1

I42 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38, NO. 2. FEBRUARY 1991
Parametric Classification of Segments in Ocular
Nystagmus
Claudio G. Rey and Henrietta L. Galiana, Member, IEEE
Abstract-A new method for nystagmus classification, using system
identification techniques, is presented. We formulate a system whose
input is head position and whose output is eye position. We approxi-
mate this system with an autoregressive with exogenous input (ARX)
model which relates the input and output (transfer function) regardless
of the temporal profile for the sensory stimulation. The system is then
identified using a least squares criteria and three indicators are pro-
duced. From these a Rag is produced that marks slow and fast phases
as well as blinks and bad data segments. Tests with simulated and real
data are presented and indicate that the segment classification is re-
markably insensitive to recording noise and that it is more robust than
previous techniques. Operator intervention is minimal. We expect the
method to be applicable for all types of ocular nystagmus. Here, how-
ever, we illustrate our results only in the context of the vestibuloocular
reRex (VOR). A discussion explains how this method can also be ap-
plied for optokinetic (OKN) or pursuit nystagmus.
I. INTRODUCTION
HIS paper addresses the problem of automated analysis of
T ocular nystagmus. Computer signal processing is essential
to handle large volumes of clinical data and to provide accurate
databases for the development of oculomotor control models.
Reflexive eye movements can be elicited by various sensory
means. They occur, for example, during pursuit of a moving
visual target, or during stabilization of the eye in space during
head movement. I n most cases, ocular responses consist of in-
termingled segments classified as slow or fast, according
to their average speed characteristics, so that a time record has
a sawtooth-like pattern called ocular nystagmus. Fig. 1 illus-
trates the nystagmus pattem generated by passive head rotation
of a human subject in the dark: the vestibuloocular reflex
(VOR). The behavioral basis for the naming of segments is
clear, whereby slow phases normally appear to have lower ve-
locities and longer durations than the fast phases. In the case of
visual stimulation, the components of the nystagmus might be
called slow-phase pursuit and saccadic, but the neural sub-
strates in all types of nystagmus are quite similar, and are often
shared by all responses. In any case, it is clear that the nomen-
clature is not always an accurate description of response seg-
ments, and has only been preserved in the literature for historical
reasons. For example, one can often observe vestibular fast
phases with peak velocities lower than nearby slow phases.
Thus, it is not surprising that previous analysis schemes which
rely on eye velocity thresholds can be prone to errors of seg-
ment classification.
Manuscript received September 5 , 1989; revised J une 19. 1990. This
work was supported by a grant fromthe National Science and Engineering
Research Council (NSERC) of Canada and the Fonds de Recherches en
SantC du QuCbec (FRSQ). C. G. Rey had an NSERC scholarship.
The authors are with the Biomedical Engineering Unit. Faculty of Med-
icine, McGill University, Montreal, P.Q. H3A 2B4, Canada.
IEEE Log Number 9041297.
EYE POSITION
20
0
(a) -20
-40
-BO
5
Fast pha e s Slow p h as es
EYE VELOCITY
500 , I
I
14 5 18.5
-500
10 5
HEAD VELOCI TY
200 I I
I I
I4 5 18 5
-200
10 5
Ti me (3)
Fig. I . Sample record of the vebtibulo-ocular reflex (VOR) (a) eye posi-
tion i n degrees, (b) eye velocity i n deg/s, (c) head velocity deg/s.
Rather than rely on velocity thresholds, we use another as-
pect of ocular reflexes: the two components of nystagmus ap-
pear to satisfy different cost functions or control strategies. The
slow phases reduce velocity error between the stimulus and the
eye, while the fast phases (or saccades) reduce position error.
In the case of the VOR, this means that slow-phase eye velocity
is nearly equal but opposite in sign to head velocity, while the
endpoint of fast phases is related to head velocity. Alterna-
tively, in the case of pursuit eye velocity attempts to match tar-
get velocity during slow phases, while position error is reduced
to nearly zero by each saccade. Similar arguments hold for all
types of ocular responses. This type of classification for the
segments of ocular nystagmus does not restrict either the tem-
poral characteristics (e.g., maximum speed) or direction of any
segment with respect to the stimulus; only the dynamic relation
with the input (transfer function) is expected to be consistent.
Thus, we propose to classify segments of nystagmus automat-
ically by examining parameters generated when specific auto-
regressive with exogenous input (ARX) models are fit to the
records. The method includes a nonlinear detector which can
mark the transition point between segments accurately to within
one sample instant, even with data low-pass filtered to 20 Hz.
11. NYSTAGMUS SEGMENT DETECTION
Several approaches have been developed for the computer-
ized analysis of ocular nystagmus [1]-[SI, [8]-[lo]. None of
these uses a system identification approach. As it will be de-
scribed below, we use three indicators which allow us to ex-
amine a nystagmus pattern, regardless of the temporal profile
0018-9294/91/0200-0142$01.00 0 1991 IEEE
REY AND GALI ANA: SEGMENT CLASSIFICATION IN NYSTAGMUS 143
used for the sensory stimulation. We expect the method to be
applicable for all types of ocular nystagmus. Here, however,
we illustrate our results only in the context of the VOR (in the
dark). The discussion explores modifications for the study of
optokinetic (OKN) or pursuit nystagmus.
Previous Approaches
Computer algorithms have been proposed for the automated
analysis of nystagmus. The analysis must solve two problems:
first, it must detect transitions between nystagmus segments,
i.e., a change in response mode; and second, it must assign
a label to each segment according to its mode: such as slow-
phase, fast-phase or bad data (e.g., blinks). Previous tech-
niques almost always rely on filtered estimates of eye velocity
in order to detect fast phases, i.e., when some velocity thresh-
old is exceeded. I n some cases, the velocity threshold is adap-
tive to recognize the fact that peak fast phase velocities can
increase with concurrent slow-phase velocity [ I ] , [ 2 ] , [4], [SI.
The one exception is J uhola [9], who first processes the eye
position record with a digital filter to find extrema (local max-
ima and minima). The segments are then classified according to
their duration: slow phases are always assumed to be of longer
length than fast phases. Of course, this assumption cannot hold
in the case of general nystagmus alternating in direction, es-
pecially with strong stimuli producing very high fast phase rates.
All of these methods assume that slow and fast phase seg-
ments must alternate and be of opposite direction. They cannot
detect fast phases of low velocity which can occur during low
speed head rotation. In addition, the expected profile in slow-
phase velocity may also be predetermined (e.g., [ 11, [SI), so
that variable responses encountered in patients are difficult to
handle. None of the methods can detect back-to-back fast phases
or saccades, or essentially zero velocity saccades as may occur
when visual errors cancel or overwhelm the VOR. As a con-
sequence, mislabeling of a given segment can cause errors to
persist for several following periods unless operator interaction
is used post-analysis. In addition, methods based on signal dif-
ferentiation will exhibit high noise levels and require heavy
smoothing. This alone will attenuate small eye movements and
obscure the exact timing of phase transitions.
Model-Based Processing
We propose to filter nystagmus records with ARX models
whose order can represent approximately both slow and fast
phase dynamics with respect to the stimulus. The parameters of
the model obtained by regression over a moving window can
then be used to classify the VOR response (see Table I V). Note
that this approach does not preset the overall form of slow-phase
velocities, and can detect phase transitions without extrema in
the velocity record (e.g., saddle points). The algorithm for seg-
ment detection should
allow for variable stimulus-response profiles, such as dur-
ing stochastic inputs;
detect fast phases, regardless of their velocities or direc-
tion with respect to the stimulus, with the possibility of back-
to-back fast phases;
detect data segment containing artifacts (not slow or fast
phases), e.g., eye blinks in electrooculograms (EOG) (see Fig.
6) or extraneous noise spikes;
be relatively insensitive to instrumentation noise in the
nystagmus record,
preserve the timing (edges) of phase transitions;
require little if any operator intervention to set up the anal-
ysis.
Our final goal is to pool nystagmus segments of a given mode
and treat them as a collection of transient responses to the forc-
ing stimulus. The intention of our research is to study the VOR
from a transient analysis viewpoint so we do not intend to re-
construct cumulative slow-phase position or a continuous
profile of slow-phase eye velocity. The transient segments will
be used in a more detailed system identification study which
will explore the dynamics, linearity and possible interactions
between nystagmus modes [ 7] , [8]. However, if so desired the
slow phase data could be interpolated to obtain a continuous
record.
111. ARX MODELS USED
VOR Models
Nystagmus analysis is described here in the context of the
VOR, so the first step is to outline expected dynamics during
vestibular nystagmus. A model for the VOR system, in the dark,
is presented in Fig. 2 . Note that the transfer functions are given
in the Laplace domain since reflex studies are normally consid-
ered in this context. In this paper, however, we will use equiv-
alent approximations in the z-domain to derive our nystagmus
filters.
From observation of vestibular nystagmus and concurrent ac-
tivities in central premotor pathways (see review in [6]), it is
clear that slow and fast control systems can be approximated by
two switching linear systems. The simplest overall transfer
functions from head to eye in Fig. 2 can be described as auto-
regressive with exogenous input ( ARX) models:
In the slow-phase mode:
E,, - E,, - I =r, , , , , (H,, - H,, - I) +k,l,,WH,, +g,,<,w E,, +e,,
(l a)
In the fast-phase mode:
E,, - E,, - I =r, a\ , (H,, - H,, ~ 1) +k,,,,, H,, +S,-,,,E,I +e,,
(1b)
where E,, and H,, are eye and head posirion samples, e,, is the
noise sequence. The noise sequence is assumed to be additive
white and uncorrelated with the head position. See Section VI
for a discussion on the noise sequence.
Note that this formulation, unlike its steady-state Laplace
equivalent, allows for transient, i.e., initial conditions are im-
plicit in the formulation. Al so, continuity is assured by the ARX
model even when switching takes place. It is for these reasons
that we have chosen an ARX formulation. The timing for
switching between modes of nystagmus and their duration is an
open research question which we will investigate with our al-
gorithm.
Reduced A R X Model
Eventually we wish to identify the systems described by (1)
and relate them to various model structures proposed in the lit-
erature. In order to achieve this goal we must detect where the
transitions between modes occur and sort the nystagmus seg-
ments for later pooling and global processing. For detection
purposes, we have discovered that a simpler ARX model is suf-
ficient. In fact, for head velocity profiles in the canal passband
I
I44 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2. FEBRUARY 1991
Physiology
canal s
Brai nst em
Syst em Trans f er Functi on
1
Fig. 2. Physiological and mathematical representation of the VOR. Sub-
scripts s andfrefer to the slow and fast phases of nystagmus, respectively:
subscript c for the semicircular canals, and e for the eye plant. Note that
the input here is assumed to be head velocity, and the response nystagmic
eye position. Hence the canals are approximately high-pass filters of head
velocity, while theeye plant low-pass filters motoneural signals. The over-
all transfer function is then of the general form El k =G s l [ ( T, s +
I)(Ts +I)] where T is very large during slow phases (e.g.. 12 s) , and
much smaller during fast phases (say 60 ms). Similarly, G is negative
during slow phases and normally positive during fast phases.
(frequencies >l / T r , see Fig. 2 caption), we can postulate a
simple first-order model for both modes within a window of the
data:
E,, - E, , - , =r( H, , - H R - , ) +e, , ;
nl I n I nz
or
AE,, =rAH, +e, ,
(2)
where [nl , n2] is the local window where the ARX parameter r
is estimated.
Iv. DESCRIPTION OF THE ALGORI THM
Estimating the ARX Paramet er
The parameter r may be obtained by the regression formula
r =@E,, * AH, J/ ( AH, , * AH, , ) (3)
where * denotes the vector inner product: AE,, is the vector
[ AE, , - ,,., . . , AE, , JT; AH,, is the vector [ AH, , ~ ,,., . . . , AH, , ] ;
w is the window size in sample numbers (60 ms in our tests).
In practice, (3) has been modified to avoid ill-conditioning
when both eye and head velocity may be near zero [O/O con-
dition in Eq. (2)]: that is, (3) uses
AE,, =AE,, - 14 At , sign ( AH, , ) ;
AH,, =AH,, +20 At,, sign ( AH, , )
(4)
where At , is the sampling period. The constants used above im-
pose a value for r of 0.7 for a slow-phase segment near the 0/0
condition.
By observing the parameter r given by the ARX local fit [(3)
and (4)], we can determine whether a segment is in slow or fast
phase, according to some preset bounds. If the parameter ex-
ceeds these bounds, it can be used to mark the occurrence of
signal artifacts such as blinks or transient noise. In summary,
the ARX model is applied within a small moving window (say
60 ms wide); the resulting parameter r and mean squared error
(MSR) for the fit are extracted as output signals called the r and
MSR indicators (see Fig. 3).
Given the r indicator alone, one can immediately classify
nystagmus segments into tentative slow, fast (directed
opposite to head velocity) and reverse fast phases (in same
direction as head velocity) (see Fig. 4). This is summarized in
Table I , where typical ranges for rare illustrated.
There are two difficulties with using the r indicator alone in
determining the phase type. First, spurious noise on the nystag-
mus trace during zero-velocity crossovers can cause erroneous
detection of fast phases or the labelling of bad data (blinks)
as fast phases. Second, since the ARX model fit is essentially
a low-pass filtering process, the actual time of phase-transition
is uncertain. It could be located anywhere in a interval of length
equal to the ARX window (e.g., 60 ms), where the r indicator
progressively changes from its values for each phase. For these
reasons, two other indicators have been developed to comple-
ment the r signal: these are the MSR of the fit and a special
nonlinear transition detector described below.
Using the MSR
The MSR or mean squared error of the ARX fit in a small
window of the signal is an indication of the presence of noise
in the signal, and/or of the presence of modeling error within
this window. We are interested in the latter. Near the points
where a transition from one phase to the other occurs, the MSR
will have a peak. Thus, it can help in confirming transitions
from one phase to another. Unfortunately, MSR peaks are not
sharp as the MSR is not a point measure of error but a neigh-
borhood measure. Thus, the MSR record is not appropriate for
localizing phase transitions. On the other hand, while the r in-
dicator may become noisy when eye velocity is near zero, the
MSR is not affected. Therefore any apparent change in phase
flagged by r can be cross-referenced with the MSR level for
confirmation. If the MSR is low, the apparent change of phase
was spurious and can be discarded. This lends robustness to the
method. The MSR can also be used to determned the location
of blinks or bad data since such data will not likely be de-
scribed by an ARX process. Thus, if the MSR level rises be-
yond values expected during normal nystagmus (see Table 111)
bad data is expected and is flagged as such.
Using the Transition Indicator
Neither the r indicator nor the MSR can be used to localize
phase transitions. This is a problem is we intend to measure the
length and timing of fast phases automatically in a study of nys-
tagmus strategies. For this reason we have developed a new
nonlinear detector of phase transitions.
The transition detector is the ratio of two local MSRs. One
is the MSR for the local ARX model described above (MSRI ).
The other (MSR2) is derived from a model which assumes the
two different ARX representations are valid for each half of the
local window. This second representation can be described by
the equations below:
E,, - E,,- I =r l ( H, , - I); n1 <n <no (5a)
E,, - E, , - =r2(H, , - H,J -l); no <n <n2 (5b)
where In,, no] and [no, n2] are the two halves of the local win-
145
I .
0.
REY AND GALI ANA: SEGMENT CLASSIFICATION I N NYSTAGMUS
EYE POSITION
p HI
I ,
r INDICATOR
MSR I NOI TATOR .... ~~ ....
12 .................................... ..I .............................. I
T H M s R ~ ~ . \ THMs R~,
(c) 06 ........... ....... i ................... i h. ........... ............ ..... 1
0
RATI O INDICATOR
3
(d)
22 23 24
Tl mC
Fig. 3 . Profiles of indicators used in the processing of nystagmus (a) eye
position, (b) r indicator fromARX regressive fi t, (c) MSR indicator, (d)
Ratio (transition) indicator. The thresholds applied fromTable I1are also
indicated.
EYE POSITION (deg)
lo. 1
-10.
;: I
r lNDICATOR
I I
i MSR I NDI CA ~OR
*TH I
0.4
l a1 Mar k er
0.2
+THMSRLO ,
0.
I I
1 PATIO I N D I C ~T ~R
SLOW SLOW
Step 1
St e p2 I j
Time ( 5 )
Fig. 4. Detection with simulated nystagmus. Eye position and the three
indicators are illustrated as i n Fig. 3 . Below., the phases estimated using
the r indicator alone are shown and the three steps used i n improving the
phase estimation: Step 1: move back half a window length; step 2: move
forward until the lower MSR threshold is exceeded (first marker); step 3:
move forward (for at the most half a window) to the maximum of the
Ratio indicator that exceeds the threshold (second marker). At the bottom
the resulting condition flag is displayed: see Table I V.
TABLE 1
CLASSI FI CATI ON WI TH THE r I NDI CATOR
APPROXI MATE RANGES FOR THE I Nl l - l AL
Phase type Approximate range for r
Slow phase [ - I , 01
Fast phase >0
Reverse fast phase < - I
dow [n,, n2]. Also note that the first MSR (MSR1) is for a con-
tinuous ARX applied to the same window.
The output ratio is assigned to the sample no:
Ratio ,,,, =MSRI,,,,/ MSR2,,,,. ( 4)
Small values of MSR2 can produce false peaks on the Ratio
output, thus we have heuristically given a floor value of MSR2
of half its mean value. The role of the Ratio indicator is that of
finding the exact location of the phase transitions based on the
estimate of this transition using the r and indicators. The next
section will explain how this is done.
Processing of Nystagmus
The detection algorithm processes three indicators: r, MSRI ,
and Ratio, and uses five thresholds (see Table 11) to determine
the state at every point of nystagmus. As a first step, the algo-
rithm looks at the three indicators and assigns a tentative state
of each point, as outlined in Table 111. Note that all of the states
form a mutually exclusive set of logical states; i.e., one and
only one is true at any point.
The algorithm then revises the flag transitions. Consider, for
example, an initial state transition from slow to fasr (or reverse
fast: Rfast). The transition detected from the r indicator is only
approximate. A two-step procedure is used to detect the exact
location of the phase transition from the Ratio indicator:
1) Start half a window before the estimated transition and
search forward until the MSRI is higher than a preset threshold
THMSRLo: set aj r s t marker around this point for the beginning
of the fast phase. See steps 1 and 2 in Fig. 4.
2) Search within half a window size after the first marker for
the local maximum of the Ratio indicator. If this local maxi-
mum exceeds a preset threshold, THRat,,, the true transition
point has been detected and a second marker is set of mark its
position. See step 3 in Fig. 4.
The first and second markers have a different significance.
Samples immediately prior to the first marker are known to be
in Slow phase. The segment between the first marker and the
second marker is also expected to be in Slow phase. If a second
marker cannot be found the exact position of the transition re-
mains unknown. This will be discussed further below.
When the state changes again (from Fast phase) a similar pro-
cedure is used to search backwards for the first and second
markers, fixing the transition point from Fast to Slow. After a
nonSlow segment is seen to end (Fast, Rfast, Bad) and markers
have been set at the beginning and the end of this segment, the
classification of the segment is reexamined as follows:
1) If there are noj rsr markers set around the nonslow seg-
ment, the segment is rejected (assumed to be Slow phase).
2 ) If there is a missing second marker, at the beginning and/
or at the end of the nonslow segment, it is marked as unknown.
3 ) If both second markers around the nonslow segment have
been set, then the segment is classified as Fast or Rfast, ac-
cording to the values of the r indicator within the segment.
146 I EEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38, NO. 2 . FEBRUARY 1991
TABLE I1
LIS1OF THRESHOLDS APPLI ED TO THE INDICATORS (SEE ALSO Fig. 3)
Threshold Significance
THMSKI c,
THMsR,,, MaximumMSR permitted anywhere
TH <,
TH,,
TH~at,c
Minimum MSR allowed during a phase transition
Lower limit for the r parameter during slow segments
Upper limit for the r parameter during slow segments
Minimumratio allowed during a phase transition
TABLE 111
RANGES FOR DI FFERENT BOOLEAN VARIABLES (INITIAL FLAG STATES)
Boolean Variable Threshold Range
Bad MSR >THwR,,
Slow
Fast
Rfast
TH,,, 5 r 5 TH,, and MSR 5 THMsR,,
r >TH,, and MSR 5 THMsRH,
TH,,,, <r and MSR 5 TH,,,,,
TABLE IV
FLAG LEVELS PRODUCED BY THE ALGORI THM TO
ENCODE NYSTAGMUS SEGMENTS
Condition or Phase Flag Output
Slow Phase 0
Fast Phase 1
Reverse Fast Phase - I
Blink or Artifact Zsign(r)
Unknown Phase sign(r)/2
As summarized in Table I V, the algorithm detectors encode
five ( 5 ) possible states in the form of a flag for the nystagmus
segments. Note that the sign of the parameter r is passed on as
a reference in the case of blinks or artifacts.
V. RESULTS
The algorithm was implemented in Fortran on DEC Vax-sta-
tions, and results were examined graphically using software
provided by the GKS library. Processing was done off-line, so
no attempt was made to optimize the code at this time. In any
case, a 30 s record of nystagmus currently requires less than
one minute for full automatic segment detection. All examples
below used similar algorithmic thresholds: window of ARX fit
=60 ms, THRdf10 =2 , THMSR,~=0.025, T HMSRHl =0.6,
TH,, =- 1 and TH,,, =-0.3.
Simuluted VOR
An example of VOR nystagmus generated by a bilateral model
of the VOR [8] is first presented in Fig. 5(a), together with the
simulation flag [Fig. 5(c)] marking the true slow or fast phase
states. The eye position was corrupted with white noise of stan-
dard deviation 0.5 deg. Higher noise levels would be easily
handled by suitable widening of the ARX window.
Note the close close correspondence of true and algorithm
flags [Fig. 5(a) and (c)]. The only difference is the indication
of a suspected reverse fast phase, corresponding to a kink
in the eye position trajectory. This could in fact be considered
Condition f l ag 1:
1 1 EAST PHASE FLAG I
LO 15 20
Ti me i s 1
Fig. 5 . Testing phase detection with simulated nystagmus (a) eye position,
(b) detection flag generated by the algorithm, (c) fast phase Ragused i n the
simulation, The ARX fit was applied over a moving window of 60 ms.
a small fast phase in a real nystagmus record without a priori
knowledge of the state of the brainstem.
Clinical Data
The nystagmus examples presented were recorded using bi-
temporal Ag-AgC1 electrodes (electrooculogram EOG). Fig. 1,
3, and 6 display the EOG records from a normal subject and
Fig. 7 displays the EOG record from a unilateral vestibular pa-
tient. Electrode drift was reduced through periodic calibrations.
The EOG records were low-pass filtered using a four-pole Bes-
se1 analog filter with a cutoff frequency of 22 Hz and sampled
at 500 Hz with 12 b quantization. The result was low-pass fil-
tered digitally to 20 Hz and decimated to 100 Hz. Records with
high nystagmus activity, and corrupting blinks, were chosen in-
tentionally to illustrate the robustness of the algorithm. Without
any human intervention post-processing, the segments ex-
tracted as putative slow phases show no corruption by missed
fast phases or noise. Note the correct classification according
to fast phase direction, and the presence of blinks. Also phase
transitions are located correctly, despite distortions caused by
the low-pass filtering to 20 Hz. In order to detect saccades or
quick phases of less than 15 ms duration, the sampling rate and
the low-pass filter cutoff frequency should be extended to at
least 150 and 30 Hz, respectively.
VI . DISCUSSION
Choosing thresholds
Good performance of the algorithm hinges on a proper selec-
tion of the thresholds used for the MSR, r, and Ratio indicators.
The results above used a window size of 60 ms (or six points
at this sampling rate). We have found that a single set of thresh-
olds can be used to analyze all nystagmus records processed by
identical analog filters and recorded at the same sampling rate
on the computer. In the case of normal stationary responses, we
have found the algorithm quite insensitive to small changes in
the thresholds, so long as a few rules of thumb are followed:
The [TH,.,,, TH,.,,] interval should contain a range of val-
ues corresponding only to slow phases, by inspection of a nys-
tagmus extract.
REY AND GALIANA: SEGMENT CLASSIFICATION I N NYSTAGMUS 147
Fig.
as in
low,
120.
I
1 sec
(b)
Sl ow Phases
I /I
I I1 / I
I1 I 1
6. Segment detection from VOR nystagmus in a normal subject (data
Fig. I ). (a) Eye position with recovered slow phase eye position be-
(b) segment detection flag, (c) recovered slow phase eye velocity.
-.. . . . ! . U*... ._ U - . ..... .
,
I sec
I
I s ec
11
11
I/
11
I1
Fig. 7. Segment detection fromVOR nystagmus in a unilateral vestibular
patient. (a) Eye position with recovered slow phase eye position below; (b)
eye velocity with recovered slow phase eye velocity below; (c) segment
detection flag. The segment detection flag levels are coded as in Fig. 6.
THMSRLo should also mark values for the MSR correspond-
ing only to (Slow, Fast, Rfast) segments without any phase tran-
sition.
THMSRHI should be higher than values for the MSR corre-
sponding transitions between Slow and Fast phases and should
be lower than MSR values corresponding to blinks.
THRario should be low enough to mark all values for the
Ratio indicator corresponding to phase transitions.
Of course, in general, we expect that choosing thresholds may
require operator intervention for certain patients, if it is in-
tended to process particularly nonlinear, or time-varying re-
sponses.
Assumptions on the Noise
Our assumption that the noise sequence e, is white can be
made given that: 1) low frequency noise from the electrodes
(drift) has been reduced through periodic calibration. 2 ) EMG
interference was avoided by using a helmet to fixate the head
instead of a bite bar, by telling the subject to be relaxed and by
using a low pass filter on the data to 20 Hz. Note that the pro-
posed extension of this filter to 30 Hz should still be acceptable.
Future Research
We have addressed the problem of nystagmus analysis during
midfrequency head oscillations. Here a simple static model was
sufficient in the ARX representation for phase detection. For
wideband (or stochastic) oscillations, the ARX model should
be extended to at least second-order to properly represent the
dynamics expected from Fig. 2. Thus, nystagmus analysis using
model-based regression should be possible for any head veloc-
ity profiles, as long as a sufficiently accurate parametric model
is used.
The same arguments could be applied to other types of nys-
tagmus, such as can be generated by optokinetic or tracking
stimuli. One need only postulate a reasonable model approxi-
mation for the reflex being studied and substitute it for the for-
mulation in ( 2 ) . This is the object of continuing studies in our
laboratory. In fact, we eventually hope to combine accurate sys-
tem identification of a reflex with parallel segment classifica-
tion: the two aspects of the problem are complementary.
CONCLUSI ON
We have presented an automated algorithm for the classifi-
cation of segments in ocular nystagmus. The algorithm is new
in that it uses ARX model parameterization to classify response
segments. It is accurate and robust through the use of cross-
referencing with multiple indicators. At the moment, it is only
for off-line analysis of nystagmus, so no attempt has yet been
made to optimize the code. However, the algorithm can easily
be programmed for on-line analysis because of the local nature
of the indicators. Furthermore, its potential to handle nonlin-
earities, noise spikes and blinks make it ideal for clinical ap-
plications.
We intend to continue refining the algorithm properties in
order to apply them to modelling studies of the oculomotor sys-
tem for all types of nystagmus. Correct detection of transitions
in ocular nystagmus is particularly important for the study of
central switching strategies which control the choice of re-
sponse mode during nystagmus. Moreover, the algorithm may
provide new automatic diagnostic tools for the evaluation of
sensory-motor processing in normals and patients.
REFERENCES
[l ] J . H. J . Allum, J. R. Tole and A. D. Weiss, MITNYS-11, A
digital program for on-line analysis of nystagmus, IEEE Trans.
Biomed. Eng., vol. BME-22, pp. 196-202, 1975.
[2] M. Azi and M. Magnin, A fuzzy set theoretical approach to
automatic analysis of nystagmic eye movements, IEEE Trans.
Biomed Eng., vol. 36, no. 9, pp. 954-963, 1989.
[3] E. Anzaldi and E. Mira, An interactive program for the analysis
of ENG tracings, Acta Ofolaryngol., vol. 80, pp. 120-127,
1975.
[4] R. W. Baloh, L. Langhofer, V. Honrubia, and R. D. Yee,
On-line analysis of eye movements using a digital computer,
Aviat. Space Environ. Med. , vol. 51, no. 6, pp. 563-567, 1980.
[5] G. R. Barnes, A procedure for the analysis of nystagmus and
other eye movements, Aviar. Space Environ. Med. , vol. 53, pp.
[6] H. L. Galiana, A bilateral model for central neural pathways in
the vestibule-ocular reflex, J . Neurophysiol., vol. 51, pp. 210-
241, 1984.
676-682, 1982.
148 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38, NO. 2, FEBRUARY 1991
[7] H. L. Galiana, Estimating VOR dynamics fromocular nystag- Henrietta L. Galiana (M87) received the
mus, inProc. XI CMBEC, Kingston, Ont , Canada, 1985, pp. B.Eng. (Electrical Honours) in 1966, followed
1-2. by the M.Eng. (EE, Biomedical) in 1968, from
[8] H. L. Galiana, The implications of structural modulation in the McGill University. After three years as a Re-
vestibulo-ocular reflex, in Proc. IEEE Montech 87, Conf search Staff Engineer at the Man-Vehicle Lab-
Biomed. Technol., Montreal, 1987. oratory of M.I.T., and several years abroad,
[9] M J uhola, Detection of nystagmus eye movements using a re- she returned to graduate research in 1977, and
cursive digital filter, IEEE Trans. Biomed. Eng. , vol. BME-35, received the Ph.D. in Electncal Engineering
pp. 389-395, 1988. (Biomedical) fromMcGill in 1981. Following
[IO] A. W. Sills, V. Honrubia, and W. E. Kumley, Algorithm for three years of post-doctoral research at the Aer-
the multiparameter analysis of nystagmus using a digital com- ospace Medical Research Unit at McGill, she
puter, Aviat. Space Environ. Med. , vol. 46, pp. 934-942, 1975. joined the staff of the Department of Biomedical Engineering where
[ I I ] C. Wall and F. 0. Black, Algorithms for the clinical analysis she currently holds the position of Associate Professor.
of nystagmus eye movements, IEEE Trans. Biomed. Eng., vol. Her main research interests are in the areas of signal processing,
BME-28, pp. 638-646, 1981. nonlinear systems and model simulation, applied to the control of eye
and head movements and sensorv-motor integration. She has held
Claudio G. Rey was born in Montevideo, Uru-
guay, in 1958. He received the B.A.Sc. and the
M.A.Sc. degrees in electrical engineering in
1983 and 1985, respectively, fromthe Univer-
sity of British Columbia, Vancouver, Canada.
He worked for two and a half years as a re-
search engineer for the University of British
Columbia working under contract for the De-
partment of National Defense, Canada. He is
currently in a doctoral programin biomedical
engineering at McGill University, Montreal,
P.Q., Canada.
Chercheur-Boursier scholarships
Santt du Qutbec since 1984.
from the Foids de Rec herches en