Temporal sorting of neural components underlying phonological processing

Cognitive Neuroscience and Neuropsychology

NeuroReport

NeuroReport 10, 2599±2603 (1999)

EVENT-RELATED haemodynamic responses (EHRs) were recorded in subjects performing phonological tasks to test whether distinguishable temporal involvement of corresponding neural components would show through. A sequence of activation leading from primary auditory cortices to premotor regions emerged in the fast repetition and the phoneme monitoring tasks used. EHRs peaked signi®cantly earlier in Wernicke's area (phonological decoding) than in Broca's area, the left supramarginal gyrus and the precentral gyrus (phonological rehearsal). Moreover, the sensitivity of within cluster temporal gradients to the nature of the tasks indicated either sensory to association cortex synchronization for fast repetition or delayed analysis for phoneme monitoring. These results are consistent with previous ®ndings on working memory and show that fMRI permits temporal tracking of cognitive activations. NeuroReport 10:2599±2603 # 1999 Lippincott Williams & Wilkins. Key words: Event-related fMRI; fMRI temporal resolution; Haemodynamic response; Phonological processing; Sequential processing

Introduction Many lesion-based, electrophysiological and functional imaging studies have been devoted to localization and characterization of regions of the brain responsible for understanding and producing language. When lexical±semantic processes are considered, a wide network emerges [1,2], whereas tasks focusing on low-level phonological processes underlying auditory language comprehension elicit activations that seem to be more circumscribed to the immediate vicinity of the left sylvian ®ssure [3±5]. Speci®c regions involved with phonology have been identi®ed using rhyming tasks [6], repetition tasks [7], syllable counting tasks [8] or meta-linguistic judgement tasks [5] on pseudo-words (pronounceable non-words) and the temporal aspects of phonological processing have been addressed in numerous electrophysiological studies [9,10]. Overall, processes taking place in such a cognitive function have been proposed to be mostly sequential [1], in contrast to lexical semantic processes that are considered mainly parallel [12]. After they have been decoded in primary sensory areas, speech sounds are transmitted to auditory associative regions such as Wernicke's area. The latter, Brodmann's area (BA) 22/42, and more generally the superior temporal gyrus bilaterally have been associated with phonetic or low-level 0959-4965 # Lippincott Williams & Wilkins

Temporal sorting of neural components underlying phonological processing Guillaume Thierry,CA Kader Boulanouar, Ferath Kherif, Jean-Philippe Ranjeva1 and Jean-FrancËois DeÂmonet FeÂdeÂration de Neurologie, INSERM U455, and 1 Neuro-Radiology Department, HoÃpital de Purpan, 31059 Toulouse Cedex 3, France

CA

Corresponding Author

phonological processing, among other functions, as they are thought to transitorily store complex sound patterns [6]. In the framework of Baddeley's model of verbal working memory [13], phonological processing is divided into a phonological store and an articulatory rehearsal module that refreshes short-lived phonological sequences. The former would involve BA 40, the left supramarginal gyrus [4,5], whereas the latter would be distributed over BA 44/45, i.e. Broca's area BA 6 (involving the supplementary motor area) bilaterally and left and right thalami [5,14]. According to Baddeley's model, auditory phonological working memory tasks involve serially organized processes starting up with phonological decoding, followed by phonological storage and articulatory rehearsal. Here, we tested whether event-related haemodynamic responses (EHRs) would reveal a measurable sequence of activation in speci®c brain regions according to the following order: primary auditory cortices, Wernicke's area, the left supramarginal gyrus, Broca's area and premotor regions. Such an approach has already led to the characterization of temporal activation patterns in working memory tasks [15]. We used two phonological tasks that were similar enough to provide comparable activation clusters involved with pseudo-word parsing and rehearsal but specialized enough (fast repetition versus phoneme monitoring) to assess the sensitivity of the Vol 10 No 12 20 August 1999

2599

NeuroReport method relative to the impact of output processes in the tasks.

Subjects and Methods Six healthy right-handed volunteers (four males and two females) aged 24±30 years (mean 25.6  2.4) gave their informed consent to participate in the experiment, which was approved by a local ethics committee. Subjects were presented pairs of disyllabic pseudo-words in the auditory modality. The ®rst task was a fast silent repetition task (FastRep) in which subjects were asked to repeat in their mind each couple of pseudo-words one time as soon as possible after stimulus onset. The second task was a sequential detection task (SeqDet) in which they had to monitor for a /b/ in the second pseudo-word if and only if there was a /d/ in the ®rst one [5,9]. FastRep required on-line analysis of the complete phonological sequence while two phonemes only had to be identi®ed in SeqDet. On the other hand, FastRep supposed a basic repetition of linguistic material while SeqDet consisted in a metalinguistic judgement on speech sounds (i.e. relating these sounds to abstract phoneme categories). In sum, each task bore an easy and a complex feature, making them equally dif®cult. We expected both these tasks to elicit activations in every region involved with auditory decoding, phonological extraction, storage and rehearsal. While FastRep relied on the manipulation of unsegmented phonological material (syllables or larger units [11]), SeqDet implied narrow segmentation of pseudo-words strings so that target phonemes could be individualized. We hypothesized that FastRep would then favour optimized transfers from auditory sensory regions towards motor output regions and SeqDet would rather enhance post-hoc manipulation of the stimuli as suggested by previous electrophysiological results [9]. EHRs were acquired using a Magnetom Vision Siemens Imager at 1.5 T in EPI mode. Image acquisition was piloted by a PC station (Dell) running SuperLab. The latter delivered auditory stimuli and triggered the imager for acquisition via electronic pulses (gating procedure). Fourteen couples of pseudo-words (for example, [roda fobu]), lasting for 1020 ms, were delivered in each of four runs at the rate of 1 every 18 s (56 stimulations overall). Acquisitions provided six contiguous transverse slices (fov 200 mm, thickness 6 mm, TE ˆ 66 ms, matrix acquisitions 96 3 128 interpolated to 128 3 128) in ,780 ms and were repeated every 2 s nine times after each stimulation. Acquisition voxel size was 2 3 1.5 3 6 mm. The bottom of the lower slice was set across the AC±PC plane. Images were 3D-realigned 2600

Vol 10 No 12 20 August 1999

G. Thierry et al. for correction of subject's motion and smoothed (Gaussian kernel of FWMH 6 mm) to account for residual inter-subject differences using SPM96 [16]. Images were then normalized (spatial and proportional scaling normalization) into Talairach space to allow group analysis, i.e. detection of activated voxels in the group. Raw signals were temporally high-pass ®ltered at 8 3 10ÿ3 Hz and modelled using a half sinusoidal waveform best ®tting stimulus rate. Voxel-based statistical comparisons were performed in the framework of the general linear model. Activations were detected with SPM96 at a threshold of p , 0.01 (corrected for multiple comparisons) within tasks (active state versus rest and reverse contrasts) and across tasks. A two-factor ANOVA was performed on individual peak latencies (Fig. 1a) in order to characterize signi®cant temporal differences across regions (Fig. 1b). Processing of individual EHRs included adjusted signal averaging, temporal correction for across slice acquisition delay, and spline interpolation for optimizing resolution. Regions corresponded to nine anatomical sub-domains clearly segregated within ®ve non-overlapping clusters commonly activated in both tasks. A reference voxel of maximum Z-score (2.5 , Z , 6) was considered in each region and adjusted signals of neighbour voxels within a distance of 3 mm were averaged together. Reference voxels (x,y,z co-ordinates in Talairach space where ÿx refers to the right) were: 36,ÿ32,12 and ÿ36,ÿ30,12 for primary auditory cortices; 46,ÿ30,12 and ÿ52,ÿ30,12 for Wernicke's area and homologous region; 10,ÿ20,6 for the left thalamus; 50,ÿ44,18 for the supra-marginal gyrus; 54,ÿ6,24 for the pre-central gyrus; and 34,14,12 and ÿ34,14,12 for Broca's area and homologous region. Grand-average EHRs were calculated from adjusted signals, temporally corrected, and spline interpolated. Activated voxels were coded according to maxima latencies of corresponding EHRs and plotted onto high-resolution anatomical slices of one of the subjects (Fig. 2). Behavioural data were acquired off-line with SuperLab at least 4 months after fMRI sessions. In the behavioural version of the tasks, subjects had to repeat pseudo-words aloud in FastRep and pronounce [debe] aloud in SeqDet so that reaction times could be obtained from voice onset time recording.

Results Mean reaction times were 1686  69 ms after stimulus onset time (SOT) for FastRep and 1709  103 ms after SOT for SeqDet. There were , 1% errors and false alarms in both tasks. The two tasks appeared to be of similar dif®culty as reaction times did not

Temporal stages of phonological processing Right PAC Left PAC Right W Left W Right B Left B Left Th Left SM Left PC

a

0.4

Normalized EHR amplitude

0.3

0.2

0.1

0

20.1

20.2 1

9

2

3

4

5

6

7

8

9

10 11 12 Time (s)

b

Time of EHR Peak (s)

8 7

Left Hemisphere

Right Hemisphere

6 5.48 5.61

5 4.75 4 3

4.11

4.04

3.70 2.98

2.73 2.85

2 1

PAC

W

Th

SM

PC

B

PAC

W

B

FIG. 1. (a) Individual EHRs of subject 1 after spline interpolation. PAC, primary auditory cortex; W, Wernicke's area and homologous regions in the right hemisphere; Th, Thalamus; SM, supra-marginal gyrus; PC, precentral gyrus; B, Broca's area and homologous region in the right hemisphere. Note the signi®cant delay of Broca's area and the supramarginal gyrus. (b) Bar plot of mean EHR peak time for nine regions signi®cantly activated in both tasks (see Methods). Mean EHR peak time after stimulus onset time (SOT) is indicated in the upper area of each bar. Double-head arrows indicate signi®cant differences ( p , 0.05) in mean peak time across regions.

differ signi®cantly across tasks (t ˆ ÿ0.422, p ˆ 0.6905) and no signi®cant differences were observed across hit rates. Focusing on the 205 voxels commonly activated in the two tasks, a left . right asymmetry was found, with 62.9% of activated voxels located in the left hemisphere. The ANOVA on latencies of EHR maxima did not reveal a signi®cant task main effect (F(1,80) ˆ 0.108, p ˆ 0.7427) but a robust region main effect (F(7,80) ˆ 2.795, p , 0.01) and no signi®cant task 3 region interaction (F(7,80) ˆ 1.286, p ˆ

NeuroReport 0.2610). Post-hoc Fisher tests yielded signi®cant differences across regions but not across tasks. Two groups of regions were signi®cantly different among subjects: left primary auditory cortex and Wernicke's area vs left supramarginal gyrus, left precentral gyrus and Broca's area (Fig. 1b). Focusing on the differences between the two tasks, there were 617 activated pixels ( p , 0.01) in FastRep and 294 in SeqDet. No region was signi®cantly more activated in SeqDet than in FastRep. The reverse contrast yielded signi®cant activations ( p , 0.01) in the primary auditory cortices bilaterally, in the left thalamus, and in the superior temporal gyrus bilaterally. Spatial-temporal results are shown in Fig. 2. The following regions progressively reached EHR peak in FastRep: the primary auditory cortex (in the region of BA 41) bilaterally (peak time (PT) < 2.5 s), the thalami bilaterally and the superior temporal gyrus bilaterally involving Wernicke's area (approximately BA 22/42; 2.5 s < PT < 4.5 s); the posterior cingular gyrus (approximately BA 29; 3 s < PT < 4.5 s); the left supramarginal gyrus (approximately BA 40), Broca's area and homologous regions on the right (approximately BA 44/45; 3 s < PT < 6 s); and the left precentral gyrus (approximately BA 6; 4.5 s < PT < 6 s). BA labelling is tentative and refers to the Talairach stereotactic atlas (see Methods for exact Talairach co-ordinates). Although less extended, activations had similar locations in SeqDet except for the right thalamus that did not show a signi®cant activation. However, EHR peaking patterns were different. As shown by the colour coding in Fig. 2, the time range of peak latencies was narrower and delayed in Wernicke's area (3.5 s < PT < 4.5 s), Broca's area (4 s < PT < 6 s) and its homologous region in the right hemisphere (4.5 s < PT < 6 s). Two characteristics differentiated FastRep temporal gradients from those of SeqDet. Gradients in FastRep ranged from earliest to latest latencies and Wernicke's clusters looked very similar to frontal clusters. In SeqDet, gradients were more condensed and delayed and the timing of activation in Wernicke's area seemed to split from those of the frontal region.

Discussion Location of activations: Beyond well-known contributions of Wernicke's area and Broca's area to phonological processes, all of the regions activated in the present experiment have been described previously in studies aiming at extracting anatomical grounds for sub-lexical auditory language processing. The thalamus has been related to attention modulation in audition [14], in language perception Vol 10 No 12 20 August 1999

2601

NeuroReport

G. Thierry et al.

FIG. 2. Spatial-temporal sequence of activations detected in four contiguous transverse slices for FastRep and SeqDet (no activation was found on the lower slice and the upper slice in both tasks). Signi®cantly activated voxels (Z . 2.2, p , 0.01, every colour but blue) are presented on transverse anatomical backgrounds (one of the subject's 3D high-resolution anatomy in blue) corresponding to the middle of functional acquisition slices. Colour coding is based on peak latency in mean EHR for each activated voxel. A temporal gradient was found in all activated clusters in both tasks and lead from primary sensory areas to primary motor regions. Note the relative heterogeneity of frontal and left temporal clusters in FastRep compared to SeqDet (white circles). L indicates the left hemisphere.

[5] and motor control of speech output [17,18]. The left temporal parietal region [19], and in particular the supramarginal gyrus [4,5], have been proposed as the site of the phonological store and, generally speaking, they are thought to be involved in the phonological loop [20,21]. Finally, the precentral gyrus is directly concerned with the programming of speech output [22]. Temporal information EHRs are delayed by several seconds compared with neurophysiological events that originate haemodynamic changes and the time-range of the delay observed here has been reported previously [23,24]. Whatever the delay, the EHR sequence has proved to be statistically robust and to be reproducible across two different tasks. One could argue that differential vasculature of brain regions could play a role in the temporal EHR shift. Although such an effect is possible, it is to our knowledge far to be ascertained by experimental data. Nevertheless, such a phenomenon would hardly be compatible with several aspects of our results. First, sensory areas reach their peak before association cortices and before pre-motor regions, which is obviously coherent with basic physiological 2602

Vol 10 No 12 20 August 1999

expectations. Second, such a vascular bias can be dismissed by across tasks differences in the time domain as it should yield similar patterns in both conditions. Indeed, input areas (primary auditory cortex and Wernicke's area) were the earliest to reach EHR peak in both tasks whereas areas involved in the articulatory loop (supra-marginal gyrus and Broca's area) and motor output regions (pre-central gyrus) were the latest. The pattern observed here is compatible with short-lived decoding processes in the temporal cortex and sustained activities in regions involved in storing and refreshing integrated phonological information. The left thalamus held an intermediate status that can be related to its role in the ampli®cation of neural signal under treatment and the on-line coupling of co-operative cortical regions [25]. These observations ®t the schematic sequence that was postulated a priori according to Baddeley's model and support the hypothesis of sequential processing for phonological mechanisms. In FastRep, activations of Wernicke's area had a time range that was comparable to those of Broca and homologous regions in the right hemisphere.

Temporal stages of phonological processing This overlap suggests a synchronized and on line load of both input and output association areas that is compatible with fast transcoding processes of unsegmented phonological strings. By contrast, temporal patterns of clusters in SeqDet suggested a temporal segregation between superior temporal regions and frontal regions that seems to con®rm sustained and delayed post-sensory activities in the latter.

Conclusion Our results are highly coherent with respect to a functional model of the phonological loop [4]. They provide a broad cut analysis of region involvement in phonological analysis that will be further improved by characterization of local haemodynamic responses. Further progress in the speed of image acquisition and the sensitivity to haemodynamic variables will be needed to permit a clari®ed analysis of reciprocal relationships between highly specialized cortical sub-domains. However, the method has proved to be sensitive enough to temporally discriminate close-related tasks. Finally, the present study offers new opportunities to improve classical cognitive models such as the phonological loop, with temporal constraints that could specify the sequential/parallel involvement of their components.

NeuroReport References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25.

Binder JR, Frost JA, Hammeke TA et al. J Neurosci 17, 353±362 (1997). Vandenberghe R, Price C, Wise R et al. Nature 383, 254±256 (1996). DeÂmonet JF, Chollet F, Ramsay S et al. Brain 115, 1753±1768 (1992). Paulesu E, Frith CD, and Frackowiak RS. Nature 362, 342±345 (1993). DeÂmonet JF, Price C, Wise R et al. Neurosci Lett 182, 25±28 (1994). Zatorre RJ, Meyer E, Gjedde A et al. Cerebr Cortex 6, 21±30 (1996). Weiller C, Isensee C, Rijntjes M et al. Ann Neurol 37, 723±732 (1995). Gabrieli JD, Poldrack RA and Desmond JE. Proc Natl Acad Sci USA 95, 906±913 (1998). Thierry G, Doyon B and DeÂmonet JF. Neuroimage 8, 391±408 (1998). Turennout M, Hagoort P and Brown CM. Science 280, 572±574 (1998). Besner D. Q J Exp Psychol 39A, 467±478 (1987). Marslen-Wilson WD. Access and integration: Projecting sound onto meaning. In: Marslen-Wilson WD, ed. Lexical Representation and Process. Cambridge: MIT press, 1989. Baddeley AD. Working Memory. Oxford: Clarendon Press, 1986: 129. Frith CD and Friston KJ. Neuroimage 4, 210±215 (1996). Cohen JD, Perlstein WM, Braver TS et al. Nature 386, 604±608 (1997). Friston KJ, Holmes AP, Worsley KJ et al. Hum Brain Mapp 2, 189±210 (1995). Murphy K, Cor®eld DR, Guz A et al. J Appl Physiol 83, 1438±1447 (1997). Chatterjee A, Yapundich R, Mennemeier M et al. Cortex 33, 419±440 (1997). Celsis P, Doyon B, Boulanouar K et al. Neuroreport 10, 1523±1527 (1999). Celsis P, Boulanouar K, Doyon B et al. Neuroimage 9, 135±144 (1999). Zubicaray GI, Williams SC, Wilson SJ et al. Cortex 34, 389±401 (1998). Wise RJS, Greene J, BuÈchel C et al. Lancet 353, 1057±1061 (1999). Buckner RL, Bandettini PA, O'Craven KM et al. Proc Natl Acad Sci USA 93, 14878±14883 (1996). McCarthy G, Luby M, Gore J et al. J Neurophysiol 77, 1630±1634 (1997). LaBerge D. Conscious Cogn 6, 149±181 (1997).

ACKNOWLEDGEMENTS: We thank Pr Isabelle Berry and Pr Claude Manelfe for giving us the opportunity to access MRI facilities of the Hospital of Purpan as well as Dominique Cardebat, Pierre Celsis, Bernard Doyon, Danielle Ibarrola, Sandra Leà and Jean-Luc Nespoulous for precious comments and discussions on the manuscript.

Received 2 June 1999; accepted 18 June 1999

Vol 10 No 12 20 August 1999

2603