The predictive value of changes in effective connectivity for human learning

Science

Volumn 283, Issue 5407, 1999, Pages 1538-1541

  Buchel C ^a   Coull, J T ^a   Friston, K J ^a  

^a UNIVERSITY COLLEGE LONDON   (United Kingdom)

Author keywords

[No Author keywords available]

Indexed keywords

ARTICLE; FUNCTIONAL ANATOMY; HUMAN; NERVE POTENTIAL; NERVE STIMULATION; NUCLEAR MAGNETIC RESONANCE IMAGING; PARIETAL LOBE; PRIORITY JOURNAL; STATE DEPENDENT LEARNING; TASK PERFORMANCE; VISUAL CORTEX;

ADULT; ASSOCIATION LEARNING; BRAIN MAPPING; CEREBRAL CORTEX; ECHO-PLANAR IMAGING; FEMALE; HIPPOCAMPUS; HUMANS; MALE; MEMORY; PARIETAL LOBE; PHOTIC STIMULATION; TEMPORAL LOBE; VISUAL CORTEX; VISUAL PATHWAYS;

EID: 17544370819     PISSN: 00368075     EISSN: None     Source Type: Journal    
DOI: 10.1126/science.283.5407.1538     Document Type: Article

References (44)

1. B. Milner, I. Johnsrude, J. Crane, Philos. Trans. R. Soc. London Ser. B 352, 1469 (1997).
2. C. Moscovitch, S. Kapur, S. Kohler, S. Houle, Proc. Natl. Acad. Sci. U.S.A. 92, 3721 (1995).
3. S. Kohler et al., Neuropsychologia 36, 129 (1998); A. M. Owen, B. Milner, M. Petrides, A. C. Evans, J. Cognit. Neurosci. 8, 588 (1996).
4. S. Kohler et al., Neuropsychologia 36, 129 (1998); A. M. Owen, B. Milner, M. Petrides, A. C. Evans, J. Cognit. Neurosci. 8, 588 (1996).
5. L. G. Ungerleider and J. V. Haxby, Curr. Opin. Neurobiol. 4, 157 (1994); M. Mishkin, L. G. Ungerleider, K. A. Macko, Trends Neurosci. 6, 414 (1983).
6. L. G. Ungerleider and J. V. Haxby, Curr. Opin. Neurobiol. 4, 157 (1994); M. Mishkin, L. G. Ungerleider, K. A. Macko, Trends Neurosci. 6, 414 (1983).
7. R. Desimone, Proc. Natl. Acad. Sci. U.S.A. 93, 13494 (1996); C. L. Wiggs and A. Martin, Curr. Opin. Neurobiol. 8, 227 (1998). Repetition suppression is mainly observed in inferior temporal cortex [G. C. Baylis and E. T. Rolls, Exp. Brain Res. 65, 614 (1987)], but it has also been observed in other regions, including parietal cortex [L. R. Squire et al., Proc. Natl. Acad. Sci. U.S.A. 89, 1837 (1992); R. L. Buckner et al., J. Neurosci. 15, 12 (1995); D. L. Schacter et al., Proc. Natl. Acad. Sci. U.S.A. 93, 321 (1996)].
8. R. Desimone, Proc. Natl. Acad. Sci. U.S.A. 93, 13494 (1996); C. L. Wiggs and A. Martin, Curr. Opin. Neurobiol. 8, 227 (1998). Repetition suppression is mainly observed in inferior temporal cortex [G. C. Baylis and E. T. Rolls, Exp. Brain Res. 65, 614 (1987)], but it has also been observed in other regions, including parietal cortex [L. R. Squire et al., Proc. Natl. Acad. Sci. U.S.A. 89, 1837 (1992); R. L. Buckner et al., J. Neurosci. 15, 12 (1995); D. L. Schacter et al., Proc. Natl. Acad. Sci. U.S.A. 93, 321 (1996)].
9. R. Desimone, Proc. Natl. Acad. Sci. U.S.A. 93, 13494 (1996); C. L. Wiggs and A. Martin, Curr. Opin. Neurobiol. 8, 227 (1998). Repetition suppression is mainly observed in inferior temporal cortex [G. C. Baylis and E. T. Rolls, Exp. Brain Res. 65, 614 (1987)], but it has also been observed in other regions, including parietal cortex [L. R. Squire et al., Proc. Natl. Acad. Sci. U.S.A. 89, 1837 (1992); R. L. Buckner et al., J. Neurosci. 15, 12 (1995); D. L. Schacter et al., Proc. Natl. Acad. Sci. U.S.A. 93, 321 (1996)].
10. R. Desimone, Proc. Natl. Acad. Sci. U.S.A. 93, 13494 (1996); C. L. Wiggs and A. Martin, Curr. Opin. Neurobiol. 8, 227 (1998). Repetition suppression is mainly observed in inferior temporal cortex [G. C. Baylis and E. T. Rolls, Exp. Brain Res. 65, 614 (1987)], but it has also been observed in other regions, including parietal cortex [L. R. Squire et al., Proc. Natl. Acad. Sci. U.S.A. 89, 1837 (1992); R. L. Buckner et al., J. Neurosci. 15, 12 (1995); D. L. Schacter et al., Proc. Natl. Acad. Sci. U.S.A. 93, 321 (1996)].
11. R. Desimone, Proc. Natl. Acad. Sci. U.S.A. 93, 13494 (1996); C. L. Wiggs and A. Martin, Curr. Opin. Neurobiol. 8, 227 (1998). Repetition suppression is mainly observed in inferior temporal cortex [G. C. Baylis and E. T. Rolls, Exp. Brain Res. 65, 614 (1987)], but it has also been observed in other regions, including parietal cortex [L. R. Squire et al., Proc. Natl. Acad. Sci. U.S.A. 89, 1837 (1992); R. L. Buckner et al., J. Neurosci. 15, 12 (1995); D. L. Schacter et al., Proc. Natl. Acad. Sci. U.S.A. 93, 321 (1996)].
12. R. Desimone, Proc. Natl. Acad. Sci. U.S.A. 93, 13494 (1996); C. L. Wiggs and A. Martin, Curr. Opin. Neurobiol. 8, 227 (1998). Repetition suppression is mainly observed in inferior temporal cortex [G. C. Baylis and E. T. Rolls, Exp. Brain Res. 65, 614 (1987)], but it has also been observed in other regions, including parietal cortex [L. R. Squire et al., Proc. Natl. Acad. Sci. U.S.A. 89, 1837 (1992); R. L. Buckner et al., J. Neurosci. 15, 12 (1995); D. L. Schacter et al., Proc. Natl. Acad. Sci. U.S.A. 93, 321 (1996)].
13. M. M. Mesulam, Ann. Neurol. 28, 597 (1990); A. R. Damasio and H. Damasio, in Large Scale Neuronal Theories of the Brain, C. Koch and J. L. Davis, Eds. (MIT Press, Cambridge, MA, 1994).
14. M. M. Mesulam, Ann. Neurol. 28, 597 (1990); A. R. Damasio and H. Damasio, in Large Scale Neuronal Theories of the Brain, C. Koch and J. L. Davis, Eds. (MIT Press, Cambridge, MA, 1994).
15. For the origins of the concept of effective connectivity, see G. L. Gerstein and D. H. Perkel, Science 164, 828 (1969); A. Aertsen and H. Preissl, Dynamics of Activity and Connectivity in Physiological Neuronal Networks (VCH, New York, 1991). For examples of the adaptation to functional neuroimaging, see K. J. Friston, L. G. Ungerleider, P. Jezzard, R. Turner, Hum. Brain Mapp. 2, 211 (1995); B. Horwitz et al., J. Cognit. Neurosci. 4, 311 (1992); A. R. McIntosh and F. Gonzalez-Lima, Hum. Brain Mapp. 2, 2 (1994). F or a recent example of the application of the concept in the context of learning, see A. R. McIntosh, R. E. Cabeza, N. J. Lobaugh, J. Neurophysiol. 80, 2790 (1998).
16. For the origins of the concept of effective connectivity, see G. L. Gerstein and D. H. Perkel, Science 164, 828 (1969); A. Aertsen and H. Preissl, Dynamics of Activity and Connectivity in Physiological Neuronal Networks (VCH, New York, 1991). For examples of the adaptation to functional neuroimaging, see K. J. Friston, L. G. Ungerleider, P. Jezzard, R. Turner, Hum. Brain Mapp. 2, 211 (1995); B. Horwitz et al., J. Cognit. Neurosci. 4, 311 (1992); A. R. McIntosh and F. Gonzalez-Lima, Hum. Brain Mapp. 2, 2 (1994). F or a recent example of the application of the concept in the context of learning, see A. R. McIntosh, R. E. Cabeza, N. J. Lobaugh, J. Neurophysiol. 80, 2790 (1998).
17. For the origins of the concept of effective connectivity, see G. L. Gerstein and D. H. Perkel, Science 164, 828 (1969); A. Aertsen and H. Preissl, Dynamics of Activity and Connectivity in Physiological Neuronal Networks (VCH, New York, 1991). For examples of the adaptation to functional neuroimaging, see K. J. Friston, L. G. Ungerleider, P. Jezzard, R. Turner, Hum. Brain Mapp. 2, 211 (1995); B. Horwitz et al., J. Cognit. Neurosci. 4, 311 (1992); A. R. McIntosh and F. Gonzalez-Lima, Hum. Brain Mapp. 2, 2 (1994). F or a recent example of the application of the concept in the context of learning, see A. R. McIntosh, R. E. Cabeza, N. J. Lobaugh, J. Neurophysiol. 80, 2790 (1998).
18. For the origins of the concept of effective connectivity, see G. L. Gerstein and D. H. Perkel, Science 164, 828 (1969); A. Aertsen and H. Preissl, Dynamics of Activity and Connectivity in Physiological Neuronal Networks (VCH, New York, 1991). For examples of the adaptation to functional neuroimaging, see K. J. Friston, L. G. Ungerleider, P. Jezzard, R. Turner, Hum. Brain Mapp. 2, 211 (1995); B. Horwitz et al., J. Cognit. Neurosci. 4, 311 (1992); A. R. McIntosh and F. Gonzalez-Lima, Hum. Brain Mapp. 2, 2 (1994). F or a recent example of the application of the concept in the context of learning, see A. R. McIntosh, R. E. Cabeza, N. J. Lobaugh, J. Neurophysiol. 80, 2790 (1998).
19. For the origins of the concept of effective connectivity, see G. L. Gerstein and D. H. Perkel, Science 164, 828 (1969); A. Aertsen and H. Preissl, Dynamics of Activity and Connectivity in Physiological Neuronal Networks (VCH, New York, 1991). For examples of the adaptation to functional neuroimaging, see K. J. Friston, L. G. Ungerleider, P. Jezzard, R. Turner, Hum. Brain Mapp. 2, 211 (1995); B. Horwitz et al., J. Cognit. Neurosci. 4, 311 (1992); A. R. McIntosh and F. Gonzalez-Lima, Hum. Brain Mapp. 2, 2 (1994). F or a recent example of the application of the concept in the context of learning, see A. R. McIntosh, R. E. Cabeza, N. J. Lobaugh, J. Neurophysiol. 80, 2790 (1998).
20. For the origins of the concept of effective connectivity, see G. L. Gerstein and D. H. Perkel, Science 164, 828 (1969); A. Aertsen and H. Preissl, Dynamics of Activity and Connectivity in Physiological Neuronal Networks (VCH, New York, 1991). For examples of the adaptation to functional neuroimaging, see K. J. Friston, L. G. Ungerleider, P. Jezzard, R. Turner, Hum. Brain Mapp. 2, 211 (1995); B. Horwitz et al., J. Cognit. Neurosci. 4, 311 (1992); A. R. McIntosh and F. Gonzalez-Lima, Hum. Brain Mapp. 2, 2 (1994). F or a recent example of the application of the concept in the context of learning, see A. R. McIntosh, R. E. Cabeza, N. J. Lobaugh, J. Neurophysiol. 80, 2790 (1998).
21. Informed consent was obtained before MRI scanning. Data were acquired with a 2 Tesla Magnetom VISION (Siemens, Erlangen, Germany) whole-body MRI system equipped with a head volume coil. We obtained contiguous multislice T2*-weighted fMRI images [echo time (TE) = 40 ms; 80.7 ms per image; 64 by 64 pixels (19.2 cm by 19.2 cm)] with echo-planar imaging using an axial slice orientation. A T2*-weighted sequence was chosen to enhance blood oxygenation level-dependent contrast. The volume acquired covered the whole brain (48 slices, each 3 mm thick, 4.1 s per volume).
22. Each object was taken from a standardized series [J. G. Snodgrass and M. Vanderwart, J. Exp. Psychol. Hum. Learn. 6, 174 (1980)] and subtended 4° by 4° visual angle (visible screen 22° by 17°). Participants were familiarized with the objects before each learning session. During ENC each of 10 line drawings was presented sequentially (the order was randomized over all eight repetitions) for 2.5 s in its location. Locations were indicated by boxes visible throughout the condition. A new object was presented every 3.2 s. The task was to name the object. All objects had monosyllabic names, allowing participants to utter the object name without jaw or head movements (estimated head motion < 1.5 mm). Participants' vocal responses were recorded by a differential microphone set-up. During RET participants were spatially cued with a nonsense shape and had to respond with the name of the object previously associated with that location. Stimulus timing was identical in all conditions.
23. -kx where 0 < k < 1 is an index of learning speed. Small values of k indicate a flatter curve and therefore slower learning.
24. The four original box-car functions were multiplied by a monotonically increasing linear function, leading to four additional regressors, which model time-by-condition interactions. [C. Büchel, J. Morris, R. J. Dolan, K. J. Friston, Neuron 20, 947 (1998)].
25. W. Singer, Science 270, 758 (1995).
26. Regions of interest (ROIs) were based on the SPM{F}. The time series, representative of a region, was defined by the first eigenvector of all voxel time series in an ROI, centered (8-mm radius) around the local F maximum (22). Time series were adjusted for confounds (for example, global mean, low-frequency components and head motion).
27. Activation at the calcarine fissure was assigned to V1 (mean ± SEM in millimeters for all participants: x = 8 ± 3.4, y = -88 ± 2, z = 0.5 ± 2). DE activation was found close to area V3a (x = 30 ± 2.1, y = -86.5 ± 1.8, z = 20.7 ± 1.5) [R. B. H. Tootell et al., J. Neurosci. 17, 7060 (1997)]. The locations of PP (x = 25 ± 3.1. y = -63.5 ± 2.8, z = 57.5 ± 2.4) and LP (x = 38 ± 2.5, y = -42 ± 3.3, z = 54 ± 2.3) were similar to previous neuroimaging studies [(2); B. Luna et al., Cereb. Cortex 8, 40 (1998)]. The location of the posterior ventral extrastriate region was found in the fusiform gyrus (ITp) (x = 34.5 ± 4, y = -68 ± 3.4, z = -17 ± 1.3); the more anterior ventral activation was in the parahippocampal gyrus (ITa) (x = 33.5 ± 0.9, y = -34.5 ± 1.3, z = -24 ± 2.3). Coordinates for these regions were comparable to previous studies (1).
28. Activation at the calcarine fissure was assigned to V1 (mean ± SEM in millimeters for all participants: x = 8 ± 3.4, y = -88 ± 2, z = 0.5 ± 2). DE activation was found close to area V3a (x = 30 ± 2.1, y = -86.5 ± 1.8, z = 20.7 ± 1.5) [R. B. H. Tootell et al., J. Neurosci. 17, 7060 (1997)]. The locations of PP (x = 25 ± 3.1. y = -63.5 ± 2.8, z = 57.5 ± 2.4) and LP (x = 38 ± 2.5, y = -42 ± 3.3, z = 54 ± 2.3) were similar to previous neuroimaging studies [(2); B. Luna et al., Cereb. Cortex 8, 40 (1998)]. The location of the posterior ventral extrastriate region was found in the fusiform gyrus (ITp) (x = 34.5 ± 4, y = -68 ± 3.4, z = -17 ± 1.3); the more anterior ventral activation was in the parahippocampal gyrus (ITa) (x = 33.5 ± 0.9, y = -34.5 ± 1.3, z = -24 ± 2.3). Coordinates for these regions were comparable to previous studies (1).
29. 2 statistic has n degrees of freedom, where n is the difference in the degrees of freedom between the null model and the one in question.
30. D. J. Felleman and D. C. Van Essen, Cereb. Cortex 1, 1 (1991).
31. M. L. Smith and B. Milner, Neuropsychologia 19, 781 (1981); S. Kohler, A. R. McIntosh, M. Moscovitch, G. Winocur, Cereb. Cortex 8, 451 (1998).
32. M. L. Smith and B. Milner, Neuropsychologia 19, 781 (1981); S. Kohler, A. R. McIntosh, M. Moscovitch, G. Winocur, Cereb. Cortex 8, 451 (1998).
33. C. Distler, D. Boussaoud, R. Desimone, L. G. Ungerleider, J. Comp. Neurol. 334, 125 (1993).
34. diff(1) = 2.3, P = 0.13]. To account for autocorrelation in the observations, we used the effective degrees of freedom in the calculation of all fit indices [K. J. Worsley and K. J. Friston, Neuroimage 2, 173 (1995)].
35. B. Jouve, P. Rosenstiehl, M. Imbert, Cereb. Cortex 8, 28 (1998).
36. -/∝ to the behavioral data (the percent of correct responses) was regressed on peak cut-off. Between-participant and between-session effects were modeled separately and both reached significance [t (15) = 3, t (15) = 2.1, P < 0.05]. Intuitively this result indicates that the temporal pattern of changes in effective connectivity not only predicted a given participant's performance but also differences in learning between sessions for an individual participant.
37. C. Büchel and K. J. Friston, Cereb. Cortex 7, 768 (1997).
38. The interpretation of changes over time in learning experiments can be difficult due to nonspecific time effects (that is, habituation, motivation, and arousal). We dissociated learning-related effects from nonspecific time effects by using three sequential learning sessions. Nonspecific time effects unrelated to learning are expressed over the duration of the whole "experimental time" (that is, over all three sessions). Conversely, learning-related effects should follow a similar pattern, but within each learning session (Figs. 1A and 2C).
39. K. J. Friston, A. P. Holmes, J. Ashburner, J.-B. Poline, "SPM Central," available at http://www.fil.ion. ucl.ac.uk/spm; K. J. Friston et al., Hum. Brain Mapp. 2, 189 (1995). All volumes were realigned to the first volume. A structural MRI, acquired with a standard three-dimensional T1-weighted sequence (1 mm by 1 mm by 1.5 mm voxel size), was coregistered to the T2* images. Finally, all the images were spatially normalized [K. J. Friston et al., Hum. Brain Mapp. 2, 1 (1995)] to a standard template [A. C. Evans et al., in proceedings of the Nuclear Science Symposium and Medical Imaging Conference, L. A. Klaisner, Ed., San Francisco, CA, 31 October to 6 November, 1993 (IEEE Service Center, Piscatawa, NJ, 1993), vols. 1-3, pp. 1813-1817]. The data were spatially smoothed with a 6-mm full width at half maximum (FWHM) Gaussian kernel. Data analysis was performed by modeling the different conditions (ENC, CO1, RET, and CO2) as reference waveforms (that is, box-cars convolved with a hemodynamic response function) in the context of multiple linear regression. The resulting F statistics for every voxel constitute a statistical parametric map SPM(F). Data were analyzed for each participant individually with a threshold of P < 0.05 (corrected for multiple comparisons). An adaptive highpass filter was added to the confound partition of the design matrix to account for low-frequency drifts [A. P. Holmes, O. Josephs, C. Büchel, K. J. Friston, Neuroimage 5, 5480 (1997)]. Voxel time series were temporally smoothed with a Gaussian filter (FWHM of 4 s).
40. K. J. Friston, A. P. Holmes, J. Ashburner, J.-B. Poline, "SPM Central," available at http://www.fil.ion. ucl.ac.uk/spm; K. J. Friston et al., Hum. Brain Mapp. 2, 189 (1995). All volumes were realigned to the first volume. A structural MRI, acquired with a standard three-dimensional T1-weighted sequence (1 mm by 1 mm by 1.5 mm voxel size), was coregistered to the T2* images. Finally, all the images were spatially normalized [K. J. Friston et al., Hum. Brain Mapp. 2, 1 (1995)] to a standard template [A. C. Evans et al., in proceedings of the Nuclear Science Symposium and Medical Imaging Conference, L. A. Klaisner, Ed., San Francisco, CA, 31 October to 6 November, 1993 (IEEE Service Center, Piscatawa, NJ, 1993), vols. 1-3, pp. 1813-1817]. The data were spatially smoothed with a 6-mm full width at half maximum (FWHM) Gaussian kernel. Data analysis was performed by modeling the different conditions (ENC, CO1, RET, and CO2) as reference waveforms (that is, box-cars convolved with a hemodynamic response function) in the context of multiple linear regression. The resulting F statistics for every voxel constitute a statistical parametric map SPM(F). Data were analyzed for each participant individually with a threshold of P < 0.05 (corrected for multiple comparisons). An adaptive highpass filter was added to the confound partition of the design matrix to account for low-frequency drifts [A. P. Holmes, O. Josephs, C. Büchel, K. J. Friston, Neuroimage 5, 5480 (1997)]. Voxel time series were temporally smoothed with a Gaussian filter (FWHM of 4 s).
41. K. J. Friston, A. P. Holmes, J. Ashburner, J.-B. Poline, "SPM Central," available at http://www.fil.ion. ucl.ac.uk/spm; K. J. Friston et al., Hum. Brain Mapp. 2, 189 (1995). All volumes were realigned to the first volume. A structural MRI, acquired with a standard three-dimensional T1-weighted sequence (1 mm by 1 mm by 1.5 mm voxel size), was coregistered to the T2* images. Finally, all the images were spatially normalized [K. J. Friston et al., Hum. Brain Mapp. 2, 1 (1995)] to a standard template [A. C. Evans et al., in proceedings of the Nuclear Science Symposium and Medical Imaging Conference, L. A. Klaisner, Ed., San Francisco, CA, 31 October to 6 November, 1993 (IEEE Service Center, Piscatawa, NJ, 1993), vols. 1-3, pp. 1813-1817]. The data were spatially smoothed with a 6-mm full width at half maximum (FWHM) Gaussian kernel. Data analysis was performed by modeling the different conditions (ENC, CO1, RET, and CO2) as reference waveforms (that is, box-cars convolved with a hemodynamic response function) in the context of multiple linear regression. The resulting F statistics for every voxel constitute a statistical parametric map SPM(F). Data were analyzed for each participant individually with a threshold of P < 0.05 (corrected for multiple comparisons). An adaptive highpass filter was added to the confound partition of the design matrix to account for low-frequency drifts [A. P. Holmes, O. Josephs, C. Büchel, K. J. Friston, Neuroimage 5, 5480 (1997)]. Voxel time series were temporally smoothed with a Gaussian filter (FWHM of 4 s).
42. K. J. Friston, A. P. Holmes, J. Ashburner, J.-B. Poline, "SPM Central," available at http://www.fil.ion. ucl.ac.uk/spm; K. J. Friston et al., Hum. Brain Mapp. 2, 189 (1995). All volumes were realigned to the first volume. A structural MRI, acquired with a standard three-dimensional T1-weighted sequence (1 mm by 1 mm by 1.5 mm voxel size), was coregistered to the T2* images. Finally, all the images were spatially normalized [K. J. Friston et al., Hum. Brain Mapp. 2, 1 (1995)] to a standard template [A. C. Evans et al., in proceedings of the Nuclear Science Symposium and Medical Imaging Conference, L. A. Klaisner, Ed., San Francisco, CA, 31 October to 6 November, 1993 (IEEE Service Center, Piscatawa, NJ, 1993), vols. 1-3, pp. 1813-1817]. The data were spatially smoothed with a 6-mm full width at half maximum (FWHM) Gaussian kernel. Data analysis was performed by modeling the different conditions (ENC, CO1, RET, and CO2) as reference waveforms (that is, box-cars convolved with a hemodynamic response function) in the context of multiple linear regression. The resulting F statistics for every voxel constitute a statistical parametric map SPM(F). Data were analyzed for each participant individually with a threshold of P < 0.05 (corrected for multiple comparisons). An adaptive highpass filter was added to the confound partition of the design matrix to account for low-frequency drifts [A. P. Holmes, O. Josephs, C. Büchel, K. J. Friston, Neuroimage 5, 5480 (1997)]. Voxel time series were temporally smoothed with a Gaussian filter (FWHM of 4 s).
43. K. J. Friston, A. P. Holmes, J. Ashburner, J.-B. Poline, "SPM Central," available at http://www.fil.ion. ucl.ac.uk/spm; K. J. Friston et al., Hum. Brain Mapp. 2, 189 (1995). All volumes were realigned to the first volume. A structural MRI, acquired with a standard three-dimensional T1-weighted sequence (1 mm by 1 mm by 1.5 mm voxel size), was coregistered to the T2* images. Finally, all the images were spatially normalized [K. J. Friston et al., Hum. Brain Mapp. 2, 1 (1995)] to a standard template [A. C. Evans et al., in proceedings of the Nuclear Science Symposium and Medical Imaging Conference, L. A. Klaisner, Ed., San Francisco, CA, 31 October to 6 November, 1993 (IEEE Service Center, Piscatawa, NJ, 1993), vols. 1-3, pp. 1813-1817]. The data were spatially smoothed with a 6-mm full width at half maximum (FWHM) Gaussian kernel. Data analysis was performed by modeling the different conditions (ENC, CO1, RET, and CO2) as reference waveforms (that is, box-cars convolved with a hemodynamic response function) in the context of multiple linear regression. The resulting F statistics for every voxel constitute a statistical parametric map SPM(F). Data were analyzed for each participant individually with a threshold of P < 0.05 (corrected for multiple comparisons). An adaptive highpass filter was added to the confound partition of the design matrix to account for low-frequency drifts [A. P. Holmes, O. Josephs, C. Büchel, K. J. Friston, Neuroimage 5, 5480 (1997)]. Voxel time series were temporally smoothed with a Gaussian filter (FWHM of 4 s).
44. We thank the departmental radiographers and the Functional Imaging Laboratory physics group for help with fMRI scanning, O. Josephs for the development of the sound pickup system in fMRI, and A. Kleinschmidt, I. Johnsrude, R. Frackowiak, and R. Henson for invaluable discussions. C.B., J.T.C., and K.J.F. were supported by the Wellcome Trust.