We have proposed the use of a hypercapnic challenge (BH task) as a method of measuring and correcting for relative hemodynamic latency differences across the whole brain. Such a task allows the interrogation of most gray-matter regions of the brain () and is thus not limited to specific regions, as is the more typically employed visual (or other sensory) task. Our results indicate that BH is a robust method for assessing non-neural, vasoreactivity-based latency differences across individual voxels. While spatial variability of the HRF shape beyond latency differences may still influence fMRI analyses, we propose that correcting for vasoreactivity latency confounds in an unbiased manner using a BH task will increase the validity of whole-brain functional connectivity and causality analyses.
The present study assumes that latency measures derived from a BH task are uncoupled from neural activation associated with either the performance of the BH task or the mild hypercapnia itself. Studies with concurrent BOLD and CBF measurements using the Davis model (Davis et al., 1998
) have shown that CMRO2 changes during a BH task are small (Kastrup et al., 1999a
), thus indicating that neural processes are unlikely to contribute greatly. BOLD signal magnitudes measured during BH were demonstrated to correlate well with those of hypercapnic challenges induced by CO2 inhalation, which requires no active participation (Kastrup et al., 2001
); furthermore, Thomason et al. found that different forms of subject feedback used to control depth of inspiration during a BH task did not alter the BH response (Thomason and Glover, 2008
), again suggesting that intentional BH control mechanisms did not significantly contribute to the overall BOLD signals. Finally, cognitive processes generally induce much smaller BOLD signals than are observed in BH (which are on the order of 2–4% or larger, even in motor and thalamic regions (Kastrup et al., 1999b
; Thomason et al., 2005
)) and thus would not add significantly to the vasoreactive response.
Previous studies have addressed latency confounds in whole-brain analysis by using cross-correlation rather than zero-lag correlation to build connectivity and activation maps (Calhoun et al., 2003
). Cross-correlation effectively reduces the sensitivity of detection power to minor latency shifts, for voxels with adequate CNR. However, cross-correlation fails to distinguish between neural and hemodynamic delays, and furthermore uses the same dataset to both estimate and correct for latency; hence, the resulting maps are biased in an unknown way. On the other hand, BH elicits a strong BOLD signal change in all vascularized regions (e.g. ), and is furthermore unbiased since the correction is applied prior to functional analysis and the correction parameters are acquired from a separate scan. BH is also simple and practical to implement; it is non-invasive, comfortable for subjects to perform, and involves a relatively short additional scan (<8 min).
Performing a latency correction using BH relies on the assumption that the timing of local vasomotor responses to a BH task is proportional to those of neural activation-induced responses. We examined the validity of this assumption by correlating voxel-wise BH latency measurements with those of a task designed to elicit simultaneous activation of sensory regions. Our results showed strong positive correlations, with best-fit slopes approaching unity in the limit of more precise BH measurements, thus supporting the hypothesis that BH latency is reflective of local vasoreactivity delays and that the transit time of blood in the major arteries may be ignored.
Because regions of the cerebellum tended to exhibit high latency, we considered the possibility that the cerebellum may be more susceptible to transit time effects than other regions. However, in a subject with high cerebellar latency in the BH task, we also observed large time lags in a cross-correlation between the PCC and the cerebellum in the (uncorrected) resting state data (). This finding helps suggest that the origin of the cerebellar delay is either vasoreactivity or another factor, such as respiratory effects.
Figure 11 Resting state cross-correlation (left) and corresponding lags (center) with a seed ROI in the PCC for one subject, thresholded at r>0.4. Right: BH latency map from the same subject. Both the BH latency map and the resting state cross-correlations (more ...)
Latency correction yielded differences in subject-specific DMN maps. Most differences were minor, perhaps owing to the fact that connectivity of the default-mode (and other resting state networks) occurs primarily in the low frequencies (<0.1Hz; (Cordes et al., 2001
)), so the spatial extent of the network may be insensitive to small latency differences. Indeed, the autocorrelation of the PCC ROI timeseries has a width of several seconds (). We expect that when examining correlations with a reference timeseries having substantial high-frequency components, latency correction may have a higher impact on the resulting networks. Indeed, the impact of latency correction on activation maps in our event-related WM task was greater. Post-correction Granger causality analysis in one subject also revealed significant causal connections across the working memory network that were absent prior to the correction. While anecdotal, this result illustrates the potential impact of vascular latency differences on quantifying causal relationships between regions.
Figure 12 Resting state time series from the PCC ROI of a representative subject, along with its frequency spectrum and autocorrelation. The time series contains mostly low frequencies (< 0.1 Hz), and the autocorrelation function has a width extending over (more ...)
Quantifying relative hemodynamic latency using BH task may contribute to a wide range of fMRI modeling studies. For example, in analyses of causality where it may be difficult or impossible to use load-dependent modulation (e.g. resting state), a causality mapping applied to time series in the BH task may serve as a “control” analysis, indicating whether an observed causal relationship in the task of interest could in fact be due to vascular delay (e.g. ). This is particularly critical for connectivity methods that do not include an explicit forward model, such as Granger causality and structural equation modeling. Latencies measured with a BH task might also be entered as parameters in deconvolution methods and biophysical models (e.g. dynamic causal modeling (Friston et al., 2003
)). DCM addresses the hemodynamic confound in neural connectivity modeling by allowing specification of a hemodynamic forward model, and simultaneously estimating both connectivity and region-specific HRF paramters (including latency) (Friston et al., 2003
; Stephan et al., 2007
). Our approach may thus be viewed as complementary, or confirmatory, to such methods. Importantly, our approach provides a means of gauging relative hemodynamic delays across regions without relying on a detailed (and suitable) specification of model parameters; indeed, unless one is certain of the model, data-driven approaches can yield misleading results, and independent measurements of latency may prove useful before DCM, ICA, and other non-hypothesis driven methods are employed.
Finally, we note that with the exception of cerebellar areas and major vessels, relative latency differences across the brain tended to be small, and the correction had only a minor impact on the resting state networks and event-related tasks examined in the current study. Therefore, although vascular latency correction is critical for studies in which the precise timing of regional responses is of interest (e.g. causality analysis), for many applications it may suffice to employ a sensory task for inferring hemodynamic latency in most other non-sensory regions.