Epilepsia, 42(8):1007–1016, 2001
Blackwell Science, Inc.
© International League Against Epilepsy
Left Vagus Nerve Stimulation with the Neurocybernetic
Prosthesis Has Complex Effects on Heart Rate and on Its
Variability in Humans
*Mark G. Frei and *†Ivan Osorio
*Flint Hills Scientific, L.L.C., Lawrence; and †Comprehensive Epilepsy Center, University of Kansas Medical Center,
Kansas City, Kansas, U.S.A.
Summary: Purpose: The purpose of this study was to deter-
mine if stimulation of the left vagus nerve (LVNS) with the
neurocybernetic prosthesis (NCP) in humans is, as claimed in
the literature, without cardiac chronotropic actions.
Methods: We analyzed 228 h of ECG recorded from five
subjects with intractable epilepsy who had not benefited from
LVNS, for effects on instantaneous heart rate (IHR) and heart
rate variability (HRV).
Results: There were two main cardiac responses: (a) brady-
cardia, and (b) tachycardia during the first half, followed by
bradycardia during the second half of stimulation (biphasic
response). Multiphasic responses characterized by alternating
bradycardia and tachycardia were rarely observed. HRV was
either increased or decreased depending on the subject and on
Recently published literature indicates a lack of effect
on heart rate of left vagus nerve stimulation (LVNS) in
humans implanted with the neurocybernetic prosthesis
(NCP). A large multicenter study on the efficacy and
safety of LVNS for treatment of intractable epilepsy (1)
and more detailed investigations of its effects on heart
rate (2) and heart period variability (3) concluded that
LVNS had no chronotropic effects. This finding, which
runs counter to the proven tenet that vagal activation
causes bradycardia (4), is explained on the basis that the
ventricle’s innervation by the left nerve is sparse com-
pared with that of the atrium (5). We had found in a
preliminary study (6) that if the periods of stimulation
were segregated from those when the device was not
active, LVNS caused bradycardia and decreased heart-
rate variability (HRV). Given the very small size of our
Revision accepted January 29, 2001.
Address correspondence and reprint requests to Dr. I. Osorio at
Comprehensive Epilepsy Center, University of Kansas Medical Center,
3901 Rainbow Boulevard, Kansas City, KS 66160-7314, U.S.A.
E-mail: iosorio@kumc.edu
1007
the stimulation parameters. HRV as a function of HR also
showed high interindividual variability, and interestingly, in
one case behaved paradoxically, increasing at higher and de-
creasing at lower heart rates.
Conclusions: LVNS at high intensities has complex effects
on IHR and HRV, which show large interindividual variability.
These spectra of cardiac responses reflect the interplay of au-
tonomic, visceral, and somatic sensory afferences and the role
of central structures in their integration. These findings also
point to the need for more comprehensive studies of cardiac
function in humans implanted with the NCP, using sensitive
methods for data processing and analysis such as those devel-
oped for this study. Key Words: Chronotropic—Cybernetic—
Cycle time—Prosthesis—Vagus nerve.
original sample, we have expanded our observations to
assess the validity of those results and to test the hypoth-
esis that LVNS causes bradycardia, but that its effect was
not uncovered in the studies we alluded to earlier (1–3),
due in part to the manner in which the data were pro-
cessed and analyzed.
METHODS
We analyzed 25 prolonged ECG recordings from five
subjects (228 h total; mean, 45.6 h per subject; range,
8–72 h) implanted with the NCP prosthesis (Table 1) and
undergoing monitoring for epilepsy surgery. These sub-
jects were included in this study, as compared with all
other subjects receiving LVNS at this Center, because
they underwent evaluation for epilepsy surgery and pro-
vided ample data for meaningful statistical analysis.
There were no changes in LVNS parameters during
monitoring, and the routine surgical evaluation protocol
at KUMC, including temporary discontinuation of anti-
seizure drugs (AEDs), was followed in all cases. The
1008 M.G. FREI AND I. OSORIO

TABLE 1. ECG recordings

NCP parameters
Total Output Pulse ON OFF Time from
Age No. of length NCP current Freq. width time time implant to
Subject (yr) Gender Diagnosis Segments (h) effect (mA) (Hz) (␮s) (s) (min) recording (mo)
S1 19 M MFE 6 43 NR 1.25 30 500 30 5.0 26
S2 20 F FLE 1 8 NR 3.25 30 750 30 5.0 8
S3(a) 23 M FLE 3 24 NR 3.25 30 500 30 5.0 25
S3(b) 6 38 NR 2.0 30 500 30 5.0 29
S4 40 M FLE 6 44 NR 3.0 30 500 30 5.0 16
S5 31 M TLE 3 72 NR 2.25 30 500 30 5.0 25
Total 25 228

NR, nonresponder; MFE, multifocal epilepsy; FLE, frontal lobe epilepsy; TLE, temporal lobe epilepsy.

data included contiguous recordings of ⱖ24 h for four of
the five subjects (a single 8-h recording was used in the
other subject) and consisted of >1 million total heart-
beats. ECG was recorded from bilateral intraclavicular
electrodes linked to each other (impedance <5 KOhm)
using a Neoped 4000 unit (BMSI, Los Gatos, CA,
U.S.A.). Data were digitized at 240 Hz with 10 bits of
precision and a ±300 ␮V dynamic range, and archived on
super VHS tapes.
Segments during which the vagus nerve was stimu-
lated (LVNS-on segments) were separated (see Appen-
dix 1) from those during which there was no stimulation
(LVNS-off segments) to (a) minimize possible “averag-
ing” effects, because on periods correspond only to
∼11% of the total data; (b) allow subjects to serve as their
own controls; and (c) account for nonstationarity due to
state changes. Instantaneous heart rate (IHR) and HRV
were computed for all segments.
Automated R-peak recognition and signals derived
from its measurement
An algorithm for automated QRS detection validated
by expert visual analysis was applied to the data (6,7).
This algorithm localizes R-wave peaks with a precision
of 1/240 s and automatically assesses quality of data,
rejecting segments contaminated with noise. Figure 1
shows a segment of raw ECG data (with LVNS-induced
artifact), together with the corresponding computed IHR.
Transformations were applied to the “time of beat” se-
quence to produce multiple time series such as the IHR
(see Table 2 for the complete list) for subsequent statis-
tical analysis.
Data manipulation: Segmentation of signals by
device period and time-locking by stimulation onset
The standard duration of an NCP cycle, which was
constant throughout our series, is 334.49 s, of which only

FIG. 1. A: The ECG signal and the left

vagus nerve stimulation (LVNS) artifact.
B: The corresponding output of the au-
tomated algorithm for heart-rate mea-
surement. Notice the intensity ramp-up
at the start, ramp-down at the end, and
the five brief “OFF gaps” during the
stimulation (annotated).

Epilepsia, Vol. 42, No. 8, 2001

 NEUROCYBERNETIC PROSTHESIS 1009

TABLE 2. Transformations applied to produce
multiple series
Time series constructed
from time-of-beat sequence Description
RR R-to-R intervals. Time (s) between
successive beats
IHR Instantaneous heart rate (beats/min).
IHR ⳱ 60/RR
DRR Time differential between adjacent
RR intervals. DRR(k + 1) ⳱
RR(k + 1) − RR(k)
Median-normalized RR A sequence of RR intervals divided
by their median value
Median-normalized IHR A sequence of IHR values divided
by their median value
HF RR We applied an FHS-proprietary
algorithm to the RR sequence,
which in this case extracts the
highest-frequency component
from the signal (i.e., it adaptively
removes a fluctuating baseline),
leaving only the highest
frequency of the RR intervals
HF IHR High-frequency IHR. Obtained by
applying the HF decomposition
algorithm to the IHR sequence
|dRR| Absolute value of dRR sequence
FHS, Flint Hills Scientific, L.L.C.
37.2 s is taken by stimulation. To facilitate comparison
between LVNS-on and LVNS-off segments, each full
NCP device cycle was treated as a row of a matrix (Fig-
ure 2). The first stimulation cycle in each recording
formed the first row, and each subsequent cycle was
added as a new row underneath the preceding one. Once
the data matrix was fully laid out, each row/cycle was
divided into nine nonoverlapping subsegments whose
length (37.2 s) was chosen based on the actual duration
of stimulation. In the resulting “structure,” the on sub-
segment is placed centrally and flanked on each side by
the four off subsegments that preceded and followed the
stimulation. Each row was aligned (“time-locked”) using
the onset of LVNS stimulus as the marker. For each
subsegment of each row, we computed the following
statistics for IHR and the other time series listed in
(Table 2): mean, median, standard deviation, and inter-
quartile range. We then computed the cumulative distri-
bution functions (CDFs) for each of these statistics for
the on versus off subsegments. Figure 3 shows an ex-
ample of the CDFs for mean and SD IHR computed for
each of the subsegments from one subject (S3). To test
the null hypothesis (identical on vs. off distributions) we
applied the two-sample Kolmogorov–Smirnov test (8)
and then examined all tests for which this hypothesis was
rejected at ␣ ⳱ 0.05, ␣ ⳱ 0.01, and ␣ ⳱ 0.001 (Ta-
ble 3).
RESULTS
The automated QRS detection system rejected 2.8 of
228 h (i.e., 1.2%) of data because of noise contamination.
Using time and frequency domain analysis, we found
that LVNS has a clear effect on IHR and HRV. These
effects, although interrelated, are discussed separately
for purposes of clarity.
Cardiac chronotropic effects
Eighty percent of prolonged ECG recordings showed a
significant difference in median normalized IHR at ␣ ⳱

FIG. 2. The procedure for creation of a

“data matrix” of signal segments and the
“phase portrait” of instantaneous heart
rate (IHR). The central column (arrow)
corresponds to the segments with left va-
gus nerve stimulation LVNS-on, flanked
on either side by four LVNS-off seg-
ments. Each row corresponds to a com-
plete device cycle (334.49 s) with
successive rows arranged in temporal
order.

Epilepsia, Vol. 42, No. 8, 2001

1010 M.G. FREI AND I. OSORIO

in magnitude and direction, which we grouped into two

main classes based on the frequency of occurrence:
1) Uniphasic, consisting of bradycardia during the en-
tire LVNS-on segment, sometimes followed by
poststimulation “rebound” tachycardia (cases S1,
S2, S3(a,b); see Figure 4A);
2) Biphasic, consisting most frequently of tachycardia
during the first half, followed by bradycardia dur-
ing the second half of stimulation (cases S4, S5;
see Figure 4B); poststimulus “rebound” tachycar-
dia also was often observed in this group’s cases.

Multiphasic responses characterized by alternating

bradycardia and tachycardia were seen infrequently (Fig-
ure 4C). The occurrence and degree of LVNS-induced
bradycardia seemed independent of the absolute IHR but
not of intensity of stimulation. One subject (S3) had a
median IHR reduction of 12% at an intensity of 3.25 mA
compared with 1.5%, after the intensity had been de-
creased to 2.0 mA 4 months later.

Effects on heart rate variability

FIG. 3. The cumulative distribution functions (CDFs) for two of
the statistical quantities [mean instantaneous heart rate (IHR)
and SD of IHR] for each IHR data matrix element for one subject
[S3(a)]. The CDFs for left vagus nerve stimulation (LVNS)-on are
graphed in red, the CDF corresponding to the LVNS-off subin-
tervals immediately after stimulation is shown in green (to illus-
trate the rebound effect), and the remaining seven OFF
subintervals are shown in blue. A: The significant decrease in
IHR during the ON period for this subject. B: The corresponding
increase in HRV as measured using SD of IHR. The significant
effects of LVNS on heart rate and HRV are evident in these
pictures.
0.05 during LVNS-on, compared with LVNS-off, and
56% showed a significant difference at ␣ ⳱ 0.001. Over
the entire series, the largest relative decrease in median
IHR during LVNS-on compared with the mean of the
same quantity over the preceding LVNS-off subsegment
medians in the same device cycle was 45.5% [S3(b)],
from 104.4 beats/min to 56.9 beats/min.
Net effect on IHR showed interindividual variability
We analyzed the data for HRV changes using both
frequency and time-domain techniques. Power spectral
density estimates of ECG, a popular measure of degree
of sympathetic and parasympathetic activation (9), were
computed separately and then compared for the IHR sig-
nal for LVNS-on versus LVNS-off subsegments for each
subject. LVNS decreased the fraction of total power in
the “HF band” (0.25–0.35 Hz) by 60%.
Using time-domain techniques, changes in HRV were
found during LVNS-on compared with LVNS-off for
every subject. These changes were consistent intra- but
not interindividually. In a preliminary study (6), we ob-
served in one subject a 59% decrease in SD of HRV
during LVNS compared with baseline (see Figure 5A,
B). Sixty percent of recordings showed a significant dif-
ference in the on versus off distributions of HRV quan-
tified as the SD of HF RR at ␣ ⳱ .05 (Table 3). The
direction of HRV changes showed interindividual differ-
ences: cases S1 and S2 had significant (␣ < 0.05) de-
creases, whereas cases S4 and S5 had significant
increases in HRV at the same ␣ levels. The remaining
subject (S3) had a parameter-dependent response: sig-
nificant (␣ < 0.05) decrease in HRV occurred at lower-

TABLE 3. Tests of the null hypothesis

Median Median
␣ level normalized normalized
.05/.01/.001 Raw RR Raw dRR Raw IHR HF RR HF IHR RR IHR |dRR|
Mean 0.24/0.2/0.2 0.4/0.24/0.2 0.28/0.2/0.2 0.48/0.32/0.2 0.56/0.32/0.2 0.76/0.68/0.48 0.8/0.72/0.56 0.36/0.32/0.24
Median 0.28/0.2/0.2 0.32/0.28/0.12 0.32/0.2/0.2 0.4/0.36/0.16 0.56/0.36/0.32 0.8/0.6/0.44 0.8/0.6/0.44 0.44/0.28/0.2
Std 0.56/0.48/0.4 0.36/0.32/0.24 0.6/0.56/0.4 0.64/0.48/0.4 0.76/0.64/0.52 0.52/0.48/0.4 0.56/0.52/0.4 0.48/0.32/0.28
Iqr 0.56/0.44/0.28 0.52/0.28/0.24 0.64/0.56/0.44 0.6/0.44/0.36 0.72/0.56/0.44 0.6/0.52/0.4 0.64/0.6/0.4 0.44/0.32/0.16

Epilepsia, Vol. 42, No. 8, 2001

 NEUROCYBERNETIC PROSTHESIS 1011

FIG. 4. The cycle-median-normalized

instantaneous heart rate (IHR) is ob-
tained by dividing the IHR signal for a
full device cycle by the median IHR for
that cycle. Time-locking the results for
multiple device cycles and overlaying
(to produce a “grand median”) reveals
left vagus nerve stimulation (LVNS) ef-
fects. A: The bradycardia caused by
LVNS over 8 h for one subject [S3(a);
86 device cycles]. During stimulation
there was a 14% reduction in baseline
IHR, followed by a poststimulus re-
bound tachycardia of 4% above base-
line. B: The biphasic response to LVNS
stimulation over 24 h for another sub-
ject (S5; 257 device cycles). There was
an initial IHR increase (4% over base-
line), followed by a 4% reduction and a
slight poststimulus rebound tachycar-
dia (1.6% over baseline). C: A multi-
phasic response to LVNS; fleeting
bradycardia (1.6% below baseline) at
LVNS onset is followed by tachycardia
(3.1% above baseline) and finally by
more sustained bradycardia (7.5% be-
low baseline), which disappears when
stimulation ends. This example also il-
lustrates poststimulation rebound
tachycardia (4% above baseline) last-
ing 15 s.

intensity settings (2.0 mA), and a significant (␣ < 0.05)
increase in HRV at higher-intensity settings (3.25 mA).
Table 3 shows the percentage of recordings with signifi-
cant differences of several relevant variables in the on
versus off distributions at these three ␣ levels (see Ap-
pendix 2).
Analysis of the interrelationship between heart rate
and heart rate variability
HRV normally varies inversely with heart rate: the
lower the IHR, the larger the potential HRV. To study
the relationship of heart rate to HRV in more detail, we
further divided each on and off subsegment into five
parts [corresponding to the NCP device timing crystal
period (4) which is 6.8 s; see Figure 1A], and computed
the median and SD of each resulting IHR sequence, and
plotted them against each other. Results varied interin-
dividually. For instance, in subject S4, HRV was higher
during VNS-on compared with VNS-off, over all ECG
recordings, whereas for subject S1, the opposite was true.
For subject S5 (see Figure 5C), LVNS caused a para-
doxic increase in HRV at higher IHR (median HRV for
LVNS-on exceeded the 65th percentile for LVNS-off at
75 beats/min) and a paradoxic decrease in HRV at lower
IHR (median HVR for LVNS-on was below the 20th
percentile for LVNS-off for IHR <60 beats/min).
DISCUSSION
Our findings demonstrate that LVNS at high intensi-
ties, in humans with intractable epilepsy, has a complex
chronotropic effect with appreciable interindividual vari-
ability. The discrepancies between this and other studies
that found no effect (1–3), are at least partially attribut-
able to methodologic differences in the processing and
analysis of the data.

Epilepsia, Vol. 42, No. 8, 2001

1012 M.G. FREI AND I. OSORIO
These differences are
1. Window size. Although we used windows of small
size to minimize the sampling or averaging effect, others
(1–3) computed heart rate from long (minute to multi-
hour) windows. Given that the on segments corresponds
only to 11% of the data generated each day, it is not
surprising that chronotropic effects caused by LVNS,
even if large, would escape detection when on and off
subsegments are pooled together. The window-size ef-
fect is so powerful that even analysis restricted exclu-
sively to the on subsegment, which is only 37.2 s in
duration, may yield false-negative results. This is most
likely to occur if the response is biphasic (e.g., tachycar-
dia → bradycardia) and each of its components is of
similar magnitude (Figure 4B), in which case, the net
change tends to zero. The window-size effect is a hidden
“statistical hurdle” in the path to meaningful analysis of
biomedical data. This is particularly true of heart activ-
Epilepsia, Vol. 42, No. 8, 2001
FIG. 5. A, B: Two typical 35-s seg-
ments for a subject analyzed in a previ-
ous study (5), one with left vagus nerve
stimulation (LVNS)-off (A) and the other
with LVNS-on (B). The ON segment ex-
hibits a biphasic response at the begin-
ning of stimulation, along with a marked
decrease in HRV as compared with the
OFF segment. C: A comparison of the
HR-to-HRV relationship over a 24-h pe-
riod for LVNS-on (red) versus LVNS-off
(blue). Plotted are the median (solid) and
the 25th and 75th percentiles (dash-dot)
for both conditions. Notice that LVNS
causes a reversal of the normal trend
and a paradoxic decrease in normal
HRV at lower heart rates.
ity, which is subject to large short-term fluctuations and
longitudinal correlation (9) requiring statistical methods
such as those used here to improve resolution of the
mean value and allow objective comparative assessment
of data samples.
2. Use of actual treatment parameters. Conclusions
based on stimulation periods with duration different from
the standard (2,3) may be invalid because heart responses
are time varying and depend on stimulus duration,
among other factors. Also, LVNS effects on IHR are not
limited to the stimulation period; poststimulation tachy-
cardia is transient but may be prominent.
3. Use of subjects as their own controls. The hetero-
geneity of the responses to LVNS found in this study
underscores the importance of this approach and the dan-
gers of data pooling.
4. Use of quantitative measures that are more relevant
for nonstationary systems such as the heart, including the
 NEUROCYBERNETIC PROSTHESIS
continuous tracking of data and the recording of the pre-
cise onset time of an event (in this case, stimulation).
Whereas SD and IQR measures of HRV are common
practice in cardiology (9), they are not useful for detec-
tion of baseline shifts due to state changes. We also
applied nonlinear procedures such as cycle-based median
normalization that improved the resolution of LVNS-
induced changes.
The HR changes associated with the sleep–wake cycle
and degree of physical or emotional activity also were
taken into consideration in our experimental design and
statistical treatment of the data. Although samples of
heart activity can be taken at times when the system’s
nonstationarity is perceived to be “relatively low,” acti-
vation of the NCP increases nonstationarity by virtue of
its direct and indirect (centrally mediated) effects on the
heart and its pattern of activation (ramp-up). Its gradual
and discontinuous (five off gaps during the on cycle; 0.3
s each) modes of stimulation force oscillations in IHR
that are temporally correlated with these activation–
deactivation cycles during the on period (Figure 6, S3).
An interesting question which the chronotropic
changes raise pertains to the site of origin of the decel-
erating influences on the heart. The placement of the
spiral electrode on the cervical trunk, which is exclu-
sively composed of axons, and the laws (11,12) govern-
ing impulse conduction along these fibers, ensure that
these impulses travel bidirectionally, that is, toward the
heart and brain. One possibility is that the chronotropic
effects are caused by the caudal spread of currents from
the site of stimulation, via the cardiac rami (13,14). The
other is that these effects are mediated through central
structures, after their activation by ascending vagal im-
pulses. The rapidity with which cardiac oscillations oc-
cur (see Figure 6) in response to the brief (0.3-s) “OFF”
gaps during stimulation, suggests that peripheral influ-
ences play a role, although the extent remains undeter-
mined.
The occurrence of tachycardia during vagal stimula-
tion in two subjects suggests an increase in sympathetic
(5) activity of sufficient magnitude to overcome the para-
sympathetically mediated bradycardia, underscoring the
complexity of the mechanisms at work. Vagal tachycar-
dia may be due to (a) selective activation of large my-
elinated vagal fibers (15), an unlikely phenomenon,
given the relatively high current intensities used for
stimulation; (b) a mechanism similar to that responsible
for poststimulation tachycardia (4); or (c) complex inter-
actions among nucleus tractus solitarius, with its feed-
forward inhibition mechanisms, dorsal motor nucleus of
the vagus, parabrachial nucleus, and the reticular system
(16–19). The complexity of these interactions is further
compounded in our cases by somatic sensory afferents:
the laryngeal discomfort caused by the NCP currents
may activate the reticular system causing arousal, which
1013
FIG. 6. The effect on instantaneous heart rate (IHR) of the brief
“OFF gaps” that occur during stimulation. The times of stimulation
start and end are annotated with grid lines, as well as the begin-
ning and end of each of the 0.3-s OFF gaps. The oscillatory
behavior of the heart dynamics is directly attributable to these
stimulation gaps; the latency to changes in HR resulting from
these gaps is on the order of five heart cycles.
is associated with increases in heart rate (13,16,19). The
spectra of cardiac responses to LVNS, described in this
study, reflect the interplay of autonomic, visceral, and
somatic sensory afferences and the role of central struc-
tures in their integration.
Patients with intractable epilepsy who responded to
LVNS were not included in this study simply because
none was available to us. Whether the findings described
here also apply to those who respond to this therapy or
receive lower-intensity stimulation is not yet known. In-
vestigation of the effects of LVNS in cardiac function in
a group of “responders” is indicated; in the event that
reproducible differences in cardiac effects between “re-
sponders” and “nonresponders” are identified, these may
serve as predictors of response and may even shed light
on mechanisms of actions and on the causes of therapeu-
tic failure. The possibility that the chronotropic effects
described herein are due to epilepsy (20) is highly un-
likely, because they were temporally and distinctly re-
stricted to the period when the NCP was turned on.
The rare occurrence of LVNS-induced cardiac asys-
tole (21,22) suggests that under certain conditions, which
may include but are not limited to misplacement of the
electrode or variation of the vagus nerve anatomy, this
form of stimulation may cause serious adverse effects.
Our observations point to the importance and need for
more comprehensive studies of cardiac function during
LVNS with the NCP. Specific aims of these studies
should be to identify factors such as stimulation param-
eters, type of epilepsy, site of electrode placement on the
cervical trunk, or conditions such as depressed level of
Epilepsia, Vol. 42, No. 8, 2001
1014 M.G. FREI AND I. OSORIO
consciousness, whether natural or drug-induced, that fa-
cilitate the occurrence or even exaggerate the changes in
IHR and HRV as a function of LVNS.
Acknowledgment: We thank Albert Merati, M.D., for his
valuable comments. This work was supported in part by grants
from Cyberonics, Inc., and by NIH SBIR grant 2R44NS34630-02.
Presented in part at AES Annual meetings (San Francisco,
1996 and San Diego, 1998).
REFERENCES
1. Handforth A, DeGiorgio CM, Schacter SC, et al. Vagus nerve
stimulation therapy for partial onset seizures: a randomized active
control trial. Neurology 1998;51:48–55.
2. Kamath MV, Upton RAM, Talalla A, et al. Neurocardiac responses
to vagoafferent electrostimulation in humans. PACE 1992;15:
1581–7.
3. Setty AB, Vaughn BV, Quint SR, et al. Heart period variability
during vagal nerve stimulation. Seizure 1998;7:213–7.
4. Levy MN, Martin P. Parasympathetic control of the heart. In:
Randall WC, ed. Nervous control of cardiovascular function. New
York: Oxford University Press, 1984:69–94.
5. Armour JA, Hopkins DA. Anatomy of the extrinsic efferent auto-
nomic nerves and ganglia innervating the mammalian heart. In:
Randall WC, ed. Nervous control of cardiovascular function. New
York: Oxford University Press, 1984:20–45.
6. Han P, Frei MG, Osorio I. Probable mechanisms of action of vagus
nerve stimulation in humans with epilepsy: is the heart a window
into the brain [Abstract]? Epilepsia 1996;37(5 suppl):83S.
7. Frei MG, Davidchack RL, Osorio I. Least squares acceleration
filtering for the estimation of signal derivatives and sharpness at
extrema. IEEE Trans Biomed Eng 1999;46:971–7.
8. Lindgren BW. Statistical theory. 3rd ed. New York: Macmillan,
1976:494–5.
9. Task Force of the European Society of Cardiology and the North
American Society of Pacing and Electrophysiology. Special report:
heart rate variability standards of measurement, physiological in-
terpretation, and clinical use. Circulation 1996;93:1043–65.
10. Sayers BMcA, Ruggiero C, Feuerlicht J. Statistical variability of
biomedical data, Part 1: the influence of serial correlation on mean
value estimates. Med Inf 1981;6:1–11.
11. Hodgkin AL. The conduction of the nervous impulse. Liverpool:
Liverpool University Press, 1967:30–55.
12. Eyzaguirre C. Physiology of the nervous system. Chicago: Year-
book Medical Publishers, 1968:29–39.
13. Brodal A. The cranial nerves Springfield: C Thomas, 1957:38–49.
14. Janes RD, Brandys JC, Hopkins DA, et al. Anatomy of human
extrinsic cardiac nerves and ganglia. Am J Cardiol 1986;57:299–
309.
15. Paintal AS. Vagal sensory receptor and their reflex effect. Physiol
Rev 1973;53:159–227.
16. Dell P, Olson R. Projections (secondaires) mesencephaliques, di-
encephaliques et amygdaliennes des afferent viscerales vagales.
CR Soc Biol Paris 1951;145:1088–91.
17. Miura M, Reis DJ. The role of the solitary and paramedian nuclei
in mediating cardiovascular reflex responses from carotid baro-
and chemoreceptors. J Physiol 1972;223:525–48.
18. Saper CB. The central autonomic system. In: Paxinos G, ed. The
rat nervous system. 2nd ed. San Diego: Academic Press,
1995:107–31.
19. Grastya’n E, Hasznos T, Lissak K, et al. Activation of the brain-
stem activating system by vegetative afferents. Acta Physiol Acad
Sci Hungary 1952;3:102–22.
20. Massetani R, Strata G, Galli R, et al. Alteration of cardiac function
in patients with temporal lobe epilepsy: different roles of EEG-
ECG monitoring and spectral analysis of variability. Epilepsia
1997;38:363–9.
21. Asconapé JJ, Moore DD, Zipes DP, et al. Bradycardia and asystole
with the use of vagus nerve stimulation for the treatment of epi-
Epilepsia, Vol. 42, No. 8, 2001
lepsy: a rare complication of intraoperative device testing. Epilep-
sia 1999;40:1452–4.
22. Tatum WO IV, Moore DB, Stecker MM, et al. Ventricular asystole
during vagus nerve stimulation for epilepsy in humans. Neurology
1999;52:1267–9.
APPENDIX 1
Identification of LVNS-on versus
LVNS-off Segments
The following characteristics unique to the NCP were
used to identify the times of left vagus nerve stimulation
(LVNS):
1. NCP activation results in a distinctive artifact (at
∼30 Hz, the programmed stimulation frequency)
that contaminates the ECG signal. This artifact co-
incides with stimulation, beginning every 334.49 s
(the programmed period of the device).
2. Activation and deactivation are gradual and appear
as ramps.
3. The device is equipped with a timing crystal that
has a cycle of 6.826 s. At the end of each crystal
cycle, stimulation ceases briefly (for ∼0.3 s), and
with it, the artifact.
These properties make the stimulation artifact distin-
guishable from other sources of signal noise (see Fig.
1A).
The following steps, in the order listed, were taken to
localize each onset of stimulation with a resulting esti-
mated precision of 0.1 s.
1. Spectral analysis of each recording was performed
to localize temporally the 30-Hz stimulation arti-
fact and daily device resetting. The trade-off be-
tween time and frequency localization inherent to
window-based Fourier analysis limited temporal
localization of stimulation onset to ∼2 s.
2. Visual analysis of samples of the raw signal was
then performed in each recording as a strategy to
improve temporal accuracy. Samples in which the
stimulation artifact was most clearly visible (i.e.,
not obscured by other sources of noise) were se-
lected for this purpose, and a few such samples
were chosen near the beginning and the end of each
recording. Each sample was initially viewed using
a 50-s window that contained the entire on interval.
The window size was then reduced to ∼2 s, and the
onset, offset, and one or more of the brief “gaps” in
stimulation were identified and marked. Because
the stimulus artifact at onset (at the beginning the
ramp-up) is of very small amplitude and may not be
visible unless the signal-to-noise ratio is optimal,
we developed a nonlinear filtering algorithm and
applied it to the data to improve precision in local-
izing time of onset. This algorithm behaves as a
 NEUROCYBERNETIC PROSTHESIS 1015

“band-pass” filter, enhancing the stimulus artifact
and reducing other unwanted noise.
3. Linear regression analysis was used to improve
precision of localization further by attaining the
best fit for on times, taking advantage of the pro-
grammed periodicity of the device. the stimulus
artifact at onset (at the beginning the ramp-up) is of
very small amplitude and may not be visible unless
the signal-to-noise ratio is optimal, we developed a
nonlinear filtering algorithm and applied it to the
data to improve precision in localizing time of on-
set. This algorithm behaves as a “band-pass” filter,
enhancing the stimulus artifact and reducing other
unwanted noise.
3) Linear regression analysis was used to improve
precision of localization further by attaining the
best fit for on times, taking advantage of the pro-
grammed periodicity of the device.
APPENDIX 2
Statistical Results
The following tables give the fraction of long record-
ings that showed a significant difference between LVNS-
on and LVNS-off at the indicated ␣ levels. Each column
shows the signal derived from the “time of beat” se-
quence, and each row corresponds to the measure applied
to that signal. Cumulative distribution functions were
then obtained for ON and OFF segments and the two-
sample Kolmogorov–Smirnov test was applied to test the
hypothesis of identical distributions.

Cumulative results for all subjects: 25 recordings, 2,411 ON, 19,283 OFF
␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 0.24/0.2/0.2 0.4/0.24/0.2 0.28/0.2/0.2 0.48/0.32/0.2 0.56/0.32/0.2 0.76/0.68/0.48 0.8/0.72/0.56 0.36/0.32/0.24
Median 0.28/0.2/0.2 0.32/0.28/0.12 0.32/0.2/0.2 0.4/0.36/0.16 0.56/0.36/0.32 0.8/0.6/0.44 0.8/0.6/0.44 0.44/0.28/0.2
Std 0.56/0.48/0.4 0.36/0.32/0.24 0.6/0.56/0.4 0.64/0.48/0.4 0.76/0.64/0.52 0.52/0.48/0.4 0.56/0.52/0.4 0.48/0.32/0.28
Iqr 0.56/0.44/0.28 0.52/0.28/0.24 0.64/0.56/0.44 0.6/0.44/0.36 0.72/0.56/0.44 0.6/0.52/0.4 0.64/0.6/0.4 0.44/0.32/0.16

S1, six recordings, 453 ON, 3,622 OFF segments

␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 0/0/0 0/0/0 0/0/0 0.5/0.33/0.17 0.67/0.33/0.33 0.5/0.5/0.17 0.5/0.5/0.33 0.17/0/0
Median 0/0/0 0.33/0.33/0.17 0/0/0 0.17/0.17/0 0.5/0.33/0.17 0.5/0.17/0 0.5/0.17/0 0.17/0/0
Std 0.17/0/0 0.17/0/0 0.17/0.17/0 0.5/0.33/0.33 0.67/0.67/0.5 0.17/0/0 0.17/0/0 0.33/0/0
Iqr 0.17/0.17/0 0.17/0.17/0.17 0.33/0.17/0 0.5/0.5/0.33 0.83/0.83/0.67 0.33/0.17/0 0.33/0.17/0 0.5/0.5/0

S2, one recording, 79 ON, 632 OFF

␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 0/0/0 1/0/0 0/0/0 1/0/0 1/0/0 1/0/0 1/1/1 1/1/1
Median 0/0/0 0/0/0 0/0/0 0/0/0 1/0/0 1/0/1 1/1/1 1/1/1
Std 0/0/0 1/1/1 0/0/0 1/1/1 1/1/1 0/1/0 0/0/0 1/1/1
Iqr 0/0/0 1/1/1 0/0/0 1/1/1 1/1/1 0/1/0 0/0/0 1/1/1

S3(a), three recordings, 249 ON, 1,992 OFF

␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 1/1/1 0.67/0.67/0.67 1/1/1 1/1/1 1/1/0.67 1/1/1 1/1/1 1/1/1
Median 1/1/1 0.67/0.67/0.67 1/1/1 1/1/0.67 0.67/0.67/0.67 1/1/1 1/1/1 1/1/1
Std 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1
Iqr 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1 1/1/1

Epilepsia, Vol. 42, No. 8, 2001

1016 M.G. FREI AND I. OSORIO

S3(b), six recordings, 396 ON, 3,165 OFF

␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 0/0/0 0.33/0/0 0/0/0 0.33/0.17/0 0.67/0.33/0 0.83/0.5/0.33 0.83/0.67/0.33 0.17/0.17/0.17
Median 0/0/0 0.17/0.17/0 0.17/0/0 0.5/0.33/0 0.83/0.33/0.33 0.83/0.5/0.17 0.83/0.5/0.17 0.33/0.33/0
Std 0.33/0.17/0 0.17/0.17/0 0.33/0.17/0 0.67/0.67/0.5 0.83/0.67/0.5 0.17/0.17/0 0.17/0.17/0 0.17/0.17/0
Iqr 0.33/0.17/0 0.33/0.17/0 0.33/0.33/0.17 0.67/0.5/0.33 0.67/0.5/0.33 0.33/0.17/0.17 0.33/0.33/0.17 0.17/0/0

S4, six recordings, 467 ON, 3,736 OFF

␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 0.17/0/0 0.33/0.17/0 0.33/0/0 0/0/0 0/0/0 0.67/0.67/0.5 0.83/0.67/0.5 0/0/0
Median 0.33/0/0 0.17/0.17/0 0.33/0/0 0/0/0 0/0/0 0.83/0.67/0.5 0.83/0.67/0.5 0.33/0/0
Std 0.83/0.83/0.67 0/0/0 1/1/0.67 0.5/0.33/0.17 0.83/0.5/0.33 0.83/0.83/0.67 1/1/0.67 0.33/0/0
Iqr 0.83/0.5/0.33 0.67/0/0 1/1/0.83 0.17/0/0 0.67/0.17/0 0.83/0.83/0.5 1/1/0.67 0.33/0/0

S5, three recordings, 767 ON, 6,136 OFF

␣ level Median Median
0.05/0.01/0.001 Raw RR Raw dRR Raw IHR HF RR HF IHR Normalized RR Normalized IHR |dRR|
Mean 0.67/0.67/0.67 1/1/1 0.67/0.67/0.67 1/0.67/0.33 0.67/0.33/0.33 1/1/1 1/1/1 1/1/0.33
Median 0.67/0.67/0.67 0.67/0.33/0 0.67/0.67/0.67 1/1/0.67 1/1/1 1/1/1 1/1/1 0.67/0.33/0.33
Std 1/1/1 1/1/0.67 1/1/1 0.67/0/0 0.33/0.33/0.33 1/1/1 1/1/1 1/1/1
Iqr 1/1/0.67 0.67/0.33/0.33 1/0.67/0.67 1/0.33/0.33 0.33/0.33/0.33 1/1/1 1/1/0.67 0.33/0.33/0

Epilepsia, Vol. 42, No. 8, 2001