Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s

Disease Detection Using EEG

Maxime Bedoin

, Bernadette Dorizzi

, Jérôme Boudy

, Kiyoka Kinugawa

and Nesma Houmani

Télécom SudParis, Institut Polytechnique de Paris, 91120 Palaiseau, France

Sorbonne Universite, CNRS, UMR Biological Adaptation and Aging, AP-HP, Charles Foix Hospital,

F-94200 Ivry-sur-Seine, France

Keywords:

EEG Signals, Functional Connectivity, Epoch Duration, Score Fusion, Frequency Bands, Alzheimer’s Disease.

Abstract:

In this work, we propose a fusion approach to analyze EEG signals for the discrimination between patients

with Subjective Cognitive Impairment (SCI) and patients suffering from Alzheimer’s Disease (AD). In this

framework, we analyze EEG signals at different epoch durations, following a multi-scale procedure, and in

different frequency bands, using Phase-Lag Index (PLI) and Dynamic Time Warping (DTW) for functional

connectivity measurement. Experiments show that our fusion approach leads to an improvement of classiﬁca-

tion results, reaching an AUC of 0.902 with PLI, and 0.894 with DTW; whereas we obtain an AUC of 0.845

with PLI and 0.801 with DTW when computing connectivity on the entire signal, as usually done in the lit-

erature. Furthermore, with the additional fusion of the scores obtained at different frequency bands, we reach

the best performance with both PLI (AUC=0.95, Accuracy=91%) and DTW (AUC=0.98, Accuracy=95%).

Finally, we investigate the generalization capability of our proposal on a test set. We found that our fusion

scheme allows obtaining better classiﬁcation results comparatively to when we consider the entire signal to

compute functional connectivity.

1 INTRODUCTION

Dementia encompasses a group of disorders caused

by the progressive dysfunction and death of brain

cells. It affects memory, language, executive func-

tions and other abilities, beyond what is considered a

normal age-related change, with a signiﬁcant impact

on daily functioning.

The diagnosis of dementia relies on a series of

clinical tests, including neurological tests and med-

ical recordings. Neuroimaging techniques, such as

Magnetic Resonance Imaging and Positron Emis-

sion Tomography Scanner are also used to assess the

brain damage. These imaging tools are costly, non-

portable, necessitate a hospital setting and the pro-

cedure can be complex and stressful for the patient.

Moreover, they are unsuitable to follow-up the ongo-

ing brain dynamics.

Electroencephalography (EEG) is a relatively

low-cost and non-invasive neuroimaging technique

that can be used either at clinical or home-based set-

ting. EEG allows capturing brain dynamics with

excellent time resolution in the millisecond range,

which provides valuable insights into the brain’s func-

tioning. EEG brainwaves reﬂect the collective electri-

cal impulses generated by the synchronized ﬁring of

neurons in the brain. These brainwaves manifest as

rhythmic oscillations, typically categorized into sev-

eral frequency bands; each associated with speciﬁc

cognitive states and mental activities.

Alzheimer’s disease (AD) is the most prevalent

form of dementia, accounting for as much as 70% of

all dementia cases (Castellani et al., 2010). It is esti-

mated that 25 million people worldwide in 2010 were

affected by AD. This ﬁgure will continue to grow, es-

pecially in Western countries as the world’s popula-

tion continues to age (Castellani et al., 2010), thus

entailing high health care costs and a considerable hu-

man toll.

AD is characterized by progressive and irre-

versible brain damage. Disease progression typically

spans over several years to decades. First, at the pre-

symptomatic stage, the person has no symptom of AD

and appears normal and unaffected, but AD-related

changes are already taking place in the brain (Bossers

et al., 2010). At the preclinical stage, the person

may go through cognitive impairment and changes,

but that are not detectable with standard tests yet.

Bedoin, M., Dorizzi, B., Boudy, J., Kinugawa, K. and Houmani, N.

Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s Disease Detection Using EEG.

DOI: 10.5220/0013169700003911

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 741-751

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

741

As there is no evidence of objective cognitive im-

pairment, this condition is called “Subjective Cogni-

tive Impairment" (SCI) (Dubois et al., 2009). Scien-

tiﬁc community shows an increasing interest to this

stage because of evidence of its association with an

increased risk of future objective cognitive decline.

Then, the patient may transition to Mild Cognitive

Impairment (MCI) stage, on which memory and cog-

nitive deﬁcits start to be noticeable and measurable

on cognitive tests. The patient then progresses to the

Mild AD stage with notable cognitive deﬁcits. These

deﬁcits continue to worsen in moderate and end-stage

disease (Alz, 2024). There can be variations in the

progression of AD and the precise symptoms can vary

between patients.

Many studies have exploited resting-state EEG for

AD diagnosis based on functional connectivity anal-

ysis, since AD is considered as a disconnection syn-

drome (Cassani et al., 2018). Various measures have

been proposed to quantify functional connectivity,

among them Phase-Lag Index (Abazid et al., 2021;

Kasakawa et al., 2016; Stam et al., 2007), Phase Co-

herence (Dauwels et al., 2010b), Mean Square Coher-

ence (Dauwels et al., 2010a), and Mutual Information

(Abazid et al., 2021; Jeong et al., 2001). Functional

connectivity values are computed pairwise between

EEG signals. To analyze them, the common paradigm

in the literature is to average the connectivity values

per EEG electrode, or to group the electrodes into re-

gions and to average the connectivity values per re-

gion (Dauwels et al., 2010b; Houmani et al., 2021).

The resulting average values are then used as EEG-

based features for AD detection, either in a statistical

analysis or a model-based classiﬁcation.

Given that the EEG signal is not stationary, it

is usually segmented into epochs to calculate con-

nectivity. The use of epochs for the extraction of

EEG markers and their subsequent averaging across

epochs has been shown to ﬁnely characterize con-

nectivity; thereby enhancing the classiﬁcation perfor-

mance. However, a review of the literature reveals

a considerable variability in the length of the EEG

epochs employed in different studies (Cassani et al.,

2018): from even less than one second (Knyazeva

et al., 2010) to 30 seconds (Lee et al., 2010). The

number of epochs used for EEG analysis also varies

substantially in the literature (Cassani et al., 2018).

Some studies analyze only one epoch (Abazid et al.,

2021) but others can analyze as much as 200 per pa-

tient (Knyazeva et al., 2010). This is especially ob-

served in research works exploiting deep learning al-

gorithms. Generally, there is no explanation about

the choice of the epoch duration and the number of

epochs. Furthermore, the length of the entire EEG

signal employed is seldom speciﬁed in the literature,

which may give the impression that this parameter has

a limited impact on the results. All theses facts col-

lectively contribute to the variability and the inconsis-

tency of the results presented across studies.

To our knowledge, no study in the literature has

investigated the impact of signal duration and varia-

tion in the epoch duration on the classiﬁcation per-

formance. In this work, we aim at ﬁlling this gap.

First, by studying the impact of varying EEG epoch

duration on the classiﬁcation of SCI and AD patients,

based on functional connectivity. Second, by propos-

ing a novel multi-scale fusion approach for EEG anal-

ysis combining classiﬁcation score outputs derived

from data analysis at different epoch durations. We

investigate the effectiveness of our proposal using two

alternative connectivity measures: Phase-Lag Index

that is frequently used in the literature for AD diagno-

sis, and Dynamic Time Warping that allows two sig-

nals to be dynamically matched following their tem-

poral variations (Karamzadeh et al., 2013). Further-

more, we propose to evaluate our fusion approach per

frequency band and then by fusing them.

2 MATERIAL AND METHODS

2.1 Database Description

The cohort used to conduct this study is composed

of resting-state EEG data of 32 SCI patients and 46

probable AD patients, acquired in a clinical setting at

Charles-Foix Hospital (Ivry-sur-Seine, France). This

retrospective study was approved by the institutional

review board of the local Ethics Committee Paris 6

and the Ethics Committee of Sorbonne University

(N°CER-2021-064).

The patients who complained of memory

impairment were referred to the outpatient memory

clinic of the hospital to undergo a battery of clinica-

l tests for brain disorders. The diagnosis was

set through a comprehensive clinical assessment,

including neuroimaging, interviews, psychometric

ﬁndings and neuro-psychological tests, in agreement

with the standard diagnostic criteria: NINDS, DSM-

IV, Jessen criteria for SCI (Jessen et al., 2014). It is

noteworthy that EEG was not included in the battery

of tests used to establish the diagnosis. Patients

with epilepsy were excluded from the study. Table

1 reports demographic and clinical characteristics of

patients included in the study.

EEG signals were recorded at rest with eyes

closed during 20 minutes at a frequency sampling

of 256 Hz, by taking care that the patients were not

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

742

Table 1: Clinical characteristics of the cohort. MMSE:

Mini-Mental State Examination; SD: Standard Deviation.

SCI (n=32) AD (n=46)

Age (mean ± SD) 68.2 ± 10.4 82.0 ± 8.6

Female (%) 81.8% 67.4%

MMSE (mean ± SD) 28.3 ± 1.6 19.0 ± 5.6

Benzodiazepin use (%) 4 (18.2%) 10 (22.7%)

Antidepressant use (%) 5 (16.0%) 18 (39.1%)

Neuroleptic use (%) 0 (0.0%) 5 (11.0%)

Hypnotic use (%) 5 (15.6%) 8 (17.4%)

falling asleep. Thirteen electrodes were used, placed

on the scalp according to the 10-20 international sys-

tem: Fp1, Fp2, F7, F3, Fz, F4, F8, FT7, FC3, FC7,

FC4, FT8, T3, C3, Cz, C4, T4, TP7, CP3, CPz, CP4,

TP8, T5, P3, Pz, P4, T6, O1, Oz, and O2.

Then, EEG signals were visually inspected to

discard the parts of the signals presenting artifacts

(eye movements, eye blinks, muscular activity, instru-

mental noise etc.). Thereby, continuous signals of 20

seconds, free from artifacts, were then kept for the

study. The obtained 20s-signals were then band-pass

ﬁltered with a third-order Butterworth ﬁlter in the fre-

quency range (1-30 Hz), as well as in the four con-

ventional frequency bands: delta (1-4 Hz), theta (4-8

Hz), alpha (4-8 Hz) and beta (12-30 Hz).

2.2 Methodology

In this work, we investigate the impact of varying

EEG epoch duration on the discrimination of AD

from SCI patients based on connectivity. To this end,

we ﬁrst segment the entire 20s EEG signal of each

patient into epochs of different durations: 10s, 5s,

and 2s. Therefore, for each patient, we obtain vari-

ous EEG signals of different durations, i.e. the entire

signal of length 20s, two epochs of 10s, four epochs

of 5s and ten epochs of 2s.

Then, for each of these four time conﬁgurations,

we compute functional connectivity (FC) between

pairs of EEG signals, aggregated into eight regions.

Figure 1 displays such regions: frontal/prefrontal

(Fp1, Fp2, Fz), frontal left (F7, F3, FT7, FC3), frontal

right (F4, F8, FC4, FT8), central (FCz, C3, CZ, C4),

temporal left (T3, TP7, CP3, T5), temporal right (T4,

CP4, TP8, T6), parietal (P3, Pz, P4), and occipital

(O1, Oz, O2) regions.

To estimate the connectivity of one region, we av-

erage the FC values computed between the pairs of

EEG signals associated to such region. For instance,

for the frontal/prefrontal region, we average the con-

nectivity values computed between Fp1 and Fp2, Fp1

and Fz, Fp2 and Fz. We also estimate the inter-

Figure 1: The 30 electrodes aggregated into 8 regions.

regions connectivity by averaging the connectivity

values computed between all pairs of EEG signals as-

sociated to such regions. For instance, for the con-

nectivity between the prefrontal/frontal region and the

occipital one, we take the mean of the connectivity

values calculated between Fp1 and O1, Fp1 and Oz,

Fp1 and O2, Fp2 and O1, Fp2 and Oz, Fp2 and O2,

Fz and O1, Fz and Oz, Fz and O2.

Finally, for each patient, each frequency band and

each time conﬁguration, we obtain a feature vector

containing 36 average connectivity values, including

8 intra-region connectivity values and 28 inter-regions

connectivity values.

To discriminate between SCI and AD patients,

we use a Support Vector Machine (SVM) classiﬁer

(Campbell and Ying, 2011), considering only the

most relevant features as inputs. These features are

selected, from the total of 36 features, using the For-

ward Orthogonal Regression (OFR) method (Stop-

piglia et al., 2003), to reduce the dimension of the in-

put feature vector for the classiﬁer. Classiﬁcation per-

formance is assessed on a development set, and then

on a test set with the objective of evaluating the ro-

bustness of the proposed multi-scale fusion approach.

To validate our fusion approach, we exploit

two FC measures relying on different mathemati-

cal concepts, namely Phase-Lag Index and Dynamic

Time Warping distance.

2.3 Connectivity Measurement

2.3.1 Phase-Lag Index (PLI)

It is computed between pairs of EEG signals as fol-

lows (Stam et al., 2007):

PLI = | < sign(∆Φ(t

)) > | (1)

where, |.| represents the absolute value, 〈.〉 indicates

the mean (over index k), “sign” denotes the signum

function that discards phase difference of 0 mod π

and ∆Φ(t

) is the phase difference between two time

series at time t

Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s Disease Detection Using EEG

743

PLI quantiﬁes the asymmetry of the distribution

of instantaneous phase differences between two sig-

nals, around zero mod π. If both signals are perfectly

phase-locked at a value of ∆Φ, the resulting PLI value

equals to 1. If both signals are either not coupled or

are coupled with a phase difference centered at ap-

proximately 0 mod π, the PLI value will be close to 0.

Intuitively, PLI allows quantifying the non-zero lags

between two signals.

2.3.2 Dynamic Time Warping (DTW)

DTW distance (Senin, 2008) is an elastic matching

metric that quantiﬁes the dissimilarity between two

time series showing a potential temporal drift. DTW

reﬂects the amount of warping required to align two

signals. It relies on ﬁnding the best warping path to

match two signals by minimizing the cumulative dis-

tance between the assigned points in the two signals.

The computation of DTW distance between two

EEG signals S

and S

of length N, consists in a re-

cursive construction of the cost matrix, as follows:

DTW [i, j] = (S

[i]−S

[ j])

min(DTW [i −1, j],

DTW [i, j − 1],

DTW [i − 1, j − 1])

(2)

where, i and j are the time points for which we

compute the DTW .

By design, the last computed value at coordinates

(N, N) corresponds to the DTW value between the

two signals. In order to optimize the aforementioned

procedure for DTW calculation, it is sufﬁcient to

search the optimal path close to the main diagonal of

the cost matrix, by applying a warping window. This

corresponds to limiting the shifting that is allowed be-

tween matched observations of the two EEG signals.

In this work, we used a Sakoe-Chiba band constraint

(Sakoe and Chiba, 1978), with a warping window size

ﬁxed to six.

2.4 Performance Assessment Protocol

Classiﬁcation was performed with a multivariate

SVM, using only the most relevant features as men-

tioned earlier. The performance is assessed according

to a consistent protocol: we generate 10 random sam-

plings of the whole database, considering 22 SCI and

28 AD patients for the development set, 10 SCI and

18 AD for the test set.

For each of the 10 samplings, we follow the

methodology outlined in Figure 2 for each frequency

band and each time conﬁguration. More precisely,

Figure 2: Experimental protocol for classiﬁcation perfor-

mance assessment.

on the development set, we ﬁrst select the most rel-

evant features among the computed 36 features, using

the OFR method and random probe technique (Stop-

piglia et al., 2003). The OFR allows ranking all the

features (the original variables) in decreasing order of

relevance. The random probe serves as a decision cri-

terion to select the most pertinent features: we dis-

card all features that are ranked after the probe, which

is a realization of random variable considered as an

additional feature. Therefore, the number of selected

features differs for each time and frequency band con-

ﬁguration.

Then, for each time conﬁguration and frequency

band, we perform a 10-folds SVM classiﬁer with

RBF kernel, and estimate the posterior probabilitie-

s of AD and SCI epochs using Platt’s estimation

method (Platt, 2000). Note that each patient has all

his/her EEG epochs in the same fold, in order to over-

come the issue of bias when evaluating the classiﬁ-

cation performance. Consequently, for each patient

in the development set, we obtain one SVM proba-

bilistic score associated to the 20s-signal, two SVM

probabilistic scores associated to two 10s-epochs,

four scores for the 5s-epochs, and ten scores of the

2s-epochs.

Classiﬁcation performance is evaluated for each

time conﬁguration, separately, and then when fusing

the SVM output scores obtained at different epoch du-

rations.

On the test set, we assess the classiﬁcation per-

formance using the SVM model trained on the

development set, considering the relevant features se-

lected with the OFR method on the whole develop-

ment set.

3 EXPERIMENTAL RESULTS

First, we present the results obtained on the develop-

ment set in a progressive manner. More speciﬁ-

cally, we present the classiﬁcation results of SCI and

AD epochs independently of the patient, for each

time conﬁguration (i.e. 20s-signal, 10s-epochs, 5s-

epochs and 2s-epochs). Hence, we analyze the per-

formance with four individual systems. Then, we ana-

lyze the classiﬁcation results of SCI and AD patients,

by averaging for each patient the SVM probabilistic

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

744

Table 2: AUC values on the development set using PLI and

DTW to discriminate AD from SCI epochs, considering the

four time conﬁgurations.

Frequency Epoch PLI DTW

[1-30]Hz

20s 0.706 0.801

10s 0.604 0.782

5s 0.603 0.765

2s 0.583 0.708

Delta

20s 0.544 0.751

10s 0.608 0.738

5s 0.579 0.746

2s 0.543 0.720

Theta

20s 0.766 0.762

10s 0.633 0.759

5s 0.610 0.744

2s 0.587 0.686

Alpha

20s 0.845 0.671

10s 0.683 0.667

5s 0.631 0.674

2s 0.590 0.623

Beta

20s 0.642 0.755

10s 0.652 0.777

5s 0.633 0.789

2s 0.563 0.770

output scores obtained on his/her EEG epochs, and

that for each time conﬁguration. After that, we eval-

uate the classiﬁcation of SCI and AD patients when

fusing with a simple average, per patient, the SVM

scores associated to his/her epochs of different dura-

tions. Finally, on the test set, we assess the general-

ization capability of our fusion scheme, after training

an SVM model on the entire development set.

3.1 Classiﬁcation of SCI and AD Epochs

Table 2 summarizes the classiﬁcation performance in

terms of AUC (Area Under The Curve) when discrim-

inating AD from SCI epochs with PLI and DTW. Re-

sults are given per frequency range, for each individ-

ual system (i.e. each time conﬁguration). Each epoch

was assigned to the class label of the corresponding

patient.

Results indicate a good classiﬁcation performance

when estimating the PLI overall the 20s-signal for all

frequency bands, except delta band. Also, we ob-

serve that the performance is degraded progressively

when computing the PLI on shorter epochs, from 10s-

epochs, 5s-epochs and then 2s-epochs.

When using DTW to quantify FC, we observe

that there is not distinctive trend in performance be-

tween the 20s-signal and shorter epochs of 10s and 5s.

Table 3: AUC values on the development set using PLI and

DTW to discriminate AD from SCI patients, considering

the four time conﬁgurations.

Frequency Epoch PLI DTW

[1-30]Hz

20s 0.706 0.801

10s 0.635 0.802

5s 0.656 0.798

2s 0.705 0.756

Delta

20s 0.544 0.751

10s 0.644 0.754

5s 0.651 0.781

2s 0.603 0.788

Theta

20s 0.766 0.762

10s 0.682 0.776

5s 0.692 0.783

2s 0.715 0.747

Alpha

20s 0.845 0.671

10s 0.742 0.677

5s 0.723 0.701

2s 0.716 0.654

Beta

20s 0.642 0.755

10s 0.689 0.786

5s 0.701 0.810

2s 0.618 0.805

However, the 2s-epoch conﬁguration leads in general

to the worst classiﬁcation results.

3.2 Classiﬁcation of SCI and AD

Patients

Table 3 reports the AUC values obtained for each indi-

vidual system using PLI and DTW to discrimate SCI

from AD patients.

We note that the fusion of SVM output scores of

epochs of a given time conﬁguration leads to better

classiﬁcation performance, compared to the obtained

results on separate epochs (see Table 2). However, no

clear trend emerges on the relationship between per-

formance and epoch duration in terms of AD and SCI

patient classiﬁcation. Based on the obtained result,

we propose to go a step forward in our fusion scheme,

by combining for each patient the SVM scores of

his/her epochs with different durations.

3.3 Merging Probabilistic Output

Scores of Different Epoch Durations

We average for each patient the SVM probabilistic

scores obtained for his/her epochs with different dura-

tions. For example, to fuse the 20s and 10s, we aver-

age the output score associated to the 20s-signal and

Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s Disease Detection Using EEG

745

Table 4: AUC values and correct classiﬁcation rates (in %)

of SCI and AD patients using PLI, when fusing SVM scores

of different epoch durations (Acc: Accuracy, Sens: Sensi-

tivity and Spec: Speciﬁcity).

Frequency Time conﬁg. AUC Acc Sens. Spec.

[1-30]Hz

20s 0.706 69.4 88.9 44.5

20s-10s-5 0.759 71.6 80.7 60.0

20s-10s-5s-2s 0.789 75.6 77.9 72.7

10s-5s-2s 0.760 72.4 78.2 65.0

Delta

20s 0.544 58.0 90.4 16.8

20s-10s-5 0.693 64.4 79.9 45.0

20s-10s-5s-2s 0.718 66.2 62.1 71.4

10s-5s-2s 0.718 66.6 56.8 79.1

Theta

20s 0.766 73.4 73.6 73.2

20s-10s-5 0.816 76.8 79.3 73.6

20s-10s-5s-2s 0.844 80.2 86.8 71.8

10s-5s-2s 0.799 74.4 81.1 65.9

Alpha

20s 0.845 77.8 82.9 71.4

20s-10s-5 0.895 81.6 82.9 80.0

20s-10s-5s-2s 0.902 83.6 89.6 75.9

10s-5s-2s 0.835 78.4 83.2 72.3

Beta

20s 0.642 64.4 74.6 51.4

20s-10s-5 0.765 71.8 72.9 70.5

20s-10s-5s-2s 0.760 71.6 74.3 68.2

10s-5s-2s 0.747 70.8 83.2 72.3

the two scores associated to the two epochs of 10s.

Then, based on the average scores computed for all

patients, we analyze the performance when discrimi-

nating AD from SCI patients.

Figures 3 and 4 display six ROC curves associated

to the 20s-signal (baseline) and to ﬁve fusion conﬁgu-

rations, considering respectively PLI and DTW as FC

measures. We report in Tables 4 and 5 the results for

the 20s conﬁguration and only three fusion conﬁgu-

rations, those leading to the best classiﬁcation perfor-

mance. Speciﬁcity is the percentage of SCI patients

well classiﬁed; sensitivity indicates the percentage of

AD patients well classiﬁed.

We observe that fusing the SVM scores of epochs

with different durations allows to highly improving

the classiﬁcation of AD and SCI patients, compared

to the baseline system (20s-signal), as well as to

individual systems that have been fused (i.e. con-

sidering each time conﬁguration separately). Indeed,

in Section 3.2, the best AUC value is obtained with

PLI in alpha considering the 20s-signal (AUC=0.845)

and with DTW in beta considering the 5s-epoch

(AUC=0.810). Nevertheless, when fusing the scores

of epochs of different durations, we notice that bet-

ter results are obtained for all fusion combinations,

reaching an AUC value of 0.902 with PLI in alpha

(see Table 4), and an AUC value of 0.894 with DTW

in beta (see Table 5).

Additionally, we clearly observe that the fusion

framework gives better results when combining the

scores of the whole signal (20s-signal) with those of

(a)

(b)

(c)

(d)

(e)

Figure 3: ROC curves for AD and SCI patient classiﬁcation

using PLI, when fusing SVM scores of different epoch du-

rations in: (a) (1-30 Hz), (b) delta, (c) theta, (d) alpha and

(e) beta.

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

746

Table 5: AUC values and correct classiﬁcation rates (in %)

of SCI and AD patients using DTW, when fusing SVM

scores of different epoch durations (Acc: Accuracy, Sens:

Sensitivity and Spec: Speciﬁcity).

Frequency Time conﬁg. AUC Acc. Sens. Spec.

[1-30]Hz

20s 0.801 73.2 73.2 73.2

20s-10s-5 0.884 81.2 83.9 77.7

20s-10s-5s-2s 0.869 79.6 88.2 68.6

10s-5s-2s 0.858 78.4 83.9 71.4

Delta

20s 0.751 70.0 73.2 65.9

20s-10s-5 0.852 76.6 87.5 62.7

20s-10s-5s-2s 0.882 79.6 88.2 68.6

10s-5s-2s 0.873 80.0 83.6 75.5

Theta

20s 0.762 71.8 75.7 66.8

20s-10s-5 0.868 79.6 83.9 74.1

20s-10s-5s-2s 0.852 79.6 78.9 80.5

10s-5s-2s 0.842 78.6 77.9 79.5

Alpha

20s 0.671 64.4 79.6 48.2

20s-10s-5 0.766 70.8 80.7 58.2

20s-10s-5s-2s 0.739 68.6 77.1 57.7

10s-5s-2s 0.730 68.2 75.0 59.5

Beta

20s 0.755 70.4 66.8 75.0

20s-10s-5 0.878 80.2 81.8 78.2

20s-10s-5s-2s 0.894 80.4 78.2 83.2

10s-5s-2s 0.887 79.8 78.9 80.9

Table 6: AUC values and correct classiﬁcation rates (in %)

of SCI and AD patients using PLI and DTW, when fusing

SVM scores of different epoch durations and different fre-

quency bands (Acc: Accuracy, Sens: Sensitivity and Spec:

Speciﬁcity).

Meas. Time conﬁg. AUC Acc. Sens. Spec.

PLI

20s 0.910 83.8 92.1 73.2

20s-10s-5 0.957 91.2 90.4 92.3

20s-10s-5s-2s 0.956 91.0 92.1 89.5

10s-5s-2s 0.929 87.0 88.2 85.5

DTW

20s 0.921 84.8 87.9 80.9

20s-10s-5 0.986 94.4 95.0 93.6

20s-10s-5s-2 0.987 95.0 94.3 95.9

10s-5s-2s 0.984 94.0 92.1 96.4

shorter epochs in a progressive manner. The best

results are obtained when fusing the scores of 20s-

signal, 10s-epochs, 5s-epochs and 2s-epochs (20s-

10s-5s-2s).

Table 6 and Figure 5 show the classiﬁcation re-

sults when fusing for each patient the SVM scores

of his/her epochs obtained at different time durations

and in the four frequency ranges (delta, theta, alpha

and beta). We observe that this double fusion scheme

allows improving signiﬁcantly the classiﬁcation per-

formance, reaching an accuracy value of 91.2% with

PLI (AUC=0.95) and of 95% with DTW (AUC=0.98).

Finally, we remark that fusing the classiﬁcation

scores of the four frequency ranges is more powerful

compared to when analyzing the signal on the whole

spectrum (1-30 Hz). Indeed, when considering the

(a)

(b)

(c)

(d)

(e)

Figure 4: ROC curves for AD and SCI patient classiﬁcation

using DTW, when fusing SVM scores of different epoch

durations in: (a) (1-30 Hz), (b) delta, (c) theta, (d) alpha

and (e) beta.

Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s Disease Detection Using EEG

747

(a)

(b)

Figure 5: ROC curves for AD and SCI patient classiﬁcation

using: (a) PLI and (b) DTW, when fusing the SVM scores

of different epoch durations and different frequency bands.

whole spectrum, Tables 4 and 5 show that we reach

an AUC value of 0.789 with PLI and of 0.884 with

DTW, for the conﬁguration (20s-10s-5s-2s). On such

conﬁguration, by processing EEG signals in each fre-

quency band separately, and then fusing the resulting

probabilistic scores, we obtain better results, reaching

an AUC value of 0.957 with PLI and of 0.987 with

DTW. This outcome also demonstrates the effective-

ness of our fusion approach.

3.4 On the Generalization Capability of

the Fusion Scheme on the Test

Subset

In this section, we investigate the effectiveness of our

fusing approach in terms of classiﬁcation prediction

on the test subset. We follow the experimental pro-

tocol explained in Section 2.4. Table 7 and Figure 6

show the results on the test set considering the whole

20s signal and the best conﬁguration obtained on the

development set, that fusing 20s, 10s, 5s and 2s.

The performance results on the test set are not

good as those obtained on the development set. Nev-

ertheless, we observe that fusing epoch durations al-

lows enhancing the performance compared to when

considering the 20s conﬁguration, in terms of AUC

and accuracy (see Table 7). This tendency is observed

Table 7: Classiﬁcation performance on the test set using

DTW, for the 20s-signal and the time conﬁguration 20s-10s-

5s-2s.

Frequency Time conﬁg. AUC Acc. Sens. Spec.

[1-30]Hz

20s 0.758 74.3 82.8 59.0

20s-10s-5s-2s 0.785 80.4 89.4 64.0

Delta

20s 0.690 66.8 97.8 11.0

20s-10s-5s-2s 0.754 73.2 91.1 41.0

Theta

20s 0.679 70.0 72.8 64.0

20s-10s-5s-2s 0.754 72.1 76.1 65.0

Alpha

20s 0.611 64.6 87.8 23

20s-10s-5s-2s 0.670 70.4 80.6 52.0

Beta

20s 0.730 72.5 83.9 52.0

20s-10s-5s-2s 0.761 75.4 89.4 50.0

Fusion

20s 0.776 74.6 92.2 43.0

20s-10s-5s-2 0.803 80.4 93.3 57.0

for all the frequency bands. Moreover, the double

fusion scheme, including both epoch durations and

frequency ranges, allows increasing the performance

even further (AUC=0.803).

Finally, note that we report only the generaliza-

tion results obtained with DTW. In fact, the obtained

results on the test set with PLI were very bad, reﬂect-

ing that there is poor stability in the selected features

with PLI from the development to the test sets, com-

paratively to DTW.

4 DISCUSSION

The purpose of this study is two folds: (i) investi-

gating the impact of varying EEG epoch duration on

the discrimination between two brain cognitive con-

ditions (SCI and AD) based on functional connectiv-

ity, (ii) evaluating the potential application of a multi-

scale fusion approach for an improved classiﬁcation

results. We have proposed both temporal and fre-

quency fusions, which consist in combining the clas-

siﬁer probabilistic scores obtained on epochs with

different time durations and in different frequency

bands. Such fusion was carried out through a sim-

ple averaging of the SVM scores which is the most

parsimonious choice.

Experiments have been conducted following a

consistent experimental protocol dividing the whole

dataset into a development set and a test set. On the

development set, results showed that when consider-

ing the whole 20s-signal (baseline system), the classi-

ﬁcation results of AD and SCI epochs are better than

when analyzing EEG signals on shorter epochs of 10s,

5s and 2s. The worst results are obtained for the 2s-

epoch duration.

By fusing, for each patient, the classiﬁer out-

put scores obtained on his/her epochs for each time

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

748

(a) (b)

(e) (f)

Figure 6: ROC curves for AD and SCI patient classiﬁcation on the test set based on DTW, using the 20s-signal and the fusion

conﬁguration (20s-10s-5s-2s) in :(a) [1-30] Hz, (b) delta, (c) theta, (d) alpha, (e) beta, and (f) considering the fusion of all

frequency bands.

conﬁguration, classiﬁcation performance of patients

was found better than that obtained when classifying

epochs. Nevertheless, no clear trend appeared on the

relationship between epoch duration and classiﬁca-

tion performance. These results may highlight the po-

tential complementarity in terms of information con-

tent, when segmenting EEG signals into epochs and

fusing the classiﬁer scores obtained per epoch.

Besides, our fusion approach showed an improve-

ment of classiﬁcation results when combining the

scores of the whole signal (20s-signal) with those of

shorter epochs in a progressive manner. The best re-

sults were obtained for the (20s-10s-5s-2s) conﬁg-

uration. Then, by fusing the classiﬁcation scores

obtained at different frequency bands, we further

improve the discrimination between SCI and AD

patients, reaching an accuracy of 91.2% with PLI

(AUC=0.95) and 95% with DTW (AUC=0.98), with a

very good balance between speciﬁcity and sensitivity.

Results also highlighted that fusing classiﬁer scores

obtained in each frequency band is more efﬁcient

than when analyzing the signal on the whole spec-

Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s Disease Detection Using EEG

749

trum. Notably, by analyzing each frequency band sep-

arately, we can retrieve more speciﬁc EEG features,

which leads to a reﬁned characterization of EEG sig-

nals and thus a better discrimination between popula-

tions.

When evaluating the results on the test set, we

found that DTW is more effective than PLI in the con-

text of this study. This can be explained by the fact

that DTW is an elastic distance that allows captur-

ing dynamic temporal-lags, which may ﬂuctuate over

time when matching two EEG signals. By contrast,

PLI assumes the temporal delay stationary.

Although the classiﬁcation results on the test set

were degraded comparatively to the development set,

our proposed multi-scale fusion approach outper-

forms the baseline system (20-s signal), reaching an

AUC value of 0.785 for the 20s-10s-5s-2s conﬁgura-

tion, and of 0.803 when additionally fusing the fre-

quency bands.

All these results ﬁrst highlight that varying EEG

signal duration has an impact on the classiﬁcation re-

sults. This can explain in part the difference of results

in the state-of-the-art. Therefore, it is important to

specify in scientiﬁc articles the duration of EEG sig-

nals and to clarify the epoching process, such as the

number and length of epochs.

Furthermore, our ﬁndings demonstrate the

effectiveness of analyzing EEG signals at different

epoch durations and fusing the classiﬁcation scores of

the extracted epochs. This framework allows a reﬁne

characterization of the brain dynamics across time by

computing the connectivity on short epochs, while

taking into account all the available information in the

whole EEG signal. Finally, combining the frequency

bands is also very pertinent in terms of classiﬁcation

results, since each frequency band conveys valuable

and complementary insights into brain function.

Our fusion scheme is based on classiﬁcation

scores. This study focuses on SVM probability output

which entails two main limitations. First, probabil-

ity estimation by Platt’s assumes the relationship be-

tween the SVM scores and the probabilities to be sig-

moidal, which might not be true in our case. However,

our fusion scheme was evaluated in the same condi-

tions as the individual systems. Second, we evalu-

ate our approach with only one classiﬁer. The results

should be conﬁrmed using other classiﬁers, leverag-

ing alternative mathematical paradigms.

5 CONCLUSION

This work points out the potential use of both tem-

poral and frequency fusion approach to improve the

characterization of EEG signals, and thus the classiﬁ-

cation results of AD and SCI patients. In addition, this

fusion approach allows obtaining good prediction per-

formance in the context of generalization of results.

In the future, we will perform further analyses to

study the extent of our fusion approach. First, we

will analyze the features that were selected for the

different durations of epochs to understand what are

the crucial variables of the region connectivity ma-

trix for the prediction. Second, we plan to investigate

the effectiveness of our approach to discriminate be-

tween SCI, AD and MCI patients. Indeed, by adding

the MCI group, the classiﬁcation would be more chal-

lenging. Our hypothesis is that our multi-scale fusion

approach can contribute to the ﬁne characterization

of these three cognitive conditions, thus enhancing

the multi-class classiﬁcation. Also, we will assess the

generalization capability of other functional connec-

tivity measures by conducting a comparative analysis

in such context.

REFERENCES

(2024). 2024 alzheimer’s disease facts and ﬁgures.

Alzheimer’s & Dementia, 20(5):3708–3821.

Abazid, M., Houmani, N., Boudy, J., et al. (2021). A Com-

parative Study of Functional Connectivity Measures

for Brain Network Analysis in the Context of AD De-

tection with EEG. Entropy, 23(11):1553.

Bossers, K., Wirz, K. T., Meerhoff, G. F., et al. (2010).

Concerted changes in transcripts in the prefrontal cor-

tex precede neuropathology in Alzheimer’s disease.

Brain, 133(12):3699–3723.

Campbell, C. and Ying, Y. (2011). Learning with Sup-

port Vector Machines. Synthesis Lectures on Artiﬁcial

Intelligence and Machine Learning. Springer Interna-

tional Publishing, Cham.

Cassani, R., Estarellas, M., San-Martin, R., et al.

(2018). Systematic Review on Resting-State EEG for

Alzheimer’s Disease Diagnosis and Progression As-

sessment. Disease Markers, 2018:5174815.

Castellani, R. J., Rolston, R. K., and Smith, M. A.

(2010). Alzheimer Disease. Disease-a-month : DM,

56(9):484–546.

Dauwels, J., Vialatte, F., and Cichocki, A. (2010a). Di-

agnosis of Alzheimer’s Disease from EEG Signals:

Where Are We Standing? Current Alzheimer Re-

search, 7(6):487–505.

Dauwels, J., Vialatte, F., Musha, T., and Cichocki, A.

(2010b). A comparative study of synchrony measures

for the early diagnosis of Alzheimer’s disease based

on EEG. NeuroImage, 49(1):668–693.

Dubois, B., Picard, G., and Sarazin, M. (2009). Early detec-

tion of Alzheimer’s disease: new diagnostic criteria.

Dialogues in Clinical Neuroscience, 11(2):135–139.

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

750

Houmani, N., Abazid, M., Santiago, K. d., et al. (2021).

EEG signal analysis with a statistical entropy-based

measure for Alzheimer’s disease detection. 2:387.

Jeong, J., Gore, J. C., and Peterson, B. S. (2001). Mu-

tual information analysis of the EEG in patients

with Alzheimer’s disease. Clinical Neurophysiology,

112(5):827–835.

Jessen, F., Amariglio, R. E., van Boxtel, M., et al. (2014).

A conceptual framework for research on subjec-

tive cognitive decline in preclinical Alzheimer’s dis-

ease. Alzheimer’s & Dementia: The Journal of the

Alzheimer’s Association, 10(6):844–852.

Karamzadeh, N. et al. (2013). Capturing dynamic patterns

of task-based functional connectivity with EEG. Neu-

roImage, 66:311–317.

Kasakawa, S., Yamanishi, T., Takahashi, T., et al. (2016).

Approaches of Phase Lag Index to EEG Signals in

Alzheimer’s Disease from Complex Network Analy-

sis. In Chen, Y.-W., Torro, C., Tanaka, S., Howlett,

R. J., and C. Jain, L., editors, Innovation in Medicine

and Healthcare 2015, Smart Innovation, Systems and

Technologies, pages 459–468, Cham. Springer Inter-

national Publishing.

Knyazeva, M. G., Jalili, M., Brioschi, A., et al. (2010). To-

pography of EEG multivariate phase synchronization

in early Alzheimer’s disease. Neurobiology of Aging,

31(7):1132–1144.

Lee, S.-H., Park, Y.-M., Kim, D.-W., and Im, C.-H. (2010).

Global synchronization index as a biological correlate

of cognitive decline in Alzheimer’s disease. Neuro-

science Research, 66(4):333–339.

Platt, J. (2000). Probabilistic Outputs for Support Vec-

tor Machines and Comparisons to Regularized Like-

lihood Methods. Adv. Large Margin Classif., 10.

Sakoe, H. and Chiba, S. (1978). Dynamic programming

algorithm optimization for spoken word recognition.

IEEE Transactions on Acoustics, Speech, and Signal

Processing, 26(1):43–49. Conference Name: IEEE

Transactions on Acoustics, Speech, and Signal Pro-

cessing.

Senin, P. (2008). Dynamic time warping algorithm review.

Stam, C. J., Nolte, G., and Daffertshofer, A. (2007). Phase

lag index: Assessment of functional connectivity

from multi channel EEG and MEG with diminished

bias from common sources. Human Brain Mapping,

28(11):1178–1193.

Stoppiglia, H., Dreyfus, G., Dubois, R., and Oussar, Y.

(2003). Ranking a random feature for variable and

feature selection. J. Mach. Learn. Res., 3(7/8):1399–

1414.

Multi-Scale Probabilistic Score Fusion for Enhancing Alzheimer’s Disease Detection Using EEG

751