We begin by summarizing the two principal findings from this study. First, adaptive filtering improves the CNR for O2
Hb when there is substantial global interference. This should remain true, at least quantitatively, even in comparison to simple subtraction [as discussed in (Zhang et al., 2007a
)] as well as low-pass filtering methods. In contrast, adaptive filtering did not help, and actually degraded, the CNR for HHb. One key factor to make the adaptive filtering approach practical is to find a way to pre-identify those measurements that will benefit most from adaptive filtering. Our study suggests that a rest period correlation coefficient can help predict the magnitude of O2
Hb CNR improvement. Second, we also demonstrated that HHb and O2
Hb differ significantly in the rest period correlation coefficient and in their response to adaptive filtering when there is visual stimulation using multi separation data collection.
Before further discussing these findings, it is important to point out that our results represent a worst case scenario. Indeed, despite the full-field, high contrast visual stimuli used in the study, some measurements may have contained no visually evoked activation, and hence could not generate a CNR improvement. While these conclusions are relatively straightforward, the reason for the qualitative difference between O2
Hb and HHb merits discussion. In principle, adaptive filtering works well when the multi separation measurements share a significant amount of common mode signal, and the statistical properties of this common mode signal change (i.e., are non stationary). The fact that adaptive filtering performs poorly for HHb means that HHb is more strongly affected by local changes than global ones, e.g. related to the heterogeneous distribution of blood vessels or blood supply. While HHb and O2
Hb are computed from the same raw NIRS data, under normal circumstances HHb is generated via oxygen utilization in tissue, and oxygen utilization by brain and scalp differ, both due to different regional demands for oxygen, and differences in cerebral blood flow regulation which are not paralleled in the scalp. These differences would lead to the prediction of poorer correspondences between scalp and brain HHb measures, as observed, and a correspondingly reduced ability for adaptive filtering to improve CNR, as also observed. In contrast, the dominant sources of physiological “noise” when trying to detect brain function are often arterial, respiratory and lower frequency vasomotor oscillations (Obrig et al., 2000
). These pulsatile components are present in the arterial compartment, which typically is greater than 95% oxygenated, and such oscillations would be expected in both cerebral and extra-cerebral tissue. Hence, one might expect better performance of our adaptive filtering method on O2
Hb, given the closer expected correspondence between reference and target measurements. Indeed, the measurement similarities were significantly closer for O2
Hb than for HHb.
In both O2
Hb and HHb, numerous data points exhibited negative changes in CNR with adaptive filtering. Decreases in CNR could be due to changes localized near the target or reference detector contact point. Any independent changes in the reference measurement, when subtracted from the target during adaptive filtering, will effectively be injected into the target measurement, thereby decreasing the target measurement CNR. In addition, if there are any systemic changes correlated with the stimulus delivery (e.g., increase in heart rate or blood pressure, more typical during motor stimulation than during visual stimulation), it is possible that this stimulus-related systemic change in the reference measurement could negatively affect the observed CNR, as found in other neuroimaging contexts (e.g., Bullmore et al., 1999
). The various potential reasons behind the CNR decreases remain to be investigated in detail.
Our reported results depend on the selection of “similarity thresholds” between reference and target measurements. In theory, an optimized threshold should be chosen, but such a threshold will change with various experimental parameters including probe geometry. Since the major goal of this study was to demonstrate how well and when our adaptive filtering approach helps in general, we did not investigate methods for optimal threshold selection and pooled the measurements from 2.6 and 4.0cm source-detector separations for analysis.
Our findings also suggest that our adaptive filtering technique is best applied when there is substantial global interference (e.g. judged by rest period correlation coefficient >0.6), and that the filtering should be exclusively applied to O2Hb signals, not HHb. We have also shown that, as anticipated, SNR of both the target and reference NIRS signals affects the adaptive filtering result: it is better if the SNR is at least 10 to 1. More work is needed to establish a complete and standardized way of predicting CNR improvement using signal processing techniques, including the reason for some CNR decreases following adaptive filtering even when interference dominates the signal (e.g., r>0.6). Further exploration could also help better understand the physiology behind the difference between HHb and O2Hb during the rest period, their differential response to adaptive filtering, and to use the resulting information to improve noninvasive, in vivo cerebral measurements.
In this study, a single short-distance detector per hemisphere was used. An important question is whether a separate short-distance detector would be needed for every optode, or whether a single such detector may be sufficient to measure global interference. For example, although the Mayer's wave is a global factor, its contribution may be different at measurements collected at different sites because of heterogeneity in blood vessel sizes, location, or geometry. In addition, since different optodes have different contact and skin conditions (e.g., due to hair, focal pigmentation, sweating), the corresponding interference may be different at different sites. Ideally, one would have one reference measurement for each target source-detector pair. However, many such reference measurements would likely be redundant. The optimal (or minimum) number of required reference measurements remains to be established.
In summary, adaptive filtering and related methods, are promising for cognitive neuroscience researchers using NIRS techniques to probe the human brain non-invasively. On the one hand, by reducing the effect of global interference on the O2Hb measurements, one could conduct shorter studies and still achieve sufficiently high CNR, thus minimizing the effects of fatigue and adaptation. On the other hand, adaptive filtering methods may make it possible to obtain better sensitivity or specificity for the time course of neural changes non-invasively, possibly even at the level of single trials, which are key to understand the relationship between brain and behavior. This approach is particularly well suited to event-related designs where the use of a low pass filter with a low frequency cutoff is not possible. This type of adaptive filtering techniques can also be applied in real-time, enabling better on-line recording evaluation or even higher sensitivity feedback paradigms for neuroscience research.