Development of a Procedure for Detecting Dementia Symptoms Using

Features in Differential Waveforms of the Pupil Light Reﬂex

Minoru Nakayama

1 a

, Wioletta Nowak

2 b

and Anna

Zarowska

2 c

Institute of Science Tokyo (Tokyo Tech.), O-okayama, Meguro-ku, Tokyo, 152–8552, Japan

Wrocław University of Science and Technology, Wrocław, 50–370, Poland

Keywords:

Pupil, Pupil Light Reﬂex, Alzheimer’s Disease, Feature Extraction, Functional Data Analysis.

Abstract:

A procedure for detecting dementia levels is developed using features of waveform shapes of the pupillary

light reﬂex (PLR) in response to chromatic light pulses on either eye. Features of waveform shapes were

extracted using a functional data analysis technique which measured the reactions of both eyes. In considering

the physiological mechanism, differential waveform shapes were also analysed. The feature was extracted as a

coefﬁcient of B-spline basis functions of the waveforms. The feature sets of the differential waveform shapes

for blue and red light pulses contributed to detection performance. Also, feature weights are used to represent

PLR reaction mechanisms and differences in response to chromatic stimuli.

1 INTRODUCTION

Conventionally, the pupil light reﬂex (PLR) is based

on ipRGC (intrinsically photosensitive retinal gan-

glion cell) systems which can be an index for the di-

agnosis of various types of diseases (Gamlin et al.,

2007; Kawasaki and Kardon, 2007; Chougule et al.,

2019; Kelbsch et al., 2019). In order to develop a

biomarker of diseases, PLRs in response to chromatic

stimuli have been observed and analysed (Nowak

et al., 2020; Nowak et al., 2021; Nakayama et al.,

2022; Nakayama et al., 2023b). In particular, the

differential responses caused by problems with sig-

nal transfer by the optic nerve of each eye may pro-

vide clinical information such as the inﬂuence of de-

mentia or Alzheimer’s disease on the synchronisa-

tion of pupil reactions (McDougal and Gamlin, 2015;

Chaitanuwong et al., 2023; Molitor et al., 2015; Nie

et al., 2020). Asynchronicity of PLRs of both eyes

shows some symptoms of the disease in the optic

nerve system, as some factors may inﬂuence on PLRs,

since most participants are elderly people (Nakayama

et al., 2024a). In order to observe the disparity of

both pupils, PLR measurements of both eyes using

a light pulse to either eye may be useful (Nakayama

et al., 2022; Nakayama et al., 2023b; Nakayama

https://orcid.org/0000-0001-5563-6901

https://orcid.org/0000-0002-4135-2526

https://orcid.org/0000-0003-4544-9082

et al., 2024b). Another issue is the methodology of

feature extraction of PLR waveforms, because the

waveform shapes of PLRs may reﬂect the dynamics

of physiological functions. However, overall wave-

form features have been used before, though detec-

tion performance for irregular responses was insufﬁ-

cient (Nowak et al., 2019). In order to improve perfor-

mance signiﬁcantly, a set of speciﬁc features of PLRs

measured during the pupil constriction phase was in-

troduced instead of the overall set of features. The

remaining part of the waveform, composed of pupil

reactions after the eye has recovered from constric-

tion are often referred to as post-illumination pupil re-

sponses (PIPRs) in the diagnostic procedure. In con-

sidering the localised features of waveform shapes,

an improved feature extraction procedure for entire

waveform is required. A functional data analysis

technique can be applied to waveforms of any length

and waveform features including PLR waveforms can

be extracted (Nakayama et al., 2024a). The beneﬁt of

this will be conﬁrmed in this paper.

This paper examines the possibility of detec-

tion for dementia levels using differential features of

waveform shapes of PLRs. In particular, the effective-

ness of extracting features from an entire PLR wave-

form and introducing features of differentials of PLR

waveforms is discussed. The following topics are ad-

dressed in this paper.

1. The features of waveform shapes, which are sum-

Nakayama, M., Nowak, W. and

Zarowska, A.

Development of a Procedure for Detecting Dementia Symptoms Using Features in Differential Waveforms of the Pupil Light Reﬂex.

DOI: 10.5220/0013244400003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 943-948

ISBN: 978-989-758-731-3; ISSN: 2184-4305

943

mations and ratios of PLRs of both eyes, are ex-

tracted using a B-spline technique while either eye

is irradiated in blue or red light pulse.

2. Classiﬁcation performance of the level of demen-

tia is examined using a combination of extracted

features of PLRs

The following section of this paper consists of re-

lated works, where the method, classiﬁcation analysis

for patients and normal control group, and summary

of the work are based on previous studies.

2 RELATED WORKS

Some clinical surveys measuring PLR reactions

to chromatic light stimuli have been conducted

since ipRGC was discovered (Gamlin et al., 2007;

Kawasaki and Kardon, 2007; Chougule et al., 2019;

Kelbsch et al., 2019). In addition to these studies,

possible levels of dementia have also been measured

using chromatic light pulse stimuli of the pupil (Moli-

tor et al., 2015; Nie et al., 2020). Since the number of

patients who can participate is limited, most analyses

have been published as case studies. In due course,

more systematic surveys were conducted, and addi-

tional analytical reports were published (Nakayama

et al., 2022; Nakayama et al., 2023b). This survey

consists of two sets of light irradiation, of the left and

right eyes, as validation tests using previously studied

data sets of left and right eye light pulses (Nakayama

et al., 2022; Nakayama et al., 2023b). The ocular

responses of each eye may differ, even though the

same stimulus is provided simultaneously. The re-

sponses of either eye can be summarised as indepen-

dent data sets when differential metrics are introduced

(Nakayama et al., 2024b). In experimental surveys,

the participation of patients is limited, and asymmet-

ric responses can be used to compensate for the dearth

of measured data.

However, in these previous studies the analyses

employed a set of ﬁxed feature points which were

extracted during a large reaction to the light pulse

(Nowak et al., 2020; Nowak et al., 2021). The over-

all features of PLR waveform shapes should be in-

troduced, as mentioned in the introduction. One ap-

proach may be the extraction of feature of waveform

shapes using a B-spline technique (Ramsay et al.,

2009; López et al., 2022; Nakayama et al., 2024a),

however an appropriate dimension reduction tech-

nique will be required in order to suppress the large

number of dimensions of the features. The features of

entire waveform shapes may be useful for the detailed

analysis of time series data, and the temporal features

can be compared for the feature selections. The pos-

sibility of patient detection using PLR waveform fea-

tures should be examined. In most cases, B-spline is

often employed as a basis function. Therefore, this

work is based on this conventional procedure as the

ﬁrst step.

3 METHOD

The set of PLR data consiting pupillary responses

to be analysed was surveyed in previous studies

(Nakayama et al., 2022; Nakayama et al., 2023b),

which measured 1 sec. chromatic light pulses (blue

or red) to either eye during a 10 sec. period of obser-

vations. The participants were elderly people and in-

cluded some patients with Alzheimer’s disease (AD),

mild cognitive impairment (MCI) and a normal con-

trol group (NC).

3.1 Experimental Procedure

3.1.1 Measurmement Protocols

Pupil diameters of each eye were measured for 10 sec.

in a temporal darkened space under the following con-

ditions (Nakayama et al., 2022). PLRs were measured

during Conditions 2-5.

1. Condition 1: Observe static pupil oscillation with-

out light pulses

2. Condition 2: Blue light pulses to the right eye

3. Condition 3: Blue light pulses to the left eye

4. Condition 4: Red light pulses to the right eye

5. Condition 5: Red light pulses to the left eye

The light stimulus is irradiated for one second dur-

ing each session. The combination of the experimen-

tal conditions is ﬁxed for all participants. In regards to

each of the four light pulse irradiation conditions, the

measured data is classiﬁed into sets for irradiation of

the left and right eyes. Condition 1 is a set of common

data for both eyes (Nakayama et al., 2024b).

The experimental procedure was approved by an

ethics committee at Osaka Kawasaki Rehabilitation

University.

3.1.2 Apparatus

Pupil diameters of both eyes were observed using a

piece of specialised measuring equipment at 60Hz

(URATANI, HITOMIRU). Light stimuli consisted of

a blue light source (469nm, 14.3cd/m

, 6.5lx) and a

red light source (625nm, 12.3cd/m

, 10.5lx).

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

944

0.75 0.80 0.85 0.90 0.95 1.00

Relative Pupil Size

70 basis

10 basis

2 3 4 5 6 7

Time (seconds)

Figure 1: A summation signal of PLRs for blue light pulses.

3.1.3 Experimental Participants

101 elderly individuals were selected to take part in

the experiment (Mean age : 78.5, SD : 8.9, F:66,

M:35) (Nakayama et al., 2022; Nakayama et al.,

2023a). Their cognitive functions were diagnosed

by a clinical doctor and sorted into three levels us-

ing MMSE (Mini-Mental State Examination) scores

(Tombaugh et al., 1996).

• AD: 32 (Mean age:83.0, SD:6.3, F:22, M:10).

• MCI: 9 (Mean age:82.1, SD:6.3, F:5, M:4).

• NC: 60 (Mean age:75.6, SD:9.2, F:38, M:22).

3.2 Analysis of Functional Data

Measured pupil sizes are summarised into sets for ir-

radiation of the left and right eyes, forming two sets

of data (the left and right eye data sets). Condition 1

is provided as a common control condition. Observed

signals from both eyes are processed as summation

of thr left and right eyes, and the ratios between the

sizes for the left and right eyes right size/left size) are

asynchronous measurements. All data are standard-

ised using the initial level measurements made during

the initial 1 sec. before irradiation. Also, the period

of the ﬁnal 2 sec. is eliminated in order to reduce the

amount of data measured. Figure 1 shows the over-

all means of summation waveforms of the left eye re-

sponses to right eye blue light irradiation. The hor-

izontal axis represents the time course, from 1 to 8

seconds. A waveform of pupil size shows pupil con-

strict, with dome delay, after 1 second blue light pulse

(1-2sec.). The summation waveform is very smooth.

This waveform can be represented using a B-spline

basis function as a functional data analysis technique

(Ramsay et al., 2009; López et al., 2022). When 70

0.85 0.90 0.95 1.00 1.05

1.10 1.15

Pupil ratio (right / left)

MCI+AD

Oveall average

1 2 3 4 5 6 7

Time (seconds)

Figure 2: Individual differential signals of PLRs for blue

light pulses.

0.996 0.998 1.000 1.002 1.004 1.006

1.008

Pupil ratio (right / left)

70 basis

10 basis

2 3 4 5 6 7

Time (seconds)

Figure 3: A differential signal of PLRs for blue light pulses.

basis functions are introduced (the solid line), the ﬁt

is much better than in the case of 10 basis functions

(the dotted line), as shonwn in Figure 1. In this pa-

per, a condition with 70 basis functions is selected for

evaluating errors of reproduction and detection per-

formances, as detailed in the following analysis.

The individual differential ratio of both eyes is il-

lustrated in Figure 2. Most differential lines are con-

nected around ratio=1, though some differences are

observed around light pulses (1-3sec.) and at the end

of the observation period (8 sec.). The red lines in-

dicate patient cases (AD and MCI). A summation

of these waveforms is illustrated in Figure 3, show-

ing that even the changes in overall summation have

not been smoothened. In this case, 70 B-spline basis

functions are required to represent the waveform.

As an FDA technique, sets of coefﬁcients of basis

functions can be used to represent features of wave-

Development of a Procedure for Detecting Dementia Symptoms Using Features in Differential Waveforms of the Pupil Light Reﬂex

945

Table 1: Contingency table: Right trained / Left tested.

Classiﬁed

Participants NC MCI+AD Total

NC 54 6 60

MCI+AD 24 17 41

Total 78 23 101

form shapes of PLRs or differential PLRs. These sets

of features of PLR waveforms are used for classiﬁca-

tion of participants in the section which follows.

4 CLASSIFICATION RESULTS

Using the extracted features from waveform shapes

and logistic regression functions, patients are clas-

siﬁed as either patients (AD or MCI) or NC partic-

ipants. A logistic function presents the probability

of diagnosing patients using the set of features men-

tioned above. The waveform features are generated

from waveform shapes, and measurement conditions

such as using blue, red, or both colours of chromatic

lights, can be combined. However, the number of di-

mensions of features increases even as the LASSO

technique (Kawano et al., 2018) is introduced in or-

der to select features which are signiﬁcant.

In an evaluation of classiﬁcation performance,

cross validation of the irradiation of the right or left

eye is introduced and another set of data of left and

right eye measurements is used, for example, when

the regression model is trained using the data set of

the right eye and tested against the data set of the

left eye. In comparison of performance using several

combination of features of waveforms, the best per-

formance obtained to date is from a combination of

using a control condition, and differential ratios for

blue and red light, for a total of 270 dimensions of

features.

A contingency table classifying performance is

summarised in Table 1. The horizontal column repre-

sents the classiﬁed results, and the vertical row shows

the type of participant. Once again, the results present

the classiﬁcation of the data set of the left eye using a

regression model trained with the data set of the right

eye. In the results, 71 out of the 101 participants were

classiﬁed correctly using a threshold 0.5, and the ac-

curacy was 70%. In another regression model classiﬁ-

cation of the right eye, the accuracy was 65%. Though

an improvement of detection performance is required,

the differential ratios of irradiations using both blue

and red light pulseses may indicate the existence of

physiological impairment, which causes dementia.

Using the data set combination, detection occurs

0.5

0.50 1

MCI+AD

Probability: Left to Right

Probability: Right to Left

Figure 4: A differential signal of PLRs for blue light pulses.

Table 2: Contingency table using both detections.

Classiﬁed

Participants NC Either(+) Both(+) Total

NC 52 6 2 60

MCI+AD 22 11 8 41

Total 74 17 10 101

twice for each participant. are produced for each par-

ticipant. These probabilities are summarised in Figure

4. The horizontal axis represents the probability of

patients trained with the right eye data set and tested

using the left eye data set, and the vertical axis repre-

sents probabilities of patients trained with the left eye

data set and tested using the right eye data set. In the

ﬁgure, the blue dots represent NC participants, and

the orange dots represent MCI or AD patients. The

threshold of 0.5 is represented by ﬁne lines along the

horizontal and vertical axes. The plots of many par-

ticipants overlap in the centre of the graph.

The two procedures produce two probabilities, so

participants are classiﬁed again using the logical OR

of the two classiﬁcation results. These results are

summarised in Table 2. Using this procedure to merge

the results, the number of patient detections increases

slightly. However, overall accuracy remains at the

same level such as 70% in the case mentioned above.

In order to evaluate the contribution of ex-

tracted waveform features, the feature values are sum-

marised, as shown in Figure 5. The horizontal axis

represents time such as a dimensional sequence, and

the vertical axis represents the weight values. Mean

weights of the three waveforms for control, blue, and

red differentials of the data sets of the right and left

eyes are summarised. The weights on both sides are

almost the same. Regarding the changes in weights,

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

946

1 2 3 4 5 6 7

Time (seconds)

Weight

Figure 5: Comparison of coefﬁcients of PLRs for control

(black), and differentials for blue and red light pulses.

the amounts are relarively large during the light pulses

and at the ends of the observations. There are some

differences between blue and red light pulses. Also,

the weights for red light pulses are maintained after

the light pulse. As a phenomenon, post-illumination

pupil response (PIPR) is observed in the case of red

light irradiation. These differences may suggest that

some weight patterns at the ends of the reaction ob-

servations show some increase. These results indicate

that the most correct detections are accomplished by

using the overall features of PLR waveform shapes.

Further feature selection and additional features of

waveforms from around the irradiation of light pulses

when constriction of the pupil does not occur may be

required in order to improve performance. The fac-

tor of the basis function should be also evaluated to

improve performance. In addition, MCI and AD de-

tection performance will be examined. As the number

of patients is limited, a different detection procedure

will be required. A detailed analysis of performance

improvements will be a subject of our further study.

5 SUMMARY

The hypothesis that pupillary light reﬂex (PLR) wave-

form shapes may provide feature metrics about pa-

tients with dementia was tested using a functional data

analysis technique to extract features of overall wave-

forms. Detection performances of a clinical survey

data consisting of PLRs of chromatic light pulses to

either eye was evaluated. The features of differential

waveform shapes of each eye contributed to their clas-

siﬁcation. The feature of waveform shapes also pre-

sented physiological features of PLRs of chromatic

light pulses.

REFERENCES

Chaitanuwong, P., Singhanetr, P., Chainakul, M., Arjkong-

harn, N., Ruamviboonsuk, P., and Grzybowski, A.

(2023). Potential ocular biomarkers for early detection

of Alzheimer’s disease and their roles in artiﬁcial in-

telligence studies. Neurology and Therapy, 12:1517–

1532.

Chougule, P. S., Najjar, R. P., Finkelstein, M. T., Kandiah,

N., and Milea, D. (2019). Light-induced pupillary re-

sponses in Alzheimer’s disease. Frontiers in Neurol-

ogy, 10(360):1–12.

Gamlin, P. D., McDougal, D. H., and Pokorny, J. (2007).

Human and macaque pupil responses driven by

melanopsin-containing retinal ganglion cells. Vision

Research, 47:946–954.

Kawano, S., Matsui, H., and Hirose, K. (2018). Statistical

Modeling via Sparse Estimation. KYORITSU Publi-

cation, Tokyo, Japan.

Kawasaki, A. and Kardon, R. H. (2007). Intrinsically pho-

tosensitive retinal ganglion cells. Journal of Neuro-

Ophthalmology, 27:195–204.

Kelbsch, C., Strasser, T., Chen, Y., Feigl, B., Gamlin, P. D.,

Kardon, R., Peters, T., Roeckeln, K. A., Steinhauer,

S. R., Szabadi, E., Zele, A. J., Wilhelm, H., and Wil-

helm, B. J. (2019). Standards in pupillography. Fron-

tiers in Neurology, 10(129):1–26.

López, O. A. M., López, A. M., and Crossa, J. (2022). Mul-

tivariate Statistical Machine Learning Methods for

Genomic Prediction. Springer, New York, USA.

McDougal, D. H. and Gamlin, P. D. (2015). Auto-

nomic control of the eye. Comprehensive Physiology,

5(1):439–473.

Molitor, R. J., Ko, P. C., and Ally, B. A. (2015). Eye move-

ments in Alzheimer’s disease. Journal of Altzheimers

Disease, 44(1):1–12.

Nakayama, M., Nowak, W., and Krecicki, T. (2024a). De-

tecting Alzheimers patients using features in differen-

tial waveforms of pupil light reﬂex for chromatic stim-

uli. EAI Endorsed Transactions on Intelligent Systems

and Machine Learning Applications, 1:1–7.

Nakayama, M., Nowak, W., and Zarowska, A. (2022). De-

tecting symptoms of dementia in elderly persons us-

ing features of pupil light reﬂex. In Proceedings of

the Federated Conference on Computer Science and

Information Systems (FedCSIS), pages 745–749.

Nakayama, M., Nowak, W., and Zarowska, A. (2023a). Pre-

diction procedure for dementia levels based on wave-

form features of binocular pupil light reﬂex. In Pro-

ceedings of ACM Eye-Tracking Research & Applica-

tions (ETRA), pages 1–6.

Nakayama, M., Nowak, W., and Zarowska, A. (2023b).

Using features of PLRs to chromatic light pulse ir-

radiations of either eye to detect dementia in elderly

persons. Communication Papers of the 18th Confer-

ence on Computer, Science and Intelligent Systems,

37:201–207.

Nakayama, M., Nowak, W., and Zarowska, A. (2024b).

Extracting dementia symptoms in elderly using ocu-

lar motor activity during pupil light reﬂexes to chro-

Development of a Procedure for Detecting Dementia Symptoms Using Features in Differential Waveforms of the Pupil Light Reﬂex

947

matic light pulses on either eye. In Proceedings

of International Conference Information Visualisation

(IV2024), pages 1–6.

Nie, J., Qiu, Q., Philips, M., Sun, L., Yan, F., Lin, X., Xiao,

S., and Li, X. (2020). Early diagnosis of mild cogni-

tive impairment based on eye movement parameters in

an aging Chinese population. frontiers in Aging Neu-

roscience, 29(Article 221):1–11.

Nowak, W., Nakayama, M., Kr˛ecicki, T., and Hachoł,

A. (2020). Detection procedures for patients of

Alzheimer’s disease using waveform features of pupil

light reﬂex in response to chromatic stimuli. EAI En-

dorsed Transactions on Pervasive Health and Tech-

nology, 6:1–11. e6.

Nowak, W., Nakayama, M., Kr˛ecicki, T., Trypka, E., An-

drzejak, A., and Hachoł, A. (2019). Analysis for ex-

tracted features of pupil light reﬂex to chromatic stim-

uli in Alzheimer’s patients. EAI Endorsed Transac-

tions on Pervasive Health and Technology, 5:1–10. e4.

Nowak, W., Nakayama, M., Trypka, E., and Zarowska, A.

(2021). Classiﬁcation of Alzheimer’s disease patients

using metrics of oculo-motors. In Proceedings of the

Federated Conference on Computer Science and In-

formation Systems (FedCSIS), pages 403–407.

Ramsay, J., Hooker, G., and Graves, S. (2009). Functional

Data Analysis with R and MATLAB. Springer, New

York, USA.

Tombaugh, T., McDowell, I., Kristjansson, B., and Hubley,

A. (1996). Mini-mental state examination (MMSE)

and the modiﬁed MMSE (3MS): A psychometric

comparison and normative data. Psychological As-

sessment, 8(1):48–59.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

948