EFFECTIVENESS FOR A SLEEPINESS TEST OF PUPIL SIZE

ESTIMATION DURING BLINK

Minoru Nakayama

CRADLE (The Center for Research and Development of Educational Technolgy)

Tokyo Institute of Technology, O-okayama 2–12–1, Meguro-ku, Tokyo, 152–8552, Japan

Keiko Yamamoto, Fumio Kobayashi

Deapartment of Health and Psychosocial Medicine, Aichi Medical University

Karimata Yazako 21, Nagakute-cho, Aichi, 480–1195, Japan

Keywords:

Pupil size, Estimation, Support Vector Regression, Sleepiness test, Pupil Unrest Index.

Abstract:

Pupillary response has been used for an index of sleepiness, but the validity of the index is not clear. In this

paper, the inﬂuence of blinks on the Pupillary Unrest Index (PUI) and the Power Spectrum Density (PSD) for

the frequency range 0.01 < f < 0.8Hz, as indices of pupil’s instability during a sleepiness test, was examined.

To estimate pupil size during blink, a procedure for collecting the clinical data was developed using Support

Vector Regression (SVR). The values of PUI increased with experimental time, and the values and deviations

of PUI for experimental observation were larger than the ones with SVR estimation. The blink time also

increased with experimental time, and there were signiﬁcant correlation relationships between the value of

PUI and blink time. The mean PSD also correlated signiﬁcantly with blink time. The relationship between

pupillary indices and a subjective sleepiness index was not signiﬁcant, as it was not in other previous works.

These results provide evidence that pupillary indices were signiﬁcantly affected by blink, and they did not

reﬂect sleepiness correctly.

1 INTRODUCTION

Temporal observation of the human eye pupil is called

as pupillography, and these observations can be used

as an index for various human activities (Kuhlmann

and B¨ottcher, 1999; Beatty, 1982). In particular,

pupillography has been used for assessment of sleepi-

ness and exhaustion using the eye sleepiness test,

which consists of measuring the magnitude of pupil-

lary change as a Pupil Unrest Index (PUI) and making

readings of the frequency power spectrum. It is often

applied to clinical observations or used in industrial

engineering situations (L¨udtke et al., 1998; Wilhelm

et al., 1998; Wilhelm et al., 1999). These indices have

been applied to the evaluation of emotional change

(Norrish and Dwyer, 2005); this analysis procedure is

recognized as a signiﬁcant measure.

Although these indices have also been applied to

diagnostic procedures, a series of research studies of

multiple sclerosis patients suggests that for healthy

people there is no signiﬁcant correlational relation-

ship between PUI and subjective sleepiness indices

such as the Stanford Sleepiness Score (Egg et al.,

2002; Frauscher et al., 2005). This means that the

evaluation procedure should be examined carefully.

A possible problem with observing the pupil is the

inﬂuence of blink (Nakayama and Shimizu, 2001),

because most methods of measuring pupil size are

based upon processing the image of the eye. Blink

can affect measurements due to the eye being ob-

scured by the eye lid during blink. Blinks are

usually discussed as an artifact in temporal obser-

vations such as mean pupil sizes or for results of

frequency analysis (Nakayama and Shimizu, 2001;

Nakayama and Shimizu, 2002). To resolve these

problems, some methods of estimating pupil size dur-

ing blink were developed (Nakayama and Shimizu,

2001; Nakayama, 2005), and the performancewas ex-

amined (Nakayama, 2006).

However, the effectiveness of the estimation pro-

cedure for a diagnostic procedure ( such as an eye

sleepiness test ) and the signiﬁcance of pupil indices

which include blinks have not been discussed sufﬁ-

ciently. In this paper, we address the inﬂuence of

blink and the validity of pupillary indices by exam-

ining the effectiveness of estimating pupil size during

558

Nakayama M., Yamamoto K. and Kobayashi F. (2008).

EFFECTIVENESS FOR A SLEEPINESS TEST OF PUPIL SIZE ESTIMATION DURING BLINK.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 558-564

DOI: 10.5220/0001058105580564

 SciTePress

Table 1: Analysis condition.

Condition Original method This paper

Sampling

Sampling rate 25Hz 60Hz

Pre-processing moving average moving average

SVR

Average window size 0.4sec. 0.4sec.

PUI

Segment length 2048(82sec.) 4096(68.3sec.)

Unit for segment 16(0.64sec.) 32(0.53sec.)

N of units 128 128

Sampling rate 1.560Hz 1.875Hz

FFT

Data length 2048(82sec.) 4096(68.3sec.)

Frequency range 0.0<f<0.8Hz 0.01<f<0.8Hz, 4Hz

Original method:(L¨udtke et al., 1998; Wilhelm et al., 1999)

blink using the support vector regression (SVR) tech-

nique.

The purposes of this paper are as follows:

1. To develop an estimation procedure using SVR

for clinical pupillary observation.

2. To examine the inﬂuence of blink on the pupillary

indices.

3. To examine the relationship between pupil indices

and subjectivesleepiness scores, and the inﬂuence

of blinks on them.

2 MEASURING PUPIL SIZE

2.1 Sleepiness Test

The most popular method to assess sleepiness is a

procedure which has been proposed by Wilhelm et

al. (L¨udtke et al., 1998; Wilhelm et al., 1999). The

measuring procedure in this paper was based on the

following method.

The measuring equipment was designed to be

worn on the eyes as goggles (Hamamatsu photon-

ics:C7364). The subjects were asked to wear this

equipment and to gaze at a small red LED light (in-

frared wave length: 890 nm) through the goggles,

with a CCD camera shooting an image of the pupil.

The subjects were instructed to sit and to remain

awake in a semi-dark room in a building during the

experiment, and were also asked to close their eyes

for one minute to promote sleepiness before starting

the experiment. The experiment lasted 12 minutes.

The pupil diameter size was measured at 60 Hz.

This experiment was conducted between 9 a.m.

and 4 p.m. in the late summer. 35 healthy males

joined the experiment, their average age was 37.9

years and the standard deviation was 4.1. They were

volunteer subjects and signed an agreement on the ex-

perimental procedure before it commenced.

Some parameters of the analyzing procedure

which were proposed by Wilhelm et al. (L¨udtke et al.,

1998) depended on the measuring equipment. One

example is the sampling rate of pupil size. The differ-

ences are summarized in Table 1.

2.2 Pre-processing

Pre-processing of pupil size during blink provides a

possible pupil size from the temporal sizes. To ex-

amine the effectiveness of pre-processing, the follow-

ing two pre-processings were created using moving

average method (MOV) and support vector regres-

sion (SVR) (Smola and Scholkopf, 1998; Nakayama,

2005). SVR and the kernel method are often used for

signal reconstruction or smoothening (Bishop, 1995;

Smola and Scholkopf, 1998).

2.2.1 Experimental Observation (Exp.)

This data set consisted of experimental observations

without any pre-processing. During the periods of

blink, the pupil diameter shows that the size was mea-

sured as 0.

2.2.2 Moving Average (MOV)

Moving average method was applied to exclude a

large deviation caused by blink and noise. Wilhelm

et al. conducted this method for every data series of

0.4 sec. (L¨udtke et al., 1998; Wilhelm et al., 1999).

This means that the sampling rate is reduced to 2.5

Hz.

2.2.3 Estimation with Svr

This processing provided estimation diameters dur-

ing blink using support vector regression (SVR) with

Gaussian kernel (Smola and Scholkopf, 1998). The

estimation function was derived from the training

data. This training data, as a prototype of pupil re-

sponse, consisted of measured pupil diameters during

the blink and estimated pupil sizes. To produce the

training data, a set of data containing 5000 data points

collected at the beginning of observations was pre-

pared. To obtain an optimized model, the dimension

n of input vector, and a precision ε(eps) and σ(std)

of Gaussian kernel needed to be calculated. A prac-

tical calculation was conducted using the SVMTorch

package (Collobert and Bengio, 2001), and parame-

ters were optimized. As a result, the following param-

eters were provided: input dimension = 45; σ = 40;

EFFECTIVENESS FOR A SLEEPINESS TEST OF PUPIL SIZE ESTIMATION DURING BLINK

559

Pupil diameter (mm)

0 5 10

Time (sec.)

Exp.

MOV

SVR

Figure 1: Example of pre-processing for pupillary change

during blink.

ε = 0.5. Estimation accuracy was examined in the

previous estimation experiments (Nakayama, 2006).

3 RESULTS

3.1 Results of Pre-processing

To examine the pupil size pre-processing perfor-

mance, an example of experimental pupil size and

processed data for 10 seconds is illustrated in Fig-

ure 1, listed from bottom to top as Exp., MOV and

SVR. The horizontal axis shows time, the vertical axis

shows pupil size with drops indicating blinks. Pre-

processing with MOV shows that temporal changes

are inﬂuenced by blink and all points are smoothened,

although there is no null point during blink periods.

On the other hand, SVR indicates the same pupil size

without large blink drops and gives possible sizes of

pupil diameters during blinks. As a result, an appro-

priate estimation procedure for clinical pupillography

can be developed from this.

3.2 Results of Pupillary Unrest Index

(PUI)

PUI as an index of instability of pupillography was

calculated following a procedure which was modiﬁed

from the original method using the parameters listed

in Table 1. According to the deﬁnition of PUI as

cumulative changes in pupil diameter (L¨udtke et al.,

1998), ﬁrstly the data were reduced by calculating the

average for periods of 32 (0.53 sec.) consecutive val-

ues, secondly the absolute values of the differences

from one 32-value average to the next one were sum-

marized for each 68.3 sec. data segment, namely 127

differences for one segment. Calculating the average

Exp.

MOV

SVR

1 2 3 4 5 6 7 8 9

Segment No.

Pupil Unrest Index (PUI)

Figure 2: Result of PUIs across segments and pre-

processing procedures.

prior to cumulation serves as a simple low pass ﬁlter-

ing and excludes high frequency noise.

Average PUIs with standard error bars across the

pre-processing procedure results were summarized in

Figure 2. PUIs for Exp. and MOV conditions are sig-

niﬁcantly higher than the ones for SVR. According to

the estimation procedure, PUI increases when pupil-

lary temporal change includes blink drop. Therefore,

PUIs for Exp. and MOV were relatively high.

Also, sleepiness may increase gradually with ex-

perimental time, so this suggests that the gradual

increase may depend on sleepiness. According to

the pre-processed PUI results and a previous work

(Nakayama, 2006), the biggest factor in PUI change

must be blink frequency, however.

3.3 Blink Time

Blink may inﬂuence a sleepiness index according to

the results of PUI. Blink time is deﬁned as the sum of

blink drop duration of measured pupil diameters. Av-

erage blink times for each segment (68.3 sec.) were

summarized as bar graphs with standard error bars in

Figure 3. The ﬁgure shows that the blink time in-

creases monotonically with the sequence number of

the segment excepting segment 5. According to blink

research, the estimated blink time may be around 1

minute per segment in the standard condition (Tada

et al., 1991). It was suggested that the blink time after

segment 5 was longer than the one for the standard

condition.

These results also indicated that blink inﬂuenced

PUIs. To examine the relationship between PUI and

blink time, correlation coefﬁcients were calculated

across all segments and preprocessing procedures.

The coefﬁcients were summarized in Table 2. Most

coefﬁcients were signiﬁcant (p < 0.05). During seg-

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

560

1 2 3 4 5 6 7 8 9

Segment No.

Blink time (sec.)

Figure 3: Average blink time for segments.

Table 2: Correlation coefﬁcient between PUI and blink

time.

Seg. No. Exp. MOV SVR

1 0.95 0.95 0.42

2 0.91 0.92 0.71

3 0.93 0.92 (0.23)

4 0.90 0.92 0.41

5 0.42 0.45 0.41

6 0.64 0.65 0.34

7 0.55 0.58 0.63

8 0.69 0.70 0.58

9 (0.16) (0.18) 0.35

N=35, ( ) not signiﬁcant

ment numbers 1 to 4, there were large correlation co-

efﬁcients for pupil diameters with blink times for Exp.

and MOV rather than for correlation coefﬁcients for

SVR. Coefﬁcients across all pre-processing stayed at

same levels after segment number 5, when the blink

times were longer than the ones in the standard con-

dition.

These results showed that blink signiﬁcantly af-

fected PUI changes in the standard condition, and the

effectiveness of pre-processing for pupils was exam-

ined while the blink times stayed at the standard level.

The relationship was also affected by the incidence of

additional blinking. Furthermore, these results sug-

gested the blink time affected PUIs despite the con-

ducting of estimations of pupil size during blink.

3.4 Frequency Analysis

Frequency power value of pupillary change, which is

given by frequency analysis over 68.3 sec., can be

used as another index of sleepiness (L¨udtke et al.,

1998; Wilhelm et al., 1999). According to the pro-

cedure, the power spectrum of pupil diameter change

was summarized. Figure 4 shows the results of Power

1 2 3 4

-10

-20

-30

Frequency (Hz)

Power Spectrum Density (dB)

Exp.

MOV

SVR

Figure 4: PSD of pupillography across pre-processing.

-40

-20

Exp.(0.01<f<0.8)

MOV(0.01<f<0.8)

SVR(0.01<f<0.8)

Exp.(1.5<f<3.5)

MOV(1.5<f<3.5)

SVR(1.5<f<3.5)

1 2 3 4 5 6 7 8 9

Segment No.

Mean PSD (dB)

Figure 5: Averaged PSD of each segment for two frequency

ranges.

Spectrum Density (PSD) for the ﬁrst segment of one

subject. PSD was estimated with pwelch function and

Parzen window function of MATLAB (MathWorks

Inc.). The vertical axis shows PSD in decibels (dB)

and the horizontal axis shows frequency (Hz) from

0.01 to 4.0 Hz. DC component as frequency power

was excepted in following analysis. The pupillary

change has a low pass ﬁlter as low as 4 Hz because

it is biological signal, and also pupilograms contain

0.05 ∼ 0.3 Hz components which are well known as

pupillary noise (Tsukahara, 1976).

During sleepiness tests, the average power value

for the frequency range (0.01 < f < 0.8Hz) is often

evaluated as the index (L¨udtke et al., 1998; Wilhelm

et al., 1999). The average PSD for frequency range

(0.01 < f < 0.8Hz) for each segment was compared

between pre-processing procedures. The results were

summarized in Figure 5. According to the results

of frequency analysis for the task evoked pupillog-

raphy, PSDs of frequency range (1.5 < f < 3.5Hz)

changed signiﬁcantly in response to the task difﬁ-

culty. Therefore, average PSDs for frequency range

EFFECTIVENESS FOR A SLEEPINESS TEST OF PUPIL SIZE ESTIMATION DURING BLINK

561

Table 3: Correlation coefﬁcients between PSDs and blink

for two frequency ranges.

Seg. 0.01 < f < 0.8Hz 1.5 < f < 3.5Hz

No. Exp. MOV SVR Exp. MOV SVR

1 -.36 -.36 (-.29) 0.81 0.86 0.44

2 -.38 -.38 (-.25) 0.86 0.87 0.65

3 -.35 -.36 (-.30) 0.65 0.77 (0.20)

4 -.53 -.53 -.51 0.57 0.63 (0.21)

5 -.42 -.42 -.37 (0.11) (0.15) (-.03)

6 -.47 -.46 -.47 (0.28) 0.37 (0.10)

7 -.52 -.52 -.46 (0.13) (0.20) (0.19)

8 -.55 -.55 -.53 (0.31) 0.45 (0.20)

9 -.62 -.62 (-.30) (-.10) (-.07) (-.07)

N=35, ( ) not signiﬁcant

(1.5 < f < 3.5Hz) were also summarized in the same

format in Figure 5. The ﬁgure shows that PSDs for

(0.01 < f < 0.8Hz) are at the same level across seg-

ments and pre-processing procedures. The PSDs for

(1.5 < f < 3.5Hz) have some differences amongst

pre-processing, but they stay at the same levels dur-

ing each experimental time point. This suggests that

PSDs are not affected by a change in blink time, such

as during a change in the level of sleepiness.

3.5 Blink Inﬂuence On Psd

To examine the inﬂuence of blink, correlation coefﬁ-

cients between blink time and average PSDs for two

frequency ranges (0.01 < f < 0.8Hz and 1.5 < f <

3.5Hz) across pre-processing procedures were calcu-

lated. The results were summarized in Table 3.

All coefﬁcients for the frequency range (0.01 <

f < 0.8Hz) were negative values and signiﬁcant ex-

cept for some coefﬁcients for SVR. There was no sig-

niﬁcant relationship between SVR and blink time dur-

ing the ﬁrst three segments, and some absolute values

of coefﬁcients for SVR were relatively smaller than

the ones for Exp. and MOV. Most PSDs for frequency

range (0.01 < f < 0.8Hz) correlated with blink time,

however. This suggests that PSDs depend on blink

time, and that the relationship is affected by the pre-

processing procedure. Also, PSDs of Exp. and MOV

for frequencyrange (1.5 < f < 3.5Hz) correlated with

blink time during segments 1-4. There were no signif-

icant relationships between them after Segment No.

5, while blink time was longer than in the standard

condition, however. These correlation relationships

seem to be caused by blinks.

3.6 Relationship with a Subjective Score

Subjective sleepiness was measured for each subject

using the Stanford Sleepiness Score (SSS) (Hoddes

et al., 1973). 33 out of 35 subjects responded to

1 2 3 4 5 6 7 8 9

Segment No.

Blink time (sec.)

High

Low

Figure 6: Averaged blink time across two sleepiness groups.

1 2 3 4 5 6 7 8 9

Segment No.

Pupil Unrest Index (PUI)

Exp.

MOV

SVR

Figure 7: PUI changes with pre-processing procedures and

two sleepiness groups.

this questionnaire. The scores were distributed from

2 to 4 on a 7 point scale. The correlation relation-

ships of SSS with both PUIs and PSDs were exam-

ined. The absolute value of the correlation coefﬁ-

cients (r) were less than 0.15 and they were not sig-

niﬁcant (p > 0.10), because all subjects were healthy

and scores were distributed in a narrow range. There-

fore, indices of pupillography do not correlate with

the subjective scores as well as in previous works

(Egg et al., 2002; Frauscher et al., 2005).

4 DISCUSSION

The two observations and suggested causes which

have been reported in this paper are examined here;

there is an inﬂuence of blink on pupillary indices

(Nakayama, 2005; Nakayama, 2006), and there is no-

correlation between pupillary indices and subjective

sleepiness (Egg et al., 2002; Frauscher et al., 2005).

There was some distribution of subjective sleepiness

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

562

Table 4: Correlation coefﬁcients between PSD and blink time across pre-processing procedures and two sleepiness groups.

0.01 < f < 0.8Hz 1.5 < f < 3.5Hz

Seg. High sleepiness Low sleepiness High sleepiness Low sleepiness

No. Exp. MOV SVR Exp. MOV SVR Exp. MOV SVR Exp. MOV SVR

1 (-.25) (-.25) (-.15) -.60 -.60 -.58 0.82 0.78 0.57 0.88 0.97 (0.18)

2 (-.33) (-.33) (-.15) (-.53) (-.53) (-.52) 0.88 0.88 0.74 0.87 0.96 (0.19)

3 (-.35) (-.35) (-.29) (-.37) (-.37) (-.31) 0.76 0.78 (0.04) 0.57 0.77 (0.38)

4 -.56 -.57 -.56 (-.51) (-.51) (-.44) 0.52 0.55 (0.15) 0.68 0.84 (0.27)

5 -.46 -.46 (-.41) (-.51) (-.51) (-.47) (0.04) (0.07) (-.12) (0.54) 0.71 (0.13)

6 -.59 -.59 -.61 (-.50) (-.50) (-.47) 0.42 (0.47) (0.01) (0.20) (0.31) (0.18)

7 -.55 -.55 -.51 (-.55) (-.55) (-.48) (0.09) (0.14) (0.31) (0.35) (0.48) (0.20)

8 -.60 -.60 -.62 -.65 -.65 -.60 (0.34) 0.48 0.55 0.60 0.71 (0.19)

9 -.68 -.68 (-.37) (-.32) (-.32) (-.31) (-.17) (-.14) (-.15) (0.45) 0.69 (0.11)

N=35, ( ) not signiﬁcant

in this experiment, such as between 2 and 5 on the

7 point scale, therefore the effect of the difference in

the subjective sleepiness on the indices was analyzed.

Firstly, 33 responded subjects were divided into two

groups; the low sleepiness group consisted of 13 sub-

jects who answered 2 on the 7 point scale of sleepi-

ness, and the high sleepiness group consisted of 20

subjects who answered 3 to 5 on the 7 point scale of

sleepiness, with an average rate of 3.4.

Blink time for each segment was summarized

across two groups in Figure 6 using the same format

as in Figure 3. In Figure 6, bars show average blink

times with error bars as standard errors. In compar-

ing the two groups, the temporal change in segment

sequences were quite different. Blink times for the

high sleepiness group increased monotonically, and

the standard errors also increased with the average.

Blink times for the high sleepiness group in the 8th

segment became 7 times the length of the time of the

1st segment. On the other hand, the average blink

times for the low sleepiness group did not change dur-

ing the experiment. There were no signiﬁcant differ-

ences in blink time between the two groups because

of the large deviation in blink time, however. If blink

time reﬂects sleepiness, the differences in blink time

between the two groups may signiﬁcantly affect the

variation in the subjective sleepiness score.

PUIs for two groups were summarized in Figure 7

using the same format as in Figure 2. Differences of

PUIs for Exp. and MOV between the two groups in-

creased with experimental time. Also, the differences

in PUI for SVR between the two groups was small

and almost constant during the experiment. There

was no signiﬁcant difference in PUI between the two

groups and three pre-processing procedures because

of the large deviation in PUI, particularly for the high

sleepiness group. Although there was no signiﬁcant

difference between the two groups, it was noted that

blink affected PUI values on pupil diameter observa-

tions when blinks were not processed appropriately.

To examine the relationship between PSDs and

subjectivesleepiness, correlation coefﬁcientsbetween

averaged PSDs and blink times for the two groups

were summarized in Table 4 using the same format

as in Table 3. Comparing correlation coefﬁcients be-

tween the two groups of subjective sleepiness ratings,

the signiﬁcance of coefﬁcientsfor the frequencyrange

0.01 < f < 0.8Hz changed across the two groups. Co-

efﬁcients in the frequency rage 1.5 < f < 3.5Hz were

relatively stable with the subjective sleepiness ratings.

According to the results, coefﬁcients between PSD

and blink time depend on both the pre-processing pro-

cedure for blink and the subjective sleepiness rating.

For coefﬁcients in the frequency range 1.5 < f <

3.5Hz, the effect of the subjective sleepiness rating

was relatively smaller than the effect in the frequency

range 0.01 < f < 0.8Hz.

The results of analyzing pupil indices during the

sleepiness test coincide with the clinical claims (Egg

et al., 2002; Frauscher et al., 2005); PUI does not

correlate with subjective sleepiness such as SSS. Al-

though it was suggested that the frequency power of

0.01 < f < 0.8Hz reﬂected the subjective sleepiness

(L¨udtke et al., 1998; Wilhelm et al., 1998; Wilhelm

et al., 1999), the relationship was not conﬁrmed in

this experiment. Additionally, even the inﬂuence of

blink when the blink drops during a temporal change

of pupil size were removed was examined. Consider-

ing the empirical evidence from these clinical obser-

vations, a sleepiness evaluation procedure should be

developed. A discussion of this will be the subject of

our further study.

5 CONCLUSIONS

In this paper, we used pupillography to examine the

inﬂuence of blink and validity of pupillary indices by

analyzing a clinical sleepiness test.

The following results were achieved:

EFFECTIVENESS FOR A SLEEPINESS TEST OF PUPIL SIZE ESTIMATION DURING BLINK

563

1. A pupil size during blink estimation procedure us-

ing the Support Vector Regression technique for

clinical pupillary observation was developed and

an appropriate level of performance was obtained.

2. The inﬂuence of blink in pupillary indices such

as Pupil Unrest Index (PUI) and Power Spectrum

Density (PSD) of pupillographywas examined. In

particular, it was shown that blink time increased

monotonically with experimental time, therefore

the inﬂuence of blink changed as the experiment

progressed.

3. The relationship between pupil indices and sub-

jective sleepiness scores such as the Stanford

Sleepiness Score (SSS) was analyzed. There was

no signiﬁcant relationship, but there were some

differences in pupil indices between high and low

SSS groups.

Development of a sleepiness test evaluation procedure

which considers blink and other factors will be the

subject of our further study.

ACKNOWLEDGEMENTS

We would like to thank Dr. Misuzu Watanabe,

Dr. Reiko Hori and Dr. Hirohito Tsuboi, Aichi Med-

ical University, who kindly provided us with the op-

portunity to analyze pupillography data collected dur-

ing their sleepiness tests.

REFERENCES

Beatty, J. (1982). Task-evoked pupillary response, process-

ing load, and the structure of processing resources.

Psychological Bulletin, 91(2):276–292.

Bishop, C. M., editor (1995). Neural Networks for Pattern

Recognition. Oxford University Press, Oxford, UK.

Collobert, R. and Bengio, S. (2001). Svmtorch: Support

vector machines for large-scale regression problems.

Journal of Machine Learning Research, 1:143–160.

Egg, R., H¨ogl, B., Glatzl, S., Beer, R., and Berger, T.

(2002). Autonomic instability, as measured by pupil-

lary unrest, is not associated with multiple sclerosis

fatigue severity. Multiple Sclerosis, 8:256–260.

Frauscher, B., Egg, R., Brandauer, E., Ulmer, H., Berger, T.,

Poewe, W., and H¨ogl, B. (2005). Daytime sleepiness

is not increased in mild to moderate multiple sclerosis:

a pupillographic study. Sleep Medicine, 6:543–547.

Hoddes, E., Zarcone, V., Smythe, H., Phillips, R., and De-

ment, W. C. (1973). Quantiﬁcation of Sleepiness. Psy-

chlophysiology, 10(4):431–436.

Kuhlmann, J. and B¨ottcher, M., editors (1999). Pupil-

lography: Principles, Methods and Applications.

W. Zuckschwerdt Verlag, Munchen, Germany.

L¨udtke, H., Wilhelm, B., Adler, M., Schaeffel, F., and Wil-

helm, H. (1998). Mathematical procedures in data

recording and processing of pupillary fatigue waves.

Vision Research, 38:2889–2896.

Nakayama, M. (2005). Estimation of eye-pupil size during

blink by support vector regression. In Modelling Nat-

ural Action and Selection: Proc. of an International

Workshop, pages 121–126.

Nakayama, M. (2006). Inﬂuence of blink on pupillary in-

dices. In IEEE BioCAS 2006: Proceeding of Biomed-

ical Circuits and Systems Conference, pages 29–32.

Nakayama, M. and Shimizu, Y. (2001). An estimation

model of pupil size for blink artifact in viewing TV

programs. IEICE Transactions, J84-A:969–977.

Nakayama, M. and Shimizu, Y. (2002). An estimation

model of pupil size for ‘blink artifact’ and it’s applica-

tions. In Proc. 10th European Symposium on Artiﬁcial

Neural Networks, (ESANN ’02), pages 251–256.

Norrish, M. I. K. and Dwyer, K. L. (2005). Preliminary

investigation of the effect of peppermint oil on an ob-

jective measure of daytime sleepiness. International

Journal of Psychophysiology, 55:291–298.

Smola, A. J. and Sch¨olkopf, Y. (1998). A tutorial on support

vector regression. NeuroCOLT2 Technical Report Se-

ries NC2-TR-1998-030.

Tada, H., Yamada, F., and Fukuda, K. (1991). Psychological

blink (in Japanese). Kita-Ouji-Shobo, Kyoto, Japan.

Tsukahara, N. (1976). Control system of vertebrate and cen-

tral nervous system (in Japanese), volume 6 of Biolog-

ical Science. Asakura Shoten, Tokyo, Japan.

Wilhelm, B., L¨udtke, H., and Wilhelm, H. (1999). Spon-

taneous pupillary oscillations, Volume 18 of Clinical

Pharmacology. W. Zuckschwerdt Verlag, Munchen,

Germany.

Wilhelm, B., Wilhelm, H., L¨udtke, H., Streicher, P., and

Adler, M. (1998). Pupillographic assessment of

sleepiness in sleep–deprived healthy subjects. SLEEP,

21(3):258–265.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

564