A Word Recognition Paradigm Through EEG Analysis: Imagined

Speech Classiﬁcation

Francesco Iacomi

, Andrea Farabbi

, Maximiliano Mollura

Edoardo Maria Polo

, Riccardo Barbieri

and Luca Mainardi

Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milano, Italy

Keywords:

Electroencephalogram (EEG), Machine Learning, Brain Computer Interface (BCI), Imagined Speech,

Biomedical Signal Processing.

Abstract:

This study presents an innovative approach for decoding imagined speech using EEG signals. The proposed

analysis aims at revealing phonetic and semantic properties of imagined words through brain activity. The ex-

perimental protocol involves presenting words to subjects while recording EEG signals via a 64-channels cap.

Each word is associated with three speciﬁc properties: length, presence of doubles, and category of meaning.

The protocol includes ﬁxation, cue presentation, thinking, and rest phases. EEG signals undergo meticulous

preprocessing stage to mitigate noise and artifacts. Features are extracted from the processed signal, including

statistical, spectral, and fractal domain measures. The dimensionality of features is reduced through statistical

means. Several classiﬁers, (e.g., MLP, KNN, LDA, QDA), are trained and evaluated to predict mentioned

properties of imagined words. An ensemble model (LDA) comprising the best 3 models mentioned above

is then employed to enhance classiﬁcation accuracy. Results illustrate the effectiveness of decoding imag-

ined word properties with average accuracies of 35.2% for ”Category”, 57.2% for ”Doubles”, and 55.8% for

”Length”. By aggregating all predictions we are able to decode each single word with a mean accuracy of

11.8% (random accuracy = 8.33%) and an average word distance of 1.54. Post-classiﬁcation studies on the

most relevant variables and on the most discriminating channels further deepen our understanding of the pro-

posed Imagined Speech cognitive process, showing different brain activations for different linguistic aspects.

1 INTRODUCTION

The central objective of this research is to decode

Imagined Speech (IS), a cognitive process where in-

dividuals mentally simulate speech production with-

out any actual vocalization or physical muscle move-

ment (Martin, 2018). The signiﬁcance of IS lies in

its potential applications, particularly in the realm of

Brain-Computer Interfaces (BCIs), where it can facil-

itate natural and efﬁcient communication among indi-

viduals (Gu et al., 2021).

The electroencephalogram (EEG) is the chosen

neurophysiological signal for this study due to its

non-invasiveness, exceptional temporal resolution,

https://orcid.org/0009-0005-2839-8414

https://orcid.org/0000-0001-5582-4654

https://orcid.org/0000-0002-2248-8145

https://orcid.org/0000-0003-0432-1314

https://orcid.org/0000-0001-9381-3833

https://orcid.org/0000-0002-6276-6314

and the minimal instrumentation required for data ac-

quisition (Yi et al., 2013). EEG’s portability and ease

of use make it a compelling choice for BCIs, aligning

with the study’s goal of creating BCIs that are acces-

sible to a wide audience (Laureys, 2005).

The proposed study also delves into the cur-

rent state of research in IS decoding. This process

can be distilled into four key steps: signal acqui-

sition, pre-processing, feature extraction, and clas-

siﬁcation (Lopez-Bernal et al., 2022). Researchers

have employed various IS prompts, including visual

and auditory cues, and have employed several tech-

niques for feature extraction, such as statistical analy-

sis (Panachakel and Ramakrishnan, 2021), frequency

domain analysis (Fitriah et al., 2022), and spatial

analysis (Lopez-Bernal et al., 2022). Classiﬁcation,

the ﬁnal step, has seen the application of both ma-

chine learning and deep learning models, each with its

strengths and areas of application. In Table 1 are sum-

marized the state of the art classiﬁcations as a term of

comparison.

622

Iacomi, F., Farabbi, A., Mollura, M., Polo, E., Barbieri, R. and Mainardi, L.

A Word Recognition Paradigm Through EEG Analysis: Imagined Speech Classiﬁcation.

DOI: 10.5220/0012372500003657

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

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 622-628

ISBN: 978-989-758-688-0; ISSN: 2184-4305

Table 1: State of the art classiﬁcation performance.

Author N of prompts Subjects Trials per word Accuracy

(Nguyen et al., 2018) 3 short words, 2 long words 6 100 50.1%(3 classes),66.2%(2 classes)

(Qureshi et al., 2018) 5 words 8 100 40.1%

(Dipti Pawar, 2020) 4 words 6 50 45.07%

(Chau, 2017) 2 words 12 180 66.35%

(Dipti Pawar, 2023) 5 words 8 180 64%

Yet, amidst these advancements, several chal-

lenges and limitations persist. These include a lim-

ited vocabulary, reduced accuracy, a predominant fo-

cus on ofﬂine approaches, and limited dataset sizes

(Panachakel and Ramakrishnan, 2021). Overcoming

these challenges need the creation of larger datasets,

innovative prompts, reﬁnements in feature extraction

and machine learning techniques, a deeper explo-

ration of connectivity during IS, and a shift toward

inter-subject approaches (Panachakel and Ramakrish-

nan, 2021).

This study seeks to tackle these challenges head-

on. It plans to do so by conducting a subject-speciﬁc

classiﬁcation of 12 carefully selected words, catego-

rizing them based on word length, the presence of

doubles, and semantic class. The goal is twofold: a)

to determine which EEG features can effectively dis-

criminate imagine words and b) expand the vocabu-

lary space for IS decoding. Additionally, we aim at

smoothing feature extraction processes, identify in-

formative EEG channels, and enhance the overall ef-

ﬁciency and performance of IS classiﬁcation.

2 MATERIALS AND METHODS

A total of 24 participants were recruited for EEG data

acquisition, comprising 14 males and 10 females aged

between 20 and 26 years. The protocol has been

approved by Politecnico di Milano Research Ethical

Committee (Opinion 80 n. 29/2021). This task pri-

marily involved the mental imaging of twelve distinct

words, carefully selected to enable unique encoding

based on semantic and grammatical Properties. The

Properties under consideration included word cate-

gory (classiﬁed into three categories: Motion, Space

Exploration and Unpleasantness), word length (cate-

gorized as short or long), and the presence of double

letters in the word. The 12 words and their properties

are reported in Table 2.

2.1 Experimental Protocol

Electroencephalogram (EEG) data were collected

from 24 participants using the 64-channel EEG sys-

tem BE Plus LTM by EBNeuro (Florence, Italy), with

electrodes positioned across the scalp following the

10-20 International System.

The acquisition protocol is structured into 5 ses-

sions, with each word presented once in a random-

ized sequence. Each word is associated with 4 dis-

tinct phases, (see Fig.1): an initial ﬁxation period, the

presentation of the word as a visual cue, a subsequent

thinking period, and last, a rest period. The entire

protocol lasts about 18 minutes, after which partici-

pants are required to ﬁll in an online form to gather

feedback on their perceptions and experiences.

2.2 Preprocessing and Feature

Extraction

Following data acquisition, rigorous preprocessing

steps were applied to the EEG recordings. These steps

were crucial for mitigating artifacts effects and safe-

guarding the data’s integrity.

In the preprocessing phase, entirely performed us-

ing EEGLAB (Delorme and Makeig, 2004), the EEG

signals were subjected to frequency ﬁltering, restrict-

ing their range to 0.05Hz to 120Hz. Notch ﬁlters were

thoughtfully applied at 50Hz and 100Hz to effectively

eliminate interference lines and their harmonics. Vi-

sual inspection was then meticulously employed to

identify and exclude any corrupted channels from the

dataset.

To further enhance data quality, the common av-

erage reference (CAR) was computed, which aids in

minimizing common noise across all electrodes. The

signals were then downsampled from their original

512Hz to 256Hz to enhance computational efﬁciency

for the subsequent procedures.

Finally we applied Independent Component Anal-

ysis (ICA). This technique played a pivotal role in

removing artifacts, particularly those arising from

blinking. After ICA, any removed channels were re-

constructed using spherical interpolation to ensure the

spatial coherence of the EEG data.

Such thorough preprocessing stage ensured that

EEG data artifacts -interference, and corruption- be

minimized for subsequent analysis with the utmost

data quality and reliability.

After preprocessing, the EEG signals underwent

an epoching process, in which a 4-second window

A Word Recognition Paradigm Through EEG Analysis: Imagined Speech Classiﬁcation

623

Table 2: Italian selected words divided by Properties (columns), in brackets their English translation.

Motion Space Exploration Unpleasantness

1 Resistenza (Endurance) Astronave (Spaceship) Delusione (Delusion)

2 Scatto (Sprint) Razzo (Rocket) Guerra (War)

3 Allenamento (Training) Navicella (Shuttle) Depressione (Depression)

4 Sforzo (Effort) Cosmo (Cosmos) Odio (Hate)

Figure 1: Experimental Protocol Pipeline.

was selected.

A comprehensive set of methods for feature ex-

traction was meticulously employed to uncover in-

tricate patterns within the data. The extracted fea-

tures included three groups of variables: statisti-

cal (Mean, Autoregressive Coefﬁcients, Variance...),

spectral (band powers, spectral skewness...) and frac-

tal domain features ( Hurst Exponent, Fractal Dimen-

sion...) amounting to a total of 34 variables. These

features were consistently derived from each channel

of the signal in the time domain, as well as from the

information obtained through the Discrete Wavelet

Transform (DWT) and the Empirical Mode Decom-

position (EMD), from which we extracted ﬁve In-

trinsic Mode Functions (IMFs). The ”Db4” wavelet

served as the mother wavelet for the DWT, and seven

details were computed, from which the features were

extracted.

Furthermore, to ensure comparability, all features

were normalized with respect to a baseline, 4-second

window preceding the start of the protocol.

This approach ensures that a consistent set of fea-

tures is extracted from different perspectives, thus

capturing a diverse range of information from the

EEG data.

Once the dataset is prepared and processed, it as-

sumes a shape of 60x26,963 [12 words*5 repetitions]

x [61 channels*(1 time-domain signal+7 DWT de-

tails+5 IMFs)*34 features + target].

2.3 Classiﬁcation

In this section, we will delve into the feature selection

and classiﬁcation procedures conducted on the train-

ing set, which comprises 80% of the original dataset.

The remaining 20% is exclusively reserved for the

blind test and subsequent evaluation.

2.3.1 Features Selection

The extracted features underwent a feature selection

process, aiming to identify the most promising ones

and their utility in classifying each Property. Addi-

tionally, topographic map of an EEG (the so called

TopoPlot) are employed to visually analyze the most

relevant channels.

It is important to note that three independent clas-

siﬁcations, one for each Property as the target, are

conducted using the same.

Firstly, the dataset undergoes a data quality check

to identify missing values. Successively, the Kruskal-

Wallis test (divergence from normality was assessed

through Kolmogorov-Smirnov test) is applied to each

variable to search for features whose medians show

the greatest difference when stratifying by the target

variable. Only features with p<0.05 are retained.

Next, each feature is standardized using a stan-

dard scaler, and a correlation analysis is performed

to eliminate variables that exhibit strong correlations

(>0.8) with the others preference is given to variables

based on the H-statistics of the Kruskal-Wallis test,

used to measure the diversity of the variable’s distri-

bution: variables with lower H-statistics are discarded

ﬁrst.

The ﬁnal step involves the application of Min-

imum Redundancy Maximum Relevance (MRMR)

feature selection algorithm. This step aims to identify

variables that are less redundant and more relevant

for the classiﬁcation task. Importantly, all the pre-

sented procedures are initially performed on the train-

ing dataset, consisting in 80% of the original dataset.

The selection of the number of variables is car-

ried out through an assessment involving ﬁve mod-

els, including Linear Discriminant Analysis (LDA),

Decision Trees, Random Forest, Support Vector Ma-

chines (SVM), and Logistic Regression. By evaluat-

ing the mean accuracy from cross-validation across

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

624

Figure 2: Classiﬁcation Pipeline.

these models, the number K of features to retain is

chosen. The range considered for K typically spans

from 1 to 4, ensuring a favorable sample-to-feature

ratio of about 10.

2.3.2 Classiﬁcation Models

Following the selection of K features, Multilayer Per-

ceptrons (MLP), k-Nearest Neighbors (KNN), Sup-

port Vector Machines (SVM), Decision Trees, Ran-

dom Forest, Naive Bayes, and Logistic Regression

Models are trained with hyperparameter tuning in

order to predict each word Property. The best-

performing model for the classiﬁcation of each Prop-

erty, as determined by the highest validation accuracy,

is then selected to carry out the ﬁnal classiﬁcation.

Furthermore, an ensemble model is employed for

classiﬁcation. This ensemble model is a Linear Dis-

criminant Analysis (LDA) that takes as input the pre-

dicted probabilities generated by the three best in-

dividual models. The ﬁnal prediction is made by

considering either the prediction from the individual

model or the ensemble model, depending on their re-

spective validation accuracies. This approach ensures

that the ﬁnal classiﬁcation is based on the most reli-

able and accurate model, as determined by rigorous

evaluation.

Once the three predictions of the individual Prop-

erties for each word have been generated, they can

be collectively analyzed to decode the original word.

This decoding process is facilitated by the unique en-

coding of each word based on its distinctive proper-

ties.

In addition, another metric known as Word Dis-

tance (D) was deﬁned. It provides a unique perspec-

tive, quantifying the distance between each word and

its predicted counterparts in the Property space.

D =

∑

i=1

∑

j=1

(δ(a

i j

, b

i j

))

(1)

Where:

• D represents the distance metric.

• N is the total number of words in the test set.

• n is the number of properties being evaluated

(3 properties in our case: Semantic Category,

Length, Doubles).

• a

i j

represents the value of the i-th word for the

j-th property in the reference dataset.

• b

i j

represents the value of the i-th word for the

j-th property in the predicted dataset.

• δ(a

i j

, b

i j

) is a function that returns 0 if a

i j

is equal

to b

i j

and 1 if they are different.

The formula calculates the average differences be-

tween the properties of words in the reference dataset

and the predicted words, with D representing the av-

erage distance across all words in the test set.

A Word Recognition Paradigm Through EEG Analysis: Imagined Speech Classiﬁcation

625

3 RESULTS

The statistics obtained from the online form showed

no signiﬁcant correlations with the accuracies of indi-

vidual Properties. The average accuracy across par-

ticipants for the three Properties is displayed in Table

The average accuracy, obtained by aggregating

predictions across participants, was 11.8% [8.1%;

15.3%] for a 12-class problem, against an 8.33% ran-

dom chance. The confusion matrix is shown in Fig.3

Furthermore, in this study, the average word dis-

tance between classiﬁed words across subjects was

1.54, in contrast to the random distance of 1.67.

Figure 3: Confusion Matrix of the 12-class problem.

For the ”Category” property the most relevant

variables among participants are: Lyapunov expo-

nents, mean, Autoregressive coefﬁcients, and Hjorth

coefﬁcients. However, for both binary problems (i.e.,

”Lenght” and ”Double” properties), Autoregressive

coefﬁcients were by far the most selected ones.

It was observed that the best results on the test

dataset were achieved when considering variables ex-

tracted from all sources or solely from those derived

from DWT or IMF. Considering only time-domain

variables did not yield favorable results.

The most active channels across participants were,

for the ”Category” property, those related to the tem-

poral lobes in both hemispheres. For ”Length,” it was

the channel on the mid-line sagittal plane, and for

”Doubles,” it was channels related to the temporal and

central lobes in both hemispheres. A visual compar-

ison between the pertinent channels for ”Category”

and ”Length” can be drawn by respectively examin-

ing Figure 4.

4 DISCUSSION

The results of this study provide valuable insights into

the classiﬁcation of word Properties from EEG data

during language processing. The protocol involved

ﬁve sessions, each with four distinct phases. The

subsequent preprocessing ensured the quality and in-

tegrity of the data, making it suitable for detailed anal-

ysis.

One notable ﬁnding is that the statistics from the

online form were not correlated with the accuracies of

individual Properties. This suggests that participants’

subjective perceptions and experiences during the ex-

periment did not align with the objective accuracy of

the property classiﬁcations. Such a lack of correlation

highlights the complexity of the relationship between

cognitive processes and self-reported experiences.

the average accuracies achieved by the devel-

oped models on data extracted from each participant

in classifying Properties varied across the different

Properties. For the ”Category” property, the accu-

racy was slightly above chance level (33%), indicat-

ing some ability to distinguish between word cat-

egories. In contrast, the binary Properties ”Dou-

bles” and ”Length” displayed accuracies of 57.3%

and 56.1%, respectively. These results highlight the

inherent challenges associated with multi-class clas-

siﬁcation tasks compared to binary classiﬁcation.

It is important to note that these lower accuracies,

particularly in the ”Category” property, are attributed

to the utilization of a dataset with limited recordings,

thus impacting on the overall classiﬁcation perfor-

mance.

However, the fact that all average accuracies are

above the random chance level supports the possibil-

ity of classifying the word Properties. Despite the rel-

atively low accuracy rates, the models points toward

the ability to discriminate between them.

Aggregating predictions from participants yielded

a mean accuracy of 11.8%. This low value can be

attributed to the increased complexity of the prob-

lem (12 classes so 8.33% random chance) and to

the low number of repetitions of each word. Never-

theless, achieving an accuracy above random chance

in a 12-class problem demonstrates the potential of

EEG-based classiﬁcation of semantic and grammati-

cal words Properties.

It is crucial to highlight that this form of word en-

coding has paved the way for the introduction of the

distance metric D. This represents one of the most in-

novative aspects of our study. In the past, each word

was treated as entirely unique, devoid of shared Prop-

erties, resulting in equidistant relationships between

all words. However, our approach has transitioned

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

626

Table 3: Classiﬁcation Accuracy, in square brackets the 95% conﬁdence intervals.

Property Mean accuracy over subjects Random Accuracy

1 Semantic Category (3-class) 35.2% [30.5%; 39.6%] 33.3%

2 Doubles (Binary) 57.2% [50.6%; 61.0%] 50.0%

3 Length (Binary) 55.8% [50.1%; 63.3%] 50.0%

Figure 4: Comparative TopoPlots between the most relevant channels in the three Properties.

from one-hot encoding and a non-dense space to a

more densely packed space, where words are encoded

based on their semantic and grammatical Properties.

Importantly, our results prove that these words can be

effectively decoded.

Analyzing the relevance of variables across par-

ticipants revealed interesting patterns. For the ”Cat-

egory” property, Lyapunov exponent, Mean, Autore-

gressive coefﬁcients, and Hjorth coefﬁcients were all

found to be equally relevant. In contrast, Autoregres-

sive coefﬁcients dominated as the most relevant vari-

ables for both binary problems. This suggests that

certain EEG features, such as Autoregressive coefﬁ-

cients, consistently play a critical role in classifying

cognitive Properties, regardless of the speciﬁc prop-

erty.

Furthermore, considering variables from different

sources—time domain, DWT, and IMF—proved to

be advantageous for achieving better classiﬁcation re-

sults. This underscores the importance of incorporat-

ing a wide range of EEG features to capture the intri-

cate dynamics of cognitive processes during language

tasks.

Channel-wise analysis highlighted speciﬁc brain

regions associated with each property. For the ”Cat-

egory” property, temporal lobes in both hemispheres

were notably active. ”Length” exhibited activity in

the mid-line sagittal plane, and ”Doubles” involved

channels related to the temporal and central lobes

in both hemispheres. These ﬁndings suggest the in-

volvement of distinct brain regions in processing dif-

ferent linguistic aspects, aligning with existing neu-

rocognitive theories (Price, 2012).

5 CONCLUSION AND FUTURE

DEVELOPMENTS

In conclusion, our study on Imagined Speech Detec-

tion has marked a signiﬁcant step forward in the chal-

lenging task of classifying imagined word Properties

using EEG data.

It has not only provided valuable insights but has

A Word Recognition Paradigm Through EEG Analysis: Imagined Speech Classiﬁcation

627

also introduced innovative concepts like word Prop-

erties and distance metrics into the realm of Imagined

Speech detection that can open up new avenues for re-

search and offers a novel direction for enhancing the

accuracy and robustness of future classiﬁcation algo-

rithms.

Additionally, our exploration into the neurophys-

iological aspects of word perception has revealed in-

triguing insights. EEG features extracted from Dis-

crete Wavelet Transform (DWT) and Intrinsic Mode

Functions (IMF) have shown promise in classiﬁca-

tion, emphasizing the importance of considering vari-

ables from multiple sources.

Our channel-wise analysis has shed light on the

brain regions associated with different Properties,

emphasizing the multifaceted nature of word percep-

tion. These insights contribute to our understanding

of the neural basis of language.

Expanding our data collection protocol to gather

more extensive and diverse datasets is a crucial step.

These richer datasets will empower machine learn-

ing models with a deeper understanding of imagined

speech recognition.

Furthermore, an exciting avenue for future re-

search lies in the fusion of Natural Language Pro-

cessing (NLP) techniques with the novel concepts

introduced in this study, such as word distance and

the comprehensive characterization of words based on

their semantic, grammatical, and phonetic properties.

In the realm of semantics, an intriguing explo-

ration involves distinguishing synonyms to uncover if

it’s feasible to capture the precise meaning of words,

thereby reﬁning the concept of semantic category.

Lastly, our study’s successful achievement in dis-

crimination word length as a classiﬁcation task opens

a new dimension for research: a regression to predict

the exact number of letters in a word. This represents

a more intricate and challenging facet of Imagined

Speech Detection, offering exciting possibilities for

future investigations.

In conclusion, we envision a future marked by

exciting developments in Imagined Speech Detection

where the convergence of artiﬁcial intelligence, neu-

roscience, and linguistics offers immense promise. In

this context, our study provides a solid foundation

for further exploration towards a more comprehensive

understanding of Imagined Speech Detection. The

path ahead promises deeper insights, increased func-

tionality, and broader applications at the intersection

of human cognition, language, and technology.

REFERENCES

Chau, A. R. S. R. T. A. B. T. (2017). Eeg classiﬁcation

of covert speech using regularized neural networks.

IEEE/ACM Transactions on Audio, Speech, and Lan-

guage Processing.

Delorme, A. and Makeig, S. (2004). Eeglab: an open source

toolbox for analysis of single-trial eeg dynamics in-

cluding independent component analysis. Journal of

Neuroscience Methods, 134(1):9–21.

Dipti Pawar, S. D. (2020). Multiclass covert speech classi-

ﬁcationusing extreme learning machine. Biomedical

Engineering Letters.

Dipti Pawar, S. D. (2023). Eeg-based covert speech decod-

ing using random rotation extreme learning machine

ensemble for intuitive bci communication.

Fitriah, N., Zakaria, H., and Rajab, T. L. E. (2022). Eeg-

based silent speech interface and its challenges: A

survey. International Journal of Advanced Computer

Science and Applications, 13(11).

Gu, X., Cao, Z., Jolfaei, A., Xu, P., Wu, D., Jung, T. P., and

Lin, C. T. (2021). EEG-Based Brain-Computer Inter-

faces (BCIs): A Survey of Recent Studies on Signal

Sensing Technologies and Computational Intelligence

Approaches and Their Applications. IEEE/ACM

Trans Comput Biol Bioinform, 18(5):1645–1666.

Laureys (2005). The locked-in syndrome : what is it like to

be conscious but paralyzed and voiceless?

Lopez-Bernal, D., Balderas, D., Ponce, P., and Molina, A.

(2022). A State-of-the-Art Review of EEG-Based

Imagined Speech Decoding. Front Hum Neurosci,

16:867281.

Martin (2018). Decoding Inner Speech Using Electrocor-

ticography: Progress and Challenges Toward a Speech

Prosthesis. Frontiers.

Nguyen, C. H., Karavas, G. K., and Artemiadis, P. (2018).

Inferring imagined speech using EEG signals: a new

approach using Riemannian manifold features. J Neu-

ral Eng, 15(1):016002.

Panachakel, J. T. and Ramakrishnan, A. G. (2021). De-

coding Covert Speech From EEG-A Comprehensive

Review. Front Neurosci, 15:642251.

Price, C. J. (2012). A review and synthesis of the ﬁrst

20 years of pet and fmri studies of heard speech,

spoken language and reading. Neuropsychologia,

50(11):2625–2641.

Qureshi, M. N. I., Min, B., Park, H.-j., Cho, D., Choi,

W., and Lee, B. (2018). Multiclass classiﬁcation

of word imagination speech with hybrid connectivity

features. IEEE Transactions on Biomedical Engineer-

ing, 65(10):2168–2177.

Yi, G., Wang, J., Bian, H., Han, C., Deng, B., Wei, X.,

and Li, H. (2013). Multi-scale order recurrence quan-

tiﬁcation analysis of EEG signals evoked by man-

ual acupuncture in healthy subjects. Cogn Neurodyn,

7(1):79–88.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

628