BAYESIAN-BASED EARLY DETECTION OF COGNITIVE

IMPAIRMENT IN ELDERLY USING fNIRS SIGNALS DURING

COGNITIVE TESTS

Shohei Kato

, Hidetoshi Endo

and Yuta Suzuki

Dept. of Computer Science and Engineering, Nagoya Institute of Technology

Gokiso-cho, Showa-ku, Nagoya 466-8555, Japan

Dept. of Comprehensive Geriatic Medicine, National Center for Geriatics and Gerontology

35 Gengo, Morioka-machi, Obu, Aichi 474-8511, Japan

Dept. of Computer Science and Engineering, Nagoya Institute of Technology, Nagoya, Japan

Keywords:

Early detection of dementia, Functional near infrared spectroscopy, fNIRS, Bayesian classiﬁer.

Abstract:

This paper presents a new trial approach to early detection of dementia in the elderly with the use of functional

brain imaging during cognitive tests. We have developed a non-invasive screening system of the elderly with

cognitive impairment. In addition of our previous research of speech-prosody based data-mining approach, we

had started the measurement of functional brain imaging for patient having a cognitive test by using functional

near-infrared spectroscopy (fNIRS). We had collected 42 CHs fNIRS signals on frontal and right and left

temporal areas from 50 elderly participants (18 males and 32 females between ages of 64 to 92) during cogni-

tive tests in a specialized medical institute. We propose a Bayesian classiﬁer, which can discriminate among

elderly individuals with three clinical groups: normal cognitive abilities (NL), patients with mild cognitive

impairment (MCI), and Alzheimer’s disease (AD). The Bayesian classiﬁer has two phases on the assumption

of screening process, that ﬁrstly checks whether a suspicion of the cognitive impairment (CI) or not (NL) from

given fNIRS signals; if any, and then secondly judges the degree of the impairment: MCI or AD. This paper

also reports the examination of the detection performance by cross-validation, and discusses the effectiveness

of this study for early detection of cognitive impairment in elderly subjects. Consequently, empirical results

that both the accuracy rate of AD and the predictive value of NL are equal to or more than 90%. This suggests

that proposed approach is adequate practical to screen the elderly with cognitive impairment.

1 INTRODUCTION

It is no doubt about abrupt increase in elderly patients

with dementia due to growing super-aging society in

developed countries. Research and development of

new dementia medications is accelerated. Develop-

ment of the early detection methods for dementia that

are both sensitive and speciﬁc is also very important

as a diagnostic tool.

To screen for dementia and cognitive impair-

ment, a questionnaire test such as Mini-Mental State

Examination (MMSE) (Folstein et al., 1975), Re-

vised Hasegawa’s Dementia Scale (HDS-R) (Imai and

Hasegawa, 1994), Clinical Dementia Rating (CDR)

(Morris, 1993), and Memory Impairment Screen

(MIS) (Buschke et al., 1999), is commonly used in

addition to a neurophysiological test (Zhang et al.,

2011) (e.g., using MRI (de Leon et al., 2004), FDG-

PET (Mosconi et al., 2010), and CSF biomark-

ers (de Leon et al., 2007)). Questionnaire tests have

some disadvantages and their use is limited in the

clinic. The MMSE, HDS-R, and CDR are more time-

consuming than a general practitioner’s consultation.

In general, the questionnaire cannot completely dis-

miss the inﬂuence of education, social class, and gen-

der difference on the results. In addition, there is a

possibility that practitioner subjectivity may affect the

scoring. Thus, we believe that the development of a

simple, non-invasive examination that is objectiveand

combined with a physiological test could enables the

early detection of dementia in a broad population.

In our previous study, we have studied a novel ap-

proach to the early detection of cognitive impairment

in the elderly (Kato et al., 2011), in which we fo-

cused on the prosodic features of speech sound dur-

ing the subject’s answers to the questionnaire. The

method had an advantage that enables everyone to

check his/her own cognitive ability anywhere because

118

Kato S., Endo H. and Suzuki Y..

BAYESIAN-BASED EARLY DETECTION OF COGNITIVE IMPAIRMENT IN ELDERLY USING fNIRS SIGNALS DURING COGNITIVE TESTS.

DOI: 10.5220/0003790001180124

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 118-124

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Table 1: A Breakdown list of participants (N=50).

Age 64-70 71-75 76-80 81-85 86-92 Total

Male 3(2,0,1) 2(1,1,0) 4(3,1,0) 7(1,4,2) 2(0,0,2) 18(7,6,5)

Female 7(4,2,1) 7(5,2,0) 8(2,5,1) 6(2,1,3) 4(1,3,0) 32(14,13,5)

Subtotal 10(6,2,2) 9(6,3,0) 12(5,6,1) 13(3,5,5) 6(1,3,2) 50(21,19,10)

Value in bracket means the number of subjects in NL, MCI, AD clinical groups.

300sec 300sec

HDS-R test

60sec

rest

60sec

Reminiscence1

60sec

rest

60sec

Reminiscence2

60sec 60sec

Working Memory1

60sec

rest

60sec

Face recall

Listening Talking

Reminiscence3

Watching

60sec

rest

60sec

rest

Category recall

rest

Working Memory2

Reading span

Working Memory3

30sec

rest

Talking about hometown,

childhood, and school

00:00:00 00:14:00

00:10:00

00:14:00 00:22:30

timeline

Figure 1: Block design task of cognitive tests.

of using speech signals only. The method is effec-

tive for the ﬁrst step of screening for dementia, but,

however, it has limitations of the reliability because

the method does not measure brain function. On the

other hand, a neurophysiological test, such as using

MRI, FDG-PET, and CSF biomarkers, imposes severe

constraint on a subject, for instance, pain at obtain-

ing cerebral spinal ﬂuid, radiation exposure, physical

restraint and so on. This is a disadvantage in early

screening, which should covers all elderlies.

In this study, we focus on functional near-infrared

spectroscopy (fNIRS) as a brain function measure-

ment system, which can eliminate physical restraint

from a subject by non-invasive procedures, and de-

velop a prototype for computer-aided diagnosis of

cognitive impairment in the elderly with the use of

fNIRS signals during cognitive tests. In this paper,

we present signal processing technique of feature ex-

traction and selection for hyper-dimensional time se-

ries data of fNIRS signals, and propose the two-phase

Bayesian classiﬁer for discriminating among elderly

individuals with three clinical groups. In addition,

we addressed the effectiveness of proposed method in

discriminating among elderly individuals with normal

cognitive abilities (NL), patients with mild cognitive

impairment (MCI), and Alzheimer’s disease (AD)

2 METHOD

2.1 Participants

Fifty Japanese subjects (18 males and 32 females be-

tween the ages of 64 and 92 years) participated in this

study. Table 1 shows the breakdown list of partici-

pants. In this study, all participants are clinically con-

ditioned that CDR of a participant in MCI group and

AD group corresponds to 0.5 and 1, respectively.

2.2 Cognitive Tests

To measure brain function of an elderly during vari-

ous cognitive tests including HDS-R, we have made

a block designed task shown in Fig. 1, and then con-

ducted simultaneous voice-fNIRS measurement dur-

ing cognitive tests. Firstly a participant talks about

the topics of hometown and childhood and answers

for an HDS-R questionnaire test for ten minutes. And

then, he/she does three reminiscence tasks (1. listen-

ing, 2. talking, 3. watching) and three working mem-

ory tasks (1. category recall, 2. reading span, 3. face

recall) for twelve minutes. These six tasks are done

for 60 seconds after rest gazing at a single point on

the display for 60 seconds interval.

2.3 fNIRS Measurement

Functional near-infrared spectroscopy (fNIRS) can

measure neural activity of the cerebral cortex using

infrared rays that are safe to living organisms (Vill-

ringer and Chance, 1997). fNIRS monitors regional

relative changes of oxy/deoxygenated hemoglobin

concentration to measure cortical activation utiliz-

ing the tight coupling between neural activity and

regional cerebral blood ﬂow (Villringer and Firnaﬂ,

1995). This measurement method requires only com-

pact experimental systems and can eliminate physical

restraint from a subject by non-invasive procedures

(Fig. 2).

We used the fNIRS topography system FOIRE-

3000 Near-Infrared Brain Function Imaging System

BAYESIAN-BASED EARLY DETECTION OF COGNITIVE IMPAIRMENT IN ELDERLY USING fNIRS SIGNALS

DURING COGNITIVE TESTS

119

fNIRS

indicator

Display

speaker

mic

partition

seated position

Figure 2: Snapshot of fNIRS measurement of an elderly

participant having a cognitive test.

(Shimadzu, Kyoto, Japan), which uses near-infrared

light with wavelengths of 780, 805, and 830 nm. We

set 16 illuminators and 15 detectors in lattice pattern

to form 42 channels (CHs) (22 CHs on frontal lobe,

10 CHs on right parietal and temporal lobe, 10 CHs

on left parietal and temporal lobe) shown in Fig. 3.

N: channel ID

: illuminator

: detector

3 4

5 6 7 8 9

10 11 12 13

14 15 16 17 18

20 21

Figure 3: Channel arrangement of fNIRS measurement.

2.4 Statistical Tests of fNIRS Signals

Preliminary to development of the screening tool, we

have conduct statistical tests of between-group signif-

icant differences using fNIRS signals of oxy-Hb dur-

ing working memory task (1. category recall). We

used two-tailed t-test with signiﬁcance level of (P<

0.001) after applying Bonferroni’s adjustment (1/42).

Fig. 4 shows the results of t-test for signiﬁcant dif-

ferences in channel-wise fNIRS signals between any

single pair from NL, MCI, and AD groups. The CHs

that exhibited signiﬁcant oxy-Hb increase are col-

ored according to the t-values, as shown in the color

bar, while those below the threshold are indicated in

gray. The results indicate the signiﬁcant difference of

fNIRS signals during cognitive test between normal

group and disease groups. This suggests that fNIRS

signals during cognitive test have potential for detec-

tion of cognitive impairment in elderly patients. Ad-

ditionally, for fNIRs signals during rest, there are no

CHs with signiﬁcant difference between any single

pair from NL, MCI, and AD groups.

-24

24-12 12

t-value

(a) NL group–MCI group

t-value

-30

30-15 15

(b) NL group–AD group

-10

10-5 5

t-value

Figure 4: Results of t-test for signiﬁcant differences in

channel-wise fNIRS signals between any single pair from

NL, MCI, and AD groups.

3 CLASSIFICATION OF NL, MCI,

AD GROUPS

The section describes a Bayesian classiﬁer, which

can discriminate among elderly individuals with three

clinical groups: normal cognitive abilities (NL), pa-

tients with mild cognitive impairment (MCI), and

Alzheimer’s disease (AD). To design algorithm for

computer-aided diagnosis of cognitive impairment in

the elderly, we consider the screening process by a

specialist in geriatrics. We thus propose a two-phase

Bayesian classiﬁer shown in Fig. 5 on the assump-

tion of screening process, that ﬁrstly checks whether

a suspicion of the cognitive impairment (CI) or not

(NL) from given fNIRS signals; if any, and then sec-

ondly judges the degree of the impairment: MCI or

AD.

3.1 Primitive Analysis of fNIRS Signals

In advance of Bayesian classiﬁcation, we make a

primitive signal processing fNIRS signals shown in

Fig. 6. Firstly, we make ﬁve fNIRS signals every

channels such that noise is reduced by channel-wise

smoothing through three low-pass ﬁlters and differ-

ence ﬁlters (Fig. 7).

F1 (cutoff freq. 1.92[Hz]): to remove noise arisen from

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

120

CH1

CH2

CH42

Fil2-3

Fil1

Fil2-3

Fil1

Fil2-3

Fil1

CH1

Fil2-3

Fil1

CH2

CH42

Fil2-3

Fil1

Fil2-3

Fil1

Fc Lr

Original Signal

Primitive Analysis

Feature Extraction

Averaging

fNIRS fitler

Figure 6: The outline of primitive analysis of fNIRS signals.

fNIRS signals

NB Classifier (1st phase)

NB Classifier (2nd phase)

NL / CI

MCI / AD

NL MCI AD

(42CH oxy-Hb)

cognitive impairment

Feature Extraction

spectoral, statistics 77 features

Primitive Analysis

CH-wise low-path and/or difference filters,

domain averaging

Figure 5: Classiﬁcation of NL/MCI/AD by two-phase

Bayesian Classiﬁer.

environmental light.

F2 (cutoff freq. 0.96[Hz]): to remove background noise

arisen from biosignal such as pulse wave and blood

pressure.

F3 (cutoff freq. 0.48[Hz]): to remove noise arisen from

body movement such as jaw, eye, neck and so on.

F1-F3: to subtract F3 from F1.

F2-F3: to subtract F3 from F2.

Secondly, we segregate 42 CHs into the following

seven brain areas and then make signal averaging that

integrates fNIRS signals within each of the areas.

Fr: 7 CHs on the right side of frontal lobe (CH:

1,5,6,10,14,15,19).

0 2 4 6 8 10 12 14 16 18

−2

−1.5

−1

−0.5

0.5

1.5

x 10

−3

0 2 4 6 8 10 12 14 16 18

−4

−3

−2

−1

x 10

−3

0 2 4 6 8 10 12 14 16 18

−1.5

−1

−0.5

0.5

1.5

x 10

−3

0 2 4 6 8 10 12 14 16 18

−3

−2

−1

x 10

−3

0 2 4 6 8 10 12 14 16 18

−4

−3

−2

−1

x 10

−3

0 2 4 6 8 10 12 14 16 18

−5

−4

−3

−2

−1

x 10

−3

Original data

Low-pass filter

Filter 2 data Filter 3 data

Filter 1 - 3 data Filter 2 - 3 data

Filter 1 data

0 20

0 20 0 20 0 20

0 20 0 20

(sec)

Difference filter

fNIRS filter

Figure 7: A ﬁlter design in fNIRS primitive analysis.

Fc: 8 CHs on the central part of frontal lobe (CH:

2,3,7,11,12,16,20,21).

Fl: 7 CHs on the left side of frontal lobe (CH:

4,8,9,13,17,18,22).

Rf: 5 CHs on the front of right parietal lobe (CH:

23,24,26,27,30).

Rr: 5 CHs on the rear of right temporal lobe (CH:

25,28,29,31,32).

Lf: 5 CHs on the front of left parietal lobe (CH:

33,34,36,37,40).

Lr: 5 CHs on the rear of left temporal lobe (CH:

35,38,39,41,42).

BAYESIAN-BASED EARLY DETECTION OF COGNITIVE IMPAIRMENT IN ELDERLY USING fNIRS SIGNALS

DURING COGNITIVE TESTS

121

Table 2: fNIRS feature candidates.

fNIRS ﬁltered Feature / Statistics

Filter 1 (F1) Mean value (mean)

Fundamental Frequency (f0)

Centroidal Frequency (fc)

Filter 3 (F3) Maximum value (max)

Minimum value (min)

Variance (var)

Mean value (mean)

Fundamental Frequency (f0)

Gradient of the linear regression

line (gr)

Filter1-3 (F1-3) Variance (var)

Filter2-3 (F2-3) Variance (var)

3.2 Extraction of fNIRS Features

We enumerate features that represents ﬂuctuations of

regional cerebral blood ﬂow if it is the slightest effec-

tive in detection of cognitive impairment, and extract

11 features shown in Table 2 from fNIRS signals in

each of the seven brain areas.

3.3 Bayesian Classiﬁer

In this paper, we adopted naive Bayes classiﬁer

(NB), (Langley et al., 1992) which is a simple

Bayesian classiﬁer with strong independence assump-

tion of attributes (Domingos and Pazzani, 1996). We

construct two classiﬁers: NB

NL/CI

, which checks

whether a suspicion of the cognitive impairment (CI)

or not (NL) at the ﬁrst phase, if any suspicion, and

MCI/AD

, which judges the degree of the impair-

ment (MCI or AD) at the second phase.

In our strategy for feature extraction, all of the 77

fNIRS features described above may not be equally

useful and important for discrimination among NL,

MCI, and AD. In this study, we conduct system-

atic feature selection by using the forward stepwise

(FSW) method (Draper and Smith, 1998), which is

the most popular form of feature selection in statistics

and consists of a combination of the forward selection

and backward elimination methods. FSW is a greedy

algorithm that adds the best feature (or deletes the

worst feature) during each round. We make a model

selection method based on the criterion of accuracy

rate of the classiﬁcation.

4 CLASSIFICATION

ASSESSMENT

We have examined discrimination performance by

-3 -2 -1 0 1 2 3 4 5

-3

-2

-1

F1 fc

F3 max

Normal cognitive ability

Cognitive Impairment

Figure 8: Distributions of NL/CI estimated by classiﬁer

NL/CI

-2 -1 0 1 2 3 4 5 6

-4

-2

-1

F1 mean

F1−3 var

Alzheimer’s Disease

Mild Cognitive Impairment

-3

Figure 9: Distributions of MCI/AD estimated by classiﬁer

MCI/AD

modeling two-phase Bayesian classiﬁers for discrim-

inating among elderly individuals with NL, MCI, and

AD, by using fNIRS signals of oxy-Hb during work-

ing memory task (1. category recall) (see Fig. 1) col-

lected from 50 participants (see Table 1). Table 3

shows the selected fNIRS features by each of NB

classiﬁers. To evaluate detection performance, we

adopted leave-one-out cross-validation.

Table 3: Selected fNIRS features.

Classiﬁer Selected Feature

NL/CI

Fr F3 max, Lr F1 fc

MCI/AD

Lf F1-3 var, Lf F1 mean, Fc F1-3 var

Fig. 8 and Fig. 9 show the distributions of NL

group / CI group and MCI group / AD group by clas-

siﬁers NB

NL/CI

and NB

MCI/AD

, respectively. In the

ﬁgures, 250 samples (145 samples in Fig. 9), such

that fNIRS signals are analyzed after divided into ﬁve

spans, are plotted. fNIRS features are mean and vari-

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

122

Table 4: Classiﬁcation results.

clinical

detection

NL MCI AD accuracy

NL 11 7 3 52.4%

MCI 1 14 4 73.7%

AD 0 1 9 90.0%

predictive value 91.7% 63.6% 56.3 68.0%

ance normalize in each variable.

Table 4 shows the confusion matrices and the

statistics of classiﬁcation results using two-phase

classiﬁers consist of NB

NL/CI

and NB

MCI/AD

. The re-

sults indicate that both the accuracy rate of AD and

the predictive value of NL are equal to or more than

90%. This means that no subject in AD groups are

misclassiﬁed into NL group (only one is misclassiﬁed

into MCI group), and that subjects classiﬁed into NL

group are not all patient with AD (only one should be

in MCI group). This suggests that proposed approach

is adequate practical to screen the elderly with cogni-

tive impairment. The results that the accuracy rate of

MCI is 73.7% and that most of subjects misclassiﬁed

are classiﬁed into AD group are both relative accept-

able performance for screening tool.

5 CONCLUSIONS

We developed a new technology for early detec-

tion of cognitive impairment in the elderly, focus-

ing on the brain activity during cognitive task. The

detection method is based on the data mining ap-

proach using Bayesian classiﬁcation and is simple and

non-invasiveprocedure using functional near-infrared

spectroscopy (fNIRS). We proposed a Bayesian clas-

siﬁer using fNIRS signals, which can discriminate

among elderly individuals with three clinical groups:

normal cognitive abilities (NL), patients with mild

cognitive impairment (MCI), and Alzheimer’s disease

(AD). This paper also reported the examination of the

detection performance by cross-validation, and the re-

sults that both the accuracy rate of AD and the predic-

tive value of NL are equal to or more than 90%. Con-

sequently, the empirical results suggested that pro-

posed approach is adequate practical to screen the el-

derly with cognitive impairment.

ACKNOWLEDGMENTS

We are grateful to SHIMADZU Corporation, Na-

tional Center for Geriatrics and Gerontology, and If-

com Co. Ltd. for fNIRS measurement system, clin-

ical data collection environment, and data collec-

tion, respectively. This work was supported in part

by SENTAN, Japan Science and Technology Agency

(JST), and part by Adaptable and Seamless Tech-

nology Transfer Program through target-driven R&D,

JST, and part by Suzuken Memorial Foundation.

REFERENCES

Buschke, H., Kuslansky, G., Katz, M., Stewart, W. F., Sli-

winski, M. J., Eckholdt, H. M., and Lipton, R. B.

(1999). Screening for dementia with the Memory Im-

pairment Screen. Neurology, 52(2):231–238.

de Leon, M. J., DeSanti, S., Zinkowski, R., Mehta, P. D.,

Pratico, D., Segal, S., Clark, C., Kerkman, D., De-

Bernardis, J., Li, J., Lair, L., Reisberg, B., Tsui, W.,

and Rusinek, H. (2004). Mri and csf studies in the

early diagnosis of alzheimer’s disease. Journal of In-

ternal Medicine, 256(3):205–223.

de Leon, M. J., Mosconi, L., De Santi, K. B. S., Zinkowski,

R., Mehta, P. D., Pratico, D., Tsui, W., Saint Louis,

L. A., Sobanska, L., Brys, M., Li, Y., Rich, K., Rinne,

J., and Rusinek, H. (2007). Imaging and CSF stud-

ies in the preclinical diagnosis of Alzheimer’s disease.

Annals of New York Academy of Sciences, 1097:114–

145.

Domingos, P. and Pazzani, M. (1996). Beyond Indepen-

dence: Conditions for the Optimality of the Simple

Bayesian Classiﬁer. In Proc. of International Confer-

ence on Machine Learning, pages 105–112.

Draper, N. and Smith, H. (1998). Applied Regression Anal-

ysis (3rd edition). John Wiley & Sons.

Folstein, M. F., Folstein, S. E., and McHugh, P. R. (1975).

“Mini-Mental State”: A practical method for grading

the cognitive state of patients for the clinician. J. Psy-

chiat. Res, 12(3):189–198.

Imai, Y. and Hasegawa, K. (1994). The revised Hasegawa’s

Dementia Scale (HDS-R): evaluation of its usefulness

as a screening test for dementia. J. Hong Kong Coll.

Psychiatr., 4(SP2):20–24.

Kato, S., Suzuki, Y., Kobayashi, A., Kojima, T., Itoh, H.,

and Homma, A. (2011). Statistical analysis of the

signal and prosodic sign of cognitive impairment in

elderly-speech: a preliminary study. In Proc. of 5th

International Jount Conference on Biomedical Engi-

neering Systems and Technologies (BIOSTEC 2011),

pages 322–327.

Langley, P., Iba, W., and Thompson, K. (1992). An anal-

ysis of Bayesian classiﬁers. In Proc. of The Tenth

National Conference on Artiﬁcial Intelligence (AAAI-

92), pages 223–228.

Morris, J. C. (1993). The Clinical Dementia Rating

(CDR): Current version and scoring rules. Neurology,

43(11):2412–2414.

Mosconi, L., Berti, V., Glodzik, L., Pupi, A., De Santi, S.,

and de Leon, M. J. (2010). Pre-clinical detection of

Alzheimer’s disease using FDG-PET, with or with-

out amyloid imaging. Journal of Alzheimers’ Disease,

20(3):843–854.

BAYESIAN-BASED EARLY DETECTION OF COGNITIVE IMPAIRMENT IN ELDERLY USING fNIRS SIGNALS

DURING COGNITIVE TESTS

123

Villringer, A. and Chance, B. (1997). Non-invasive optical

spectroscopy and imaging of human brain function.

Trends Nurosci., 20:435–442.

Villringer, A. and Firnaﬂ, U. (1995). Coupling of brain

activity and cerebral blood ﬂow: basis of functional

nuroimaging. Cerebrovasc. Brain Metab. Rev., 7:240–

276.

Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., and the

Alzheimer’s Disease Neuroimaging Initiative (2011).

Multimodal classiﬁcation of alzheimer’s disease and

mild cognitive impairment. Journal of Neuroimage,

55(3):856–867.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

124