Beyond noise: reply to Laufs et al.
1Max Planck Institute of Psychiatry, Munich, Germany, 2Department of Psychology, Humboldt-Universität zu Berlin, Germany, 3Academic Radiology, University of Nottingham, Queen's Medical Centre, Nottingham, UK and 4Centre of Mental Health, Klinikum Ingolstadt, Ingolstadt, Germany
Correspondence to: C. Kaufmann, Humboldt-Universität zu Berlin, Rudower Chaussee 18, 12489 Berlin, Germany E-mail: christian.kaufmann{at}psychologie.hu-berlin.de
Received March 13, 2007. Accepted March 18, 2007.
Sir, We have read with interest the letter by Laufs et al. on correlations between EEG spectral frequencies and BOLD signals. The aim of the authors is to provide an alternative approach to a study published last year in Brain where we showed sleep stage specific BOLD signal variations in a group of subjects (Kaufmann et al., 2006
). In essence, the authors claim that the results of our study are incongruent to previous imaging studies on sleep, and that physiological as well as technical noise are sufficient reasons for our findings. They criticize our data analysis, but some of their assumptions are incorrect while others do not apply to our data. However, we thank the authors for their interest in our work, and reply to their comments.
Given the fact that the sleep stage scoring rules by Rechtschaffen and Kales (1968) represent the cornerstone both for basic and clinical human sleep research, and to continue the line of research of previous PET studies, we argue that a sleep stage-related approach represents the most concise analysis of our data. This is particularly true because of the sampling rate, and hence we refrained from referencing BOLD to graphoelements of short duration that indeed would benefit from shorter TR.
Laufs et al. choose to ignore the consistencies of our results with previous imaging studies on sleep, selectively overstating inconsistencies from which they infer methodological inadequacy. In fact, the majority of our data is well in line with both metabolic and blood flow PET studies. But of course, and we have discussed this in our paper, there are quite expectedly differences that can be attributed to the higher spatio-temporal resolution of FMRI compared with H215O-PET; there is an almost 10-fold increase in acquisition time per volume (10 versus 90 s) and a 2-fold larger smoothing kernel (8 versus 16 mm), and these ratios are even higher for metabolic PET data. Accordingly, differences between FMRI and PET are very likely since, for example, 90 s of sleep stage 1 supposedly contain a waxing and waning in vigilance. Nevertheless, our data show a substantial overlap to many PET results including the study of Kjaer et al. (2002) reporting on signal decreases in the thalamic area with stronger decreases in the left thalamic area, as well as BA 40 and 6. We also found an increased activation in occipital regions. To exemplify, the respective coordinates of our results at –17, –59, 31 and Kjaer's at –26, –64, –14 or –18, –88, 12 should be considered. Our data showed a successive decrease in thalamic areas from wakefulness towards deeper sleep stages, together with decreases in regions like the cingulate gyrus, precuneus and several cortical areas. This is well in agreement to the overall decreased activity in deeper NREM sleep stages in comparison to wakefulness as shown in other imaging studies (Maquet, 2000).
We thank Laufs et al. for a detailed discussion of technical MRI scanner noise that we also carefully considered including measurements of the scanner specific drift over time. However, we did not include most of these considerations in the methods section of our paper because of space restrictions, and as these considerations are well known. We agree that drifts are scanner specific, but must warn against the unjustified and unreferenced assumption that "for most devices a characteristic period of around 2 min can be assumed". This statement is certainly not true for the scanner used in our study (General Electric, 1.5 T), nor for another scanner (Siemens, 1.5 T) we tested. For these two scanners we could not observe short-time periodicities within the time scale of half an hour. Interestingly, arterial spin labelling studies, insensitive to scanner drift (Wang et al., 2003) confirm that BOLD-FMRI do not reveal false positive results in direct contradiction to Laufs et al.'s claim that our activation maps "contain extra false negatives and positives". Some of the criticism may stem from Laufs et al.'s misinterpretation of our method to remove low-frequency drifts (Macey et al., 2004). Laufs et al. erroneously state that we tried "to remove low-frequency drift by including a global mean regressor in the[ir] design matrix". Thus, the illustration in their Fig. 2 does in fact not apply to our data analysis. Figure 1 demonstrates the validity of our approach in achieving white noise, a prerequisite of statistical inference using statistical parametric mapping (SPM).
|
|
Next, Laufs et al. correctly point out that physiological noise from respiratory and cardiac oscillations must be carefully considered. We evaluated cardiac noise for the NREM sleep study (introducing 20 additional regressors representing temporal distance of each slice to the last heart beat), but then decided for the most parsimonious model without cardiac regressors as there was no significant benefit by including them. In other studies where relevant effects could be detected such as during REM sleep, we have chosen to model cardiac regressors (Wehrle et al., 2007
). The experimental setting chosen by Laufs et al. to demonstrate the invalidity of our approach is unfortunately very misleading. Using a much shorter repetition time at a higher field strength sensitizes the data to cardiac noise, and hence the need to correct for them. In contrast, in our experimental design cardiac oscillations become blurred due to the long TR. Figure 2 shows data from one subject during wakefulness with different times of repetition and the resulting connectivity maps to demonstrate the influence of repetition time.
Due to high autonomic nervous system activity, heart beat and blood pressure variability is highest during wakefulness (as well as during REM sleep) (Parmeggiani, 2005). Therefore, cardiac artefacts should be most pronounced during wakefulness due to the increased variance in the data. The approach suggested by Laufs et al. (Figs 1 and 4) introduces additional regressors in order to model cardiac pulsation artefacts. Unfortunately, detailed methods or corresponding specific results are not given, and thus adequate evaluation of the supposedly neuronal correlations to spectra or sleep graphoelements remains elusive. However, we do not suggest for subsequent studies to apply long TR to attenuate noise effects, as such technical restrictions are fortunately obsolete with modern MR equipment. Rather, we explicitly agree with the approach by Lund et al. (2006) or Deckers et al. (2006). Given our observations (Fig. 2), the results of hypothalamic connectivity patterns are qualitatively different from noise patterns, also indicated by independent remote brain areas (parietal and lateral prefrontal) correlating to hypothalamic region activity. Whether this pattern is characteristic for the process of falling asleep has to be either verified or falsified empirically on the basis of a group of subjects, not on anecdotal evidence.
Another issue relates to the frequency of events that should, from the perspective of optimal design planning, appear randomized with the same frequency of occurrence, and not coincide with frequencies of events of no interest. Here, Laufs et al. supposedly argue on the assumption of usually obtained extended duration of different sleep stages. However, falling asleep is a constant waxing and waning of vigilance, and this variability appears increased in environments like a MR scanner. This is reflected in the rather short average duration of sleep stages (but not total amount) in our data, which was 70 s for wakefulness, 38 s for sleep stage 1, 49 s for sleep stage 2 and 131 s for slow wave sleep, respectively.
Lastly, the authors present data derived from a single measurement. As Laufs et al. are very parsimonious in giving details about methodology, especially for their EEG analysis, it is difficult to comment on the results they exemplify. Generally, the differentiation of signal and noise in higher EEG frequency bands is a demanding task. It is illustrated by the high amount of alpha activity during slow wave sleep as mentioned and depicted in Fig. 3 of their measurement. This raises the question whether that alpha-delta EEG anomaly may be due to a medical CNS-condition of the subject (Moldofsky and MacFarlane, 2005), or just represents residual gradient artefacts. Similarly, the lack of correlates to alpha activity during wakefulness, or to theta activity during sleep stage 1, is very surprising. As transition between these two frequency bands are the hallmark of the transition from wakefulness to sleep stage 1, some results resembling data from Kjaer et al. (2002) or our data (Kaufmann et al., 2006) might be expected, even from a central derivation as always used in sleep analysis. Finally, most aspects presented are not discussed in the framework of existing neuroimaging sleep studies. The interpretation of a pineal gland BOLD related activation being connected to spindle modulating effects of melatonin is a bold statement and raises the question why the effect is not observable when directly correlating spindle activity. In their conclusion, the authors claim that no study has ever investigated haemodynamic brain activity in and across sleep stages. Besides our own studies a number of group studies should be mentioned (Hofle et al., 1997; Czisch et al., 2004; Dang-Vu et al., 2005; Fukunaga et al., 2006; conference abstracts: Horovitz et al., 2006; Miyauchi et al., 2006; Phillips et al., 2006).
In summary, we provide further evidence to refute that our recently published results represent noise by showing them to be qualitatively different from noise patterns, and by demonstrating that the chosen approach was suitable to remove low-frequency noise in our data. The arguments put forward by Laufs et al. are appreciated as they highlight the need for careful evaluation of technical and physiological noise, and that using different parameters and scanners may well require alternate post-processing methods.
 |
References |
|---|
 Top References |
|---|
Czisch M, Wehrle R, Kaufmann C, Wetter TC, Holsboer F, Pollmacher T, et al. Functional MRI during sleep: BOLD signal decreases and their electrophysiological correlates. Eur J Neurosci (2004) 20:566–74.[CrossRef][ISI][Medline]
Dang-Vu TT, Desseilles M, Laureys S, Degueldre C, Perrin F, Phillips C, et al. Cerebral correlates of delta waves during non-REM sleep revisited. Neuroimage (2005) 28:14–21.[CrossRef][ISI][Medline]
Deckers RH, van Gelderen P, Ries M, Barret O, Duyn JH, Ikonomidou VN, et al. An adaptive filter for suppression of cardiac and respiratory noise in MRI time series data. Neuroimage (2006) 33:1072–81.[CrossRef][ISI][Medline]
Fukunaga M, Horovitz SG, van Gelderen P, de Zwart JA, Jansma JM, Ikonomidou VN, et al. Large-amplitude, spatially correlated fluctuations in BOLD fMRI signals during extended rest and early sleep stages. Magn Reson Imaging (2006) 24:979–92.[CrossRef][ISI][Medline]
Hofle N, Paus T, Reutens D, Fiset P, Gotman J, Evans AC, et al. Regional cerebral blood flow changes as a function of delta and spindle activity during slow wave sleep in humans. J Neurosci (1997) 17:4800–8.
Horovitz SG, Fukunaga M, Zwart JA, van Gelderen P, Fulton SC, Balkin TJ, et al. The default-mode network connectivity during awake and early sleep: a simultaneous EEG-BOLD-fMRI study. Human Brain Mapping Conference (2006) Vol. 31. Florence: Neuroimage. S67.
Kaufmann C, Wehrle R, Wetter TC, Holsboer F, Auer DP, Pollmacher T, et al. Brain activation and hypothalamic functional connectivity during human non-rapid eye movement sleep: an EEG/FMRI study. Brain (2006) 129:655–67.
Kjaer TW, Law I, Wiltschiotz G, Paulson OB, Madsen PL. Regional cerebral blood flow during light sleep–a H(2)(15)O-PET study. J Sleep Res (2002) 11:201–7.[ISI][Medline]
Lund TE, Madsen KH, Sidaros K, Luo WL, Nichols TE. Non-white noise in fMRI: does modelling have an impact? Neuroimage (2006) 29:54–66.[CrossRef][ISI][Medline]
Luo WL, Nichols TE. Diagnosis and exploration of massively univariate neuroimaging models. Neuroimage (2003) 19:1014–32.[CrossRef][ISI][Medline]
Macey PM, Macey KE, Kumar R, Harper RM. A method for removal of global effects from fMRI time series. Neuroimage (2004) 22:360–366.[CrossRef][ISI][Medline]
Maquet P. Functional neuroimaging of normal human sleep by positron emission tomography. J Sleep Res (2000) 9:207–31.[CrossRef][ISI][Medline]
Miyauchi S, Misaki M, Kan S, Abe T. Brain activation and its time course accompanying rapid eye movements during REM sleep revealed by simultaneous polysomnographic recording with fMRI. Human Brain Mapping Conference (2006) Vol. 31. Florence: Neuroimage. S132.
Moldofsky H, MacFarlane JG. Fibromyalgia and chronic fatigue syndromes. In: Principles and practice of sleep medicine—Kryger MH, Roth T, Dement WC, eds. (2005) Philadelphia: Saunders. 1225–36.
Parmeggiani PL. Physiological regulation in sleep. In: Principles and practice of sleep medicine—Kryger MH, Roth T, Dement WC, eds. (2005) Philadelphia: Saunders. 185–91.
Phillips CL, Tielen G, Maquet P. Iteratively applied ICA to reject the ballisto-cardiogram artifact of EEG data acquired during sleep fMRI. Human Brain Mapping Conference (2006) Vol. 31. Florence: Neuroimage. S85.
Rechtschaffen A, Kales A, eds. A manual of standardized terminology, techniques and scoring system for sleep stages of human subjects. (1968) Washington, DC: U.S. Public Health Service, U.S. Government Printing Office.
Wang J, Aguirre GK, Kimberg DY, Roc AC, Li L, Detre JA. Arterial spin labeling perfusion fMRI with very low task frequency. Magn Reson Med (2003) 49:796–802.[CrossRef][ISI][Medline]
Wehrle R, Kaufmann C, Wetter TC, Holsboer F, Auer DP, Pollmacher T, et al. Functional microstates within human REM sleep: first evidence from fMRI of a thalamocortical network specific for phasic REM periods. Eur J Neurosci (2007) 25:863–71.[CrossRef][ISI][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||