Temporal encoding in auditory evoked neuromagnetic fields ...

AUDITORY AND VESTIBULAR SYSTEMS NEUROREPORT

Temporal encoding in auditory evoked
neuromagnetic ®elds: stochastic resonance

Steven M. Stuf¯ebeam, David Poeppel, and Timothy P.L. RobertsCA

UCSF, Neuroradiology, Box 0628, Department of Radiology, 513 Parnassus Ave, San Francisco, CA 94143, USA
CACorresponding Author

Received 12 September 2000; accepted 20 October 2000

Recent investigations have demonstrated that temporal pat- jitter in latency associated with the M100 of auditory evoked
terns of sensory neural activity detected by magnetoencephalo- ®elds was reproducibly measured. Speci®cally, higher intensity
graphy (MEG) re¯ect features of the stimulus. In this study, sounds were associated with an increased reliability. The
neuromagnetic activity was investigated using an event detec- technique was also applied to the noise-enhanced evoked
tion algorithm based on the correlation coef®cient. The results auditory response, producing an objective demonstration of a
of the technique are compared with widely used methods of cortical manifestation of the phenomenon of stochastic reso-
analysis in two experimental conditions and are shown to nance-the paradoxical enhancement in the measurement of the
identify features in the single-trial MEG response that are not signal-to-noise ratio (SNR) induced by optimal addition of noise
apparent in the response obtained by averaging across
repeated trials. As an example of the technique, the physiologic to system input. NeuroReport 11:4081±4085 & 2000 Lippincott

Williams & Wilkins.

Key words: Auditory evoked ®eld; Correlation; M100; Magnetoencephalography; N1; N1m; Single trial analysis; Stochastic resonance

INTRODUCTION stimuli and is not accounted for in the averaged waveform,
Most recent functional imaging studies have focused on which in fact is intentionally employed in part to over-ride,
the spatial organization of brain function. However, there or minimize, the impact of such latency variation. In
is considerable evidence that information encoding may general, the use and interpretation of averaged waveforms
exist in the temporal patterns of neuronal activity. This makes implicit and critical demands on latency reliability
study utilizes magnetoencephalography (MEG), which and precision of determination.
combines high temporal resolution with some spatial
localization ef®cacy, to explore the ®ne temporal nature of Single trials analysis has been successfully applied pre-
the auditory evoked neuromagnetic ®eld. Speci®cally, viously in electrical event-related potential (ERP) studies,
responses are collected during presentation of low sensa- including Puce et al. [3], who used the maximum like-
tion level signals, in conjunction with background acoustic lihood estimation to estimate the latency jitter in the P3
noise. ERP. Other methods are available and are reviewed else-
where [4,5]. Indeed event-related designs are also becom-
Evidence exists that temporal encoding of auditory ing more commonplace in functional MRI brain mapping
stimulus information exists at a variety of levels: sound studies [6±8]. We propose a single trial analysis of MEG
properties are encoded into the temporal activity of action recorded neuromagnetic activity, which quanti®es the
potentials in the auditory nerve [1], which are relayed to reproducible timing and in morphology (reliability) of
the primary auditory cortex. Recently, the precise latency components of the neuromagnetic ®eld. We suggest that
of events detected in the neuromagnetic activity of audi- use of a quantity such as reliability will allow the elucida-
tory cortex was shown to encode pitch and timbre [2]. Such tion of a cortical manifestation of stochastic resonance.
non-invasive measurement of the timing of human cortical
electrical events has relied on measurements of the latency In our study, the correlation coef®cient is used for the
of component peaks of evoked electrical and neuromag- analysis because it measures, in a general sense, the
netic ®elds, such as the electrical N1, or its magnetic similarity in a data set [9±11]. This method presented here
counterpart, N1m or M100. Such analysis is usually per- differs from the Woody method [12] in that it is exquisitely
formed on the averaged response of multiple individual sensitive to variations in the timing of a neuromagnetic
trials of a repeated stimulus where the signal-to-noise ratio waveform [10]. The Woody method used the cross-correla-
(SNR) is improved. As successful as this type of analysis tion between a sliding template and the single trial
has been over the past decades, it is limited. For example, response. The cross-correlation is very sensitive to changes
a physiological measure such as latency jitter relies on the in single trial response magnitude, but is relatively insensi-
precise timing of the neural evoked response to individual tive to timing [9,10], whereas the correlation coef®cient is
insensitive to changes in response signal magnitude, but is

0959-4965 & Lippincott Williams & Wilkins Vol 11 No 18 18 December 2000 4081

NEUROREPORT S. M. STUFFLEBEAM, D. POEPPEL AND T. P. L. ROBERTS

quite sensitive to temporal variations. We propose a varia- tion and delivery) of 80, 160, 320 and 640 mV, correspond-
tion of the Woody technique which eliminates the sliding ing to noise-to-signal ratios (of the input acoustic signal) of
template and uses the correlation coef®cient to compute $0.22, 0.44, 0.88 etc. in addition to the case of zero
the reliability of the neuromagnetic response. background noise. In addition, a 40 mV level was used in
the ®nal two subjects to broaden the noise level range to
Two experiments on auditory evoked ®elds were in- include extremely low noise levels. We chose to employ
cluded to illustrate potential advantages of the technique. voltage measures of relative noise level, rather than noise
The ®rst experiment assessed the relationship between sensation levels per se, due to the time-varying nature of
evoked response reliability and stimulus presentation level, the threshold for noise detection, which we measured in
for simple sinusoidal stimuli. each subject, and which we found to increase progressively
during sustained noise presentation. To allow for habitua-
Many functional imaging studies (including fMRI and tion in the experimental studies, the background noise was
MEG) are ultimately limited by the SNR of the detected played to the subjects for at least two minutes prior to the
responses. Ingenious approaches to addressing this con- onset of stimulus tones.
cern include hardware improvements (e.g. higher ®eld
magnets for MRI), or methodological (e.g. signal aver- Data acquisition: A 37-channel biomagnetometer sensor
aging), or analysis (e.g. principal components analysis) array (MAGNES, Biomagnetic Technologies Inc., San Die-
strategies. A rather different approach may exist by ex- go, CA) was positioned over the left temporal area contral-
ploiting a phenomenon recognized in information theory ateral to the stimulus presentation, so as to optimally
and some single cell preparations, the paradoxical im- record the M100 component of the evoked neuromagnetic
provement in output signal-to-noise ratio that results from ®eld elicited by a reference 1 kHz tone at $40 dB SL, in the
addition of an optimum amount of noise to the input absence of background noise. This is standard set up
signal. This phenomenon is known as stochastic resonance procedure in our laboratory and others. MEG epochs of
(SR). 600 ms duration were collected around each stimulus
(including 100 ms pre-trigger baseline). In the case of the
Thus, in a second component of the reported study we continuous background noise stimuli, the epoch collection
demonstrate a cortical manifestation of the stochastic was triggered by the onset of the sinusoidal tone. Data
resonance effect in which the output measure (the relia- sampling rate was 1.047 kHz/channel with a 1 Hz real-time
bility of the MEG recorded neuromagnetic evoked ®eld) is high-pass ®lter.
improved by the addition of acoustic noise, of the appro-
priate level, to a stimulus signal. Data analysis: Two methods of analysis were used for
each stimulus. First, the averaged response was computed,
MATERIALS AND METHODS time-locked to stimulus onset. The averaged epoch for each
All research procedures were performed in accordance stimulus condition was then ®ltered with a 1±20 Hz band-
with the institutional committee on human research. In- pass ®lter to remove high frequency noise in the evoked
formed consent was obtained from the volunteers. response signals. The M100 component was de®ned as
occurring at the maximal r.m.s. evoked ®eld in the post-
Intensity variation stimuli: A single male subject (age 30 stimulus window 80±180 ms, subject to a single equivalent
right-handed) underwent MEG recording while presented dipole model correlation exceeding r ˆ 0.95, and the M100
with auditory sinusoidal tone stimuli of varying sensation latency and r.m.s. evoked ®eld magnitude across 37
level. Stimuli were generated using an Apple Macintosh detector channels were determined.
computer driving a Digidesign Audiomedia II sound card
delivered via Eartone ER3A transducers (Etymotic, Elk Second, a single trial analysis of the raw, unaveraged
Grove, Il USA) to ear inserts. The stimuli were presented data was performed: the correlation coef®cient was com-
independently to the right ear at interstimulus intervals of puted between the averaged waveform (described above)
0.9±1.1 s. Pure sinusoidal tones of 1000 Hz (400 ms dura- and each individual MEG epoch waveform. The averaged
tion, 5 ms linear on/offset ramps) were presented under waveform served as a reference template for single trial
®ve blocked conditions: at intensity levels ranging from response matching. Each individual epoch was also ®ltered
peri-threshold to normal conversation levels: 5, 10, 20, 30, with a 20 Hz cut-off low-pass ®lter to increase the robust-
and 40 dB sensation level (SL). The stimulus was presented ness of the technique. The single sensor channel with the
100 times in each condition. All measurements were largest magnitude averaged response was used for calcula-
repeated on a total of four occasions to assess reproduci- tion of the correlation coef®cient because it had the highest
bility. signal-to-noise ratio. In principle, any sensor channel, or
combination of sensor channels could be analyzed simi-
Background noise stimuli: In six healthy right-handed larly. The normalized correlation coef®cient (mean re-
volunteers, 200 Hz pure tones (400 ms duration with 5 ms moved) was computed for each epoch over a 100 ms time
linear on/offset ramps) at slightly suprathreshold levels window centered on the averaged M100 peak as follows:
(6 dB SL) were presented to the right ear in the presence of
various levels of continuous auditory white noise (band- latencyˆ‡windowa2
width 50±10 000 Hz) with an interstimulus interval of 1.5± (EiTian À ET)
1.7 s. Noise levels were kept constant during presentation
of > 100 repetitions (trials) of the tone. Different noise level C ˆ iˆ latencyÀwindowa2 ,
conditions were applied in a random order, including a
condition without background noise. Noise levels were ó EóT
de®ned by r.m.s. output voltages (prior to ®nal ampli®ca-
where C was the correlation coef®cient, n was the number

4082 Vol 11 No 18 18 December 2000

TEMPORAL ENCODING IN AUDITORY EVOKED NEUROMAGNETIC FIELDS NEUROREPORT

of points, Ei and Ti were the magnitude of the magnetic (a)
®eld at time i for the epoch and template respectively, and 160
óE and óT were the standard deviations of the epoch and
template waveforms. Thus, 100 correlation coef®cients 150
were obtained for each stimulus condition block. To
investigate encoding accuracy, a threshold correlation va- M100 latency (ms) 140
lue was used on the correlation coef®cients obtained on the
single trial basis. A match between the individual epochs 130
and the averaged response was termed a 'hit' when the
correlation coef®cient exceeded a threshold value. The 120
choice of threshold correlation coef®cient is somewhat
arbitrary, but should be chosen so as to optimize effective 110
dynamic range and to minimize ¯oor and ceiling effects in
the experimental conditions. Thus, for the stimulus inten- 100
sity variation experiment a correlation coef®cient threshold 0 5 10 15 20 25 30 35 40 45
value of 0.8 was used, since responses were on the whole
similar, and a lower threshold for acceptance would lead Stimulus intensity (dB SL)
quickly to a saturation, or ceiling limit. For the background
noise series, a lower value (0.5) was used, commensurate (b)
with the weaker evoked response. To control for the 200
possibility that the background noise itself increases the
number of hits in the absence of an auditory stimulus, the M100 magnitude (fT rms) 180
correlation coef®cient analysis was performed using the
template around the center of the À100 ms to 0 ms presti- 160
mulus window (i.e. 50 ms prior to the onset of the
stimulus). This is a similar control used by others [3], by 140
using the prestimulus window as the control rather than
the later portion of the evoked response. The number of 120
hits used in this sense measures the number of random
hits with the background brain noise, most notably the 100
background alpha activity.
80
RESULTS 0 5 10 15 20 25 30 35 40 45
Stimulus intensity variation: Parameters derived from
the averaged response and the correlation coef®cient of Stimulus intensity (dB SL)
individual MEG epochs yielded complementary results
(Fig. 1). In agreement with previous studies [13,14], Fig. Hits (%)(c)
1a,b show that the averaged M100 evoked ®eld magnitude 65
increases with more intense stimuli; correspondingly, the 60 5 10 15 20 25 30 35 40 45
averaged M100 latency decreases. Figure 1c shows the 55 Stimulus intensity (dB SL)
results of the correlation coef®cient analysis. In terms of 50
hits, or matches to a template, a reproducible linear 45
(r ˆ 0.9) relationship exists between the number of hits and 40
the intensity of the stimulus. The latency decrease effect is 35
non-linear with dB SL, plateauing above ~30 dB SL, as 30
reported previously [13]. Interestingly, the hits score con- 25
tinues to improve linearly up to 40 dB SL. The plots for the 20
magnitude of the M100 detected magnetic ®eld and the 15
number of hits de®ned by the correlation coef®cient both 0
yield a near linear relationship in the 5±40 dB SL range,
but are measures of different quantities. The number of Fig. 1. (a) Latency of M100 averaged response versus stimulus intensity
hits can be thought to re¯ect primarily the reduction of of a 400 ms 1 kHz pure tone. (b) Magnitude of the averaged M100
physiologic jitter in the M100 or the reliability of individual response versus the intensity for the same tone. (c) Number of hits of
evoked responses. This no doubt contributes to an in- the single trial data that have a correlation coef®cient . 0.8 vs intensity
creased peak evoked ®eld amplitude, since time-locking is of the tone.
tighter.
levels of noise the number of hits decreased as might be
Background noise: All subjects demonstrated an increase expected intuitively. With the correlation threshold set to
in the number of hits at an optimal level of noise compared 0.5 the responses were classi®ed as hits 30 Æ 2% of the time
to the no noise condition (Fig. 2a). Using a ÷2 analysis, all in the absence of added stimulus noise. This represented
reach statistical signi®cance, assuming the chance of a hit our baseline. With the addition of an optimal, small
is 30% at baseline (see legend to Fig. 2a). At the highest

Vol 11 No 18 18 December 2000 4083

NEUROREPORT S. M. STUFFLEBEAM, D. POEPPEL AND T. P. L. ROBERTS

(a) Subject #1 Subject #2 Subject #3
50 50 50
45 45
Reliability (% Hits) 45 Reliability (% Hits) 40 Reliability (% Hits) 40
35 35
40 30 30
25 25
35 200 0.5 1 1.5 2 200 0.5 1 1.5 2

30 noise-to-signal ratio noise-to-signal ratio

25 Subject #5 Subject #6
50 50
200 0.5 1 1.5 2 45 45
noise-to-signal ratio 40 40
35 35
Reliability (% Hits) Subject #4 Reliability (% Hits) 30 Reliability (% Hits) 30
50 25 25
45 200 0.5 1 1.5 2 20
40
35 noise-to-signal ratio 0 0.5 1 1.5 2
30 noise-to-signal ratio
25
200 0.5 1 1.5 2

noise-to-signal ratio

(b) Reliability (% hits) ** Fig. 2. (a) Reliability of the neuromagnetic response from all subjects.
50 All stimuli are 6 dB SL pure tone 200 Hz burst (400 ms duration) during
45 0.5 1 1.5 2 the presence of continuous background acoustic noise (and quiet state).
40 noise-to-signal ratio Acoustic background noise is quanti®ed as ratio of mV r.m.s. relative to
35 the mV r.m.s. of the tone burst stimulus. A hit is de®ned as correlation
30 coef®cient . 0.5 for single trial waveform when compared to the
25 averaged waveform over all repetitions. The dotted line represents 95%
20 con®dence level determined by maximum likelihood estimation based on
0 the assumption that the chances of a match is 0.30. This probability was
determined as the result of several subjects undergoing the quiet, (no
noise) condition ®ve times, which resulted in an average of 30 Æ 3.1
matches. (b) Averaged reliability (circles) of neuromagnetic response
from all six subjects (mean Æ s.e.m.); asterisks indicate signi®cance for
MANOVA repeated measure . 95%. Lower curve (squares) indicates
(false positive) matches to the background activity averaged across all six
subjects (mean Æ s.e.m.).

amount of background white acoustic noise, this hit rate was not evident in the averaged data, and revealed the
rose progressively to a maximum of up to 50% at a objective manifestation of stochastic resonance in the audi-
stimulus input signal-to-noise ratio of between 4 and 6 tory evoked neuromagnetic ®eld.
(noise-to-signal ratio 0.167±0.25), and reaching statistical
signi®cance (MANOVA repeated measure . 95%). The DISCUSSION
mean (Æ s.d.) increase in hits was found to be 9.1 Æ 2 hits/ We have outlined a data analysis method, based on the
100 trials, representing a mean effect size of 30% (from correlation coef®cient, during a selected time-window of
baseline of 30 hits). Maximizing a measured response evoked activity, that relies on the similarity of the averaged
through the addition of an optimal amount of background evoked neuromagnetic response to the single trial data set.
noise, known as the stochastic resonance peak, is the It has several important qualities that make it a valuable
signature of the stochastic resonance phenomenon (Fig. addition to MEG evoked response analysis. In some
2b). Notice that the number of false positive hits remained circumstances it provides a novel measurement not found
at the same level, despite the addition of acoustic noise (no in the averaged data. Importantly, it is sensitive to the
statistically signi®cant difference with repeated measure reduction of physiologic jitter in the latency of the M100. It
ANOVA ( p . 0.65) compared with the no-noise condition). is also sensitive to variations in single trial evoked
Interestingly, there was no signi®cant change in the aver- response morphology. The background noise experiments
aged response latency or peak magnitude of the evoked demonstrate a peak in the number of matches to the
magnetic ®eld for the same range of background noise. averaged response template, indicating decreased variabil-
Thus, single trial analysis demonstrated an obvious varia- ity in the M100 auditory evoked ®eld component.
tion between conditions (input stimulus noise levels) that Although there was decreased variability in the response

4084 Vol 11 No 18 18 December 2000

TEMPORAL ENCODING IN AUDITORY EVOKED NEUROMAGNETIC FIELDS NEUROREPORT

to individual trials, interestingly there was no discernible sign of cochlear implants [17], which may in fact offer an
change in the magnitude or latency of the averaged opportunity to supply input noise that is uncorrelated
response. across neuronal populations, as mentioned above. As
suggested by our results and others' [17] adding noise to
In the absence of random noise and morphologic varia- cochlear implant signals may, in some circumstances, be
tion of the evoked response waveform, the dominant preferred. It is also possible that other functional neuroi-
contribution to reduced correlation between single trials maging techniques, such as fMRI, might bene®t from the
and the mean (number of matches) is the latency shift, or augmentation of brain activity achieved through SR. Our
jitter, itself. Speci®cally, in the auditory evoked responses, results suggest that the addition of background noise
latency shifts can arise from either stimulus intensity might enhance the evoked response elicited by a weak
differences [14] or stimulus frequency differences [2]. stimulus. Although MEG was used for this study, the
Spectral and timbral properties, including contributions to principle might be applicable to, for example, evoked
phonetic character, may analogously be encoded in the electrical responses and event-related functional magnetic
evoked ®eld latency [2]. Thus, de®ning a tolerance for resonance imaging.
latency shifts can be interpreted as de®ning an insensitivity
range for stimulus intensity, frequency and other attri- CONCLUSION
butes. Adjusting the correlation coef®cient threshold corre- Precise neural timing is important in neural systems. We
spondingly alters this degree of tolerance and thus the described an increase in the neuromagnetic reliability of
insensitivity range. the auditory evoked ®eld by the addition of external
acoustic noise. This phenomenon of stochastic resonance
The results of the background noise experiment are was not discernible in the averaged response, and required
intriguing. A manifestation of stochastic resonance in the a single-trial analysis of the ®ne temporal structure of
auditory cortex is, perhaps, not an increase in the magni- neuromagnetic ®eld.
tude of the neuromagnetic response, but rather a decrease
in the variability of single trial responses. The correlation- REFERENCES
based analysis may be a method of approaching the study
of stochastic resonance with evoked magnetic ®elds. Intra- 1. Rieke F. Spikes. Exploring the Neural Code. Cambridge, Mass: MIT Press;
cellular recordings of the auditory nerve show that infor- 1997.
mation ¯ow rate (measured by signal-to-noise ratio of the
spike rate) of peri-threshold stimuli is increased with the 2. Roberts TPL and Poeppel D. Neuroreport 7, 1138±1140 (1996).
addition of an optimal amount of noise. It follows that the 3. Puce A, Berkovic SF, Cadusch PJ et al. EEG Clin Neurophysiol 92,
evoked cortical ®elds might occur at a more reproducible
latency, rather than necessarily showing an increase in the 352±364 (1994).
magnitude of the evoked ®eld. 4. Childers DG, Perry NW, Fischler IA et al. Crit Rev Biomed Eng 4,

The increase in the reliability of the neuromagnetic 185±200 (1987).
response reported here is modest, and not detectable with- 5. Childers DG, Fischler IS, Boaz TL et al. IEEE Trans Biomed Eng 33,
in the averaged neuromagnetic response. This may re¯ect
the fact that the noise in this experiment is externally- 1069±1075 (1986).
supplied acoustic noise. Theoretical analysis has suggested 6. Glover GH. Neuroimage 9, 416±29 (1999).
that the stochastic resonance effect is greatest when noise 7. Pinel P, Le Clec'H G, van de Moortele PF et al. Neuroreport 10,
is uncorrelated across many cells, which is dif®cult to
achieve with external acoustic noise [15,16]. Theoretically, 1473±1479 (1999).
the stochastic resonance effect may be greatest when the 8. D'Esposito M, Zarahn E and Aguirre GK. Psychol Bull 125, 155±164
noise is generated internally, or at least is uncorrelated
across neurons. Additionally, the stochastic resonance (1999).
effect of slightly supra-threshold stimuli might be less 9. Smulders FIY, Kenemans JL and Kok A. EEG Clin Neurophysiol 92,
pronounced than with sub-threshold stimuli. We did not
use stimuli below the hearing threshold, as these stimuli 107±114 (1994).
did not evoke a measurable neuromagnetic response. 10. Ford JM, White P, Lim KO et al. Biol Psychiatry 35, 96±103 (1994).
11. Kachigan SK. Multivariate Statistical Analysis: A Conceptual Introduction.
Other applications taking advantage of SR include de-
New York: Radius Press; 1991.
12. Woody CD. Med Biol Eng 5, 539±553 (1967).
13. Elberling C, Bak C, Kofoed B et al. Scand Audiol 10, 203±207 (1981).
14. Stuf¯ebeam SM, Poeppel D, Rowley HA et al. Neuroreport 9, 91±94

(1998).
15. Parnas BR. IEEE Trans Biomed Eng 43, 313±318 (1996).
16. Collins JJ, Imhoff TT and Grigg P. Nature 383, 770 (1996).
17. Morse RP and Evans EF. Nature Med 2, 928±932 (1996).

Acknowledgements: The authors thank Biomagnetic Technologies Inc. for equipment support and Susanne Honma R.T. and Mary
Mantle R., EEG.T., for excellent technical assistance. Frederic Theunessen, PhD and Howard Rowley, MD are gratefully
acknowledged for thoughtful discussions and insightful comments. This work was supported by the UCSF Radiology Pilot
Research program.

Vol 11 No 18 18 December 2000 4085