatenation diagram applied to fMRI data (top) Figure 2. Multi-Session Tensor-ICA diagram applied to fMRI data (bottom) FMRI data 1 FMRI data 2 Spatial maps Components Time Time Time Time Components Space Space Time Space FMRI data Components Time Components Space Spatial maps Figure 3. The first PICA component of pre-exercise data (top: threshold map, bottom: z-transformed map) Figure 4. The first PICA component of post-exercise data (top: threshold map, bottom: z-transformed map) Figure 5. The 1 st PICA component signal of pre-exercise Figure 6. The 1 st PICA component signal of post-exercise Figure 7. 3D repeated measures ANOVA comparing post-exercise rs-BOLD vs. pre-exercise rs-BOLD Materials and Methods - study approved by St. Joseph’s Healthcare REB Protocol: 1) Anatomical Scan: 3D IR-prepped fSPGR T1-weighted sequence (24cm FOV, TE/TR/TI/flip= 2.1/7.5/450/12 o , 512x512, 1.5mm thick/ 0mm skip, 140 acquired slices). 2) Pre-exercise rs-BOLD: (24cm FOV, TE/TR=35/2000ms, flip=90 o , 64x64, 4mm thick/0mm skip, 35 slices, 180 phases; 6 minutes). 3) Exercise Paradigm: Subject pedaled for 30s as a warm-up and to familiarize them with the ergometer motion (MRI compatible, Lode, The Netherlands; see figure below). During this 30s the power was gradually increased from 0 to 5 Watts. The subject then pedaled for 3 minutes, at 5W. This power was chosen because power above 5W was difficult for subjects to maintain. 4) Post-exercise rs-BOLD (same parameters as pre-exercise) Data Analysis - motion correction and spatial registration of all images (pre and post exercise) to the first pre-exercise image - Probabilistic Independent Component Analysis (PICA) applied for classification. This is described by a generative linear latent variables model [2]. Equation (1) explains the PICA model for classification: Where, x i represents the p-dimensional vector of individual measurements at voxel i, s i is the q-dimensional vector of non-Gaussian source signals (q<p) and finally, η i denotes Gaussian noise which is explained by [2]: The vector μ defines the mean of the observations x i where the index i is over the set of all voxel locations ʋ. Nonlinear decorrelation is the basic ICA method. Independent components are the maximally non-Gaussian components which are calculated by Maximum Likelihood estimator Equation (2): Where λ l is the eigenvalue of voxel-wise pre-whitened data vectors. The PICA approach was performed using FSL-MELODIC (Model-free ICA-based analysis for resting-state fMRI) [5]. The multi-session temporal concatenation method Fig. 1 and multi-session tensor-ICA Fig. 2 were applied to data. Individual subject results from FSL-MELODIC were spatially transformed to the Talairach coordinate system and analyzed with the 3D-ANOVA command in AFNI [4], comparing after exercise vs. before. (1) (2) Photo (left) showing 3T MRI and Lode ergometer. On right is a schematic of the ergometer and a subject (http://lode.nl/en/). View publication stats View publication stats