The effects of hemodynamic lag on functional connectivity and behavior after stroke

Abstract

Stroke disrupts the brain's vascular supply, not only within but also outside areas of infarction. We investigated temporal delays (lag) in resting state functional magnetic resonance imaging signals in 130 stroke patients scanned two weeks, three months and 12 months post stroke onset. Thirty controls were scanned twice at an interval of three months. Hemodynamic lag was determined using cross-correlation with the global gray matter signal. Behavioral performance in multiple domains was assessed in all patients. Regional cerebral blood flow and carotid patency were assessed in subsets of the cohort using arterial spin labeling and carotid Doppler ultrasonography. Significant hemodynamic lag was observed in 30% of stroke patients sub-acutely. Approximately 10% of patients showed lag at one-year post-stroke. Hemodynamic lag corresponded to gross aberrancy in functional connectivity measures, performance deficits in multiple domains and local and global perfusion deficits. Correcting for lag partially normalized abnormalities in measured functional connectivity. Yet post-stroke FC-behavior relationships in the motor and attention systems persisted even after hemodynamic delays were corrected. Resting state fMRI can reliably identify areas of hemodynamic delay following stroke. Our data reveal that hemodynamic delay is common sub-acutely, alters functional connectivity, and may be of clinical importance.

Keywords: Acute stroke; brain ischemia; cerebral blood flow; cerebral hemodynamics; cognitive impairment; functional MRI.

Figure 1.

Using temporal cross-correlation to measure lag in resting state fMRI. (a) The global reference signal (black line, top panel) is measured by averaging all non-lesioned gray matter voxels. Each voxel timecourse (red line, top panel) is then compared to the reference signal. The exemplar voxel is circled in (c). (b) The voxel timecourse is shifted forward and backward 8 s (±4 TR) and shift correlation is computed (equation 1). 5.08 s is the optimal shift determined by parabolic interpolation. (c) Voxelwise hemodynamic lag image. Orange/yellow indicates a lag behind the reference signal, cyan indicates a lead. Lesioned areas are shown in black. A small caudal infarct in left posterior cerebral artery (PCA) territory shows associated lag in the entire left PCA distribution. A full lag map for this individual is shown in row 5 of Figure 2.

Figure 2.

Delay in the resting hemodynamic response (lag) is observed after stroke. (a) Lag maps for five sub-acute stroke patients overlaid on MP-RAGE. Lesion locations are shown in black. (b) Histogram of lag value frequencies measured in a set of 169 ROIs in both patients and controls. Both contain some regions that appear to lead the gray matter signal, but the patients show more regions that lag behind it than controls (shaded gray – 1.7% vs 0.2% of ROIs > +4 s). (c) To generate a group average, subjects were grouped by stroke arterial territory and the contralesional hemisphere was subtracted from the ipsilesional hemisphere. Average lag maps for two largest groups, middle cerebral artery (MCA), and posterior cerebral artery (PCA), are shown. (d) Lag maps at the one-year follow-up from the same patients shown in (a). Lag appears to largely resolve by 1year post-stroke onset in many but not all cases. (e) Lag laterality scores for all patients included in the study at all timepoints. Patients with only one timepoint are shown as a single circle. Patients are divided by lesion side to show the frequent correspondence with lag side.

Figure 3.

Areas of lag show reduced blood flow. (a) A sample patient showing left hemisphere lag and corresponding areas of hypoperfusion on ASL. (b) The average blood flow as a percent of healthy controls is measured for ROIs and plotted as a function of lag in all 107 sub-acute scans. Areas with greater than 2 s of lag show a significant decrement in rCBF (relative to zero lag) and hypoperfusion increases as lag magnitude increased. (c) Doppler imaging of the internal carotid artery (ICA) on the affected side is categorized as ≤50%, 51–79%, or ≥80% occlusion and compared lag in the affected hemisphere.

Figure 4.

BOLD power spectra. ROIs with lag show significantly decreased power relative to ‘no lag’ ROIs in the upper part of the resting state range (0.046–0.09 Hz). Power in this range is closer to that of infarcted regions. Transparent boundaries depict SEM in average power.

Figure 5.

Lag disrupts functional connectivity but can be partially corrected. (a) Seed-based correlation maps for four ROIs in different locations that showed >2 s of lag. The leftmost column show the four ROIs overlaid on the lag maps. The second column shows the correlation map for each ROI. The third column shows the correlation map for each ROI after the courses have been shifted to align optimally with the global signal. The rightmost column shows the average correlation map for the same four ROIs in the controls. Note improved similarity between patient correlation maps and controls following correction. (b) Relationship between lag and FC aberrancy (before and after correcting for lag) across all 169 ROIs and 107 subjects. The red line demonstrates the average aberrancy and SEM for ROIs at different shifts. Correcting timecourses for lag makes the functional connectivity in those regions less aberrant (black line).

Figure 6.

Stroke FC–behavior relationships persist after lag correction. (a/e) Comparison of lag laterality measured within motor ROIs versus motor deficit or within dorsal attention network ROIs versus neglect reveal a significant relationship between lag and behavior. Patients with motor lag laterality > 1 s are shaded gray. (b/f) Interhemispheric connectivity and function are highly correlated (Pearson’s correlation). (c/g) This relationship persists after lag has been corrected and (d/h) this relationship also persists when all subjects with within-network between-hemisphere lag > 1 s (21/107 for motor, 21/101 for dorsal attention) are excluded. Relationships in (a) and (e) were computed Spearman’s nonparametric rank test; all other relationships were measured with Pearson’s correlation.