Prediction of Task-Related BOLD fMRI with Amplitude Signatures of Resting-State fMRI

Abstract

Blood oxygen contrast-functional magnetic resonance imaging (fMRI) signals are a convolution of neural and vascular components. Several studies indicate that task-related (T-fMRI) or resting-state (R-fMRI) responses linearly relate to hypercapnic task responses. Based on the linearity of R-fMRI and T-fMRI with hypercapnia demonstrated by different groups using different study designs, we hypothesized that R-fMRI and T-fMRI signals are governed by a common physiological mechanism and that resting-state fluctuation of amplitude (RSFA) should be linearly related to T-fMRI responses. We tested this prediction in a group of healthy younger humans where R-fMRI, T-fMRI, and hypercapnic (breath hold, BH) task measures were obtained form the same scan session during resting state and during performance of motor and BH tasks. Within individual subjects, significant linear correlations were observed between motor and BH task responses across voxels. When averaged over the whole brain, the subject-wise correlation between the motor and BH tasks showed a similar linear relationship within the group. Likewise, a significant linear correlation was observed between motor-task activity and RSFA across voxels and subjects. The linear rest-task (R-T) relationship between motor activity and RSFA suggested that R-fMRI and T-fMRI responses are governed by similar physiological mechanisms. A practical use of the R-T relationship is its potential to estimate T-fMRI responses in special populations unable to perform tasks during fMRI scanning. Using the R-T relationship determined from the first group of 12 healthy subjects, we predicted the T-fMRI responses in a second group of 7 healthy subjects. RSFA in both the lower and higher frequency ranges robustly predicted the magnitude of T-fMRI responses at the subject and voxel levels. We propose that T-fMRI responses are reliably predictable to the voxel level in situations where only R-fMRI measures are possible, and may be useful for assessing neural activity in task non-compliant clinical populations.

Keywords: BOLD; breath hold; fMRI; hypercapnia; motor cortex; prediction; resting-state fluctuations; vascular.

Figures

Figure 1

Estimation of the BOLD amplitude change during the resting state (top) and task (bottom) using the temporal SD of the time series.

Figure 2

Spatiotemporal structure of the BOLD amplitude change (SD) from the motor-task region of interest (ROI). (A–C) Spatial and (D–I) temporal. (A) Resting state, (B) motor task, and (C) BH task. Voxels with large RSFA also tend to have large BOLD amplitude change during the motor and BH tasks. Such a spatial correspondence between R-fMRI, T-fMRI, and hypercapnia was observed over all subjects. (Different color scales have been used to visually normalize the color maps across the different experimental conditions). (D–F) BOLD signal time courses of the R-fMRI, T-fMRI, and hypercapnia respectively from a typical gray matter voxel within the motor-task ROI. (G–I) BOLD signal time courses of the R-fMRI, T-fMRI, and hypercapnia respectively from a typical white matter voxel within the motor-task ROI.

Figure 3

(A) Voxel-level relationship between motor vs BH (r = 0.91; P < 10−7) and (B) motor vs RSFA (r = 0.84; P < 10−7). Voxels were derived after cross correlating the BOLD response time courses with the gamma-variate convolved motor-task reference vector. Voxels with cross-correlation coefficient (cc) ≥0.7 corresponding to a Bonferroni corrected P < 10−5 was used to generate the correlations. Plots are representative of a typical subject.

Figure 4

Voxel-wise relationship between R-fMRI, T-fMRI, and hypercapnic (BH) responses in a typical subject. (A) Low frequency RSFA vs motor task (r = 0.82; P < 10−8), (B) high frequency RSFA vs motor task (r = 0.67; P < 10−8), (C) low frequency RSFA vs BH (r = 0.88; P < 10−8), and (D) high frequency RSFA vs BH (r = 0.66; P < 10−8).

Figure 5

Subject-wise relationship between R-fMRI, T-fMRI and hypercapnic (BH) responses. (A) T-fMRI vs BH (r = 0.57; P < 0.04), (B) T-fMRI vs low frequency RSFA (r = 0.50; P < 0.05), and (C) T-fMRI vs high frequency RSFA (r = 0.58; P < 0.04). A strong linear relationship was observed from the voxel-averaged BOLD signals at the subject level. The relationships shown in (B) and (C) were subsequently used to predict T-fMRI responses from a second group of subjects scanned at a different site.

Figure 6

Subject-wise relationship of the low and high frequency components of RSFA between themselves and BH. (A) High frequency RSFA vs low frequency RSFA (r = 0.67; P < 0.02), (B) BH vs high frequency RSFA (r = 0.51; P < 0.05), and (C) BH vs low frequency RSFA (r = 0.70; P < 0.01). A strong linear relationship was observed from the voxel-averaged BOLD signals at the subject level. The relationship indicates a significantly higher vascular component in the low frequency RSFA compared to high frequency RSFA.

Figure 7

Predicted volume of the motor-task activated ROI with an accuracy of (A) 90% and above, (B) 75% and above. T-fMRI responses across voxels were predicted using the subject-wise R–T relationship in the low and high frequency RSFA derived from the first group of subjects. *Significantly different with respect to low frequency RSFA; paired t-test, P < 0.02. **Significantly different with respect to high frequency RSFA; paired t-test, P < 0.001.